A core loss prediction method based on empirical state space model

Abstract

The present invention relates to the technical field of electromagnetics and magnetic material loss prediction, and specifically to a method for predicting magnetic core loss based on an empirical state-space model. This method combines magnetic flux density (Bt) and magnetic field intensity (Ht) with temperature and frequency as key physical variables and incorporates them into a neural network model. This not only broadens the model's input feature space but also considers the cross-coupling effects between variables over a wider range, enabling the model to demonstrate strong predictive capabilities under different materials and operating conditions, while also improving its applicability in a variety of environments. Furthermore, by introducing multi-state attribute constraints and considering their mutual coupling and constraint effects, multiple parallel neural network modules are designed to simulate the superposition effects of core loss, breaking through the limitations of traditional single-formula models and more comprehensively reflecting the complex variations in core loss, thereby effectively improving the model's performance under a variety of operating conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetism and magnetic material loss prediction, and in particular to a magnetic core loss prediction method based on an empirical state space model. Background Art

[0002] In existing technologies, core loss prediction relies primarily on empirical formulas and neural network models. The Steinmetz Equation (SE) is a representative empirical formula, which calculates core loss using experimentally extracted factors k, α, and β along with frequency and flux density. Subsequent improvements, including the Modified Steinmetz Equation (MSE), Generalized Steinmetz Equation (GSE), and Improved Generalized Steinmetz Equation (iGSE), have expanded its applicability to non-sinusoidal waveforms. However, these formulas still suffer from issues such as inapplicability of self-simulated factors and significant discrepancies in calculated results. This is particularly problematic with complex core materials and cross-coupling effects.

[0003] Neural network models demonstrate strong generalization capabilities by automatically fitting the nonlinear relationship between characteristic variables and core loss. For example, the two-stage machine learning model proposed by Li H et al., the transfer learning technique used by Dogariu E et al., and the machine learning algorithm combined with small sample training by L Zhang et al. have all achieved good results in core loss prediction, but they also suffer from poor generalization. In recent years, researchers have begun to attempt to incorporate empirical formulas into neural network models. For example, Bar'yakhtar VG et al. combined the Landau-Lifshitz-Gilbert (LLG) equation with the application of microscopic magnetic theory, and Tanaka H et al. incorporated the LLG equation into the magnetic circuit model to calculate iron loss. Saeed S et al. determined core loss using the Jiles-Atherton hysteresis model and combined it with Dowell to calculate eddy current loss, further improving the model's predictive performance. However, existing technologies still have shortcomings in the synergy of multiple variables and the fusion of multimodal features, and are unable to fully reflect the complex variations of core loss. Summary of the Invention

[0004] To achieve the above objectives, this application provides the following technical solutions:

[0005] According to a first aspect of the present invention, the present invention claims a method for predicting core loss based on an empirical state space model, comprising:

[0006] Taking ESSM Block as the main operation model, each ESSM Block contains a Mamba block and a proxy attention block. The Mamba block dynamically calculates the synergistic effect between multiple variables through the state space model and the selective state update mechanism. The proxy attention block combines the Softmax attention mechanism and the linear attention mechanism to balance global information capture and computational efficiency. Through the parallel design of multiple ESSM Blocks, it scans the synergistic effects of factors at different scales, independently processes input features, optimizes the parameter space and improves prediction accuracy.

[0007] Furthermore, the method comprises:

[0008] Magnetic flux density Bt, magnetic field intensity Ht, temperature, and frequency are used as input features. The data is processed using the maximum and minimum normalization method to avoid gradient explosion and balance feature weights.

[0009] The state-space fitting capability of the Mamba model circumvents the linear assumption limitations of traditional empirical formulas. Combined with the Agent Attention mechanism, it dynamically adjusts the degree of attention paid to different input features, achieving efficient information extraction and global dependency capture.

[0010] The superposition effect of core loss is simulated through multi-state attribute constraints and parallel neural network module design.

[0011] Furthermore, the method comprises:

[0012] The state space model of the Mamba block is represented by the discrete time state space equation xk+1=Axk+Buk;

[0013] Where xk is the state variable, uk is the input variable, A and B are the state transfer matrices;

[0014] By combining the ESSM Block model, the Mamba block works together with the agent attention block;

[0015] The multi-state space model and Agent Attention mechanism are used to dynamically adjust the degree of attention to different input features.

[0016] Furthermore, the method further comprises:

[0017] The linear attention mechanism of the proxy attention block is implemented by calculating the query matrix Q, the key matrix K, and the value matrix V. The softmax function is used to normalize the attention weights to ensure that the relative importance of different features can be accurately captured.

[0018] The ESSM Block model combines the Mamba block and the proxy attention block to achieve a balance between global information capture and computational efficiency;

[0019] Dynamically adjust the degree of attention to different input features through the multi-state space model and Agent Attention mechanism;

[0020] The parallel design of ESSM Block introduces multiple adjustable channels, each of which processes input features of different scales;

[0021] Combining the Mamba block and the Agent Attention mechanism, the ESSM Block dynamically adjusts the degree of attention according to the importance of the input features, achieving a balance between global information capture and computational efficiency.

[0022] Furthermore, the method further comprises:

[0023] Magnetic flux density, magnetic field intensity, temperature and frequency are used as input features. The data is processed through the maximum and minimum normalization method. The state space fitting capability of the Mamba model is utilized, combined with the Agent Attention mechanism to dynamically adjust the degree of attention to different input features. The superposition effect of core loss is simulated through multi-state attribute constraints and parallel neural network module design.

[0024] Furthermore, the method further comprises:

[0025] The maximum and minimum normalization method performs a linear transformation on the original data and scales the data to the interval [0, 1]. The normalization process eliminates the dimensional differences between different features.

[0026] Combined with the ESSM Block model, the Mamba block and the proxy attention block are used to dynamically capture the synergistic effects among multiple variables. The Mamba block is responsible for extracting local features.

[0027] The agent attention block dynamically adjusts the degree of attention to different input features through the Agent Attention mechanism.

[0028] Furthermore, the method further comprises:

[0029] The Mamba model optimizes the state space fitting capability by minimizing the mean square error (MSE) between the predicted value and the actual value, ensuring that the error is within the preset range.

[0030] The Mamba model combines the ESSM Block, fusing the Mamba block and the proxy attention block;

[0031] Introducing a multi-state space model and Agent Attention mechanism to dynamically adjust the degree of attention paid to input features;

[0032] The Agent Attention mechanism calculates the attention weight of each input feature, with a weight value between 0 and 1, and dynamically adjusts the degree of attention paid to different input features;

[0033] Multi-state attribute constraints are achieved by setting constraints on multiple state variables, and the number of constraints is adjustable within a preset range.

[0034] Furthermore, the method further comprises:

[0035] Use multiple parallel neural network layers to process different input features at different levels;

[0036] The system can meet the needs of tasks of different complexity through the adjustability of the number of layers;

[0037] Combining the Mamba block and the proxy attention block in the ESSM Block model, the model dynamically calculates the synergistic effects between multiple variables and finds the optimal balance between different input features;

[0038] The multi-state space model and Agent Attention mechanism are introduced to adaptively adjust the degree of attention to different features.

[0039] The present invention relates to the technical field of electromagnetics and magnetic material loss prediction, and specifically to a method for predicting magnetic core loss based on an empirical state-space model. This method combines magnetic flux density (Bt) and magnetic field intensity (Ht) with temperature and frequency as key physical variables and incorporates them into a neural network model. This not only broadens the model's input feature space but also considers the cross-coupling effects between variables over a wider range, enabling the model to demonstrate strong predictive capabilities under different materials and operating conditions, while also improving its applicability in a variety of environments. Furthermore, by introducing multi-state attribute constraints and considering their mutual coupling and constraint effects, multiple parallel neural network modules are designed to simulate the superposition effects of core loss, breaking through the limitations of traditional single-formula models and more comprehensively reflecting the complex variations in core loss, thereby effectively improving the model's performance under a variety of operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a structural diagram of an Empirical State-Space Model of a magnetic core loss prediction method based on an empirical state-space model claimed in an embodiment of the present application;

[0041] Figure 2 A Mamba model structure diagram of a core loss prediction method based on an empirical state space model as claimed in an embodiment of the present application;

[0042] Figure 3A comparison diagram of the proxy attention and multi-head attention structures of a core loss prediction method based on an empirical state space model claimed in an embodiment of the present application;

[0043] Figure 4 A core loss prediction flow chart of a core loss prediction method based on an empirical state space model as claimed in an embodiment of the present application. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0045] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features specified as "first," "second," or "third" may explicitly or implicitly include at least one of such features. In the description of this application, "plurality" means at least two, for example, two, three, etc., unless otherwise specifically defined. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are intended only to illustrate the relative positional relationships and movement of components in a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly. Furthermore, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements and may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to such process, method, product, or apparatus.

[0046] References to "embodiments" herein mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0047] According to a first embodiment of the present invention, the present invention claims protection. According to a first aspect of the present invention, the present invention claims protection for a method for predicting core loss based on an empirical state space model, comprising:

[0048] Taking ESSM Block as the main operation model, each ESSM Block contains a Mamba block and a proxy attention block. The Mamba block dynamically calculates the synergistic effect between multiple variables through the state space model and the selective state update mechanism. The proxy attention block combines the Softmax attention mechanism and the linear attention mechanism to balance global information capture and computational efficiency. Through the parallel design of multiple ESSM Blocks, it scans the synergistic effects of factors at different scales, independently processes input features, optimizes the parameter space and improves prediction accuracy.

[0049] Furthermore, the method comprises:

[0050] Magnetic flux density Bt, magnetic field intensity Ht, temperature, and frequency are used as input features. The data is processed using the maximum and minimum normalization method to avoid gradient explosion and balance feature weights.

[0051] The state-space fitting capability of the Mamba model circumvents the linear assumption limitations of traditional empirical formulas. Combined with the Agent Attention mechanism, it dynamically adjusts the degree of attention paid to different input features, achieving efficient information extraction and global dependency capture.

[0052] The superposition effect of core loss is simulated through multi-state attribute constraints and parallel neural network module design.

[0053] Furthermore, the method comprises:

[0054] The state space model of the Mamba block is represented by the discrete time state space equation xk+1=Axk+Buk;

[0055] Where xk is the state variable, uk is the input variable, A and B are the state transfer matrices;

[0056] By combining the ESSM Block model, the Mamba block works together with the agent attention block;

[0057] The multi-state space model and Agent Attention mechanism are used to dynamically adjust the degree of attention to different input features.

[0058] Furthermore, the method further comprises:

[0059] The linear attention mechanism of the proxy attention block is implemented by calculating the query matrix Q, the key matrix K, and the value matrix V. The softmax function is used to normalize the attention weights to ensure that the relative importance of different features can be accurately captured.

[0060] The ESSM Block model combines the Mamba block and the proxy attention block to achieve a balance between global information capture and computational efficiency;

[0061] Dynamically adjust the degree of attention to different input features through the multi-state space model and Agent Attention mechanism;

[0062] The parallel design of ESSM Block introduces multiple adjustable channels, each of which processes input features of different scales;

[0063] Combining the Mamba block and the Agent Attention mechanism, the ESSM Block dynamically adjusts the degree of attention according to the importance of the input features, achieving a balance between global information capture and computational efficiency.

[0064] Furthermore, the method further comprises:

[0065] Magnetic flux density, magnetic field intensity, temperature and frequency are used as input features. The data is processed through the maximum and minimum normalization method. The state space fitting capability of the Mamba model is utilized, combined with the Agent Attention mechanism to dynamically adjust the degree of attention to different input features. The superposition effect of core loss is simulated through multi-state attribute constraints and parallel neural network module design.

[0066] Furthermore, the method further comprises:

[0067] The maximum and minimum normalization method performs a linear transformation on the original data and scales the data to the interval [0, 1]. The normalization process eliminates the dimensional differences between different features.

[0068] Combined with the ESSM Block model, the Mamba block and the proxy attention block are used to dynamically capture the synergistic effects among multiple variables. The Mamba block is responsible for extracting local features.

[0069] The agent attention block dynamically adjusts the degree of attention to different input features through the Agent Attention mechanism.

[0070] Furthermore, the method further comprises:

[0071] The Mamba model optimizes the state space fitting capability by minimizing the mean square error (MSE) between the predicted value and the actual value, ensuring that the error is within the preset range.

[0072] The Mamba model combines the ESSM Block, fusing the Mamba block and the proxy attention block;

[0073] Introducing a multi-state space model and Agent Attention mechanism to dynamically adjust the degree of attention paid to input features;

[0074] The Agent Attention mechanism calculates the attention weight of each input feature, with a weight value between 0 and 1, and dynamically adjusts the degree of attention paid to different input features;

[0075] Multi-state attribute constraints are achieved by setting constraints on multiple state variables, and the number of constraints is adjustable within a preset range.

[0076] Furthermore, the method further comprises:

[0077] Use multiple parallel neural network layers to process different input features at different levels;

[0078] The system can meet the needs of tasks of different complexity through the adjustability of the number of layers;

[0079] Combining the Mamba block and the proxy attention block in the ESSM Block model, the model dynamically calculates the synergistic effects between multiple variables and finds the optimal balance between different input features;

[0080] The multi-state space model and Agent Attention mechanism are introduced to adaptively adjust the degree of attention to different features.

[0081] In this embodiment, the model mainly uses ESSM Block (Empirical State-Space Model) as the main operation model, and its structure is as follows: Figure 1 As shown in the figure, each ESSM Block contains a Mamba block and an agent attention block. The SSM state space contained in Mamba can maximize the cross-coupling effect between simulation factors. Therefore, it is more suitable for describing the dynamic system modeling method that describes the results by state changes. Combined with the agent attention, it can achieve a better fitting effect. The overall expression of the model is shown in Formula 1.

[0082] (Formula 1);

[0083] in, Represents the input core loss characteristics, including Bt, Ht, temperature and frequency, is the corresponding core loss value, Represents the ESSM block, is the proxy attention block, For parallel The number of blocks.

[0084] The model as a whole adopts a parallel design of multiple ESSM Blocks. By scanning the synergistic effects of factors at different scales, each ESSM Block independently processes the input features. The advantage of this parallel structure design is that it can simultaneously capture dynamic features at multiple scales, thereby more comprehensively characterizing the dynamic behavior of the system.

[0085] In the Mamba module, the state space model design and the selective state update mechanism design can dynamically calculate the fitting results of a large number of scales. Its structure is as follows: Figure 2 As shown in the figure, traditional state-space models are often difficult to converge in terms of model fitting due to the coordination problem of multiple variables. The selective mechanism and parallel computing strategy in Mamba, combined with the parameter design of automatic update convergence, can better fit the coordination problem between multiple variables.

[0086] Proxy attention combines the traditional Softmax attention mechanism with the linear attention mechanism. It balances the contradiction between the ability to capture global information and computational efficiency. Since core loss is affected by multiple factors, the complex synergy between these factors is mostly highly nonlinear. Therefore, from the perspective of model construction, a module that can efficiently capture global dependencies and has low computational complexity is required. Therefore, proxy attention can more efficiently realize information extraction when processing core loss prediction tasks.

[0087] Secondly, when Mamba is combined with the agent attention module, it can actively adjust parameters and efficiently capture the parts of Mamba's state space with higher information entropy. Therefore, the combination of the two integrates multi-scale and multi-dimensional information, which helps the model maintain high prediction accuracy and high generalization ability when facing unknown working conditions.

[0088] like Figure 3 As shown in the figure, we show the intuitive difference between proxy attention and multi-head attention, and dynamically adjust the attention distribution by introducing proxy variables, thereby improving the expressiveness and computational efficiency of the model.

[0089] By introducing proxy vectors, proxy attention can dynamically adjust the attention distribution, enhance the model's ability to capture global context information, and to some extent alleviate the parameter redundancy problem caused by too many heads in multi-head attention. This enables the model to more effectively utilize computing resources while maintaining or improving performance, especially in practical application scenarios with limited resources. When the number of input features is When setting The number of , we compare the computational complexity of proxy attention with that of multi-head attention, by reducing The number of parameters is greatly reduced, and the complexity is limited to the linear range, while improving the fitting effect and reducing the complexity redundancy problem.

[0090] The core loss modeling process is as follows Figure 4 As shown in the figure, the input features used include Ht, Bt, frequency and temperature. The four features are all processed using the maximum and minimum normalization method. By reducing the data to [0,1], the gradient explosion problem is avoided, and the weight problem between each input feature can be balanced. The preliminary data enhancement is performed by controlling the proportion of each part in the data set. Then, after model training, under ideal conditions, the model is verified and evaluated, and each indicator will reach a relatively ideal situation. At this time, the model can obtain a more accurate core loss by using Ht, Bt, frequency and temperature. If the index gap is large, the model will be retrained until the indicators of the model reach a more reasonable level.

[0091] Error avg is used to measure the average squared difference between the predicted value and the true value. The formula is shown in Equation 2.

[0092] (Formula 2);

[0093] in, It is The measured values ​​of the samples, It is The predicted value of the sample, Represents the total number of samples.

[0094] Error RMS is the square root of Error avg, which provides an error scale with the same units as the original data, as shown in Equation 3.

[0095] (Formula 3);

[0096] in, It is The measured values ​​of the samples, It is The predicted value of the sample, Represents the total number of samples.

[0097] R² measures the accuracy of the model's predictions. It represents the proportion of the variation explained by the model to the total variation. The formula is shown in Equation 4.

[0098] (Formula 4);

[0099] in, It is The measured values ​​of the samples, It is The predicted value of the sample, is the average of the measured values, Represents the total number of samples.

[0100] The Error Rate gives the relative error percentage between the predicted value and the actual measured value, as shown in Equation 5.

[0101] (Formula 5);

[0102] in, It is The measured values ​​of the samples, It is The predicted value of the sample, Represents the total number of samples.

[0103] We compared the ESSM with representative models from these two categories in various aspects. We tested four waveform recordings from ten materials. Since the formula parameters and sampling data used in the ciGSE experiment were not clearly specified, and the test samples were only approximately 10% of each sample, resulting in non-reproducible results, we used the test results from the paper as a reference and tested all samples in the dataset. We used Error RMS as the primary metric and Error 95-Prct as a secondary metric, and the results are presented here. The experimental data demonstrates that our model demonstrates significant overall superiority, maintaining a stable low error level for most test samples, with an average error rate generally between 1% and 3%. This result contrasts sharply with the traditional iGSE model, which exhibits an error as high as 37.63% under certain conditions. While the ciGSE model offers improvements over the iGSE, it still exhibits significant performance degradation at high temperatures, while our model demonstrates superior predictions at high temperatures. Furthermore, the model's excellent generalization across different materials supports its application in diverse application scenarios.

[0104] Furthermore, we list the existing classic neural network models and the test results. The ESSM model is far superior to the FNN and FANN in both Error RMS and Error 95-Prct. In the tests of ten materials, the ESSM is the optimal solution among the existing models. Taking 3C94 material as an example, FNN and FANN show large Error 95-Prct errors, while the ESSM model performs well. This is attributed to its unique state space structure, which shows high versatility.

[0105] We compared the ESSM with representative models from these two categories in various aspects. We tested four waveform records from ten materials. Since the formula parameters and sampling data in the ciGSE experiment were not clearly specified, and the test samples were only approximately 10% of each material, resulting in non-reproducible results, we used the test results from the paper as a reference and tested all samples in the dataset. We used Error RMS as the primary metric and Error 95-Prct as a secondary metric. The results are summarized in Table 1. The experimental data show that our model demonstrates significant overall superiority. Our model maintains a stable low error level for most test samples, with an average error rate generally between 1% and 3%. This result stands in stark contrast to the traditional iGSE model, which exhibits an error as high as 37.63% under certain conditions. Although the ciGSE model improves upon the iGSE, it still suffers from significant performance degradation at high temperatures. Our model, however, demonstrates superior prediction results at high temperatures. Furthermore, the model's excellent generalization across different materials supports its application in diverse application scenarios.

[0106] Table 1 Comparison between ESSM model and mainstream empirical formula

[0107]

[0108] Furthermore, we list the existing classic neural network models and list the test results as shown in Table 2. The ESSM model is far superior to the FNN and FANN in both Error RMS and Error 95-Prct. In the tests of ten materials, the ESSM is the optimal solution among the existing models. Taking 3C94 material as an example, the FNN and FANN show large Error 95-Prct errors, while the ESSM model performs well. This is attributed to its unique state space structure, which shows high versatility.

[0109] Table 2 Comparison between ESSM model and mainstream neural network models

[0110]

[0111] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0112] In addition, the functional units in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units. The above is only an implementation method of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of this application, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

[0113] The above detailed description of the specific embodiments of the invention is intended only as an example, and the present application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions of the invention are also within the scope of the present application. Therefore, equivalent changes, modifications, and improvements made without departing from the spirit and scope of the present application should be included within the scope of the present application.

Claims

1. A core loss prediction method based on an empirical state space model, characterized in that: include: The ESSM Block is the main computational model. Each ESSM Block contains a Mamba Block and a Proxy Attention Block. The Mamba Block dynamically calculates the synergistic effects between multiple variables through a state space model and a selective state update mechanism. The Proxy Attention Block combines the Softmax Attention Mechanism with the Linear Attention Mechanism to balance global information capture and computational efficiency. Through the parallel design of multiple ESSM Blocks, it scans the synergistic effects of factors at different scales, independently processes input features, optimizes the parameter space, and improves prediction accuracy. Magnetic flux density Bt, magnetic field intensity Ht, temperature, and frequency are used as input features. The data is processed using the maximum and minimum normalization method to avoid gradient explosion and balance feature weights. The state-space fitting capability of the Mamba model is used to circumvent the linear assumption limitations of traditional empirical formulas. The AgentAttention mechanism is combined to dynamically adjust the degree of attention paid to different input features, achieving efficient information extraction and global dependency capture. Through multi-state attribute constraints and parallel neural network module design, the superposition effect of core loss is simulated; The overall expression of the model is shown in Formula 1: (Formula 1); in, Represents the input core loss characteristics, including Bt, Ht, temperature and frequency, is the corresponding core loss value, Represents the ESSM block, is the proxy attention block, For parallel The number of blocks.

2. The method for predicting core loss based on an empirical state space model according to claim 1, wherein: include: The state space model of the block is obtained by discrete time state space equation To express; in is the state variable, is the input variable, A and B are the state transfer matrices By combining the ESSM Block model, the Mamba block works together with the agent attention block; The multi-state space model and Agent Attention mechanism are used to dynamically adjust the degree of attention to different input features.

3. The method for predicting core loss based on an empirical state space model according to claim 1, wherein: Also includes: The linear attention mechanism of the proxy attention block is implemented by calculating the query matrix Q, the key matrix K, and the value matrix V. The softmax function is used to normalize the attention weights to ensure that the relative importance of different features can be accurately captured. The ESSM Block model combines the Mamba block and the proxy attention block to achieve a balance between global information capture and computational efficiency; Dynamically adjust the degree of attention to different input features through the multi-state space model and Agent Attention mechanism; The parallel design of ESSM Block introduces multiple adjustable channels, each of which processes input features of different scales; Combining the Mamba block and the Agent Attention mechanism, the ESSM Block dynamically adjusts the degree of attention according to the importance of the input features, achieving a balance between global information capture and computational efficiency.

4. The method for predicting core loss based on an empirical state space model according to claim 1, wherein: Also includes: The maximum and minimum normalization method performs a linear transformation on the original data and scales the data to the interval [0, 1]. The normalization process eliminates the dimensional differences between different features. Combined with the ESSM Block model, the Mamba block and the proxy attention block are used to dynamically capture the synergistic effects among multiple variables. The Mamba block is responsible for extracting local features. The agent attention block dynamically adjusts the degree of attention to different input features through the Agent Attention mechanism.

5. The method for predicting core loss based on an empirical state space model according to claim 1, wherein: Also includes: The Mamba model optimizes the state space fitting capability by minimizing the mean square error (MSE) between the predicted value and the actual value, ensuring that the error is within the preset range. The Mamba model combines the ESSM Block, fusing the Mamba block and the proxy attention block; Introducing a multi-state space model and Agent Attention mechanism to dynamically adjust the degree of attention paid to input features; The Agent Attention mechanism calculates the attention weight of each input feature, with a weight value between 0 and 1, and dynamically adjusts the degree of attention paid to different input features; Multi-state attribute constraints are achieved by setting constraints on multiple state variables, and the number of constraints is adjustable within a preset range.

6. The method for predicting core loss based on an empirical state space model according to claim 1, wherein: Also includes: Use multiple parallel neural network layers to process different input features at different levels; The system can meet the needs of tasks of different complexity through the adjustability of the number of layers; Combining the Mamba block and the proxy attention block in the ESSM Block model, the model dynamically calculates the synergistic effects between multiple variables and finds the optimal balance between different input features; The multi-state space model and Agent Attention mechanism are introduced to adaptively adjust the degree of attention to different features.

Images