Precise electric power theft detection method based on multi-modal Mama

Abstract

The invention discloses an accurate electric power theft detection method based on multi-modal Mamba. The method comprises the following steps: S1, carrying out adaptive multi-scale data preprocessing, and carrying out denoising and normalization processing on input power time sequence data by adopting an adaptive score standardization method and a local abnormal factor technology so as to improve data quality; s2, a multi-mode Mama feature extraction layer is adopted; s3, a federated learning training layer: constructing a secure federated learning framework, realizing cooperative training among a plurality of power supply areas, and introducing a differential privacy mechanism at the same time to protect privacy of model parameters and training data; s4, an intelligent confrontation defense layer; s5, an adaptive classification decision system; s6, a lightweight optimization strategy; efficient and safe cooperative detection is realized through innovative fusion of a multi-modal Mamba model and a federated learning framework; through an intelligent confrontation system and lightweight optimization, the robustness and practicability of the model are improved; according to the method, the detection accuracy, the calculation efficiency, the privacy protection and the like are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of smart grid technology, and specifically relates to a smart grid electricity theft detection method and system based on multimodal Mamba. The present invention realizes high-precision, safe and reliable electricity theft detection by innovatively integrating federated learning, multimodal Mamba model and intelligent confrontation system. Background Art

[0002] With the rapid development of smart grids, electricity theft has become increasingly complex and hidden, causing huge economic losses to power companies. Existing electricity theft detection methods mainly rely on data analysis from a single institution and traditional deep learning models. These methods face multiple problems in practical applications, such as data silos, low computational efficiency, and poor model generalization. At the same time, due to the limitations of data privacy protection, it is difficult for multiple institutions to conduct effective data sharing and collaborative analysis.

[0003] At present, domestic and foreign scholars have conducted extensive research on smart grid electricity theft detection. Traditional methods are mainly based on rule models and statistical analysis, which cannot effectively handle complex electricity theft patterns. In recent years, deep learning methods have been widely used in this field, such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). However, these methods have high computational complexity when processing long sequence data and lack the ability to adaptively learn new electricity theft patterns.

[0004] Therefore, there is an urgent need for a smart grid electricity theft detection method that can achieve multi-institutional collaboration, protect data privacy, and have adaptive learning capabilities. Summary of the invention

[0005] The purpose of the present invention is to provide a smart grid electricity theft detection method and system based on multimodal Mamba, and to combine federated learning to solve the problems existing in the prior art. Specifically, the present invention proposes a multi-level electricity theft detection framework, which includes a distributed data processing layer, a multimodal feature extraction layer, a federated learning training layer, and an intelligent adversarial defense layer. The framework implements distributed data preprocessing and privacy protection at the bottom layer, uses an improved multimodal Mamba model for feature extraction and fusion in the middle layer, and implements model training and optimization through federated learning at the top layer, while introducing an intelligent adversarial system to improve the robustness of the model.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] In the distributed data processing layer, the present invention first introduces an adaptive multi-scale data preprocessing method. An accurate power theft detection method based on multimodal Mamba, the specific steps include: S1: Adaptive multi-scale data preprocessing: First, an adaptive score normalization method is introduced to dynamically adjust the normalization parameters according to the data distribution characteristics of different time periods:

[0008] Z(t)=(X(t)-μ(w)) / (σ(w)*α(t)), where μ(w) and σ(w) are the mean and standard deviation in the sliding window w, respectively, and α(t) is an adaptive weight factor, which is calculated as follows: α(t)=exp(-β*|S(t)-S(t-1)|), S(t) is the seasonal index at time t, and β is a smoothing parameter. An improved local anomaly factor Enhanced LOF is designed: First, the time series correlation weight is introduced to make the samples at adjacent time points have a higher reference value; second, multi-scale decomposition is performed in combination with wavelet transform, and anomaly detection is performed in different frequency domains; finally, the anomaly judgment standard is adaptively adjusted through dynamic thresholds; at the same time, the time series segmentation technology is used to adaptively determine the segmentation points based on the change point detection algorithm, and the overlapping sliding window technology is used to ensure the continuity of information between segments for subsequent analysis;

[0009] S2: Multimodal Mamba feature extraction layer: In the feature extraction layer, the present invention designs an innovative multimodal Mamba model; for the input sequence x(t), the state update equation is:

[0010] h(t)=Ah(t-1)+Bx(t)

[0011] y(t)=Ch(t)+Dx(t)

[0012] Among them, the state matrix A adopts a block diagonal structure to achieve efficient parallel computing; B is the input weight matrix; C is the output weight matrix; D is the direct mapping matrix. h(t-1) is the hidden state at time t-1, h(t) is the hidden state at the current time t, and x(t) is the input feature vector at time t.

[0013] The spatial feature extraction branch adopts a two-dimensional state space model, and its state update equation is:

[0014] H(t)=A s H(t-1)+B s X(t)

[0015] Y(t)=C s H(t)+D s X(t)

[0016] Among them, H(t) is the two-dimensional state tensor, X(t) is the input feature map, and As , B s , C s , D s is a learnable parameter matrix; H(t-1) is the hidden state at time t-1, H(t) is the hidden state at the current time t, and X(t) is the input feature vector at time t; it increases the focus on key features and forms a rich multimodal feature representation;

[0017] S3: Federated learning training layer: In the model training phase, the federated learning framework is used; the framework includes two main processes: local training and global aggregation. In the local training process, each participant uses local data to train the multimodal Mamba model, while protecting the model parameters through the differential privacy mechanism. In the global aggregation process, a secure aggregation algorithm is used to perform weighted averaging on the model updates of each party, ensuring the security and effectiveness of model training. For the local model parameter w of client i, i , the global model update equation is:

[0018] w g =∑(η i ·w i ) / ∑η i , where w g is the model parameter after global aggregation, w i represents the local model parameters of each participant (client i), η i is the client weight, q i is the data quality score of client i, which is protected by differential privacy mechanism:

[0019] η i =exp(ε·q i +Lap(1 / λ)), where q i is the data quality score of the client, ε is the privacy protection parameter (epsilon), which is used to control the privacy protection strength, λ is the scaling parameter of Laplace noise, and Lap(1 / λ) is the Laplace noise used for differential privacy protection;

[0020] S4: Intelligent adversarial defense layer: In order to improve the robustness and adaptability of the model, the present invention introduces an intelligent adversarial system; the system includes two core components: an adversarial sample generator and a defense optimizer; the adversarial sample generator can simulate various types of electricity theft modes, including intermittent electricity theft, concealed electricity theft and distributed electricity theft; the defense optimizer analyzes the attack characteristics and dynamically adjusts the defense strategy of the model, thereby improving the model's ability to identify new electricity theft modes; the present invention proposes an adaptive adversarial sample generation method; for the original sample x, the optimization goal of generating the adversarial sample x' is: min||x'-x|| 2 +λ·L adv(f(x'),y), where x represents the original sample, x' represents the adversarial sample, ||x'-x|| 2 represents the square of the Euclidean distance between the original sample and the adversarial sample, and λ is a balance parameter used to adjust the weight L of the perturbation size and the adversarial loss. adv is the adversarial loss function, f(x') is the model's predicted output for the adversarial sample, and y represents the true label to ensure the rationality of the generated sample; defense optimization is achieved by dynamically adjusting the model parameter θ*: θ * =argmin E[L(f θ(x) ,y)+β·L(f θ(x') ,y)], θ* represents the optimized model parameters, E is the expectation operator, L is the loss function, and f θ represents the model function with parameter θ. Finally, β is the balance coefficient, which is used to adjust the weight of the loss of the original sample and the adversarial sample. Defense optimization is achieved by dynamically adjusting the model parameter θ* by adjusting the λ and β values ​​according to the detection results of the adversarial sample.

[0021] S5: Adaptive classification decision system: The fused features are passed to the fully connected layer, and the category prediction is performed through the Softmax function. In order to improve the classification accuracy, an ensemble learning-based method is introduced to combine the prediction results of multiple models, thereby improving the stability and accuracy of the final decision;

[0022] S6: Lightweight optimization strategy: Design lightweight network structure to ensure the efficiency of the model and scalability in practical applications;

[0023] For a given time series X = {x1, x2, ..., x T}, where each x i is at timestamp t i The specific steps in S1 are as follows: Introduce an adaptive fractional normalization method to dynamically adjust the normalization parameters according to the data distribution characteristics of different time periods:

[0024] Z(t)=(X(t)-μ(w)) / (σ(w)*α(t)),

[0025] Where μ(w) and σ(w) are the mean and standard deviation within the sliding window w, respectively, and α(t) is an adaptive weight factor, which is calculated as follows:

[0026] α(t)=exp(-β*|S(t)-S(t-1)|),

[0027] Among them, S(t) is the seasonal index at time t, and β is the smoothing parameter. The time series data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate individual differences between different users and make the data more comparable. The local outlier factor LOF is also used for data denoising to identify and remove outliers to improve data quality. At the same time, the time series segmentation technology is used to adaptively determine the segmentation points based on the change point detection algorithm, and the overlapping sliding window technology is used to ensure the continuity of information between segments for subsequent analysis.

[0028] Preferably, the specific operations in S5 are as follows:

[0029] The fused features are passed to the fully connected layer and the category prediction is performed through the Softmax function; the formula of the Softmax function is:

[0030]

[0031] Among them, z k is the score of the kth category, where k is the number of categories. In order to improve the classification accuracy, an ensemble learning-based method is introduced to combine the prediction results of multiple models, thereby improving the stability and accuracy of the final decision.

[0032] Preferably, the specific optimization measures in S6 include:

[0033] Reduce model parameters: By using a selective state space model with linear time complexity in the MMM Layer, the computational burden and memory usage of the RNN model on long time series are reduced;

[0034] Reduce convolution layer parameters: In the process of image feature extraction, use smaller convolution kernels and depth-wise separable convolution to reduce the computational cost of convolution operations;

[0035] Optimize the training process: Use the AdamW optimizer, combined with the learning rate decay strategy, to accelerate the convergence of the model and avoid overfitting;

[0036] Parallel processing architecture: In the MMM Layer, by processing time series and two-dimensional image features in parallel, the model can simultaneously extract key information from both types of data, reduce the time consumption of step-by-step calculations, and enhance overall computing efficiency.

[0037] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0038] First, the federated learning framework enables secure collaboration among multiple institutions, significantly improving the detection performance of the model while protecting data privacy. Experimental results show that compared with traditional methods, the detection accuracy of the present invention is improved by more than 25%.

[0039] Second, the innovative multimodal Mamba model greatly improves the efficiency and effect of feature extraction. When processing long sequence data, the computational complexity of the model is only at the linear level of traditional recurrent neural networks, while maintaining a high feature extraction quality.

[0040] Third, the intelligent adversarial system significantly improves the robustness and adaptability of the model. Through continuous adversarial training and defense optimization, the system can effectively identify and respond to new electricity theft patterns, reducing the false alarm rate by 30%. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the process of the present invention;

[0042] Figure 2 It is a schematic diagram of multimodal Mamba feature extraction in the present invention;

[0043] Figure 3 Schematic diagram of the federated learning training layer. DETAILED DESCRIPTION

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

[0045] Example 1

[0046] See also Figure 1 , provides a technical solution: an accurate power theft detection method based on multimodal Mamba, the specific steps include:

[0047] S1: Adaptive multi-scale data preprocessing: First, an adaptive score normalization method is introduced to dynamically adjust the normalization parameters according to the data distribution characteristics of different time periods:

[0048] Z(t)=(X(t)-μ(w)) / (σ(w)*α(t)),

[0049] Where μ(w) and σ(w) are the mean and standard deviation within the sliding window w, respectively, and α(t) is an adaptive weight factor, which is calculated as follows:

[0050] α(t)=exp(-β*|S(t)-S(t-1)|),

[0051] Among them, S(t) is the seasonal index at time t, and β is the smoothing parameter. Design an improved local anomaly factor (Enhanced LOF): First, introduce the time series correlation weight so that the samples at adjacent time points have a higher reference value; second, combine the wavelet transform for multi-scale decomposition, and perform anomaly detection in different frequency domains; finally, adjust the anomaly judgment standard adaptively through the dynamic threshold. At the same time, the time series segmentation technology is used to adaptively determine the segmentation points based on the change point detection algorithm, and the overlapping sliding window technology is used to ensure the continuity of information between segments for subsequent analysis.

[0052] S2: Multimodal Mamba feature extraction layer (see Figure 2 ): For the input time series x(t), the state update equation is:

[0053] h(t)=Ah(t-1)+Bx(t)

[0054] y(t)=Ch(t)+Dx(t)

[0055] Among them, the state matrix A adopts a block diagonal structure to achieve efficient parallel computing; B is the input weight matrix; C is the output weight matrix; D is the direct mapping matrix. h(t-1) is the hidden state at time t-1, h(t) is the hidden state at the current time t, and x(t) is the input feature vector at time t. The spatial feature extraction branch adopts a two-dimensional state space model, and its state update equation is:

[0056] H(t)=A s H(t-1)+B s X(t),

[0057] Y(t)=C s H(t)+D s X(t)

[0058] Among them, H(t) is the two-dimensional state tensor, X(t) is the input feature map, and A s , B s , C s , D s is a learnable parameter matrix; H(t-1) is the hidden state at time t-1, H(t) is the hidden state at the current time t, and X(t) is the input feature vector at time t; it increases the focus on key features and forms a rich multimodal feature representation.

[0059] S3: Federated Learning Training Layer (see Figure 3): In the model training phase, the federated learning framework is used. This framework includes two main processes: local training and global aggregation. In the local training process, each participant uses local data to train the multimodal Mamba model, while protecting the model parameters through the differential privacy mechanism. In the global aggregation process, a secure aggregation algorithm is used to perform weighted averaging on the model updates of each party, ensuring the security and effectiveness of model training. For the local model parameter w of client i, i , the global model update equation is:

[0060] w g =Σ(η i ·w i ) / Ση i ,

[0061] Among them, w g is the model parameter after global aggregation, w i represents the local model parameters of each participant (client i), η i is the client weight, q i is the data quality score of client i, which is protected by differential privacy mechanism:

[0062] η i =exp(ε·q i +Lap(1 / λ)),

[0063] where q i is the data quality score of the client, ε is the privacy protection parameter (epsilon), which is used to control the privacy protection strength, λ is the scaling parameter of the Laplace noise, and Lap(1 / λ) is the Laplace noise used for differential privacy protection.

[0064] S4: Intelligent adversarial defense layer: In order to improve the robustness and adaptability of the model, an intelligent adversarial system is introduced. The system consists of two core components: an adversarial sample generator and a defense optimizer. The adversarial sample generator can simulate various types of electricity theft modes, including intermittent electricity theft, concealed electricity theft, and distributed electricity theft. The defense optimizer analyzes the attack characteristics and dynamically adjusts the model's defense strategy, thereby improving the model's ability to recognize new electricity theft modes. The present invention proposes an adaptive adversarial sample generation method. For the original sample x, the optimization goal of generating the adversarial sample x' is:

[0065] min||x'-x|| 2 +λ·L adv (f(x'),y),

[0066] Among them, x represents the original sample, x' represents the adversarial sample, ||x'-x|| 2represents the square of the Euclidean distance between the original sample and the adversarial sample, and λ is a balance parameter used to adjust the weight L of the perturbation size and the adversarial loss. adv is the adversarial loss function, f(x') is the model's predicted output for the adversarial sample, and y represents the true label to ensure the rationality of the generated sample; defense optimization is achieved by dynamically adjusting the model parameter θ*: θ * =argmin E[L(f θ(x) ,y)+β·L(f θ(x') ,y)], θ* represents the optimized model parameters, E is the expectation operator, L is the loss function, and f θ represents the model function with parameter θ. Finally, β is the balance coefficient, which is used to adjust the weight of the loss of the original sample and the adversarial sample. Defense optimization is achieved by dynamically adjusting the model parameter θ* by adjusting the λ and β values ​​according to the detection results of the adversarial sample.

[0067] S5: Adaptive classification decision system: The fused features are passed to the fully connected layer and the category prediction is performed through the Softmax function. The formula of the Softmax function is:

[0068]

[0069] Among them, z k is the score of the kth category, where k is the number of categories. In order to improve the classification accuracy, an ensemble learning-based method is introduced to combine the prediction results of multiple models, thereby improving the stability and accuracy of the final decision.

[0070] S6: Lightweight optimization strategy: Design a lightweight network structure to ensure the model’s high efficiency and scalability in practical applications. Optimization measures include:

[0071] Reduce model parameters: By using a selective state space model with linear time complexity in the MMM Layer, the computational burden and memory usage of the RNN model on long time series are reduced.

[0072] Reduce convolution layer parameters: In the process of image feature extraction, smaller convolution kernels and depth-wise separable convolutions are used to reduce the computational cost of convolution operations.

[0073] Optimize the training process: Adopt AdamW optimizer and combine learning rate decay strategy to accelerate the convergence of the model and avoid overfitting.

[0074] Parallel processing architecture: In the MMM Layer, by processing time series and two-dimensional image features in parallel, the model can simultaneously extract key information from both types of data, reduce the time consumption of step-by-step calculations, and enhance overall computing efficiency.

[0075] Although embodiments of the present invention have been shown and described (see the above detailed description for details), it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An accurate power theft detection method based on multimodal Mamba, characterized by: The specific steps include: S1: Adaptive multi-scale data preprocessing: First, introduce the adaptive fractional normalization method, and dynamically adjust the normalization parameters according to the data distribution characteristics of different time periods: Z(t) = (X(t)-μ(w)) / (σ(w)*α(t)), where μ(w) and σ(w) are the mean and standard deviation in the sliding window w, respectively, and α(t) is the adaptive weight factor, which is calculated in the following way: α(t) = exp(-β*|S(t)-S(t-1)|), S(t) is the seasonal index at time t, and β is the smoothing parameter; Design an improved local anomaly factor Enhanced LOF: First, introduce the time series correlation weight so that the samples at adjacent time points have a higher reference value; second, combine the wavelet transform for multi-scale decomposition, and perform anomaly detection in different frequency domains; finally, adaptively adjust the anomaly judgment standard through dynamic thresholds; at the same time, adopt the time series segmentation technology, adaptively determine the segmentation points based on the change point detection algorithm, and adopt the overlapping sliding window technology to ensure the continuity of information between segments for subsequent analysis; S2: Multimodal Mamba feature extraction layer: In the feature extraction layer, the present invention designs an innovative multimodal Mamba model; for the input sequence x(t), the state update equation is: h(t)=Ah(t-1)+Bx(t) y(t)=Ch(t)+Dx(t) Among them, the state matrix A adopts a block diagonal structure to achieve efficient parallel computing; B is the input weight matrix; C is the output weight matrix; D is the direct mapping matrix; h(t-1) is the hidden state at time t-1, h(t) is the hidden state at the current time t, and x(t) is the input feature vector at time t; The spatial feature extraction branch adopts a two-dimensional state space model, and its state update equation is: H(t)=A s H(t-1)+B s X(t) Y(t)=C s H(t)+D s X(t) Among them, H(t) is the two-dimensional state tensor, X(t) is the input feature map, and A s , B s , C s , D s is a learnable parameter matrix; H(t-1) is the hidden state at time t-1, H(t) is the hidden state at the current time t, and X(t) is the input feature vector at time t; it increases the focus on key features and forms a rich multimodal feature representation; S3: Federated learning training layer: In the model training phase, the federated learning framework is used; the framework includes two main processes: local training and global aggregation. In the local training process, each participant uses local data to train the multimodal Mamba model, while protecting the model parameters through the differential privacy mechanism. In the global aggregation process, a secure aggregation algorithm is used to perform weighted averaging on the model updates of each party, ensuring the security and effectiveness of model training. For the local model parameter w of client i, i , the global model update equation is: w g =∑(η i ·w i ) / ∑η i , where w g is the model parameter after global aggregation, w i represents the local model parameters of each participant (client i), η i is the client weight, q i is the data quality score of client i, which is protected by differential privacy mechanism: η i =exp(ε·q i +Lap(1 / λ)), where q i is the data quality score of the client, ε is the privacy protection parameter (epsilon), which is used to control the privacy protection strength, λ is the scaling parameter of Laplace noise, and Lap(1 / λ) is the Laplace noise used for differential privacy protection; S4: Intelligent adversarial defense layer: In order to improve the robustness and adaptability of the model, the present invention introduces an intelligent adversarial system; the system includes two core components: an adversarial sample generator and a defense optimizer; the adversarial sample generator can simulate various types of electricity theft modes, including intermittent electricity theft, concealed electricity theft and distributed electricity theft; the defense optimizer analyzes the attack characteristics and dynamically adjusts the defense strategy of the model, thereby improving the model's ability to identify new electricity theft modes; the present invention proposes an adaptive adversarial sample generation method; for the original sample x, the optimization goal of generating the adversarial sample x' is: min||x'-x|| 2 +λ·L adv (f(x'),y), where x represents the original sample, x' represents the adversarial sample, ||x'-x|| 2 represents the square of the Euclidean distance between the original sample and the adversarial sample, and λ is a balance parameter used to adjust the weight L of the perturbation size and the adversarial loss. adv is the adversarial loss function, f(x') is the model's predicted output for the adversarial sample, and y represents the true label to ensure the rationality of the generated sample; defense optimization is achieved by dynamically adjusting the model parameter θ*: θ * =argmin E[L(f θ(x) ,y)+β·L(f θ(x') ,y)], θ* represents the optimized model parameters, E is the expectation operator, L is the loss function, and f θ represents the model function with parameter θ. Finally, β is the balance coefficient, which is used to adjust the weight of the loss of the original sample and the adversarial sample. Defense optimization is achieved by dynamically adjusting the model parameter θ* by adjusting the λ and β values ​​according to the detection results of the adversarial sample. S5: Adaptive classification decision system: The fused features are passed to the fully connected layer, and the category prediction is performed through the Softmax function. In order to improve the classification accuracy, an ensemble learning-based method is introduced to combine the prediction results of multiple models, thereby improving the stability and accuracy of the final decision; S6: Lightweight optimization strategy: Design lightweight network structure to ensure the efficiency of the model and scalability in practical applications; For a given time series X = {x1, x2, ..., x T }, where each x i is at timestamp t i The specific steps in S1 are as follows: Introduce an adaptive fractional normalization method to dynamically adjust the normalization parameters according to the data distribution characteristics of different time periods: Z(t)=(X(t)-μ(w)) / (σ(w)*α(t)), Where μ(w) and σ(w) are the mean and standard deviation within the sliding window w, respectively, and α(t) is an adaptive weight factor, which is calculated as follows: α(t)=exp(-β*|S(t)-S(t-1)|), Among them, S(t) is the seasonal index at time t, and β is the smoothing parameter. The time series data is converted into a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate individual differences between different users and make the data more comparable. The local outlier factor LOF is also used for data denoising to identify and remove outliers to improve data quality. At the same time, the time series segmentation technology is used to adaptively determine the segmentation points based on the change point detection algorithm, and the overlapping sliding window technology is used to ensure the continuity of information between segments for subsequent analysis.

2. According to claim 1, a precise power theft detection method based on multimodal Mamba is characterized by: The specific operations in S5 are as follows: The fused features are passed to the fully connected layer and the category prediction is performed through the Softmax function; the formula of the Softmax function is: Among them, z k is the score of the kth category, where k is the number of categories. In order to improve the classification accuracy, an ensemble learning-based method is introduced to combine the prediction results of multiple models, thereby improving the stability and accuracy of the final decision.

3. According to claim 1, a precise power theft detection method based on multimodal Mamba is characterized by: Specific optimization measures in S6 include: Reduce model parameters: By using a selective state space model with linear time complexity in the MMM Layer, the computational burden and memory usage of the RNN model on long time series are reduced; Reduce convolution layer parameters: In the process of image feature extraction, use smaller convolution kernels and depth-wise separable convolution to reduce the computational cost of convolution operations; Optimize the training process: Use the AdamW optimizer, combined with the learning rate decay strategy, to accelerate the convergence of the model and avoid overfitting; Parallel processing architecture: In the MMM Layer, by processing time series and two-dimensional image features in parallel, the model can simultaneously extract key information from both types of data, reduce the time consumption of step-by-step calculations, and enhance overall computing efficiency.

Images