A method for constructing a decoding model of a motor imagery brain-computer interface signal and a decoding method
By constructing the knowledge data fusion network KDFNet and combining prior knowledge and deep learning, efficient decoding of motor imagery brain-computer interface signals was achieved, solving the problem of insufficient generalization performance in existing technologies and improving decoding accuracy and classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2025-01-21
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for decoding motor imagery brain-computer interface signals have insufficient generalization performance across different devices and subjects, and require a large amount of labeled EEG data to achieve high classification accuracy.
A knowledge data fusion network KDFNet is constructed, which combines a temporal convolution module, a spatial convolution module, a feature engineering module, and a classification module. The convolution kernels are initialized through bandpass filters and co-spatial pattern filters. By combining prior knowledge and EEG data, log-variance feature extraction and traditional classifier parameter initialization are adopted to achieve efficient end-to-end decoding.
It improves the decoding accuracy and generalization performance of motor imagery brain-computer interface signals, enabling efficient decoding with fewer training samples and significantly improving the accuracy of MI classification.
Smart Images
Figure CN120045995B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of precise decoding of motor imagery brain-computer interface signals, and more specifically, relates to a method for constructing a decoding model and a decoding method for motor imagery brain-computer interface signals. Background Technology
[0002] Brain-computer interfaces (BCIs) establish a direct communication path between the human brain and external devices (computers, wheelchairs, robots, etc.). Electroencephalography (EEG) records electrical activity on the scalp and is the most widely used input signal in non-invasive BCIs due to its ease of use and low cost. EEG-based BCIs have been applied to robot control, speech decoding, stroke rehabilitation, and consciousness assessment.
[0003] Motor imagery (MI) refers to a user imagining movement of a part of a limb (or muscle) without actual movement. This process induces changes in the sensory-motor rhythms (SMR) in corresponding areas of the cerebral cortex, primarily involving μ rhythms (8-12 Hz) and β rhythms (14-30 Hz). ERD and ERS phenomena appear in the corresponding brain regions. For example, when imagining movement of the left hand, the energy of the two rhythmic signals in the right sensory cortex decreases, while the energy increases in the left sensory cortex; when imagining movement of the right hand, the energy decreases in the two rhythmic signals in the left sensory cortex and increases in the right sensory cortex. Therefore, based on this phenomenon, detecting SMR patterns in specific areas of the cerebral cortex can be used to identify the body parts imagined by the user, thereby generating various control commands.
[0004] Motor imagery brain-computer interfaces refer to interfaces where a person generates specific EEG patterns while performing a particular imagery task. Through preprocessing, feature extraction, and classification of these patterns, the EEG is decoded, and activation in different brain regions is detected to determine the user's intent. This enables direct communication and control between the human brain and external devices. Common areas for motor imagery include the left hand, right hand, both feet, and tongue.
[0005] Traditional EEG signal analysis methods typically rely on prior knowledge of ERD / ERS and follow three steps: signal processing, feature extraction, and feature classification. Signal processing aims to improve the signal-to-noise ratio (SNR) of the EEG signal. Since ERD / ERS occur within specific frequency bands, corresponding bandpass filters are usually applied to MI EEG. Basic feature extraction techniques focus on time-domain or frequency-domain analysis to extract discriminative features. Due to the spatial characteristics of MI EEG, Common Spatial Pattern (CSP) and its variants are widely used for feature extraction in MI EEG classification. It transforms the original multi-channel EEG signal into a more separable spatial pattern by designing spatial filters to maximize the variance ratio of the filtered signals from different classes. Initially, CSP was proposed for binary classification problems, and its extension to multi-class classification has been reported. Further, Filter Bank CSP (FBCSP) was proposed, which divides the EEG signal into multiple frequency bands through bandpass filters, extracts CSP features from each band, and then selects the most useful features for classification. After CSP filtering, the average energy of the signal in each channel is typically calculated as a feature for classification. Various classifiers can then be used, such as logistic regression (LR), support vector machine (SVM), and linear discriminant analysis (LDA). Although traditional EEG signal analysis methods incorporate some prior knowledge, significant differences in the distribution of EEG signal data across different devices and subjects often limit their generalization performance on diverse datasets.
[0006] Convolutional neural networks (CNNs) have also achieved excellent results in EEG signal decoding. ShallowCNN and DeepCNN have been reported for raw EEG classification. ShallowCNN is inspired by the FBCSP, with its temporal convolution, spatial convolution, log-variance calculation, and classifier network modules corresponding to specific computational steps in FBCSP. DeepCNN is similar but contains more convolutional and pooling layers. A more compact EEGNet has also been reported, using separable convolutions to reduce model parameters and demonstrating good performance on various brain-computer interface tasks.
[0007] Deep learning methods integrate feature extraction and classification into end-to-end neural networks, aiming to learn the optimal classification model directly from raw EEG data. Convolutional neural networks (CNNs) are among the most popular deep learning models for EEG decoding, demonstrating excellent performance in various MI EEG classification tasks. However, these models require large amounts of labeled EEG data to achieve high classification accuracy, and EEG data collection is time-consuming and inefficient. Summary of the Invention
[0008] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method for constructing a decoding model of motor imagination brain-computer interface signals and a decoding method. The purpose is to propose a decoding model construction method with high generalization performance and efficient decoding of motor imagination brain-computer interface signals.
[0009] To achieve the above objectives, according to one aspect of the present invention, a method for constructing a decoding model of motor imagery brain-computer interface signals is provided, comprising:
[0010] A knowledge data fusion network, KDFNet, is constructed, comprising a temporal convolution module, a spatial convolution module, a feature engineering module, and a classification module. The temporal convolution module contains m 1×1 temporal convolution kernels, where m represents the number of frequency bands required for the motor imagery brain-computer interface signal. Different temporal convolution kernels correspond to different frequency bands, and each kernel performs temporal convolution on each channel signal. The spatial convolution module contains m*n C×1 spatial convolution kernels. The output of one temporal convolution kernel is further processed by n spatial convolution kernels to obtain features for the corresponding frequency band that include both frequency domain and spatial knowledge. The feature engineering module calculates features representing spectral power on the output of each spatial convolution and concatenates the results as spectral power knowledge features. The classification module performs motor imagery classification based on these spectral power knowledge features.
[0011] The FBCSP model is trained in a supervised manner to obtain the parameter values of the classifier in the model; the number of bandpass filters in the FBCSP model, the length of each bandpass filter, and the corresponding frequency band simultaneous domain convolution kernel; the number of co-spatial pattern filters corresponding to each frequency band and the number of spatial domain convolution kernels corresponding to that frequency band; the parameters of the fully connected layer in the classification module are initialized with the parameter values of the classifier so that KDFNet can acquire classification knowledge; KDFNet is trained in a supervised manner to obtain additional knowledge from the training data and complete the construction of the decoding model of motor imagery brain-computer interface signals.
[0012] Furthermore, the feature engineering module is specifically used to calculate the log-variance features of the features output by each spatial convolution and to concatenate all log-variances as spectral power knowledge features.
[0013] Furthermore, the parameters of the corresponding spatial convolution kernel are initialized using the parameter values of each co-space mode filter, so that KDFNet can obtain spatial prior knowledge.
[0014] Furthermore, the parameters of the corresponding frequency domain convolution kernel are initialized using the parameter values of each bandpass filter, so that KDFNet can obtain prior knowledge in the frequency domain.
[0015] According to another aspect of the present invention, a method for decoding motor imagery brain-computer interface signals is provided, which uses a decoding model constructed by the decoding model construction method described above to achieve decoding.
[0016] According to another aspect of the invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above and / or the steps of the method described above.
[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed by a processor, it controls the device in which the storage medium is located to perform the steps of the method described above and / or the steps of the method described above.
[0018] According to another aspect of the present invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the method described above and / or the steps of the method described above.
[0019] In summary, compared with the prior art, the technical solutions conceived by this invention have the following main advantages:
[0020] 1. This invention proposes a method for constructing a Knowledge Data Fusion Network (KDFNet), bridging CSP and CNN by combining prior knowledge with EEG data. Specifically, firstly, the Knowledge Data Fusion Network KDFNet is constructed. Based on the hyperparameters of each module in KDFNet, an FBCSP model is built and trained. The parameters of the fully connected layers in the classification module of KDFNet are initialized using the parameter values of the classifier in the FBCSP model, enabling KDFNet to acquire classification knowledge. Then, KDFNet is trained under supervised conditions, and the gradient descent method for the classification loss is further improved to obtain additional knowledge from the training data, achieving knowledge and data fusion. This is a decoding model construction method with high generalization performance and efficient decoding of motor imagery brain-computer interface signals. Experimental tests verify that the KDFNet model has significant and stable performance in MI classification and can serve as a practical decoding method for EEG-based brain-computer interfaces (BCIs).
[0021] 2. This invention further proposes using log-variance activation to capture power spectrum information, which helps the neural network capture spectral power features related to motion imagery and improves decoding accuracy.
[0022] 3. The present invention further proposes to initialize the temporal convolutional layer and the spatial convolutional layer by using a bandpass filter and a common spatial mode filter, respectively, to mine the spatial and frequency prior knowledge of the signal. This enables the KDFNet model to integrate sensory-motor rhythm and event-related desynchronization and synchronization modes, thereby enabling the training of a decoding model with high generalization performance and high efficiency with fewer training samples. Attached Figure Description
[0023] Figure 1 This is a flowchart of a method for constructing a decoding model of motor imagery brain-computer interface signals provided in an embodiment of the present invention;
[0024] Figure 2 This is a schematic diagram illustrating the knowledge source, knowledge integration, and knowledge representation of KDFNet provided in this embodiment of the invention;
[0025] Figure 3 This is a schematic diagram of the KDFNet structure provided in an embodiment of the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0027] Example 1
[0028] A method for constructing a decoding model for motor imagery brain-computer interface signals, such as... Figure 1 As shown, it includes:
[0029] A knowledge data fusion network, KDFNet, is constructed, comprising a temporal convolution module, a spatial convolution module, a feature engineering module, and a classification module. The temporal convolution module contains m 1×1 temporal convolution kernels, where m represents the number of frequency bands required for the motor imagery brain-computer interface signals. Different temporal convolution kernels correspond to different frequency bands, and each kernel performs temporal convolution on each channel signal. The spatial convolution module contains m*n C×1 spatial convolution kernels. The output of one temporal convolution kernel is further processed by n spatial convolution kernels to obtain features for the corresponding frequency band that include both frequency domain and spatial knowledge. The feature engineering module calculates features representing spectral power on the output of each spatial convolution and concatenates the results as spectral power knowledge features. The classification module performs motor imagery classification based on these spectral power knowledge features.
[0030] Supervised training of the FBCSP model was conducted to obtain the parameter values of the classifier in the model; the number of bandpass filters in the FBCSP model, the length of each bandpass filter, and the corresponding frequency band simultaneous domain convolution kernel; the number of co-space mode filters corresponding to each frequency band and the number of spatial domain convolution kernels corresponding to that frequency band;
[0031] The parameters of the fully connected layer in the classification module are initialized using the parameter values of the classifier so that KDFNet can acquire classification knowledge; KDFNet is trained in a supervised manner to obtain additional knowledge from the training data, thus completing the construction of the decoding model for motor imagery brain-computer interface signals.
[0032] This embodiment is a method for deep integration of CSP and CNN. It designs a CSP-enhanced CNN model and a CNN initialization method that incorporates expert knowledge. It is an efficient decoding method that deeply combines expert knowledge and neural networks in brain-computer interface MI classification.
[0033] As a preferred implementation, the feature engineering module is specifically used to calculate the log-variance features of the features output by each spatial convolution and to concatenate all the log-variances as spectral power knowledge features.
[0034] As a preferred implementation, the parameters of the corresponding spatial convolution kernel are initialized using the parameter values of each co-space mode filter, so that KDFNet can obtain spatial prior knowledge.
[0035] As a preferred implementation, the parameters of the corresponding frequency domain convolution kernel can be initialized using the parameter values of each bandpass filter, so that KDFNet can obtain frequency domain prior knowledge.
[0036] Combination Figure 2 This paper introduces the knowledge sources, knowledge fusion, and knowledge representation of KDFNet.
[0037] KDFNet's knowledge source is the traditional knowledge-driven machine learning process in MI classification, namely the classic FBCSP process. Each module utilizes corresponding prior knowledge:
[0038] Time-domain filtering stage: Prior knowledge of the ERD / ERS frequency distribution is provided through multiple bandpass filters.
[0039] Frequency domain filtering stage: CSP filtering is used to provide prior knowledge of the spatial distribution of ERD / ERS in each frequency band.
[0040] Feature engineering stage: Extract energy-based log-variance features from the spatially filtered data.
[0041] Classification stage: Feature classification is performed using an interpretable traditional classifier.
[0042] Regarding the integration of knowledge, the overall structure of KDFNet remains consistent with the knowledge-driven FBCSP process, such as... Figure 3 As shown, it mainly consists of four parts: a temporal convolution module, a spatial convolution module, a feature engineering module, and a classification module. In one specific implementation, temporal and spatial filtering are implemented using convolutional layers, and prior knowledge is integrated by initializing the temporal and spatial convolution kernels with bandpass and CSP filters, respectively. In the feature engineering module, KDFNet uses log-variance calculation along the time dimension to merge prior knowledge of power spectrum-related features. The classification module consists of fully connected layers and a softmax layer for classification, integrating prior knowledge by initializing the fully connected layer parameters to those in a traditional classifier.
[0043] One specific implementation method for representing knowledge in KDFNet is as follows:
[0044] (a) Encode frequency domain knowledge into a temporal convolution kernel.
[0045] (b) Spatial domain knowledge is reflected in the spatial convolution kernel.
[0046] (c) Feature-related knowledge is represented by logarithmic variance calculation.
[0047] (d) Classifier knowledge is reflected in the parameters of the fully connected layer.
[0048] Therefore, the construction of KDFNet mainly includes two steps: 1) obtaining prior knowledge from the traditional FBCSP process; 2) constructing and training the network.
[0049] Next, we will introduce the structure of each module.
[0050] 1. Temporal Convolution Module
[0051] To integrate frequency domain knowledge, temporal convolution is a natural method for bandpass filtering based on the temporal convolution theorem. Fixed-size CNN convolutional kernels slide across the input signal, calculating the output through dot product operations, similar to FIR filtering. Therefore, by embedding bandpass FIR filters in the temporal convolutional layers of KDFNet, the frequency features of EEG signals can be effectively captured. Considering that ERD / ERS mainly occur in the 8-32Hz range, similar to FBCSP, multiple convolutional kernels are used to extract information from different frequency bands. Figure 3As shown, the frequency range is divided into 4Hz sub-bands, resulting in m = 6 temporal convolutional kernels, each with a size of 1×l, and each kernel corresponding to a feature map. Therefore, the input-output data format is m×C×(T-l+1), where C is the number of EEG signal channels and T is the number of time sampling points for each channel of the EEG signal. To accelerate convergence and enhance generalization, two-dimensional batch normalization is applied along the feature map dimension.
[0052] 2. Spatial Convolution Module
[0053] The spatial convolution module enables the model to capture spatial discriminative features, which is crucial for MI classification. This module uses depthwise convolution, where the feature map of each temporal filter is processed by a different spatial filter. To integrate prior knowledge, we first train CSP filters on m frequency bands on the training data, obtaining n CSP filters in each band, and then use them to initialize the spatial convolution kernels. This initialization ensures that the spatial convolution kernels start from a good position and can be further optimized during training to better fit the data, while leveraging both prior knowledge and the data.
[0054] 3. Feature Engineering Module
[0055] In most CNN networks, classic activation functions such as GELU and ELU are used for nonlinear transformations, and average pooling or max pooling is used to reduce feature dimensionality for classification. However, different MI (Motion Imagination) categories have different spectral power patterns. Variance-based operations that capture the spectral power of time series may be more suitable for MI classification. Therefore, we use a log-variance layer that computes the log-variance of the time series to extract energy-type discriminative features (spectral power knowledge feature output). After computing the log-variance features of the feature output containing frequency domain and spatial domain knowledge (obtained by computing the log-variance of the feature maps after time-domain and spatial convolutions), we obtain m×n features, which are then concatenated and processed using one-dimensional batch normalization.
[0056] 4. Classification Module
[0057] The features extracted by the log-variance layer are input into the fully connected layer for classification. The fully connected layer has m×n input units (feature dimension) and k output units (number of classification categories). Typically, the parameters of the fully connected layer are randomly initialized. To improve this, we initialize them using the parameters of a traditional classifier from the FBCSP (Filter Bank Common Space Pattern) procedure, giving the classification layer a better start. Finally, the output of the fully connected layer is passed through a softmax layer to obtain the class probabilities.
[0058] In this implementation, KDFNet comprises two supervised training phases: a traditional FBCSP classification learning phase and a network learning phase. In the traditional FBCSP classification process, m FIR filters with a bandwidth of length l are first computed. Then, n CSP filters are trained on the frequency-filtered signal for each frequency band. The spatially filtered signal is used to compute the log-variance feature vector, which serves as the input for training the LR classifier. In the KDFNet training phase, the convolutional kernels in the temporal and spatial filtering blocks are initialized with FIR and CSP filters, respectively. The parameters of the fully connected components are initialized using the parameters from the LR classifier. Then, KDFNet is trained by minimizing the cross-entropy loss, enabling it to learn from the data and supplement prior knowledge.
[0059] Experiments demonstrate that, on the MI dataset, the method proposed in this invention achieves higher classification accuracy compared to traditional classification methods (CSP-LR, FBCLP-LR, MDRM) and deep models (without prior knowledge: CNN+LSTM, HybridNet, LMDA-Net, EEGConformer; with prior knowledge: EEGNet, ShallowCNN, FBCNet, CSP-Retraining, FBCSP-Retraining), as shown in Table 1. Table 1 displays the classification accuracy of different methods on the MI dataset for different users. The highest result for each user is highlighted in bold, and the second highest result is indicated by an underline.
[0060] Table 1
[0061]
[0062] In summary, this invention proposes KDFNet, which integrates knowledge-driven and data-driven approaches into EEG-based MI classification. Prior knowledge of the sensorimotor rhythm frequency band is incorporated by initializing the temporal convolutional kernel using a bandpass FIR filter. Similarly, spatial convolutional kernels with CSP filters are initialized using spatial distribution knowledge of ERD and ERS. Differentiable log-variance calculations are then applied to capture power spectrum information. Furthermore, the parameters of a conventional LR classifier trained on these knowledge-based features are used to initialize the parameters of the KDFNet classification layers. Finally, KDFNet is fine-tuned using labeled training data.
[0063] In this implementation, KDFNet incorporates prior knowledge from different stages of the motor imagery brain-computer interface system. In the temporal filtering stage, frequency knowledge is incorporated by initializing the temporal convolution kernel using a bandpass finite impulse response (FIR) filter. In the spatial filtering stage, spatial knowledge is integrated by initializing the spatial convolution kernel using a CSP filter. For feature engineering, knowledge of handcrafted features is introduced through power spectrum calculation based on logarithmic variance. Finally, in the classification stage, the parameters of the classification layer are initialized using a conventional classifier. After embedding prior knowledge, KDFNet further performs data-driven end-to-end optimization of the trained EEG. In summary, KDFNet integrates prior knowledge at multiple stages and adapts to the data, thereby achieving better MI EEG decoding performance.
[0064] It should be noted that the various optimization methods, such as log-variance features, spatial prior knowledge, and frequency prior knowledge, can be arbitrarily combined and applied in conjunction with classification knowledge to construct the decoding model.
[0065] Example 2
[0066] A method for decoding motor imagery brain-computer interface signals, which uses a decoding model constructed by the decoding model construction method described above to achieve decoding.
[0067] The relevant technical solutions are the same as in Embodiment 1, and will not be repeated here.
[0068] Example 3
[0069] This application also relates to an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.
[0070] The electronic device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The memory can be used to store computer programs and / or modules. The processor performs various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory.
[0071] The relevant technical solutions are the same as above, and will not be repeated here.
[0072] Example 4
[0073] This application also relates to a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.
[0074] Specifically, the memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.
[0075] The relevant technical solutions are the same as above, and will not be repeated here.
[0076] Example 5
[0077] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the method described in the above embodiments of this application.
[0078] The relevant technical solutions are the same as above, and will not be repeated here.
[0079] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for constructing a decoding model of motor imagery brain-computer interface signals, characterized in that, include: A knowledge data fusion network, KDFNet, is constructed, comprising a temporal convolution module, a spatial convolution module, a feature engineering module, and a classification module. The temporal convolution module contains m features of size [missing information]. The temporal convolution kernels, where m is the number of frequency bands required for the motor imagery brain-computer interface signal, with different kernels corresponding to different frequency bands, and each kernel performs temporal convolution on each channel signal; the spatial convolution module contains m*n kernels of size m. The spatial convolution kernels are used to further process the output of one temporal convolution kernel into n spatial convolution kernels, resulting in features containing frequency domain knowledge and spatial knowledge for the corresponding frequency band. The feature engineering module is used to perform feature calculations to characterize the spectral power of the output of each spatial convolution and concatenate the results as spectral power knowledge features. The classification module is used to perform motion image classification based on the spectral power knowledge features. The FBCSP model is trained in a supervised manner to obtain the parameter values of the classifier in the model; the number of bandpass filters in the FBCSP model, the length of each bandpass filter, and the corresponding frequency band simultaneous domain convolution kernel; the number of co-spatial pattern filters corresponding to each frequency band and the number of spatial domain convolution kernels corresponding to that frequency band; the parameters of the fully connected layer in the classification module are initialized with the parameter values of the classifier so that KDFNet can acquire classification knowledge; KDFNet is trained in a supervised manner to obtain additional knowledge from the training data and complete the construction of the decoding model of motor imagery brain-computer interface signals; In addition, the parameter values of the bandpass filters in the FBCSP model are obtained, and the parameters of the corresponding time-domain convolution kernel are initialized using the parameter values of each bandpass filter, so that KDFNet can obtain frequency domain prior knowledge. It also obtains the parameter values of the cospace mode filter in the FBCSP model, and initializes the parameters of the corresponding spatial convolution kernel with the parameter values of each cospace mode filter so that KDFNet can obtain spatial prior knowledge. The feature engineering module is specifically used to calculate the logarithmic variance of the features output by each spatial convolution and to concatenate all logarithmic variances as spectral power knowledge features.
2. A method for decoding signals from a motor imagery brain-computer interface, characterized in that, Decoding is achieved by constructing a decoding model using the decoding model construction method described in claim 1.
3. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in claim 1 and / or the steps of the method as described in claim 2.
4. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed by a processor, it controls the device on which the storage medium is located to perform the steps of the method as claimed in claim 1 and / or the steps of the method as claimed in claim 2.
5. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in claim 1 and / or the steps of the method as described in claim 2.