Brain-controlled motion intention recognition method based on global optimization ANFIS

Abstract

The invention discloses a brain-controlled motion intention recognition method based on global optimization ANFIS, which utilizes an ANFIS model to perform brain-controlled intention classification decision, introduces a global spatio-temporal expression module, fully considers global spatio-temporal information of multi-input EEG data, introduces a rule attention module, and allows the model to recognize the most active intention-related fuzzy rule through an attention mechanism. And important intention features are selected to carry out accurate EEG motion intention classification decision. According to the method, the interpretability characteristic of the fuzzy inference system and the learning ability of the adaptive network are comprehensively utilized, system parameters can be changed according to priori knowledge, the extracted intention characteristics are explained layer by layer, and the final decision output is closer to the real output. According to the method, the performance of a brain-controlled intention recognition model applied to the field of brain-computer interaction is effectively improved, and the method can be suitable for assisting brain-controlled motion intention classification decision making based on EEG under a lower limb rehabilitation task and has high robustness and high efficiency.

Description

Technical Field

[0001] The present application belongs to the technical field of lower limb assisted rehabilitation of brain-computer interfaces, and in particular, to a method for identifying brain-controlled movement intention based on a global optimized ANFIS. Background Art

[0002] As a cutting-edge rehabilitation technology, brain-computer interface is gradually replacing traditional therapies and aims to promote the recovery of neurological function. Electroencephalogram (EEG) is the most common electrophysiological signal in brain-computer interface applications. It has the advantages of non-invasiveness, low cost, compatibility, portability, etc., and has a higher temporal resolution than other brain wave measurement methods (such as magnetoencephalography, functional magnetic resonance imaging and near-infrared spectroscopy). EEG can be defined as the measurement of brain electrical activity caused by the current induced by neurons in the brain. It can reflect the brain's neural activity in real time. Its active participation has been shown to induce corticospinal plasticity and increase the possibility of motor recovery. At present, the EEG signal rhythm corresponding to the intention of movement can be monitored. The movement-related cortical potential is the brain activity processed during the planning and preparation of the movement, which can decode the intention of movement from the pre-movement state. Therefore, by decoding the EEG signal, the individual's intended action can be identified, helping patients control physical or virtual devices and enhancing their ability to communicate and interact with the real world.

[0003] In the field of cognitive computing, neural networks, as a powerful class of machine learning models, have been widely used to identify EEG signals, with superior classification performance and can effectively improve cognitive recognition capabilities. This advantage can be attributed to the deep hierarchical structure of neural networks, which can effectively detect complex patterns and well represent training data in a layer-by-layer manner. Neural networks are able to effectively learn internal representations, but it is not clear how this effective representation is distributed layer by layer and how it arises from learning. The lack of transparency in the training process often leads to critical trust-related issues in key application areas such as medical rehabilitation, where it is crucial to clearly link inputs such as patient characteristics to predicted results. Explainability is a key component of artificial intelligence-based medical decision support systems and an effective means to enhance the robustness of clinical decisions. Therefore, explainability has become a core issue of concern in the medical field.

[0004] In recent years, research on interpretable systems has shown that simple rule-based methods, such as decision lists and IF-THEN rules, can provide more understandable explanations for the decision-making process. Rule extraction is a method to reveal hidden knowledge in the network. Fuzzy inference systems autonomously learn fuzzy membership functions to derive fuzzy rules from a large amount of training data, making the decision-making process easier for humans to interpret. Adaptive neuro-fuzzy inference system (ANFIS) is a hybrid intelligent system that uses neural networks to adjust rule-based fuzzy systems and enhance the ability of autonomous learning and adaptation. Although ANFIS has shown significant improvements in classification accuracy, interpreting a large number of fuzzy rules and parameters for high-dimensional application problems remains a challenge that needs to be addressed. In addition, when analyzing EEG data with temporal characteristics, the potential temporal correlation between multiple inputs is ignored. Summary of the invention

[0005] The purpose of this application is to provide a method for brain-controlled movement intention recognition based on global optimization ANFIS, referred to as GA-ANFIS, which uses an interpretable adaptive fuzzy inference system (ANFIS) to make EEG-based brain-controlled intention classification decisions, introduces a global spatiotemporal expression module to explore the global intention information of time-series EEG data, and uses the attention mechanism to identify relevant important fuzzy rules in ANFIS to grasp the easily overlooked but important EEG brain-controlled intention features, while improving the robustness and effectiveness of EEG brain-controlled intention recognition.

[0006] In order to achieve the above purpose, the technical solution of this application is as follows: A method for identifying brain-controlled motion intention based on global optimization ANFIS, comprising: Obtain EEG data, perform preprocessing, and extract EEG features; The EEG features are combined into an input vector and input into the global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features; The EEG features and their spatiotemporal multidimensional features are input into the ANFIS model, and after passing through the fuzzification layer, rule layer and normalization layer of the ANFIS model, the triggering ratio of the rules in the rule base during the reasoning process is obtained, and then the expression strength of the rules is obtained after passing through the defuzzification layer of the ANFIS model. The trigger ratio of the rule in the rule base is passed through the rule attention module to obtain the attention score vector; In the output layer of the ANFIS model, the attention score vector is multiplied by the expression strength of each rule and then summed up. The final expressed motion intention is determined based on the actual output value obtained by the sum.

[0007] Furthermore, the step of acquiring EEG data, performing preprocessing, and extracting EEG features includes: First, preprocessing was performed, i.e., using a bandpass filter to filter the EEG data and remove visible artifacts; Then, the time window of interest is intercepted to split the preprocessed data stream into several segments; The EEG features were extracted using power spectral density.

[0008] Furthermore, the step of combining the EEG features into an input vector and inputting it into a global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features includes: Capture the instantaneous changes of the input vector through time convolution to obtain the temporal pattern expression; Through deep convolution and full connection operations, spatiotemporal multi-dimensional features are further extracted.

[0009] Furthermore, the rule attention module implements the operation shown in the following formula: in, and is the learned attention parameter, represents the attention score vector, Indicates the trigger ratio.

[0010] This application proposes a method for identifying brain-controlled movement intention based on global optimization ANFIS, which is used to improve the accurate decision-making of brain-controlled intention for assisting lower limb rehabilitation. The ANFIS model is used to classify and decide on brain-controlled intention, and the interpretability characteristics of the fuzzy reasoning system and the learning ability of the adaptive network are comprehensively utilized. The system parameters can be changed according to prior knowledge, and the extracted intention features can be interpreted layer by layer, so that the final decision output is closer to the real output. The global spatiotemporal expression module is introduced to fully consider the global spatiotemporal information of multi-input EEG data, learn and mine the spatial distribution characteristics and long-range dependencies (i.e., spatiotemporal multidimensional features) of the temporal EEG movement intention to improve global interpretability. The rule attention module is introduced, and its attention mechanism allows the model to identify the most active intention-related fuzzy rules to select important intention features for accurate EEG movement intention classification decisions, while reducing the rule dimension and improving the model efficiency. It effectively improves the performance of the brain-controlled intention recognition model applied in the field of brain-computer interaction, and can be applied to the classification and decision-making of brain-controlled movement intention based on EEG under the task of assisting lower limb rehabilitation. It has strong robustness and efficiency, and provides technical support for lower limb rehabilitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a schematic diagram of the ANFIS model structure based on global optimization in this application.

[0012] Figure 2 This is a flow chart of the brain-controlled movement intention recognition method based on global optimization ANFIS. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0014] In order to achieve explainable and efficient brain-controlled movement intention classification decisions, this application designs an ANFIS model based on global optimization, namely GA-ANFIS, for movement intention recognition for brain-controlled assisted lower limb rehabilitation. The fuzzy rule method of ANFIS is used to mine potential intention representations and enhance the interpretability of hidden layer knowledge expression. The global spatiotemporal expression module is introduced to explore the spatial distribution characteristics of EEG data and mine long-range dependencies. The attention mechanism is used to select important rules, allowing ANFIS to pay more attention to relevant fuzzy rules and characterize the intention features that are easily overlooked but important to accurately identify movement intentions.

[0015] like Figure 1 As shown in the figure, the GA-ANFIS model includes a global spatiotemporal expression module, ANFIS and a rule attention module, wherein ANFIS has a 5-layer structure, the first 3 layers are rule antecedents (i.e., the IF condition part of the fuzzy rule), and the last 2 layers are rule consequents (i.e., the THEN result part of the fuzzy rule). This application first introduces a global spatiotemporal expression module in the rule antecedent part, fully considering the global spatiotemporal correlation between multiple inputs. Then, the attention weight is calculated based on the output features of the rule antecedent, which is used to constrain the rule consequent, so as to select important rules related to motion intention to achieve robust, accurate and efficient classification decisions.

[0016] An embodiment of the present application, such as Figure 2 As shown in the figure, a method for brain-controlled motion intention recognition based on global optimization ANFIS is proposed, including: Step S1, obtain EEG data, perform preprocessing, and extract EEG features.

[0017] The collection of electroencephalogram (EEG) signals generally requires wearing a brain-computer interface device, such as an EEG cap. The EEG cap is a non-invasive brain-computer interface device that collects EEG signals from the scalp through the electrodes on the EEG cap to monitor the discharge activity of a group of neurons. Compared with invasive brain-computer interfaces, the advantage of the EEG cap is that it is easy to operate and does not require surgical implantation, so it is relatively safe. However, its disadvantage is that it is not accurate enough and cannot monitor the discharge activity of a single neuron like an invasive brain-computer interface.

[0018] Aiming at the lower limb movement task, a corresponding experimental paradigm was designed, including two situations: Resting state experiment: When the human brain does not receive any external stimulation, there is regular neural activity. In the resting state laboratory, EEG data is collected with eyes open and eyes closed when the subjects are relaxed, still, and try to avoid thinking activities.

[0019] Task-based experiment: The task-based experiment is divided into two levels: gait phase motor imagery and movement pattern motor imagery. The experimental subjects involved two categories: patients with lower limb movement disorders and normal control groups (i.e. healthy people). During the experiment, the subjects were required to imagine four gait phases and five movement patterns 20 times each, with a 2-second rest in between each. During the experiment, the EEG data of the subjects were collected in real time.

[0020] Since the collected raw EEG data contains abnormal data such as noise and artifacts, the EEG data is preprocessed and signal features are extracted. First, a wired impulse response type bandpass filter (0.05-40 Hz) is used to filter the EEG data to eliminate power frequency interference other than spontaneous EEG signals, namely high-frequency noise, and then the independent component analysis method is used to remove visible artifacts such as eye movements. Subsequently, the time window of interest is intercepted and the preprocessed data stream is divided into several segments, each lasting 13 seconds. Finally, the power spectral density is used to extract EEG features to provide stable signal features for subsequent data analysis. All extracted EEG signal feature samples are divided into segments of length n, that is, each EEG segment contains n EEG feature samples, expressed as , each subject’s EEG data contains segments, where N is the total number of EEG feature samples.

[0021] Step S2: Combine the EEG features into an input vector and input it into the global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features.

[0022] The ANFIS model is jointly constructed by an adaptive network and a fuzzy inference system. It integrates the interpretable characteristics of the fuzzy inference system and the autonomous learning ability of the adaptive network. It can change the system parameters according to prior knowledge to make the output of the system model closer to the real output. Since the EEG signal is a time series data with multiple channels, in order to fully explore the spatiotemporal feature information of EEG, this embodiment introduces a global spatiotemporal expression module in the ANFIS model to characterize the spatiotemporal multidimensional information based on EEG.

[0023] EEG signals have multi-channel spatial distribution characteristics and long-range dependent temporal characteristics. In order to fully explore the spatial and temporal information in EEG signals, this embodiment embeds a global spatiotemporal expression module. Based on the n inputs of the ANFIS model, the input of the global spatiotemporal expression module is constructed. = Express the global spatiotemporal information of EEG multi-input data, where , , C is the number of EEG signal channels, and T is the window length of the input EEG signal.

[0024] In a specific embodiment, the EEG features are combined into an input vector and input into a global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features, including: Step 2.1: Capture the instantaneous changes of the input vector through temporal convolution to obtain the temporal pattern expression.

[0025] This step uses the temporal convolution layer to capture the instantaneous changes of the EEG signal, thereby constructing the temporal pattern of the EEG signal. The formal expression is as follows:

[0026] in, Indicates that 2D convolution filtering is performed in the time dimension, and the kernel size is , the number is 4, Represents the Gaussian Error Linear Unit GELU function.

[0027] Step 2.2: Through deep convolution and full connection operations, further extract spatiotemporal multi-dimensional features.

[0028] Subsequently, a depthwise convolution operation is used to learn the spatial filter, whose convolution kernel size is (C, 1), which does not change the number of channels. Deep convolution is used to capture the spatiotemporal multi-dimensional features of EEG signals from different channels. , which helps the model fully understand the brain's activity patterns under various motor intentions. The formal statement is as follows:

[0029] Finally, a linear function (i.e., fully connected operation) is used to transform the spatiotemporal multidimensional features. Dimension, that is , the converted spatiotemporal multidimensional features The dimension is consistent with the dimension of the first n inputs. This spatiotemporal multidimensional feature is used as the n+1th input and input into the ANFIS model for subsequent fuzzification operations.

[0030] This embodiment can effectively extract the spatial distribution characteristics of multi-channel EEG, mine long-range dependencies, and fully explore the global cognitive intention information of multi-input EEG.

[0031] Step S3, input the EEG features and their spatiotemporal multidimensional features into the ANFIS model, and pass through the fuzzification layer, rule layer and normalization layer of the ANFIS model to obtain the triggering ratio of the rules in the rule base during the reasoning process, and continue to pass through the defuzzification layer of the ANFIS model to obtain the expression strength of the rules.

[0032] The adaptive neural fuzzy inference system (ANFIS model) combines the adaptive network to build a multi-layer feedforward network consisting of nodes and directed links connecting the nodes, including a 5-layer computing structure. The ANFIS model has n inputs And 1 output, and K fuzzy rules, the fuzzy rule form is as follows:

[0033] Rule k: IF , THEN .

[0034] in, is the input variable corresponding to rule k The fuzzy set membership degree of Used to calculate the output result corresponding to rule k.

[0035] The first layer, the fuzzification layer: the input layer.

[0036] Each node is a fuzzy transformation node corresponding to an input feature. Each node has the same function, such as using the Gaussian membership function The input variables are fuzzified and converted into membership of different fuzzy sets. The formal description is as follows:

[0037] in, and represent the mean and variable of Gaussian fuzzy set respectively.

[0038] Layer 2, Rules Layer: Each node represents a fuzzy logic rule, and the number of nodes is equal to the number of rules K. The node obtains input from the corresponding fuzzified neuron and calculates the trigger strength of the rule, which is obtained by multiplying the membership of different fuzzy sets. The formula is as follows:

[0039] in, represents the credibility of the kth rule (i.e., the trigger strength), Represents the input sample The degree of belonging to the kth fuzzy set (i.e., membership degree), Represents the product operator.

[0040] Layer 3, Normalization layer: The nodes in this layer normalize the trigger strength from the previous layer to the trigger ratio of the kth rule in the rule base, that is, the degree to which the kth rule is used in the entire reasoning process. The calculation formula is as follows:

[0041] Layer 4, defuzzification layer: In each node of this layer, the linear expression of the kth rule is compared with the trigger ratio Multiply them together to get the weighted output, that is, calculate the expression strength of rule k , the formula is as follows: Where p and q are the conclusion parameters of the corresponding rule. The expression strength of rule k That is, the strength of the expression of the movement intention under this rule.

[0042] Layer 5, output layer: Combine each rule and convert all fuzzy rule strength values ​​into actual output values ​​through weighted average method, that is: The actual output value is the corresponding motion intention.

[0043] Step S4: Pass the trigger ratio of the rule in the rule base through the rule attention module to obtain an attention score vector.

[0044] This application introduces a rule attention module. The attention-based fusion operator first weights multiple input rules and then aggregates them into a representation. This attention mechanism allows the model to dynamically adjust its attention to rules that are more task-informative. In order to calculate the attention weight of each rule, all rule features are input into a multi-layer perceptron (MLP), whose output has values ​​close to 0 for all elements except the selected rule whose value is close to 1.

[0045] The attention score vector output by the rule attention module is calculated as follows: in, and is the learned attention parameter, As a nonlinear activation function, this module enables the model to mine the fuzzy adjacent relations of input features while reducing the dimension of rules and improving model efficiency.

[0046] Step S5: In the output layer of the ANFIS model, the attention score vector is multiplied by the expression strength of each rule and then summed up. The final expressed motion intention is determined based on the actual output value obtained by the summation.

[0047] This step multiplies the expression strength of the rules output by the defuzzification layer of the ANFIS model by the attention score vector, and only the expression strength of important rules will be used as the final output. The calculation formula is as follows:

[0048] in, It represents the MLP-based rule attention module, which is used to select important rules related to motion intention. The output layer mainly converts the rule expression strength into the actual output value, that is, the actual motion intention.

[0049] In this embodiment, in the attention score vector, except for the element corresponding to the selected important rule whose value is close to 1, the values ​​of all elements are close to 0, so the actual movement intention can be obtained more accurately.

[0050] This application can be applied to auxiliary rehabilitation tasks of brain-controlled lower limbs, accurately making decisions based on EEG's movement intentions, enhancing the robustness and effectiveness of brain-controlled intention recognition, and has the advantages of high efficiency, accuracy, and explainability, further promoting the development of the field of brain-computer collaborative intelligence.

[0051] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for identifying brain-controlled motion intention based on global optimization ANFIS, characterized in that: The method for identifying brain-controlled motion intention based on global optimization ANFIS includes: Obtain EEG data, perform preprocessing, and extract EEG features; The EEG features are combined into an input vector and input into the global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features; The EEG features and their spatiotemporal multidimensional features are input into the ANFIS model, and after passing through the fuzzification layer, rule layer and normalization layer of the ANFIS model, the triggering ratio of the rules in the rule base during the reasoning process is obtained, and then the expression strength of the rules is obtained after passing through the defuzzification layer of the ANFIS model. The trigger ratio of the rule in the rule base is passed through the rule attention module to obtain the attention score vector; In the output layer of the ANFIS model, the attention score vector is multiplied by the expression strength of each rule and then summed up. The final expressed motion intention is determined based on the actual output value obtained by the sum.

2. The method for identifying brain-controlled motion intention based on global optimization ANFIS according to claim 1, characterized in that: The step of obtaining EEG data, performing preprocessing, and extracting EEG features includes: First, preprocessing was performed, i.e., using a bandpass filter to filter the EEG data and remove visible artifacts; Then, the time window of interest is intercepted to split the preprocessed data stream into several segments; The EEG features were extracted using power spectral density.

3. The method for identifying brain-controlled movement intention based on global optimization ANFIS according to claim 1, characterized in that: The step of combining the EEG features into an input vector and inputting it into a global spatiotemporal expression module to extract the spatiotemporal multi-dimensional features of the EEG features includes: Capture the instantaneous changes of the input vector through time convolution to obtain the temporal pattern expression; Through deep convolution and full connection operations, spatiotemporal multi-dimensional features are further extracted.

4. The method for identifying brain-controlled movement intention based on global optimized ANFIS according to claim 1, characterized in that: The rule attention module implements the operation shown in the following formula: in, and is the learned attention parameter, represents the attention score vector, Indicates the trigger ratio.

Images