Method and system for removing artifacts of mixed signal, terminal and storage medium

Abstract

The invention discloses a mixed signal artifact removal method and system, a terminal and a storage medium, and relates to the technical field of bio-electricity signal processing. The method comprises the following steps: performing transcranial alternating current stimulation on a user through a transcranial electrical stimulation device; a mixed signal generated by a user based on transcranial alternating current stimulation is acquired through an electroencephalogram acquisition module; the mixed signal is a composite signal generated based on an electroencephalogram signal and a transcranial alternating current stimulation signal; and inputting the mixed signal into an artifact removal model to remove artifacts generated by the transcranial alternating current stimulation signal so as to obtain a reconstructed clean electroencephalogram signal. According to the method, a deep learning algorithm is adopted to establish an artifact removal model, and transcranial alternating current stimulation artifacts in mixed signals are removed through the model. The deep learning algorithm can autonomously learn artifact features through a plurality of training samples, does not need to manually define the artifact features, and can adapt to a plurality of application scenes.

Description

Technical Field

[0001] This invention relates to the field of bioelectric signal processing technology, and in particular to a method, system, terminal, and storage medium for artifact removal of mixed signals. Background Technology

[0002] Transcranial electrical stimulation (tES) is a non-invasive technique that applies a weak electrical current to the cerebral cortex via scalp electrodes to modulate neuronal excitability. Closed-loop neuromodulation, by acquiring real-time biological signals such as electroencephalograms (EEGs), allows for dynamic analysis and adjustment of stimulation parameters, thus improving safety.

[0003] In existing technologies, electroencephalography (EEG) instruments and transcranial electrical stimulation (TCS) devices are mostly designed separately, requiring the installation of two independent devices on the subject. This increases operational complexity and limits the convenience and application scenarios of the technology. Regarding signal processing, EEG signal amplitudes are typically 10-100 microvolts, while electrical stimulation signal amplitudes can reach millivolts or even volt levels—a difference of several orders of magnitude. This leads to artifacts contaminating or overwhelming the EEG signal when electrical stimulation is activated, making it impossible to effectively observe brain activity during stimulation, resulting in data loss and incomplete information utilization. Existing TCS artifact removal algorithms mainly include template subtraction and adaptive filtering algorithms. The former, based on a fixed template, cannot adapt to all scenarios, while the latter requires manually setting a reference signal, making it difficult to handle complex nonlinear artifacts. Therefore, existing artifact removal algorithms have limited effectiveness in removing tCS artifacts.

[0004] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, system, terminal and storage medium for artifact removal of mixed signals, in view of the above-mentioned defects of the prior art, so as to solve the problem that the existing artifact removal algorithms have limited removal effect on tACS artifacts.

[0006] The technical solution adopted by this invention to solve the problem is as follows: In a first aspect, embodiments of the present invention provide a method for removing artifacts from mixed signals, the method comprising: Transcranial alternating current stimulation is applied to the user using a transcranial electrical stimulation device; The EEG acquisition module acquires a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signals and transcranial alternating current stimulation signals. The mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, thereby obtaining a reconstructed clean EEG signal.

[0007] In one embodiment, the EEG acquisition module and the transcranial electrical stimulation device are integrated into one unit, and the circuits corresponding to the EEG acquisition module and the transcranial electrical stimulation device are independent of each other and do not interfere with each other.

[0008] In one implementation, the mixed signal is input into an artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, resulting in a reconstructed clean EEG signal, including: The mixed signal is input into the preprocessing module to obtain the preprocessed mixed signal; The preprocessed mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, thereby obtaining the reconstructed clean EEG signal.

[0009] In one implementation, the preprocessing module is used to perform one or more of the following operations: Power supply interference in the input signal is removed by using a Butterworth notch filter; Low-frequency drift of the input signal is removed by using a high-pass filter; Remove unstable signal segments of preset duration at the beginning and end of the input signal; The input signal is sliced ​​using a sliding window in units of preset duration to obtain several signal segments. The step size of each window sliding forward is half the length of the window. Each of the aforementioned signal segments is normalized to a specified signal amplitude range.

[0010] In one implementation, the artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

[0011] In one embodiment, the encoder module includes several cascaded convolutional modules, each of which includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, with the number of channels in each one-dimensional convolutional layer decreasing sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

[0012] In one implementation, the artifact removal model is pre-trained, and the method for obtaining training samples includes: Treat a clean EEG signal and its corresponding actual mixed signal as a signal pair; The preprocessing module performs signal processing on the signal pairs to obtain several training sample pairs; wherein, during the training phase, the preprocessing module is also used to perform time alignment on the signal pairs; each training sample pair includes a preprocessed clean signal segment segmented from the clean EEG signal, and a corresponding preprocessed mixed signal segment segmented from the actual mixed signal.

[0013] In one implementation, the training method for the artifact removal model includes: Select one of the unprocessed training sample pairs and input the mixed signal segment from the training sample pair into the initial artifact removal model; The initial artifact removal model is used to remove artifacts from the mixed signal segment, and the corresponding reconstructed clean signal segment is output. The clean signal segment in the training sample pair is used as a supervision signal, and the parameters of the initial artifact removal model are updated according to the supervision signal and the reconstructed clean signal segment. Determine whether the training termination condition is met. If not, continue to select an unprocessed training sample pair until the training termination condition is met, and obtain the trained artifact removal model.

[0014] In one implementation, the training termination condition includes: The root mean square difference between the monitoring signal and the reconstructed clean signal segment is minimized.

[0015] Secondly, embodiments of the present invention also provide a mixed-signal artifact removal system, the system comprising: A transcranial electrical stimulation device is used to provide transcranial alternating current stimulation to a user. The EEG acquisition module is used to acquire a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signals and transcranial alternating current stimulation signals. The artifact removal module is used to input the mixed signal into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals and obtain a reconstructed clean EEG signal.

[0016] In one embodiment, the artifact removal module includes: The preprocessing module is used to preprocess the mixed signal to obtain a preprocessed mixed signal; An artifact removal model is used to remove artifacts generated by transcranial alternating current stimulation signals from the preprocessed mixed signal to obtain the reconstructed clean EEG signal.

[0017] In one implementation, the artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

[0018] In one embodiment, the encoder module includes several cascaded convolutional modules, each of which includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, with the number of channels in each one-dimensional convolutional layer decreasing sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

[0019] Thirdly, embodiments of the present invention also provide a terminal, the terminal including a memory and one or more processors; the memory stores one or more programs; the programs include instructions for performing the artifact removal method for mixed signals as described above; the processor is used to execute the programs.

[0020] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having stored thereon a plurality of instructions adapted to be loaded and executed by a processor to implement the steps of the artifact removal method for mixed signals as described above.

[0021] The beneficial effects of this invention are as follows: In this embodiment, a transcranial electrical stimulation (TCS) device is used to apply TCS to a user; an EEG acquisition module acquires a mixed signal generated by the user based on the TCS; this mixed signal is a composite signal generated based on both EEG and TCS signals; the mixed signal is input into an artifact removal model to remove artifacts generated by the TCS signals, resulting in a reconstructed clean EEG signal. This invention employs a deep learning algorithm to establish an artifact removal model, which removes TCS artifacts from the mixed signal. The deep learning algorithm can autonomously learn artifact features using various training samples, eliminating the need for manual definition of artifact features and adapting to various application scenarios. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the artifact removal method for mixed signals provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the artifact removal model provided in the embodiments of the present invention.

[0025] Figure 3 This is a schematic diagram of the mixed signal artifact removal system provided in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the terminal provided in the embodiment of the present invention. Detailed Implementation

[0027] This invention discloses a method, system, terminal, and storage medium for artifact removal of mixed signals. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0028] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0029] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0030] To address the aforementioned deficiencies in existing technologies, this invention provides a method for artifact removal from mixed signals. The method includes: applying transcranial alternating current (AC) stimulation to a user using a transcranial electrical stimulation (TCS) device; acquiring a mixed signal generated by the user based on the AC stimulation using an EEG acquisition module; wherein the mixed signal is a composite signal generated based on EEG signals and AC stimulation signals; and inputting the mixed signal into an artifact removal model to remove artifacts generated by the AC stimulation signals, thereby obtaining a reconstructed clean EEG signal. The TCS device of this invention can perform either direct current (DC) or alternating current (AC) stimulation, and the artifact removal model effectively removes electrical stimulation artifacts, ultimately obtaining a clean EEG signal.

[0031] like Figure 1 As shown, the method includes: Step S100: Transcranial alternating current stimulation is applied to the user using a transcranial electrical stimulation device.

[0032] Specifically, the transcranial electrical stimulation device in this embodiment, through a dual-loop design with two independent current loops, can generate various electrical stimulation waveforms, including transcranial direct current stimulation (tDCS), transcranial alternating current stimulation (tACS), and time-domain coherent stimulation (TIS). tDCS modulates the cortical resting membrane potential with a constant current, making it suitable for regulating long-term neural plasticity; while tACS drives synchronous oscillation of neurons with sinusoidal current, enabling targeted enhancement of specific cognitive functions. This multi-mode compatibility allows the device to adapt to different application scenarios; for example, tDCS can be used to modulate prefrontal cortex excitability, and switching to tACS in Alzheimer's disease intervention can enhance theta wave synchronization.

[0033] For example, the chip in the transcranial electrical stimulation (TCS) device employs a miniaturized, programmable design, ensuring the overall electronic circuitry remains small and portable. The TCS channel is designed with two current loops, allowing simultaneous connection of up to four stimulation electrodes. This design supports both single-channel stimulation modes (one anode + one cathode), such as transcranial direct current stimulation, transcranial alternating current stimulation, transcranial random noise stimulation, and spurious stimulation; and two-channel stimulation modes (two anodes + two cathodes), such as temporal interference (TI).

[0034] In one implementation, the time-domain coherent stimulation in transcranial alternating current stimulation is achieved by generating a high-frequency wave through two current loops.

[0035] Specifically, the dual-loop design in this embodiment supports temporal coherent stimulation: two pairs of high-frequency currents generate a low-frequency envelope through phase interference, which penetrates the skull and forms a focused electric field in deep brain regions. Compared with traditional single-loop transcranial electrical stimulation (tES), the electric field distribution of dual-loop temporal coherent stimulation is more concentrated, which can reduce non-specific activation of surface tissues and improve the modulation precision of deep targets (such as the thalamus and hippocampus). Specifically, temporal coherent stimulation is achieved by generating a high-frequency wave in each current loop.

[0036] In one implementation, the transcranial electrical stimulation channel is connected to up to four electrical stimulation electrodes simultaneously.

[0037] Specifically, the transcranial electrical stimulation channel in this embodiment has a flexible site design, allowing for the simultaneous connection of four electrical stimulation electrodes, and the four transcranial electrical stimulation electrodes can be placed at any location on the scalp of the brain according to usage requirements.

[0038] In one implementation, the EEG acquisition module and the transcranial electrical stimulation device are designed as an integrated unit, and the circuits corresponding to the EEG acquisition module and the transcranial electrical stimulation device are independent of each other and do not interfere with each other.

[0039] Specifically, to improve ease of use, the EEG acquisition module and transcranial electrical stimulation (TCS) device in this embodiment are integrated into a single unit, resulting in a miniaturized multi-channel EEG + TCS integrated system and a universal bidirectional brain-computer interface research device. Secondly, to reduce the impact of TCS circuit noise on EEG signals, the circuits for EEG recording and TCS generation in this embodiment are independent and can be activated separately as needed. For example, only EEG acquisition can be activated, allowing it to be used as a conventional multi-channel EEG device without relying on the stimulation function; or only TCS can be activated, allowing it to be used as a standalone stimulation device without activating the EEG acquisition module; or both can be activated simultaneously for real-time recording of EEG changes after TCS is applied. In practical applications, the EEG acquisition channels allow for flexible placement, enabling users to flexibly change the EEG acquisition sites according to experimental needs.

[0040] Step S200: The EEG acquisition module acquires a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signal and transcranial alternating current stimulation signal.

[0041] Specifically, this embodiment records electroencephalogram (EEG) signals under transcranial alternating current stimulation (tACS). TACS is a non-invasive brain stimulation method that can be used to modulate brain neural activity at specific frequencies. However, simultaneously acquiring EEG signals during stimulation may introduce strong artifacts, which are several orders of magnitude larger than neural signals, severely affecting real-time brain state monitoring and hindering the research and application of closed-loop neuromodulation systems. In this embodiment, EEG signals containing tACS artifacts are defined as mixed signals. The tACS artifacts in the mixed signals are characterized by strong periodicity and high amplitude, and traditional filtering methods easily cause distortion of the EEG signals.

[0042] In one implementation, the EEG acquisition module allows for flexible electrode placement, covering acquisition channels corresponding to the frontal, parietal, temporal, and occipital lobes as needed. Multiple channels can cover multiple brain regions, capturing functional connectivity between brain regions and enabling a more comprehensive analysis of the neural effects of stimuli.

[0043] For example, the EEG instrument in this embodiment is designed to use low-density EEG recording channels, such as an 8-channel design, which can meet the needs of most research. The signal acquisition chip of the EEG instrument uses a high-performance biosensor chip with a signal sampling rate of up to 16000Hz. The EEG acquisition channel design is flexible and can cover the acquisition channels corresponding to the frontal lobe, parietal lobe, temporal lobe, and occipital lobe as needed.

[0044] Step S300: Input the mixed signal into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signal and obtain a reconstructed clean EEG signal.

[0045] Specifically, this embodiment pre-constructs an artifact removal model that can accurately identify and remove stimulus-related components, i.e., remove artifacts generated by transcranial alternating current stimulation signals from mixed signals, thereby reconstructing clean EEG signals. The artifact removal model does not rely on stimulus reference waveforms or multi-channel data, and can effectively identify and remove tACS artifacts from single-channel mixed signals, preserving real neural activity waveforms, for use in closed-loop neuromodulation and brain-computer interface systems.

[0046] For example, a new deep learning algorithm can be developed based on existing datasets to obtain an effective algorithm for handling electrical stimulation artifacts. An artifact removal model can then be established based on this algorithm, along with corresponding software programs to achieve product-level functionality. The system corresponding to this embodiment will provide a reliable, stable, and advanced research tool for the field of closed-loop electrical stimulation neuromodulation, offering more flexible and efficient solutions for brain science research and clinical applications. For instance, it can be used for real-time brain state perception and feedback modulation in closed-loop tACS systems; or deployed on wearable brain-computer interface devices, neuromodulation instruments, and mobile terminals.

[0047] In one implementation, the artifact removal model, after being trained using multiple training samples, can handle various artifacts such as transcranial direct current stimulation (tDCS), transcranial alternating current stimulation (tACS), transcranial random noise (tRNS), and temporal coherent stimulation (TI). It can be applied to real-time brain state perception and feedback regulation in closed-loop tACS systems.

[0048] In one implementation, the mixed signal is input into an artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, resulting in a reconstructed clean EEG signal, including: The mixed signal is input into the preprocessing module to obtain the preprocessed mixed signal; The preprocessed mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, thereby obtaining the reconstructed clean EEG signal.

[0049] Specifically, the preprocessing module performs initial purification of the mixed signal to reduce noise other than tACS artifacts, while optimizing signal quality to provide better input for the subsequent artifact removal model. After preprocessing, the main interference in the mixed signal is reduced to only tACS artifacts, which are then input into the artifact removal model. The artifact removal model utilizes the characteristic differences between tACS artifacts and EEG signals to separate and remove tACS artifacts from the mixed signal, ultimately reconstructing a clean EEG signal.

[0050] In one implementation, the preprocessing module is used to perform one or more of the following operations: Power supply interference in the mixed signal is removed by using a Butterworth notch filter; The low-frequency drift of the mixed signal is removed by using a high-pass filter; Remove unstable signal segments of preset duration at the beginning and end of the mixed signal; The mixed signal is sliced ​​using a sliding window in units of preset duration to obtain several signal segments. The step size of each window sliding forward is half the length of the window. Each of the aforementioned signal segments is normalized to a specified signal amplitude range.

[0051] Specifically, the preprocessing module performs a series of targeted operations to initially clean and standardize the input signal, providing high-quality input for the subsequent artifact removal model. These operations can be used individually or in combination, and their main purpose is to reduce interference, standardize the format, and highlight the effective signal.

[0052] Butterworth notch filters are used to remove power supply interference from mixed signals, preventing power supply noise from being misidentified as tACS artifacts or EEG features by subsequent models.

[0053] High-pass filters are used to remove low-frequency drift in mixed signals, preventing low-frequency drift from causing the overall signal to rise or fall, ensuring the baseline stability of EEG signals, and reducing the interference of drift on signal amplitude analysis.

[0054] Signal trimming is used to remove unstable signal segments of a preset duration at the beginning and end of a mixed signal. The purpose is to eliminate the worst quality parts at both ends of the signal, retain the signal during the middle stable period, and improve the reliability of subsequent analysis.

[0055] Signal segmentation processing is used to slice the mixed signal into discrete signal segments in units of preset duration. Furthermore, each window slides forward by a step size equal to half the window length, and overlapping steps prevent signal features from being truncated at the window edges.

[0056] Normalization is used to normalize each signal segment to a specified signal amplitude range in order to improve the generalization ability of the model.

[0057] By combining these operations, the preprocessing module can significantly improve the quality of the input signal, providing a data foundation for the artifact removal model to accurately separate tACS artifacts and EEG signals, and ensuring the reliability of the subsequently obtained clean EEG signals.

[0058] For example, in practical applications, the preprocessing module performs the following steps on the input signal: using a Butterworth notch filter (Q=10, 50Hz) to remove power supply interference; using a high-pass filter (0.5Hz) to remove low-frequency drift; removing segments at the beginning and end that may contain transition artifacts for 2 seconds each; slicing the signal in 2-second increments (50% overlap); and performing Min-Max Normalization on all segments to normalize them to the [0,1] signal amplitude range.

[0059] In one implementation, the artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

[0060] Specifically, the preprocessed mixed signal consists of several signal segments obtained after filtering, cropping, segmentation, and normalization. Each signal segment is then input into the artifact removal model. The artifact removal model is a deep learning model based on an encoder-decoder structure. Its working principle involves compressing signal features through dimensionality reduction operations (e.g., downsampling) by the encoder, followed by dimensionality increase operations (e.g., upsampling) by the decoder to reconstruct the signal. Ultimately, this achieves the goal of automatically removing tACS artifacts from the mixed signal and reconstructing a clean EEG signal. Specifically, the encoder module is responsible for extracting key features and compressing data from the mixed signal. Through dimensionality reduction, it effectively distinguishes the features of the EEG signal and tACS artifacts, while filtering out redundant information in the mixed signal. After these two key steps of dimensionality reduction and feature processing, the target feature data output by the encoder module belongs to the low-dimensional features of the mixed signal. These low-dimensional features retain the essential features of the EEG signal while significantly weakening the interference features of tACS artifacts, laying a good foundation for subsequent signal reconstruction. The decoder module's task is to reconstruct a clean EEG signal with the same length as the input signal based on the target feature data extracted by the encoder. In this process, the decoder's dimensionality increase operation plays a crucial role: it not only restores the low-dimensional features output by the encoder to the original signal's dimensions, but also further removes residual tACS artifacts, ensuring that the final output is a pure EEG signal.

[0061] For example, in the model inference stage, signal acquisition is first performed to obtain a real-time or offline EEG+tACS mixed signal. The original mixed signal is preprocessed to obtain several signal segments, and each signal segment is input into a trained artifact removal model. The artifact removal model outputs an artifact-removed EEG signal with an amplitude range of [0,1]. The artifact-removed EEG signals corresponding to each signal segment are concatenated into a continuous signal stream, thus obtaining the reconstructed clean EEG signal. Further, after obtaining the reconstructed clean EEG signal, post-processing can be performed, specifically including: for each signal segment in the reconstructed clean EEG signal, using the pre-stimulation EEG segment as a reference segment, and performing inverse normalization on the signal segment based on the reference segment; or, uniformly applying Z-score normalization (a data normalization method) for brain-computer interface (BCI) feature extraction or brain state classification analysis.

[0062] In one implementation, the encoder module includes several cascaded convolutional modules. Each convolutional module in the encoder module includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, and the number of channels in each one-dimensional convolutional layer decreases sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

[0063] Specifically, the encoder module's function is to extract features and compress information from the input mixed signal. It progressively simplifies the signal dimensions, retains key features, and filters out redundant information. The overall structure of the encoder module consists of multiple convolutional modules connected sequentially, with the output of one convolutional module serving as the input to the next. Each convolutional module comprises a cascaded structure of one-dimensional convolutional layers, normalization layers, and activation layers. The one-dimensional convolutional layers are designed for the time-series characteristics of EEG signals, extracting local temporal features through convolutional kernels. The normalization layers standardize the output of the convolutional layers, preventing signal values ​​from being too large or too small, which could lead to model training instability and accelerate model convergence. The activation layers introduce nonlinear transformations, enabling the model to learn complex nonlinear relationships. The number of channels in each one-dimensional convolutional layer of the encoder module decreases progressively: the initial number of channels is higher to capture various detailed features of the signal; as the layers deepen, the number of channels decreases to compress information, retain core features, and prepare for subsequent decoding and reconstruction.

[0064] The decoder module reconstructs a clean EEG signal based on the target feature data output by the encoder. It needs to progressively restore the signal's dimensionality and details, forming a symmetrical logic with the encoder module from compression to reconstruction. The overall structure of the decoder module corresponds to that of the encoder module, consisting of multiple convolutional modules connected sequentially, but with the opposite function: reconstructing the signal from compressed features. Each convolutional module in the decoder module contains a cascaded structure of convolutional layers, normalization layers, and activation layers. The convolutional layers can be one-dimensional or one-dimensional transposed convolutional layers (also called deconvolutional layers). The main role of the one-dimensional transposed convolutional layer in the decoder module is to perform dimensionality enhancement: mapping and amplifying the low-dimensional features compressed by the encoder to higher dimensions, which is crucial for signal reconstruction. The normalization and activation layers serve the same purpose as in the encoder module: stabilizing training and introducing nonlinear transformations. The number of channels in each convolutional layer in the decoder module increases sequentially, forming a symmetry with the decrease in the number of channels in the encoder module. This allows the compressed low-dimensional features to be gradually restored to a dimension that matches the original signal after passing through the decoder module. By increasing the number of channels, the detailed information of the signal is supplemented, and finally a complete and clean EEG signal is reconstructed.

[0065] For example, this embodiment uses a symmetric one-dimensional convolutional autoencoder neural network architecture to establish a deep learning algorithm for removing transcranial alternating current stimulation (tACS) artifacts from single-channel electroencephalogram (EEG) signals, thereby obtaining an artifact removal model.

[0066] The core functionalities of the artifact removal model include: Using convolutional neural networks to extract temporal features: such as Figure 2 As shown, one-dimensional convolutional layers (Conv1D) and one-dimensional transposed convolutional layers (Conv1DTranspose) are used for encoding and decoding; the network can automatically identify stimulus artifact structures in the spectrum and time domain, thereby effectively separating real EEG signals from stimulus components.

[0067] Adaptable to single-channel or multi-channel EEG input: The network is designed as a channel-independent type, suitable for real-time processing; it can be applied to multi-channel scenarios, and has strong versatility and scalability by processing each channel in parallel.

[0068] The core modules of the artifact removal model include: The encoder module specifically includes: 5 one-dimensional convolutional layers with channels of 64, 32, 16, 8, and 4 respectively; each convolutional layer is followed by a normalization layer and an activation function; and a regularization strategy layer is added at the end of the encoding. Figure 2 In this context, k3n64s1 indicates that the kernel size k is 3, the number of channels n is 64, and the stride s is 1.

[0069] Specifically, the encoder module uses five one-dimensional convolutional layers for two reasons: firstly, to avoid too many layers, which would be unsuitable for real-time signal processing and prevent slow processing speeds and overfitting; and secondly, to avoid too few layers, which would result in underfitting and incomplete signal feature extraction. Secondly, the decreasing number of channels is to continuously compress data from the previous layer and extract the most essential signal features. Furthermore, the stride can be set to 1 or 2 to achieve progressive downsampling. A stride of 1 does not change the feature length; a stride of 2 reduces the feature length, allowing for the extraction of higher-level abstract features. A batch normalization (BN) layer is used to accelerate training, stabilize the model, and prevent gradient explosion or vanishing. An activation function (LeakyReLU activation function) is used to avoid the dead neuron problem of ReLU. A dropout regularization layer is added at the end of the encoder to enhance generalization ability and prevent overfitting; the dropout rate can be limited to 0.2.

[0070] The decoder module specifically includes: several Conv1D layers, a one-dimensional transposed convolutional layer for dimension increase operation, and an output layer; the number of channels is 4, 8, 16, 32, and 64 respectively; each layer is equipped with a normalization layer and an activation function.

[0071] Specifically, the decoder module is the reverse process of the encoder module. In the decoder module, a one-dimensional convolutional layer is used to restore temporal resolution; a one-dimensional transposed convolutional layer is used to increase the dimensionality to restore temporal resolution; the output layer is activated using the sigmoid function, with an output amplitude range of [0,1]. Finally, the reconstructed clean EEG signal is obtained through the output layer.

[0072] In one implementation, the artifact removal model is pre-trained, and the method for obtaining training samples includes: Treat a clean EEG signal and its corresponding actual mixed signal as a signal pair; The preprocessing module performs signal processing on the signal pairs to obtain several training sample pairs; wherein, during the training phase, the preprocessing module is also used to perform time alignment on the signal pairs; each training sample pair includes a preprocessed clean signal segment segmented from the clean EEG signal, and a corresponding preprocessed mixed signal segment segmented from the actual mixed signal.

[0073] Specifically, during the training phase, data preparation and preprocessing are required. The data preparation process includes: using a publicly available dataset containing known clean EEG signals and their paired actual mixed signals (EEG+tACS); and preserving valid EEG signal channels. Taking a clean EEG signal and its corresponding actual mixed signal as an example, the preprocessing process for this signal pair includes: removing power supply interference from both the clean EEG signal and the actual mixed signal using a Butterworth notch filter; removing low-frequency drift from both the clean EEG signal and the actual mixed signal using a high-pass filter; time-aligning the clean EEG signal and the actual mixed signal; removing unstable signal segments of preset duration at the beginning and end of both the clean EEG signal and the actual mixed signal; performing sliding window slicing on both the clean EEG signal and the actual mixed signal in preset duration units to obtain several signal segments, with each window sliding forward by half the window length; and normalizing each signal segment to a specified signal amplitude range to obtain several training sample pairs. Each training sample pair includes a clean signal segment generated from a clean EEG signal and a mixed signal segment generated from an actual mixed signal. The clean signal segment and the mixed signal segment are time-aligned and have high signal quality after preprocessing, making them suitable for model training.

[0074] For example, the preprocessing module performs the following steps on the clean EEG signal and the actual mixed signal from the training phase: It uses a Butterworth notch filter (Q=10, 50Hz) to remove power supply interference; it uses a high-pass filter (0.5Hz) to remove low-frequency drift; it aligns the clean EEG signal and the actual mixed signal in time; it removes segments with potential transition artifacts at the beginning and end (2 seconds each); it slices the signal in 2-second sliding window units (50% overlap); and it performs Min-Max Normalization on all segments to normalize them to the [0,1] signal amplitude range. Finally, a total of 2196 pairs of input-output samples are obtained.

[0075] In one implementation, the training method for the artifact removal model includes: Select one of the unprocessed training sample pairs and input the mixed signal segment from the training sample pair into the initial artifact removal model; The initial artifact removal model is used to remove artifacts from the mixed signal segment, and the corresponding reconstructed clean signal segment is output. The clean signal segment in the training sample pair is used as a supervision signal, and the parameters of the initial artifact removal model are updated according to the supervision signal and the reconstructed clean signal segment. Determine whether the training termination condition is met. If not, continue to select an unprocessed training sample pair until the training termination condition is met, and obtain the trained artifact removal model.

[0076] During the training phase, this embodiment models the problem as a supervised learning task, defining the incompletely trained model as the initial artifact removal model. Specifically, mixed signal segments are used as input data for the initial artifact removal model, and clean signal segments time-aligned with them are used as training targets. The initial artifact removal model is trained end-to-end, enabling it to automatically learn a nonlinear mapping for recovering clean EEG signals from artifact-laden signals. The initial artifact removal model is iteratively trained using a large number of training samples until the training termination condition is met, thereby improving the accuracy of the model in reconstructing clean signal segments.

[0077] For example, in this embodiment, the loss function of the artifact removal model is established based on the training objective. The training objective is to minimize the mean squared error (MSE) between the model's output layer and the clean signal segment used as the supervision signal.

[0078] The mean squared error (MSE) loss function is related to the original underlying neural signal. The definition is as follows: ; in, This represents the total number of samples in the training set; The first part of the real clean EEG training set One sample, The first step in reconstructing brainwave signals One sample.

[0079] During the training phase, the training set, validation set, and test set are divided into three parts based on an 8:1:1 ratio, and the Adam optimizer is used for training.

[0080] In one implementation, the training termination condition includes: The root mean square difference between the monitoring signal and the reconstructed clean signal segment is minimized.

[0081] Specifically, in supervised training of the artifact removal model, the supervision signal refers to the real, clean EEG signal without tACS artifacts. Because the model's goal is to output a clean EEG signal, a supervision signal is needed to determine whether the model's output is acceptable. Reconstructing a clean signal segment refers to the clean EEG signal output by the decoder module during model training, based on the features extracted by the encoder, and to be verified. Root mean square error (RMSE) is a key metric for measuring the similarity between two signals; the smaller the RMSE value, the smaller the difference between the two signals, and the closer the artifact removal model's output is to the real, clean signal.

[0082] In one implementation, at least one metric, such as relative root mean square error (RRMSE) or correlation coefficient (e.g., Pearson correlation coefficient), can be used to evaluate the performance of the artifact removal model.

[0083] Specifically, the performance of the artifact removal model can be evaluated based on the relative root mean square error (RMSE). The RMSE quantifies the difference between the reconstructed clean signal and the monitoring signal; a smaller value indicates that the model output is closer to the real signal. Alternatively, the performance can be evaluated based on the correlation coefficient. The correlation coefficient quantifies the similarity between the reconstructed signal and the monitoring signal; a value closer to 1 indicates more consistent waveform and frequency characteristics. This embodiment uses either a dual-index or single-index approach to accurately evaluate model performance. In practical applications, when both the monitoring signal and the reconstructed clean signal segment are time-domain signals, the RMSE and / or correlation coefficient are calculated using their time-domain amplitudes. When both are frequency-domain signals, the power spectral density of the monitoring signal and the reconstructed clean signal segment is calculated using Welch transform, and the RMSE and / or correlation coefficient are calculated based on these two power spectral densities.

[0084] For example, to evaluate the artifact removal capability and reconstruction accuracy of the artifact removal model, the following evaluation method is set: Relative Root Mean Square Error (RRMSE): Calculation of clean EEG signals and reconstructing brain signals The smaller the RRMSE value, the smaller the signal reconstruction error and the better the effect.

[0085] Calculating RRMSE in the time domain: RRMSE is calculated directly using the amplitude in the time domain. The relative root mean square error at the t-th iteration. The calculation formula is as follows: ; Calculating RRMSE in the frequency domain requires first calculating the power spectral density (PSD) of the two signals using Welch transform, and then calculating the RRMSE. The relative root mean square error at the f-th iteration... The calculation formula is as follows: ; In the formula, RMS represents the root mean square value.

[0086] Calculating the correlation coefficient (CC) in the time domain: CC is calculated directly using the amplitude in the time domain. The correlation coefficient at the t-th iteration. The calculation formula is as follows: ; In the formula, Represents covariance; Indicates variance.

[0087] Calculating the correlation coefficient (CC) in the frequency domain requires first calculating the power spectral density (PSD) of the two signals using Welch transform, and then calculating the CC. The correlation coefficient at the f-th iteration... The calculation formula is as follows: .

[0088] Experiments show that the method in this embodiment achieves a correlation coefficient of 0.639 in the time domain and 0.880 in the frequency domain on the test set. It performs best under high-frequency artifacts (e.g., 40Hz), with a correlation coefficient as high as 0.955, significantly outperforming comparative algorithms such as the Smoothed Moving Average (SMA) and Adaptive Filtering (AF). This demonstrates that the method in this embodiment performs excellently under high-frequency stimuli or nonlinear artifact scenarios.

[0089] The advantages of this invention are: In terms of equipment and scenarios: the equipment is miniaturized, wearable, and integrated. It covers three application scenarios: EEG signal acquisition, transcranial electrical stimulation, and EEG signal + transcranial electrical stimulation.

[0090] In terms of electrical stimulation functions: Electrical stimulation supports multiple stimulation waveforms (tDCS, tACS, tRNS, TI) and allows for customizable electrical stimulation parameters; among them, TI stimulation uses two high-frequency AC stimulation waveforms to precisely target the deep brain in a non-invasive manner, achieving deep brain stimulation comparable to implantable deep brain stimulation (DBS) technology, achieving the effect of invasive, precise target brain stimulation in a non-invasive manner; electrical stimulation supports temporal coherent stimulation (TI) with a stimulation frequency of 1KHz-2KHz.

[0091] Regarding signal acquisition: When using TI stimulation, the EEG amplifier can acquire data at a sampling rate of 4000Hz with 8 channels; the 8-channel, EEG+2-loop electrical stimulation can realize the integrated circuit design of sensing and electrical stimulation, and the sampling rate of the EEG amplifier can be as high as 8 channels at 4KHz, acquiring EEG and stimulation waveforms.

[0092] Regarding algorithms and artifact removal: Based on a deep convolutional autoencoder, a deep learning algorithm for artifact removal of transcranial alternating current stimulation (tACS) is established; it can also remove artifacts of transcranial direct current stimulation (tDCS), transcranial random noise stimulation (tRNS), and temporal coherent stimulation (TI), allowing researchers to see the EEG signals during electrical stimulation.

[0093] Based on the above embodiments, the present invention also provides a mixed-signal artifact removal system, such as... Figure 3 As shown, the system includes: Transcranial electrical stimulation device 01 is used to provide transcranial alternating current stimulation to the user; The EEG acquisition module 02 is used to acquire a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signals and transcranial alternating current stimulation signals. The artifact removal module 03 is used to input the mixed signal into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals and obtain a reconstructed clean EEG signal.

[0094] In one implementation, the EEG acquisition module and the transcranial electrical stimulation device are designed as an integrated unit, and the circuits corresponding to the EEG acquisition module and the transcranial electrical stimulation device are independent of each other and do not interfere with each other.

[0095] In one implementation, the artifact removal module includes: The preprocessing module is used to preprocess the mixed signal to obtain a preprocessed mixed signal; An artifact removal model is used to remove artifacts generated by transcranial alternating current stimulation signals from the preprocessed mixed signal to obtain the reconstructed clean EEG signal.

[0096] In one implementation, the preprocessing module is used to perform one or more of the following operations: Power supply interference in the input signal is removed by using a Butterworth notch filter; Low-frequency drift of the input signal is removed by using a high-pass filter; Remove unstable signal segments of preset duration at the beginning and end of the input signal; The input signal is sliced ​​using a sliding window in units of preset duration to obtain several signal segments. The step size of each window sliding forward is half the length of the window. Each of the aforementioned signal segments is normalized to a specified signal amplitude range.

[0097] In one implementation, the artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

[0098] In one implementation, the encoder module includes several cascaded convolutional modules. Each convolutional module in the encoder module includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, and the number of channels in each one-dimensional convolutional layer decreases sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

[0099] In one implementation, the artifact removal model is pre-trained, and the system further includes a training sample acquisition module, which is used to: Treat a clean EEG signal and its corresponding actual mixed signal as a signal pair; The preprocessing module performs signal processing on the signal pairs to obtain several training sample pairs; wherein, during the training phase, the preprocessing module is also used to perform time alignment on the signal pairs; each training sample pair includes a preprocessed clean signal segment segmented from the clean EEG signal, and a corresponding preprocessed mixed signal segment segmented from the actual mixed signal.

[0100] In one implementation, the system further includes a training module, which is used for: Select one of the unprocessed training sample pairs and input the mixed signal segment from the training sample pair into the initial artifact removal model; The initial artifact removal model is used to remove artifacts from the mixed signal segment, and the corresponding reconstructed clean signal segment is output. The clean signal segment in the training sample pair is used as a supervision signal, and the parameters of the initial artifact removal model are updated according to the supervision signal and the reconstructed clean signal segment. Determine whether the training termination condition is met. If not, continue to select an unprocessed training sample pair until the training termination condition is met, and obtain the trained artifact removal model.

[0101] In one implementation, the training termination condition includes: The root mean square difference between the monitoring signal and the reconstructed clean signal segment is minimized.

[0102] Based on the above embodiments, the present invention also provides a terminal, the principle block diagram of which can be as follows: Figure 4 As shown, the terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for removing artifacts from mixed signals. The display screen can be a liquid crystal display (LCD) or an e-ink display.

[0103] Those skilled in the art will understand that Figure 4 The schematic diagram shown is merely a partial structural diagram related to the present invention and does not constitute a limitation on the terminal to which the present invention is applied. A specific terminal may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0104] In one implementation, the terminal's memory stores one or more programs, and these programs are configured to be executed by one or more processors, and the programs contain instructions for performing a method for removing artifacts from mixed signals.

[0105] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0106] In summary, this invention discloses a method, system, terminal, and storage medium for artifact removal from mixed signals, relating to the field of bioelectrical signal processing technology. The method includes: applying transcranial alternating current (TCD) stimulation to a user via a transcranial electrical stimulation device; acquiring a mixed signal generated by the user based on the TCD stimulation via an EEG acquisition module; the mixed signal being a composite signal generated from EEG signals and TCD stimulation signals; and inputting the mixed signal into an artifact removal model to remove artifacts generated by the TCD stimulation signals, resulting in a reconstructed clean EEG signal. This invention employs a deep learning algorithm to establish an artifact removal model, which removes TCD stimulation artifacts from the mixed signal. The deep learning algorithm can autonomously learn artifact features through various training samples, eliminating the need for manual definition of artifact features and adapting to various application scenarios.

[0107] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for artifact removal in mixed signals, characterized in that, The method includes: Transcranial alternating current stimulation is applied to the user using a transcranial electrical stimulation device; The EEG acquisition module acquires a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signals and transcranial alternating current stimulation signals. The mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, thereby obtaining a reconstructed clean EEG signal.

2. The artifact removal method for mixed signals according to claim 1, characterized in that, The EEG acquisition module and the transcranial electrical stimulation device are designed as an integrated unit, and the circuits corresponding to the EEG acquisition module and the transcranial electrical stimulation device are independent of each other and do not interfere with each other.

3. The artifact removal method for mixed signals according to claim 1, characterized in that, The mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, resulting in a reconstructed clean EEG signal, including: The mixed signal is input into the preprocessing module to obtain the preprocessed mixed signal; The preprocessed mixed signal is input into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals, thereby obtaining the reconstructed clean EEG signal.

4. The artifact removal method for mixed signals according to claim 3, characterized in that, The preprocessing module is used to perform one or more of the following operations: Power supply interference in the input signal is removed by using a Butterworth notch filter; Low-frequency drift of the input signal is removed by using a high-pass filter; Remove unstable signal segments of preset duration at the beginning and end of the input signal; The input signal is sliced ​​using a sliding window in units of preset duration to obtain several signal segments. The step size of each window sliding forward is half the length of the window. Each of the aforementioned signal segments is normalized to a specified signal amplitude range.

5. The artifact removal method for mixed signals according to claim 3, characterized in that, The artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

6. The artifact removal method for mixed signals according to claim 5, characterized in that, The encoder module includes several cascaded convolutional modules. Each convolutional module in the encoder module includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, and the number of channels in each one-dimensional convolutional layer decreases sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

7. The artifact removal method for mixed signals according to claim 3, characterized in that, The artifact removal model is pre-trained, and the methods for obtaining training samples include: Treat a clean EEG signal and its corresponding actual mixed signal as a signal pair; The preprocessing module performs signal processing on the signal pairs to obtain several training sample pairs; wherein, during the training phase, the preprocessing module is also used to perform time alignment on the signal pairs; each training sample pair includes a preprocessed clean signal segment segmented from the clean EEG signal, and a corresponding preprocessed mixed signal segment segmented from the actual mixed signal.

8. The artifact removal method for mixed signals according to claim 7, characterized in that, The training method for the artifact removal model includes: Select one of the unprocessed training sample pairs and input the mixed signal segment from the training sample pair into the initial artifact removal model; The initial artifact removal model is used to remove artifacts from the mixed signal segment, and the corresponding reconstructed clean signal segment is output. The clean signal segment in the training sample pair is used as a supervision signal, and the parameters of the initial artifact removal model are updated according to the supervision signal and the reconstructed clean signal segment. Determine whether the training termination condition is met. If not, continue to select an unprocessed training sample pair until the training termination condition is met, and obtain the trained artifact removal model.

9. The artifact removal method for mixed signals according to claim 8, characterized in that, The training termination conditions include: The root mean square difference between the monitoring signal and the reconstructed clean signal segment is minimized.

10. A mixed-signal artifact removal system, characterized in that, The system includes: A transcranial electrical stimulation device is used to provide transcranial alternating current stimulation to a user. The EEG acquisition module is used to acquire a mixed signal generated by the user based on transcranial alternating current stimulation; the mixed signal is a composite signal generated based on EEG signals and transcranial alternating current stimulation signals. The artifact removal module is used to input the mixed signal into the artifact removal model to remove artifacts generated by transcranial alternating current stimulation signals and obtain a reconstructed clean EEG signal.

11. The artifact removal system for mixed signals according to claim 10, characterized in that, The artifact removal module includes: The preprocessing module is used to preprocess the mixed signal to obtain a preprocessed mixed signal; An artifact removal model is used to remove artifacts generated by transcranial alternating current stimulation signals from the preprocessed mixed signal to obtain the reconstructed clean EEG signal.

12. The artifact removal system for mixed signals according to claim 11, characterized in that, The artifact removal model includes: The encoder module is used to extract features from the preprocessed mixed signal through a dimensionality reduction operation to obtain target feature data; The decoder module is used to reconstruct the target feature data through a dimension-increasing operation to obtain the reconstructed clean EEG signal.

13. The artifact removal system for mixed signals according to claim 12, characterized in that, The encoder module includes several cascaded convolutional modules. Each convolutional module in the encoder module includes a cascaded one-dimensional convolutional layer, a normalization layer, and an activation layer, and the number of channels in each one-dimensional convolutional layer decreases sequentially. The decoder module includes several cascaded convolutional modules. Each convolutional module in the decoder module includes cascaded convolutional layers, normalization layers, and activation layers. The convolutional layers in the decoder module are one-dimensional convolutional layers or one-dimensional transposed convolutional layers, and the number of channels in each convolutional layer increases sequentially.

14. A terminal, characterized in that, The terminal includes a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for performing the artifact removal method for mixed signals as described in any one of claims 1 to 9; the processor is used to execute the programs.

15. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are applicable to be loaded and executed by a processor to implement the steps of the artifact removal method for mixed signals as described in any one of claims 1 to 9.

Images