A method and system for identifying a take-over load based on multi-channel electromyography

Abstract

The application discloses a kind of based on multi-channel electromyogram's takeover load identification method and system, belong to man-machine interaction field field.The present application is to solve the technical problems of insufficient utilization of multi-channel electromyogram information, insufficient description of takeover load time evolution and the influence of training process by uneven sample distribution, by pre-processing and time reconstruction of multi-channel electromyogram signal, time modeling, feature abstraction and adaptive feature fusion, and based on reinforcement learning training takeover load identification model.The present application can realize the dynamic identification and classification prediction of takeover load.

Description

Technical Field

[0001] This invention belongs to the field of human-computer interaction, specifically relating to a method and system for identifying control load based on multi-channel electromyography. Background Technology

[0002] In conditional automated driving systems, drivers need to take over during the transition between automated and manual driving. The takeover load is closely related to the safety of the takeover. Excessive takeover load can lead to delayed reactions and decreased operational stability, increasing the risk of takeover failure. Existing driver takeover load identification technologies typically collect information such as electrocardiogram (ECG), electroencephalogram (EEG), electrical conductance analysis (EDA), eye movements, heart rate variability, steering wheel angle, pedal operation, or electromyography (EMG), and use this information to construct multi-channel time-series data as model input. At the algorithmic level, early methods often employed traditional machine learning models such as support vector machines and random forests to classify manually extracted statistical features or time-frequency domain features. With the development of deep learning, convolutional neural networks, recurrent neural networks, long short-term memory networks (LSTM), and their variants have been increasingly used for driver load identification to enhance the ability to model the dynamic features of time-series signals.

[0003] Among existing related technologies, one invention patent proposes a method for assessing driving load based on a Hidden Markov Model (HMM). This method uses ECG as the primary input, extracts artificial features in the time, frequency, and nonlinear domains through a sliding time window, and employs an HMM for driving load classification. However, its observation end relies on artificial features, making it difficult to perform end-to-end deep representation learning on high-dimensional multi-channel raw time-series signals. Another invention patent proposes a method for personalized quantification of driver cognitive load through co-evolution, achieving continuous quantification of cognitive load through regression modeling and collaborative pseudo-labeling iteration. However, this method is mainly based on structured feature vectors and does not highlight the temporal dependence and channel contribution differences of multi-channel EMG during takeover. A third invention patent proposes a method, system, device, medium, and product for determining driver cognitive load, based on multimodal physiological data and convolutional neural networks. However, its temporal dependence is mainly modeled indirectly through convolutional receptive fields, failing to adequately characterize the differences in contribution of different channels over time.

[0004] Furthermore, the invention patent for a real-time driver workload identification method based on multimodal data fusion and its mapping method to behavioral and physiological features proposes using multimodal features such as driving behavior, physiology, and eye movement as inputs, and employing GBDT, neural networks, KNN, random forests, and logistic regression for stacked integration. However, its overall framework belongs to feature engineering and model-level fusion, and temporal information mainly relies on time window featureization for indirect representation. The invention patent for a real-time driver workload detection method and system based on lightweight equipment proposes collecting physiological detection data using lightweight devices such as pressure-sensing cushions and telemetry eye trackers, and outputting workload results through clustering, ensemble learning, or supervised learning. However, its fusion method leans more towards model integration or voting, making it difficult to characterize the dynamic changes in the contribution of multi-channel signals at different times with fine granularity.

[0005] In summary, existing technologies for driver takeover load identification still have the following problems: First, multi-channel modeling methods are relatively crude. Some schemes rely only on single-channel physiological signals or single behavioral features. Even when using multi-channel or multi-source input, they mostly remain at the level of feature splicing, statistical weighting, or model-level fusion, lacking adaptive modeling of the structural relationships, temporal characteristics, and contribution differences of different channels. This leads to the accumulation of redundant information and difficulty in highlighting key information. Second, existing methods mainly rely on static supervised classification or regression mapping, which is insufficient in characterizing the continuity, stages, and long-term dependencies of load states during takeover. It is difficult to optimize the identification strategy at the sequence level, and the stability of the identification results is limited under complex takeover conditions. Third, actual takeover load samples often have problems such as unbalanced class distribution, insufficient minority class samples, and high difficulty in distinguishing boundary samples. Existing training methods mostly use random sampling or equal weight updates, which fail to fully distinguish the importance of samples. This can easily lead to the model biasing towards the majority class or low-difficulty load categories, reducing the ability to identify key risk load levels and easily confused load states. Summary of the Invention

[0006] The purpose of this invention is to propose a method and system for identifying dominance load based on multi-channel electromyography (EMG) to solve the technical problems of insufficient utilization of multi-channel EMG information, inadequate characterization of the temporal evolution of dominance load, and susceptibility of the training process to uneven sample distribution, thereby achieving stable, accurate, and dynamic identification of dominance load.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] A method for identifying control load based on multichannel electromyography includes the following steps: The multi-channel electromyography signals during the takeover process are preprocessed and reconstructed to obtain multi-channel time-series samples, and the load labels corresponding to the multi-channel time-series samples are determined. The multi-channel time-series samples are subjected to time-series modeling, feature abstraction, and adaptive feature fusion to obtain multi-channel time-series features; Using the multi-channel time-series features as state input, the takeover load level discrimination as the identification action, and training an offline reinforcement learning-based takeover load identification model based on the load labels; The trained overhaul load identification model is identified based on an independent test set to obtain the overhaul load identification result, and the overhaul load identification model is saved.

[0009] Furthermore, the multi-channel electromyography signals during the takeover process are preprocessed and reconstructed to obtain multi-channel time-series samples, and the load labels corresponding to the multi-channel time-series samples are determined, including: Based on the muscle channel configuration, the multi-channel electromyographic signals during the takeover process are aligned to obtain aligned electromyographic signals; The aligned electromyographic signals are preprocessed to obtain preprocessed electromyographic signals; The preprocessed electromyographic signals are reconstructed temporally and associated with load labels to obtain multi-channel temporal samples and their corresponding load labels.

[0010] Further, the aligned electromyographic signals are preprocessed to obtain preprocessed electromyographic signals, including: The aligned electromyographic signal is filtered to obtain a denoised electromyographic signal; The amplitude of the denoised electromyographic signal is normalized to obtain a preprocessed electromyographic signal.

[0011] Further, the preprocessed electromyographic signals are temporally reconstructed and associated with load labels to obtain multi-channel temporal samples and their corresponding load labels, including: The temporal signals of each electromyographic channel in the preprocessed electromyographic signal are sampled at equal intervals according to a preset step size to obtain the channel reconstruction sequence. The channel reconstruction sequences are combined according to the channel correspondence to obtain multi-channel time series samples; The workload self-assessment score generated after the takeover is completed will be used as the load label corresponding to the multi-channel time series sample.

[0012] Furthermore, the multi-channel time-series samples are subjected to time-series modeling, feature abstraction, and adaptive feature fusion to obtain multi-channel time-series features, including: Temporal dependency extraction is performed on the multi-channel time series samples to obtain temporal modeling features; The time-series modeling features are abstracted to obtain abstract time-series features; Adaptive feature fusion is performed on the abstract temporal features to obtain multi-channel temporal features.

[0013] Furthermore, temporal dependency extraction is performed on the multi-channel time-series samples to obtain temporal modeling features, including: Temporal dependency extraction is performed on the multi-channel temporal samples to obtain candidate temporal features; The candidate temporal features are integrated temporally to obtain temporal modeling features.

[0014] Furthermore, temporal dependency extraction is performed on the multi-channel time-series samples to obtain temporal modeling features, including: The multi-channel time series samples are subjected to forward and reverse time series modeling to obtain bidirectional time series features. The bidirectional time series features are modeled unidirectionally to obtain the time series modeling features.

[0015] Furthermore, adaptive feature fusion is performed on the abstract temporal features to obtain multi-channel temporal features, including: Based on the abstract temporal characteristics, determine the channel fusion weights or channel relationship data; The abstract time series features are fused across channels based on the channel fusion weights or the channel relationship data to obtain multi-channel time series features.

[0016] Furthermore, adaptive feature fusion is performed on the abstract temporal features to obtain multi-channel temporal features, including: Determine channel attention weights or channel gating weights based on the aforementioned abstract temporal features; The abstract temporal features are modulated by channel based on the channel attention weights or the channel gating weights to obtain multi-channel temporal features.

[0017] Furthermore, using the multi-channel time-series features as state input, and takeover load level discrimination as the identification action, a takeover load identification model based on offline reinforcement learning is trained based on the load labels, including: The multi-channel temporal features are input into the reinforcement learning interactive environment to obtain the environment state; Based on the environmental conditions, output load identification actions, and generate reward data based on the load identification actions and the load tags; The takeover load identification model is trained based on the environmental state, the load identification action, the reward data, and the next environmental state.

[0018] Further, based on the environmental state, a load identification action is output, and reward data is generated according to the load identification action and the load label, including: Based on the aforementioned environmental conditions, output load level action, load scoring action, or risk estimation action; Reward data is generated based on the correspondence between the load level action, the load scoring action, or the risk estimation action and the load label.

[0019] Further, training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Based on the environmental state, the load identification action, the reward data, and the next environmental state, generate action value update data, strategy update data, or actor commentator update data; The network parameters of the takeover load identification model are updated based on the action value update data, the strategy update data, or the actor commentator update data.

[0020] Further, training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Experience samples are generated based on the environmental state, the load identification action, the reward data, and the next environmental state; Target experience samples are sampled from the experience replay pool based on the importance data of the experience samples. The takeover load identification model is trained based on the target experience samples.

[0021] Further, training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Experience samples are generated based on the environmental state, the load identification action, the reward data, and the next environmental state; The importance data of the empirical samples are determined based on time difference error, sample uncertainty, sample difficulty, or misclassification frequency. Based on the importance data, target experience samples are weighted and sampled from the experience replay pool, and a takeover load identification model is trained based on the target experience samples.

[0022] Further, training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Based on the environmental state, the load identification action, the reward data, and the next environmental state, predictive action value data and target action value data are generated. The online network parameters of the takeover load identification model are updated based on the predicted action value data and the target action value data. The target network parameters of the takeover load identification model are softly updated based on the updated online network parameters.

[0023] Furthermore, the trained overhaul load identification model is used to identify overhaul loads based on an independent test set to obtain overhaul load identification results, and the overhaul load identification model is saved, including: Preprocessing and time-series reconstruction of multi-channel electromyography signals from an independent test set yields test time-series samples. The test time series samples are subjected to time series modeling, feature abstraction, and adaptive feature fusion to obtain test time series features; The test timing features are input into the trained overhaul load identification model to obtain the overhaul load identification result, and the trained overhaul load identification model is saved.

[0024] A multi-channel electromyography-based takeover load identification system includes: The sample construction module is used to preprocess and reconstruct the multi-channel electromyography signals during the takeover process to obtain multi-channel time-series samples and determine the load labels corresponding to the multi-channel time-series samples. The feature extraction module is used to perform time series modeling, feature abstraction, and adaptive feature fusion on the multi-channel time series samples to obtain multi-channel time series features; The model training module is used to take the multi-channel time series features as state input, take the takeover load level discrimination as the recognition action, and train an offline reinforcement learning-based takeover load recognition model based on the load labels. The model evaluation module is used to identify the trained overhaul load identification model based on an independent test set, obtain the overhaul load identification result, and save the overhaul load identification model.

[0025] The present invention has achieved the following beneficial effects.

[0026] 1. This invention reduces the impact of noise interference, amplitude differences, and individual differences on subsequent modeling by preprocessing and reconstructing multi-channel electromyographic signals during the takeover process, and expands the multi-channel time-series samples, making the input data more suitable for the time-series feature learning of the takeover load identification model.

[0027] 2. This invention uses multi-channel electromyography signals as input for takeover load identification, and obtains physiological load information of multiple muscle channels during the takeover process through channel alignment and joint modeling. This can improve the completeness of the representation of changes in takeover load state, avoid the problem that single-channel schemes are easily affected by individual differences and noise disturbances, and improve the shortcomings of simple splicing or shallow fusion in that it is difficult to characterize the intrinsic correlation of multi-channel signals.

[0028] 3. By performing time-series modeling and feature abstraction on multi-channel time-series samples, this invention can extract the time-dependent features of electromyographic signals from continuous takeover processes, thereby improving the ability to characterize the continuity, stages, and dynamic changes of takeover load. In an optional embodiment, a composite time-series modeling structure combining bidirectional long short-term memory networks and long short-term memory networks is adopted, which can enhance the ability to characterize the dependencies before and after the takeover process and the staged evolution of load state.

[0029] 4. This invention learns the contribution of different electromyographic channels in the identification of the control load through adaptive feature fusion, and can dynamically adjust the features of each channel to overcome the problems of redundant information accumulation and key information submersion caused by fixed weight fusion or simple feature splicing. In an optional embodiment, key electromyographic channels are enhanced and redundant or noisy channels are suppressed by channel attention weight or channel gating weight, thereby improving the discriminativeness and stability of multi-channel temporal features.

[0030] 5. This invention uses multi-channel temporal features as the state input of the reinforcement learning interactive environment and the takeover load level identification result as the action, which can transform the takeover load identification from static supervised classification into a sequential decision-making process consisting of state, action, reward and next state; by generating reward data through the correspondence between load identification action and load label, the identification strategy can be optimized based on reward and punishment feedback during the training process, thereby improving the model's adaptability to continuous takeover scenarios and the overall decision stability.

[0031] 6. This invention reuses experience samples consisting of environmental states, load identification actions, reward data, and the next environmental state through an experience replay pool, and performs weighted sampling based on the importance data of the experience samples. This can improve sample utilization efficiency, enabling the model to prioritize key samples with high error, high uncertainty, high difficulty, or high misclassification frequency. Combined with the priority experience replay mechanism and optional class-aware training strategy, it can reduce the impact of class imbalance on model training and improve the ability to identify minority class load levels and high-difficulty load states. Attached Figure Description

[0032] Figure 1 This is an overall framework diagram of a multi-channel electromyography-based overload identification method in the embodiments. Figure 2 A comparison chart of the model confusion matrix under the normal experience replay and the priority experience replay mechanisms; Figure 3 A comparison chart of performance metrics for different ablation models. Detailed Implementation

[0033] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, detailed descriptions are provided below through embodiments.

[0034] This invention provides a method for identifying control load based on multi-channel electromyography, the framework of which is as follows: Figure 1 As shown, taking a driver as an example, the method can be implemented based on the acquisition of multi-channel electromyography (EMG) signals during driver takeover, and can rely on computing devices to complete data preprocessing, model training, offline inference, and model storage. The computing devices can be servers equipped with GPUs or high-performance workstations, with typical hardware configurations including multi-core CPUs, at least 16GB of system memory, dedicated graphics cards, and solid-state drive storage. The software environment can include a general-purpose operating system, a Python interpreter and its ecosystem, and can rely on NumPy, Pandas, and Scikit-learn data processing and analysis libraries for data preprocessing, and use the PyTorch deep learning framework to build and train the takeover load recognition model. In an optional embodiment, the method can also be deployed on an in-vehicle controller, edge computing unit, or cloud-edge-device collaborative computing environment after model compression and optimization.

[0035] Step S1: Preprocess and reconstruct the multi-channel electromyography signals during the takeover process to obtain multi-channel time-series samples, and determine the load labels corresponding to the multi-channel time-series samples.

[0036] Specifically, during the driver's takeover operation, surface electromyography (EMG) devices are used to simultaneously collect EMG signals from key muscle areas of the driver. Multi-channel EMG signals are used to characterize changes in the driver's physiological load during the takeover process. The load label can be determined by the driver's self-assessment based on the NASA-TLX scale after each takeover, and is used for training and validation of the takeover load identification model.

[0037] In an optional embodiment of the present invention, step S1 may include: Step S11: According to the muscle channel configuration, the multi-channel electromyographic signals during the driver takeover process are aligned to obtain aligned electromyographic signals.

[0038] Specifically, the muscle channel configuration can be determined based on sensor configuration conditions or engineering requirements. In one specific embodiment, nine muscle channels related to the takeover action are selected, covering the main force-generating muscle groups of the upper and lower limbs. Examples of the names of the nine muscle channels and their corresponding muscle locations are shown in Table 1.

[0039] Table 1. Correspondence between channels and muscles In one optional embodiment, the number of electromyographic channels can be reduced or increased according to the application scenario. For example, only upper limb-related muscle channels can be selected, or lower limb and trunk-related muscle channels can be added. When the number of electromyographic channels changes, the input dimensions and channel alignment methods required for subsequent time-series modeling can be adjusted accordingly.

[0040] Step S12: Preprocess the aligned electromyographic signals to obtain preprocessed electromyographic signals.

[0041] Specifically, preprocessing is used to reduce noise interference in the raw electromyographic signals and to reduce amplitude differences between different electromyographic channels and different subjects.

[0042] In an optional embodiment of the present invention, step S12 may include: Step S121: Filter the aligned electromyographic signal to obtain a denoised electromyographic signal.

[0043] Specifically, filtering is used to suppress the effects of power frequency interference and high-frequency noise on the quality of electromyographic signals.

[0044] Step S122: The amplitude of the denoised electromyographic signal is normalized to obtain the preprocessed electromyographic signal.

[0045] Specifically, amplitude normalization is used to keep the signal amplitude distribution consistent across different electromyographic channels and among different subjects.

[0046] Step S13: Perform time-series reconstruction on the preprocessed electromyographic signals and associate them with load labels to obtain multi-channel time-series samples and their corresponding load labels.

[0047] Specifically, temporal reconstruction is used to recombine the sampling points in the preprocessed electromyographic signal into a temporal subsequence according to time intervals. Let the preprocessed temporal signal of a certain electromyographic channel be represented as: in, This represents the preprocessed timing signal of a specific electromyographic channel. These represent the sampling points of the electromyographic channel; Indicates the number of sampling points.

[0048] In an optional embodiment of the present invention, step S13 may include: Step S131: The temporal signals of each electromyographic channel in the preprocessed electromyographic signal are sampled at equal intervals according to a preset step size to obtain the channel reconstruction sequence.

[0049] Specifically, the temporal segments of each electromyographic channel are processed according to a preset step size. By performing equidistant sampling, sampling points spaced at fixed time intervals in the original sequence are recombined into new time-series subsequences. The channel reconstruction sequence can be represented as: in, Indicates the first Channel reconstruction sequence; Indicates the preset step size; Indicates the sequence number of the channel reconstruction sequence.

[0050] In one specific embodiment, the original sequence is rearranged by combining the first sampling point and the eleventh sampling point in sequence, and by combining the second sampling point and the twelfth sampling point in sequence, to obtain multiple enhanced sequences with different time span characteristics.

[0051] Step S132: Combine the channel reconstruction sequences according to the channel correspondence to obtain multi-channel time series samples.

[0052] Step S133: The workload self-assessment score generated after the driver completes the takeover is used as the load label corresponding to the multi-channel time series sample.

[0053] Step S2 involves performing time series modeling, feature abstraction, and adaptive feature fusion on the multi-channel time series samples to obtain multi-channel time series features.

[0054] Specifically, multi-channel time-series samples can be represented as: in, Represents multi-channel time-series samples; Indicates time Nine-channel electromyography observation vector; This indicates the number of time steps. In embodiments where the number of electromyographic channels varies, The dimensions and muscle channel configuration are adjusted accordingly.

[0055] In an optional embodiment of the present invention, step S2 may include: Step S21: Extract time-series dependencies from multi-channel time-series samples to obtain time-series modeling features.

[0056] Specifically, a bidirectional long short-term memory network can be used to perform bidirectional temporal modeling on multi-channel time-series samples. This allows for the learning of dynamic changes in electromyographic signals during driver takeover from both forward and backward temporal perspectives. The output of the bidirectional temporal modeling can be represented as: in, Indicates time Bidirectional time-series modeling output; This represents the hidden state obtained from forward temporal modeling; This represents the hidden state obtained from reverse time series modeling.

[0057] In an optional embodiment of the present invention, step S21 may include: Step S211: Extract time-series dependencies from multi-channel time-series samples to obtain candidate time-series features.

[0058] Specifically, temporal dependencies are extracted from multi-channel temporal samples based on recurrent neural networks, temporal convolutional networks, one-dimensional convolutional networks, or attention-based temporal encoders. The recurrent neural network can include a long short-term memory network, a bidirectional long short-term memory network, a gated recurrent unit (GRU), or a bidirectional GRU. The temporal convolutional network can be a TCN. The attention-based temporal encoder can be a Transformer-based temporal encoder. These network structures are used for temporal dependency modeling and feature extraction of multi-channel electromyography (EMG) temporal signals.

[0059] Step S212: Perform temporal integration on the candidate temporal features to obtain temporal modeling features.

[0060] In another optional embodiment of the present invention, step S21 may include: Step S211: Perform forward and reverse time series modeling on the multi-channel time series samples to obtain bidirectional time series features.

[0061] Step S212: Perform unidirectional time series modeling on the bidirectional time series features to obtain time series modeling features.

[0062] Specifically, after obtaining the bidirectional temporal features, the bidirectional temporal features can be input into a unidirectional long short-term memory network for temporal integration.

[0063] Step S22: Perform feature abstraction on the time series modeling features to obtain abstract time series features.

[0064] Specifically, feature abstraction is used to further integrate the temporal modeling features to obtain feature representations for subsequent adaptive feature fusion. .

[0065] Step S23: Perform adaptive feature fusion on the abstract temporal features to obtain multi-channel temporal features.

[0066] Specifically, adaptive feature fusion can fuse abstract temporal features by learning the importance of different electromyographic channels in the process of taking over the load identification.

[0067] In an optional embodiment of the present invention, step S23 may include: Step S231: Determine channel fusion weights or channel relationship data based on abstract temporal features.

[0068] Specifically, the channel fusion weights can be channel attention weights, vector-based gating weights, or channel selection data determined based on learnable weights; the channel relationship data can be determined by a graph-based cross-channel relationship modeling method.

[0069] Step S232: Perform cross-channel fusion of abstract time series features based on channel fusion weights or channel relationship data to obtain multi-channel time series features.

[0070] In another optional embodiment of the present invention, step S23 may include: Step S231: Determine channel attention weights or channel gating weights based on abstract temporal features.

[0071] Specifically, channel attention weights can be used to scalar weight different electromyographic channels; channel gating weights can be vector-form weights learned for each electromyographic channel and used to modulate channel features element-wise.

[0072] Step S232: Modulate the abstract temporal features according to the channel attention weights or channel gating weights to obtain multi-channel temporal features.

[0073] Specifically, when channel attention weights are used, the weighted fusion process can be expressed as: in, Indicates multi-channel timing characteristics; This represents the channel attention weight vector corresponding to each electromyographic channel; Represents abstract temporal characteristics; This indicates a channel-by-channel weighted operation. Multi-channel timing characteristics. This serves as the state input for the takeover load identification model in step S3.

[0074] Step S3: Using multi-channel time-series features as state input, taking the takeover load level discrimination as the identification action, and training an offline reinforcement learning-based takeover load identification model based on load labels.

[0075] Specifically, the takeover load identification model can employ a reinforcement learning framework based on action value functions to model the driver takeover load identification process, and the action value function can be approximated using deep neural networks. The driver takeover load representation result can be a discrete load level.

[0076] In an optional embodiment of the present invention, step S3 may include: Step S31: Input the multi-channel temporal features into the reinforcement learning interactive environment to obtain the environment state.

[0077] Specifically, in a reinforcement learning interactive environment, the state space is composed of multi-channel temporal features, at time... The environmental state can be represented as: in, Indicates time The environmental conditions; Indicates time Multi-channel timing characteristics.

[0078] Step S32: Output load identification action based on environmental status, and generate reward data based on load identification action and load label.

[0079] Specifically, the action space is used to describe the output characteristics of the takeover load. Let the driver takeover load be divided into... If there are several levels, then the action space can be represented as: in, Represents the action space; These represent load identification actions corresponding to different load levels. In this embodiment, each load identification action corresponds to a control load level determination result.

[0080] Specifically, the reinforcement learning interactive environment generates reward data based on the matching relationship between the load recognition action and the load label. The reward data can be represented as: in, Represents reward data; Indicates positive reward; This indicates a penalty.

[0081] In an optional embodiment of the present invention, step S32 may include: Step S321: Output load level action, load scoring action, or risk estimation action based on environmental conditions.

[0082] Specifically, load level actions correspond to discrete category decisions, load scoring actions correspond to continuous load scoring, and risk estimation actions correspond to driver takeover risk estimation.

[0083] Step S322: Generate reward data based on the correspondence between load level actions, load scoring actions, or risk estimation actions and load labels.

[0084] Step S33: Train the takeover load identification model based on environmental status, load identification actions, reward data, and the next environmental status.

[0085] Specifically, at time The intelligent agent, based on the current environmental state Select load identification action The action value function can be expressed as: in, Indicates the environmental state Execute the load identification action below The value of the action; Expressing expectations; Indicates an immediate reward; Indicates the discount factor; Indicates the next environmental state; Indicates a candidate action.

[0086] In an optional embodiment of the present invention, step S33 may include: Step S331: Based on the environmental state, load identification actions, reward data, and the next environmental state, generate action value update data, strategy update data, or actor commentator update data.

[0087] Specifically, the takeover load identification model can employ a reinforcement learning framework based on action value functions, or it can employ a policy gradient-based method, an actor-commentator architecture, or an improved form thereof.

[0088] Step S332: Update the network parameters of the takeover load identification model based on action value update data, strategy update data, or actor commentator update data.

[0089] In another optional embodiment of the present invention, step S33 may include: Step S331: Generate an experience sample based on the environmental status, load identification action, reward data, and the next environmental status.

[0090] Specifically, in the reinforcement learning interaction process, experience samples can be represented as ,in, Indicates the state of the environment; Indicates a load identification action; Represents reward data; Indicates the next environmental state.

[0091] Step S332: Sample target experience samples from the experience replay pool based on the importance data of experience samples.

[0092] Specifically, the experience replay pool can adopt a priority experience replay structure, or it can adopt a hierarchical replay, multi-buffer replay, or online-offline hybrid replay structure.

[0093] Step S333: Train the takeover load identification model based on the target experience samples.

[0094] In another optional embodiment of the present invention, step S33 may include: Step S331: Generate an experience sample based on the environmental status, load identification action, reward data, and the next environmental status.

[0095] Step S332: Determine the importance data of the empirical samples based on time difference error, sample uncertainty, sample difficulty, or misclassification frequency.

[0096] Specifically, when determining the importance of data based on time difference error, a sampling probability is assigned to each empirical sample according to the time difference error corresponding to the empirical sample. The sampling probability can be expressed as: in, Indicates the first The sampling probability of an empirical sample; Indicates the first Time difference error of empirical samples; Indicates the first Time difference error of empirical samples; This indicates the priority adjustment factor.

[0097] Step S333: Based on the importance data, target experience samples are weighted and sampled from the experience replay pool, and the takeover load identification model is trained based on the target experience samples.

[0098] In another optional embodiment of the present invention, step S33 may include: Step S331: Based on the environmental state, load identification action, reward data, and the next environmental state, generate predicted action value data and target action value data.

[0099] Specifically, the target action value data can be represented as: in, This represents the value data of the target action; Indicates an immediate reward; Indicates the discount factor; Indicates candidate actions; Indicates based on target network parameters The value of a definite action.

[0100] Step S332: Update the online network parameters of the takeover load identification model based on the predicted action value data and the target action value data.

[0101] Specifically, the online network parameters of the takeover load identification model can be updated by minimizing the mean squared error between the predicted action value data and the target action value data. The loss function can be expressed as: in, Represents the loss function; Expressing expectations; This represents the value data of the target action; Indicates based on online network parameters Determined predictive action value data.

[0102] Step S333: Soft update the target network parameters of the takeover load identification model based on the updated online network parameters.

[0103] Specifically, the soft update form of the target network parameters can be expressed as: in, Indicates online network parameters; Indicates the target network parameters; This represents the update coefficient.

[0104] Step S4: Based on the independent test set, identify the trained overhaul load identification model, obtain the overhaul load identification result, and save the overhaul load identification model.

[0105] Specifically, the independent test set is used to evaluate the recognition performance of the trained takeover load recognition model, and the driver takeover load recognition result can be the driver takeover load level.

[0106] In an optional embodiment of the present invention, step S4 may include: Step S41: Preprocess and reconstruct the timing of the multi-channel electromyography signals in the independent test set to obtain the test timing samples.

[0107] Specifically, the method of preprocessing and reconstructing the multi-channel electromyography (EMG) signals in the independent test set can be consistent with the method of preprocessing and reconstructing the multi-channel EMG signals during the driver takeover process in step S1.

[0108] Step S42: Perform time series modeling, feature abstraction, and adaptive feature fusion on the test time series samples to obtain test time series features.

[0109] Specifically, the method of performing time series modeling, feature abstraction, and adaptive feature fusion on the test time series samples can be consistent with the method of performing time series modeling, feature abstraction, and adaptive feature fusion on multi-channel time series samples in step S2.

[0110] Step S43: Input the test time sequence features into the trained takeover load identification model to obtain the driver takeover load identification result, and save the trained takeover load identification model.

[0111] This invention also provides a multi-channel electromyography-based overload identification system for performing the above method, comprising: The sample construction module is used to preprocess and reconstruct the multi-channel electromyography signals during the driver takeover process to obtain multi-channel time-series samples and determine the load labels corresponding to the multi-channel time-series samples. The feature extraction module is used to perform time series modeling, feature abstraction, and adaptive feature fusion on multi-channel time series samples to obtain multi-channel time series features; The model training module is used to take multi-channel time-series features as state input, driver takeover load level discrimination as recognition action, and train an offline reinforcement learning-based takeover load recognition model based on load labels. The model evaluation module is used to identify the trained takeover load identification model based on an independent test set, obtain the driver takeover load identification result, and save the takeover load identification model.

[0112] Comparative experiment: This experiment used 7810 samples, of which 2350 were test sets; each sample consisted of 9 channels, with each channel having a signal length of 4800. This experiment compared the performance of the method of this invention with the following existing typical time-series classification models on a multi-channel electromyography signal workload identification task: (1) 1D-CNN: Kiranyaz S, Avci O, Abdeljaber O, et al. 1D convolutional neural networks and applications: A survey[J]. Mechanical systems and signalprocessing, 2021, 151: 107398. (2) BiGRU: Liu H, Zhang F, Tan Y, et al. Multi-scale quaternion CNN andBiGRU with cross self-attention feature fusion for fault diagnosis of bearing[J]. Measurement Science and Technology, 2024, 35(8): 086138. (3)BiLSTM:Graves A, Schmidhuber J. Framewise phoneme classificationwith bidirectional LSTM networks[C] / / Proceedings. 2005 IEEE InternationalJoint Conference on Neural Networks, 2005. IEEE, 2005, 4: 2047-2052. (4)CNN-LSTM:Pan H, He X, Tang S, et al. An improved bearing faultdiagnosis method using one-dimensional CNN and LSTM[J]. Strojniski Vestnik-Journal of Mechanical Engineering, 2018, 64(7-8): 443-453. (5)TCN:An Empirical Evaluation of Generic Convolutional and RecurrentNetworks for Sequence Modeling. (6)Inception Time:Ismail Fawaz H, Lucas B, Forestier G, et al.Inceptiontime: Finding alexnet for time series classification[J]. Data Miningand Knowledge Discovery, 2020, 34(6): 1936-1962. (7)MiniROCKET:Dempster A, Schmidt D F, Webb G I. Minirocket: A veryfast (almost) deterministic transform for time series classification[C] / / Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & datamining. 2021: 248-257. (8) Transformer: Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30. (9) Mamba: Gu A, Dao T. Mamba: Linear-time sequence modeling with selective state spaces[C] / / First conference on language modeling. 2024. (10) RIE: Jiang B, Wu H, Xia Q, et al. An efficient surfaceelectromyography-based gesture recognition algorithm based on multiscalefusion convolution and channel attention[J]. Scientific Reports, 2024, 14(1):30867. The test results of this experiment are shown in Table 2.

[0113] Table 2 Comparison of the overall performance evaluation results of each model As shown in Table 2, traditional structures (1D-CNN, BiGRU, BiLSTM) exhibit weak overall performance, with accuracies ranging from only 59.96% to 66.47%, and low Macro-F1 and Macro-AUC. With enhanced model representation capabilities, the recognition performance of CNN-LSTM, TCN, InceptionTime, and MiniROCKET significantly improves. Novel sequence models such as Transformer and Mamba demonstrate advantages in long-dependency modeling, increasing accuracies to 71.36% and 72.12%, respectively. Our proposed method achieves optimal results across all evaluation metrics, with an accuracy of 80.04%, a Macro-F1 score of 0.775, and a Macro-AUC of 0.867, significantly outperforming all comparative models. Compared to the representative temporal model Mamba and the recently proposed RIE method, our method achieves test accuracies improvements of 7.92% and 4.76%, respectively, fully demonstrating its superior performance and advantages in multi-channel EMG signal workload recognition tasks.

[0114] In one experiment, to verify the impact of the priority experience replay mechanism on the performance of easily confused category recognition, the ordinary experience replay mechanism was used as a control scheme, and the priority experience replay mechanism was used as the scheme of this invention for comparison. Figure 2 The confusion matrix results for the test set under the ordinary experience replay mechanism and the priority experience replay mechanism are shown. Under the ordinary experience replay mechanism, there are misclassifications between adjacent load levels. Specifically, 14.8% of low-load samples are misclassified as medium load, 25% of medium-load samples are misclassified as low load, and 19.3% of high-load samples are misclassified as medium load. After adopting the priority experience replay mechanism, the identification accuracy of low load, medium load, and high load increases by 4.9%, 5.8%, and 7.1%, respectively, and the overall classification accuracy increases by 5.5%. This indicates that the priority experience replay mechanism can increase the sampling weight of difficult-to-classify samples and samples with high misclassification frequency, and enhance the learning ability of the overhaul load identification model to the boundary features of adjacent load levels.

[0115] In one experiment, to verify the role of the composite LSTM structure and channel attention mechanism in multi-channel EMG feature modeling, two comparison models were set up. Comparison model 1 used a regular MLP instead of the composite LSTM, while comparison model 2 removed the channel attention mechanism and adopted a simple channel fusion structure. Figure 3 The ablation experiment results of the feature extraction network are shown. The accuracy of the proposed solution is improved by 6.64% compared to the comparative model 1, indicating that the composite LSTM structure can capture the temporal dynamic information of multi-channel EMG signals; the accuracy of the proposed solution is improved by 3.66% compared to the comparative model 2, indicating that the channel attention mechanism can enhance the features of key muscle channels and suppress redundant channel information. These results demonstrate that the combination of composite temporal modeling and channel selectivity enhancement can improve the discriminative ability of multi-channel EMG overhaul load identification.

[0116] Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A method for identifying control load based on multi-channel electromyography, characterized in that, Includes the following steps: The multi-channel electromyography signals during the takeover process are preprocessed and reconstructed to obtain multi-channel time-series samples, and the load labels corresponding to the multi-channel time-series samples are determined. The multi-channel time-series samples are subjected to time-series modeling, feature abstraction, and adaptive feature fusion to obtain multi-channel time-series features; Using the multi-channel time-series features as state input and takeover load level discrimination as action, a takeover load identification model based on offline reinforcement learning is trained based on the load labels. The trained overhaul load identification model is identified based on an independent test set to obtain the overhaul load identification result, and the overhaul load identification model is saved.

2. The method as described in claim 1, characterized in that, The multi-channel electromyography (EMG) signals during the takeover process are preprocessed and reconstructed to obtain multi-channel time-series samples, and the load labels corresponding to the multi-channel time-series samples are determined, including: Based on the muscle channel configuration, the multi-channel electromyographic signals during the takeover process are aligned to obtain aligned electromyographic signals; The aligned electromyographic signals are preprocessed to obtain preprocessed electromyographic signals; The preprocessed electromyographic signals are reconstructed temporally and associated with load labels to obtain multi-channel temporal samples and their corresponding load labels.

3. The method as described in claim 2, characterized in that, The preprocessed electromyographic signals are temporally reconstructed and associated with load labels to obtain multi-channel temporal samples and their corresponding load labels, including: The temporal signals of each electromyographic channel in the preprocessed electromyographic signal are sampled at equal intervals according to a preset step size to obtain the channel reconstruction sequence. The channel reconstruction sequences are combined according to the channel correspondence to obtain multi-channel time series samples; The workload self-assessment score generated after the takeover is completed will be used as the load label corresponding to the multi-channel time series sample.

4. The method as described in claim 1, characterized in that, The multi-channel time-series samples are subjected to time-series modeling, feature abstraction, and adaptive feature fusion to obtain multi-channel time-series features, including: Temporal dependency extraction is performed on the multi-channel time series samples to obtain temporal modeling features; The time-series modeling features are abstracted to obtain abstract time-series features; Adaptive feature fusion is performed on the abstract temporal features to obtain multi-channel temporal features.

5. The method as described in claim 4, characterized in that, Temporal dependency extraction is performed on the multi-channel time series samples to obtain temporal modeling features, including: Temporal dependency extraction is performed on the multi-channel temporal samples to obtain candidate temporal features; The candidate temporal features are integrated temporally to obtain temporal modeling features.

6. The method as described in claim 4, characterized in that, Adaptive feature fusion is performed on the abstract temporal features to obtain multi-channel temporal features, including: Determine channel fusion weights or channel relationship data based on the aforementioned abstract temporal characteristics; The abstract time series features are fused across channels based on the channel fusion weights or the channel relationship data to obtain multi-channel time series features.

7. The method as described in claim 1, characterized in that, Using the multi-channel time-series features as state input, and takeover load level discrimination as action, a takeover load identification model based on offline reinforcement learning is trained based on the load labels, including: The multi-channel temporal features are input into the reinforcement learning interactive environment to obtain the environment state; Based on the environmental conditions, output load identification actions, and generate reward data based on the load identification actions and the load tags; The takeover load identification model is trained based on the environmental state, the load identification action, the reward data, and the next environmental state.

8. The method as described in claim 7, characterized in that, Training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Based on the environmental state, the load identification action, the reward data, and the next environmental state, generate action value update data, strategy update data, or actor commentator update data; The network parameters of the takeover load identification model are updated based on the action value update data, the strategy update data, or the actor commentator update data.

9. The method as described in claim 7, characterized in that, Training a takeover load identification model based on the environmental state, the load identification action, the reward data, and the next environmental state includes: Experience samples are generated based on the environmental state, the load identification action, the reward data, and the next environmental state. The importance data of the empirical samples are determined based on time difference error, sample uncertainty, sample difficulty, or misclassification frequency. Based on the importance data, target experience samples are weighted and sampled from the experience replay pool, and a takeover load identification model is trained based on the target experience samples.

10. A multi-channel electromyography-based load identification system, characterized in that, include: The sample construction module is used to preprocess and reconstruct the multi-channel electromyography signals during the takeover process to obtain multi-channel time-series samples and determine the load labels corresponding to the multi-channel time-series samples. The feature extraction module is used to perform time series modeling, feature abstraction, and adaptive feature fusion on the multi-channel time series samples to obtain multi-channel time series features; The model training module is used to take the multi-channel time series features as state input, take the takeover load level discrimination as the recognition action, and train an offline reinforcement learning-based takeover load recognition model based on the load labels. The model evaluation module is used to identify the trained overhaul load identification model based on an independent test set, obtain the overhaul load identification result, and save the overhaul load identification model.

Images