Stepping search type artificial neural electrode positioning method based on machine learning

Abstract

The invention relates to a step search type artificial neural electrode positioning method based on machine learning, and belongs to the field of neuroscience. Specifically, in a micro-driver control and execution module, a central control chip converts a digital feedback signal into an analog signal, the analog signal is amplified by a stack driving and coil current chip to generate a driving signal, and a micro-driver drives and controls an electrode to execute corresponding independent target actions under excitation of the micro-driver; in the signal acquisition and processing module, an electrode is implanted into a target brain region based on a stepping search type positioning method to perform searching and positioning of neurons and preprocessing after signal acquisition; in the machine learning model analysis and discrimination module, the quality of the neuron signal is discriminated in real time by using an integrated learning model, and a discrimination result is combined with the electrode state parameters to generate a digital feedback signal, and the digital feedback signal is transmitted to the central control chip for the next control cycle. Compared with a traditional random search type artificial neural electrode positioning method, the method can achieve a more accurate, more efficient and safer positioning effect.

Description

Technical Field

[0001] The present invention belongs to the field of neuroscience, and particularly relates to a step-by-step search artificial neural electrode positioning method based on machine learning. Background Art

[0002] In recent years, the brain-computer interface technology has developed rapidly in the fields of brain science, neuroscience, bioscience, information science, and medicine. The brain-computer interface technology, also known as the brain-computer fusion perception technology, is a new control technology that establishes information interaction between the brain of a human or animal and a computer or other electronic devices, and has been widely used in scenarios such as neuroscience research, motor rehabilitation of paralyzed patients, identity security verification, virtual reality interaction, and overcoming learning disabilities. The invasive artificial neural electrode is a typical representative of the brain-computer interface. The accurate positioning of neurons during its working implantation process plays a crucial role in the acquisition of brain electrical signals, directly affecting the communication quality between the brain and external devices. The traditional positioning method of invasive artificial neural electrodes is mainly based on a random search strategy, that is, using a microdriver to control the neural electrode to randomly search up and down for neurons in the target brain area to capture brain electrical signals. The traditional positioning method has problems such as poor accuracy, low efficiency, incomplete search range, and the possibility of damaging neurons and causing damage to brain tissue during the search process. To solve the above problems, the present invention proposes a new step-by-step search artificial neural electrode positioning method. The basic idea of this method is that the microdriver uses the step-by-step search method to control each electrode in the high-density probe-type neural microelectrode array to independently step-detect neuron signals, uses machine learning algorithms to evaluate and judge the quality of neuron signals, and then generates feedback signals and drive signals to circularly feedback and control the execution actions of the piezoelectric microdriver, realizing the rapid and accurate positioning of the electrode array to neuron cells, and thus realizing multi-brain area, large-scale, and high-precision brain electrical signal measurement. This method has the significant advantages of high positioning accuracy, good accuracy, fast search speed, large range, and good safety. Summary of the Invention

[0003] (I) Technical Problems to be Solved The technical problem to be solved by the present invention is to provide an artificial neural electrode positioning method to overcome the defects of low positioning accuracy, poor accuracy, slow search speed, incomplete search range, and risk of damage to brain tissue existing in the prior art during the process of positioning neurons.

[0004] (II) Technical Solutions To achieve the above object, the present invention provides a step-by-step search artificial neural electrode positioning method based on machine learning. The method mainly includes three parts: a microdriver control and execution module, a signal acquisition and processing module, and a machine learning model analysis and discrimination module. The specific implementation steps are as follows: S1: The central control chip converts the digital feedback signal into an analog signal; S2: The analog signal is pre - processed by the chip to generate a driving signal; S3: The micro - driver drives and controls each electrode in the artificial neural electrode array to perform corresponding independent target actions under the excitation of the driving signal; S4: The micro - driver drives and controls multiple electrodes to be implanted into the target brain region of the mouse by a step - by - step search - type positioning method for neuron search, positioning, and signal acquisition; S5: The collected neuron signals are filtered, rectified, and amplified by the pre - processing chip, and are subjected to analog - to - digital conversion to convert the neuron signals into data; S6: Input the said data into the integrated learning model; S7: Use the integrated learning model to perform real - time discrimination on the quality of neuron signals, and generate a digital feedback signal by combining the discrimination result with state parameters such as the preset signal acquisition duration t0 and the total travel distance h0 of the target area, and finally transmit it to the central control chip for the next control cycle.

[0005] Preferably, in step S2, the analog signal piezoelectric stack driving chip and the coil current chip are amplified and then become the driving signals of the piezoelectric stack and the coil respectively.

[0006] Preferably, in step S4, taking advantage of the fact that the micro - driver can flexibly and independently drive and control the movement of multiple electrodes, each electrode in the electrode array can simultaneously perform a step - by - step search for neurons in the target brain region. The specific process of neuron positioning of the electrode is as follows: The electrode is quickly implanted into the target brain region with a large step distance ∆h1, and steps down for search with a step size of the neuron scale Δh of the brain region. If the electrode locates an effective neuron, the electrode remains in place for neuron signal acquisition. If the acquisition time t is less than the preset signal acquisition duration t0 of the experiment, the electrode remains in place to collect neuron signals. Otherwise, it means that the signal acquisition task of the preset duration has been completed, and the electrode steps up continuously until it exits the target brain region; If the electrode fails to successfully locate an effective neuron, it is judged at each step whether the total travel distance h0 of the target area has been completed. If the step travel distance h is less than the total travel distance h0 of the target area, the electrode continues to step down for search. Otherwise, it means that the neuron search and positioning task of the preset total travel distance has been completed, and the electrode steps up continuously until it exits the target brain region; If there are no neurons in a specific brain region within the target brain region, and the electrode fails to locate an effective neuron after stepping through the total travel distance h0 of the target area, the electrode steps up continuously until it exits the target brain region.

[0007] Preferably, in step S6, the integrated learning model has been trained and constructed based on neuron signal characteristics before the entire control cycle, and specifically includes the following steps: A: Obtain neuron signal sample data; B: Perform gain pre - processing on the obtained neuron signal sample data to obtain multiple training samples, forming a training sample set; C: Divide the training sample set according to a ratio, where 70% is used for the training set for generating model training, and 30% is used for the test set for generating model testing; D: Train the deep residual shrinkage neural network through the training set, and save the trained deep residual shrinkage neural network model. Then, input the training set and the test set into the trained deep residual shrinkage neural network model for feature extraction, thereby obtaining a new training set and a new test set; E: Repeatedly train the ensemble learning model through the new training set to obtain the trained ensemble learning model. Test the neuron signal features in the new test set through the trained ensemble learning model to judge the accuracy of the model. By repeatedly adjusting the parameters and optimizing the model parameters, finally construct an ensemble learning model based on neuron signal features.

[0008] Preferably, in step B, the gain preprocessing includes a missing data processing process and a normalization processing process.

[0009] Preferably, in step D, the deep residual shrinkage neural network adds an attention mechanism and soft thresholding on the basis of the deep residual neural network.

[0010] Furthermore, the deep residual shrinkage neural network generally includes a convolutional layer, a certain number of residual shrinkage modules, a batch normalization, a ReLU activation function, a global average pooling layer, and a fully connected output layer. Each residual shrinkage module replaces the "reweighting" in the attention mechanism in the residual mode with "soft thresholding", and includes a batch normalization, a convolutional layer, a pooling layer, a soft thresholding module, an attention mechanism module, and a fully connected layer; among them, the soft thresholding function plays a non-linear role, and by embedding it into the sub-network, the automatic setting of the soft thresholding can be realized.

[0011] Even further, during data training, first input the neuron signal feature data into the convolutional layer for processing, and then pass through a certain number of residual shrinkage modules. In each residual shrinkage module, first take the absolute value of the processed neuron signal feature data, and then obtain a set of feature data through global average pooling and averaging; in another path, input the feature data after global average pooling into a small fully connected network, process it with a Sigmoid activation function in the fully connected network, then multiply and soft threshold the feature data in the two paths, then input it into the convolutional layer for processing and perform global average pooling, and finally obtain the final output data through the fully connected layer.

[0012] Preferably, in step S7, the real-time discrimination of the neuron signal quality by using the ensemble learning model is specifically as follows: If the electrode successfully locates a valid neuron, a digital signal sequence is generated for the electrode to remain in place, and the electrode collects neuron signals. If the collection time t is less than the preset signal collection time t0, a digital signal sequence is generated for the electrode to remain in place, and the electrode continues to collect neuron signals. Otherwise, it means that the signal collection task of the preset time has been completed, and a digital signal sequence is generated for the electrode to continue to step upward until it exits the target brain area. If the electrode fails to successfully locate a valid neuron, it determines at each step whether the total distance h0 of the target area has been completed. If the stepping distance h is less than the total distance h0 of the target area, a digital signal sequence is generated for the electrode to continue to step downwards and search; otherwise, it represents that the task of searching and locating neurons in the total distance of the target brain area has been completed, and a digital signal sequence is generated for the electrode to continue to step upwards until it exits the target brain area; If there are no neurons in a specific brain area within the target brain region, and no valid neurons are located after the electrode steps to search the target area for a total stroke of h0, a digital signal sequence is generated that continuously steps upward until exiting the target brain region.

[0013] (III) Beneficial effects The present invention proposes a step-by-step search method for locating artificial neural electrodes based on machine learning, which enables the electrodes to step-by-step search for neurons in the target brain area, avoiding damage to brain tissue caused by random search for neurons, and greatly improving the efficiency and accuracy of locating artificial neural electrodes. This method has great application and development prospects in the fields of neurology and brain science, and provides new ideas and new approaches for further research on artificial neural electrodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the installation of the micro-actuator on the mouse head; Figure 2 Working principle diagram of a microactuator integrating a single piezoelectric stack with multiple artificial neural electrodes; Figure 3 This is a schematic diagram of a step-by-step search artificial neural electrode positioning method proposed by the present invention; Figure 4 Operational flow chart for micro-actuator driven electrode positioning, collection, and discrimination of neuron signals; Figure 5 Schematic diagram of the integrated learning model for neuronal signal feature recognition. DETAILED DESCRIPTION

[0015] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0016] This embodiment relates to a step - search - type artificial neural electrode positioning method based on machine learning. First, based on the step - search - type positioning strategy, a micro - driver integrating a single piezoelectric stack and multiple artificial neural electrodes is used to control the simultaneous and independent implantation of multiple electrodes into the target brain region. Then, based on the machine - learning algorithm, the detected neuron signals are discriminated and the micro - driver is feedback - controlled, so as to achieve neuron positioning and signal acquisition, and finally achieve a neuron - positioning effect with high precision, high efficiency, and high safety.

[0017] Refer to Figure 1 It is a schematic diagram of the installation of the micro - driver on the head of a mouse in this embodiment. The micro - driver is implanted into the head of the mouse to make full preparations for the subsequent positioning, acquisition, and discrimination of neuron signals.

[0018] The micro - gripper is mainly composed of a resistive heating coil, a printed circuit board (PCB), and a thermally - controlled phase - change material (PCM). The PCM is a type of material that absorbs or releases a large amount of heat during melting or solidification while maintaining a constant temperature. When the resistive heating coil does not carry current, the PCM is in a solid state; when the heating coil carries current, the PCM melts and becomes liquid. Refer to Figure 2 It is the working principle diagram of the micro - driver integrating a single piezoelectric stack and multiple artificial neural electrodes in this embodiment. The specific working principle is as follows: Step 1: In the initial state, the piezoelectric stack is under a low - level excitation signal; Coil 1 is at a low level, the heating coil does not carry current, the PCM is solidified and plays a clamping role; Coil 2 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping role; Coil 3 is at a low level, the heating coil does not carry current, the PCM is solidified and plays a clamping role; Coil 4 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping role; Coil 5 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping role; Coil 6 is at a low level, the heating coil does not carry current, the PCM is solidified and plays a clamping role; Coil 7 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping role; Coil 8 is at a low level, the heating coil does not carry current, the PCM is solidified and plays a clamping role.

[0019] Step 2: Apply a high - level excitation signal to the piezoelectric stack. The piezoelectric stack elongates and drives the lower PCB to move downward; Coils 1 - 8 maintain the state in Step 1; Under the clamping action, Electrodes 1 and 3 remain stationary, and Electrodes 2 and 4 move downward by Δ h distance.

[0020] Step 3: The piezoelectric stack maintains the high - level excitation signal and remains stationary; Coils 1 - 8 are all in a low - level state, the heating coil does not carry current, the PCM is solidified and plays a clamping role.

[0021] Step 4: The piezoelectric stack maintains a high-level excitation signal and remains stationary; Coil 1 is at a low level, the heating coil does not carry current, the PCM solidifies and acts as a clamp; Coil 2 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping effect; Coil 3 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping effect; Coil 4 is at a low level, the heating coil does not carry current, the PCM solidifies and acts as a clamp; Coil 5 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping effect; Coil 6 is at a low level, the heating coil does not carry current, the PCM solidifies and acts as a clamp; Coil 7 is at a low level, the heating coil does not carry current, the PCM solidifies and acts as a clamp; Coil 8 is at a high level, the heating coil carries current, the PCM melts into a liquid state and has no clamping effect.

[0022] Step 5: Apply a low-level excitation signal to the piezoelectric stack, the piezoelectric stack recovers and drives the lower PCB to move upward; Coils 1 - 8 maintain the state in Step 4; Under the clamping effect, Electrodes 1 and 4 remain stationary, and Electrodes 2 and 3 move upward by Δ h distance together with the lower PCB.

[0023] Step 6: The piezoelectric stack maintains a low-level excitation signal and remains stationary; Coils 1 - 8 are all in a low-level state, the heating coils do not carry current, the PCM solidifies and acts as a clamp, and the 4 electrodes remain stationary.

[0024] By applying different excitation signals, each independent electrode realizes different step motions: 1) remain stationary, 2) step downward, 3) step upward. Compared with traditional micro-drivers, the micro-driver integrating a single piezoelectric stack and multiple artificial neural electrodes enables multiple electrodes to move independently in parallel, theoretically improving the overall positioning success rate and positioning efficiency of the electrode array.

[0025] Refer to Figure 3 This is the schematic diagram of a step-by-step search type artificial neural electrode positioning method proposed in this embodiment, including a micro-driver control and execution module, a signal acquisition and processing module, and a machine learning model analysis and discrimination module.

[0026] I. Micro-driver control and execution module The central control chip converts the digital feedback signal into an analog signal, which is then amplified by the piezoelectric stack driver chip and the coil current chip to become the driving signals for the piezoelectric stack and the coil respectively. The micro-driver drives and controls each electrode in the artificial neural electrode array to perform corresponding independent target actions under the excitation of the driving signals.

[0027] II. Signal acquisition and processing module The micro-driver drives and controls multiple electrodes to be implanted into the target brain region of a mouse for neuron search and positioning and signal acquisition by using the step-by-step search type positioning method. Refer to Figure 4This is the operation flow chart for the positioning, acquisition, and discrimination of neuron signals by the micro-driver in this embodiment. Utilizing the advantage that the micro-driver can flexibly and independently drive the movement of multiple electrodes, multiple electrodes can simultaneously perform step-by-step searches for neurons on one brain region. The specific process of neuron positioning for the electrodes is as follows: The electrodes are rapidly implanted into the target brain region with a large step size ∆h1 and then step downward for searches with the step size of the neuron scale ∆h in the brain region. If an effective neuron is located by the electrode, the electrode remains in place for neuron signal acquisition. If the acquisition time t is less than the preset acquisition signal duration t0 of the experiment, the electrode continues to remain in place for neuron signal acquisition. Otherwise, it means that the signal acquisition task with the preset duration has been completed, and the electrode steps upward continuously until it exits the target brain region. If the electrode fails to successfully locate an effective neuron, it is judged at each step whether the total travel h0 of the target area has been completed. If the step travel h is less than the total travel h0 of the target area, the electrode continues to step downward for searches. Otherwise, it means that the neuron search and positioning task for the total travel of the target area has been completed, and the electrode steps upward continuously until it exits the target brain region. If there are no neurons in a specific brain region within the target brain region and no effective neuron is located after the electrode steps through the total travel h0 of the target area, the electrode steps upward continuously until it exits the target brain region. Then, the acquired neuron signals are filtered, rectified, and amplified by a preprocessing chip, and analog-to-digital conversion is performed to convert the neuron signals into data.

[0028] III. Machine Learning Model Analysis and Discrimination Module The data is input into an ensemble learning model trained and constructed based on neuron signal characteristics before the entire control loop. Refer to Figure 5 This is the schematic diagram of the ensemble learning model for neuron signal feature recognition. The overall construction idea is as follows: First, neuron signal sample data is obtained, and the obtained neuron signal data is preprocessed with gain to obtain multiple training samples, forming a training sample set. Among them, the preprocessing includes the missing data processing process and the normalization processing process. The training sample set is divided according to a ratio, where 70% is used to generate the training set for model training, and 30% is used to generate the test set for model testing. The deep residual shrinkage neural network is trained with the training set, and the trained deep residual shrinkage neural network model is saved. Then, the training set and the test set are input into the trained deep residual shrinkage neural network model for feature extraction, thereby obtaining a new training set and a new test set. The ensemble learning model is trained with the new training set to obtain the trained ensemble learning model, and the trained ensemble learning model is used to test the neuron signal features in the new test set to judge the accuracy of the model. By repeatedly adjusting the parameters and optimizing the model parameters, an ensemble learning model based on neuron signal characteristics is finally constructed.

[0029] Then, the integrated learning model based on neuron signal features is used to discriminate the quality of neuron signals multiple times: If the electrode successfully locates valid neurons, a digital signal sequence for keeping the electrode in place is generated, and the electrode collects neuron signals. If the acquisition time t is less than the preset acquisition signal duration t0, a digital signal sequence for keeping the electrode in place is generated, and the electrode continues to collect neuron signals; otherwise, it represents that the signal acquisition task of the preset duration has been completed, and a digital signal sequence for the electrode to continuously step upward until it exits the target brain region is generated; If the electrode fails to successfully locate valid neurons, it is judged at each step whether the total travel h0 of the target area has been completed. If the step travel h is less than the total travel h0 of the target area, a digital signal sequence for the electrode to continue stepping downward to search is generated; otherwise, it represents that the neuron search and positioning task of the total travel of the target brain region has been completed, and a digital signal sequence for the electrode to continuously step upward until it exits the target brain region is generated; If there are no neurons in a specific brain region within the target brain region and no valid neurons are located after the electrode steps and searches the total travel h0 of the target area, a digital signal sequence for continuously stepping upward until it exits the target brain region is generated. Then, the discrimination result is combined with state parameters such as the preset acquisition signal duration t0 and the total travel h0 of the target area to generate a digital feedback signal, which is finally transmitted to the central control chip for the next control cycle.

[0030] As can be seen from the above embodiments, the present invention proposes a step-by-step search artificial neural electrode positioning method based on machine learning. This method can effectively utilize the step-by-step search strategy and machine learning algorithm to quickly and accurately locate neurons in the target brain region, greatly improving the positioning accuracy, accuracy, efficiency, overall success rate and safety of the artificial neural electrode array. The implementation of the present invention will provide new ideas and methods for the further development of artificial neural electrode technology, have great application potential, and are expected to be widely used in the fields of neuroscience, brain science, biomedicine, etc.

[0031] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A step-by-step search artificial neural electrode positioning method based on machine learning, characterized in that The method includes a micro - driver control and execution module, a signal acquisition and processing module, and a machine - learning model analysis and discrimination module. The specific steps are as follows: S1: The central control chip converts the digital feedback signal into an analog signal; S2: The analog signal is amplified by a piezoelectric stack driving chip and a coil current chip respectively to generate driving signals for the piezoelectric stack and the coil; S3: The micro - driver drives and controls each electrode in the artificial neural electrode array to perform corresponding independent target actions under the excitation of the driving signal; S4: The micro - driver drives and controls multiple electrodes to be implanted into the target brain region of a mouse for neuron search, positioning and signal acquisition by a step - by - step search - type positioning method; S5: The collected neuron signals are filtered, rectified and amplified by a pre - processing chip, and are converted into data through analog - to - digital conversion; S6: The data is input into an ensemble learning model trained and constructed based on neuron signal characteristics before the entire control loop; S7: The ensemble learning model is used to perform real - time discrimination on the quality of neuron signals, and the discrimination result is combined with state parameters such as a preset signal acquisition duration t0 and a total travel distance h0 of the target area to generate a digital feedback signal, which is finally transmitted to the central control chip for the next control loop.

2. The method for positioning a step-search artificial neural electrode based on machine learning according to claim 1, wherein In step S7, the specific steps of using the ensemble learning model to perform real - time discrimination on the quality of neuron signals are as follows: If the electrode successfully locates an effective neuron, a digital signal sequence for the electrode to remain in place is generated, and the electrode collects neuron signals. If the acquisition time t is less than the preset signal acquisition duration t0, a digital signal sequence for the electrode to remain in place is generated, and the electrode continues to collect neuron signals; otherwise, it means that the signal acquisition task of the preset duration has been completed, and a digital signal sequence for the electrode to step upward continuously until it exits the target brain region is generated; If the electrode fails to successfully locate an effective neuron, it is judged at each step whether the total travel distance h0 of the target area has been completed. If the step travel distance h is less than the total travel distance h0 of the target area, a digital signal sequence for the electrode to continue to step downward for search is generated; otherwise, it means that the neuron search and positioning task of the total travel distance of the target area has been completed, and a digital signal sequence for the electrode to step upward continuously until it exits the target brain region is generated.

Images