An intelligent control method and system based on in-vitro cultured biological neuron networks

Abstract

The application provides an intelligent control method and system based on in-vitro cultured biological neuron network, and the steps of the method include: extracting and culturing neurons in-vitro of organisms, and constructing a biological chip through a microelectrode array chip; selecting an electrode area of the biological chip based on a firing rate scanned out by a channel of the biological chip; determining a relative distance based on a position of a subject and a position of a target in a real-time task picture to be judged, selecting an electrode area one or an electrode area two for applying stimulation to the biological chip and a stimulation intensity based on the relative distance; obtaining feedback of the biological chip to the stimulation and constructing an initial feedback vector, calculating a significance parameter for each channel in the initial feedback vector, selecting each channel based on the significance parameter of each channel, and constructing a channel vector based on the selected channels; and inputting the channel vector into a pre-trained classification model to output an action through the classification model.

Description

Technical Field

[0001] This invention relates to the field of biological neuron technology, and in particular to an intelligent control method and system based on in vitro cultured biological neuron networks. Background Technology

[0002] In recent years, with the rapid development of brain-inspired intelligence and neural engineering, in vitro cultured biological neural networks, as a novel information processing carrier, have gradually demonstrated enormous potential in neural signal modeling, intelligent decision-making and control, and biocomputing, becoming one of the important directions in brain-inspired system research. In vitro biological neural networks refer to neural networks formed by the self-organization and connection of biological neurons in a non-physiological environment through in vitro culture technology. These networks not only possess the nonlinear computing capabilities and synaptic plasticity unique to biological nervous systems, but can also be precisely manipulated and observed under experimental conditions. Compared to traditional artificial neural networks, in vitro biological neural networks exhibit unique advantages in energy efficiency, parallel processing, real-time learning, and autonomous adaptation, thus becoming an important breakthrough in the intersection of brain science research and artificial intelligence.

[0003] In constructing real-time interactions between in vitro biological neural networks (BNNs) and external systems, electrical stimulation technology is one of the most mature and widely used methods. This method applies electrical signals to neurons through external electrodes, inducing changes in membrane potential and thus generating a response in the network. In particular, multi-electrode array (MEA) technology, with its high temporal and spatial resolution and multi-channel synchronous stimulation and recording capabilities, has become an indispensable core interface tool in BNN research. It not only supports high-precision observation and quantitative analysis of neural activity but also provides an efficient channel for information interaction between BNNs and virtual systems, enabling neural networks to possess dynamic processing capabilities in the perception-feedback closed loop.

[0004] In recent years, leveraging the MEA platform, Brain Neural Networks (BNNs) have demonstrated excellent plasticity and intelligent behavior in various typical interactive tasks. For example, the Brainoware system, by culturing human brain organoids on MEAs and constructing a hardware architecture based on reservoir computing, successfully achieved speech recognition and nonlinear equation prediction tasks, proving the memory and mapping capabilities of BNNs in temporal data processing. However, existing in vitro neurons struggle to locate representative neurons in the final actual control, leading to a decrease in control accuracy. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide an intelligent control method and system based on in vitro cultured biological neural networks to eliminate or improve one or more defects existing in the prior art.

[0006] One aspect of the present invention provides an intelligent control method based on in vitro cultured biological neural networks, the method comprising the following steps:

[0007] Neurons from organisms are extracted and cultured in vitro, and biochips are constructed using microelectrode array chips;

[0008] The biochip is scanned, and the electrode regions of the biochip are selected based on the firing rate obtained from the channel scanning of the biochip. The electrode regions of the biochip include electrode region one and electrode region two.

[0009] The real-time task image to be judged is acquired, and the relative distance between the position of the body and the position of the target body in the real-time task image to be judged is determined. Based on the relative distance, electrode region one or electrode region two for applying stimulation to the biochip, as well as the stimulation intensity, are selected.

[0010] The feedback of the biochip to the stimulus is obtained and constructed as an initial feedback vector. For each channel in the initial feedback vector, a significance parameter is calculated. Each channel is selected based on the significance parameter of each channel. A channel vector is constructed based on the selected channels.

[0011] The channel vector is input into a pre-trained classification model, and the classification model outputs the action.

[0012] This scheme, employing the aforementioned approach, induces in vitro biological neural networks (BNNs) to generate specific responses to external inputs, achieving dynamic control of the neural state space and providing stable, distinguishable high-dimensional state vectors for intelligent control systems. In this scheme, BNNs are modeled as a physical implementation of a reservoir computing (RC) system. Reservoir computing is a structure oriented towards time-series data processing. Its core mechanism lies in mapping inputs to a high-dimensional state space through a reservoir with nonlinear dynamic characteristics, while the output is extracted by a linearly trainable readout layer. Traditional RC systems often rely on recurrent neural networks (RNNs) as reservoirs, while in vitro BNNs, due to their inherent biophysical properties, become an ideal carrier for constructing physical reservoirs. In vitro BNNs possess two key characteristics: nonlinear computational capability and network plasticity. Neurons integrate spatial and temporal information through complex synaptic networks, exhibiting highly nonlinear signal processing capabilities. At the same time, the neural state undergoes temporary changes after receiving input and gradually returns to the ground state without continuous stimulation, demonstrating typical short-term memory characteristics. Furthermore, the saliency parameters of each channel can be calculated through feedback from the biochip, and the channel vector can be selected to locate representative neurons and ensure control precision.

[0013] In some embodiments of the present invention, in the step of obtaining the feedback of the biochip to the stimulus and constructing it as an initial feedback vector, and calculating the significance parameter for each channel in the initial feedback vector:

[0014] The channel number of each Spike response that occurs within the first preset time period after the biochip is stimulated is used to construct the initial feedback vector;

[0015] For each channel in the initial feedback vector, a significance parameter is calculated based on the number of Spike responses for that channel in the initial feedback vector and the number of Spike responses for that channel during a second preset time period before it is stimulated.

[0016] In some embodiments of the present invention, in the step of calculating the significance parameter based on the number of Spike responses of the channel in the initial feedback vector and the number of Spike responses of the channel during a second preset time period before stimulation, the significance parameter is calculated using the following formula:

[0017]

[0018] in, Indicates channel The number of Spike responses in the initial feedback vector; Indicates channel The number of Spike responses in the second preset time period before being stimulated. Indicates channel The significance parameter.

[0019] In some embodiments of the present invention, in the step of selecting each channel based on the significance parameter of each channel and constructing a channel vector based on the selected channel, the significance parameter is compared with a preset threshold to determine the selected channel, and the channel number of the selected channel is used as the value of each dimension in the channel vector to obtain the channel vector.

[0020] In some embodiments of the present invention, the steps of extracting and culturing neurons from an organism in vitro, and constructing a biochip using a microelectrode array chip include:

[0021] After peeling off the fetal rat cortical tissue and cutting it into small pieces, the tissue was washed in DMEM culture medium and then transferred to a digestion solution containing 0.125% trypsin and deoxyribonuclease I. The tissue was digested at 37°C for 20 minutes.

[0022] After complete dissociation by pipetting, a single-cell suspension was obtained, and the cells were separated by centrifugation at 80×g for 5 minutes. The supernatant was then removed.

[0023] The obtained cells were resuspended in a complete medium containing Neurobasal-Plus medium, B27-Plus, GlutaMax, and penicillin-streptomycin antibiotics;

[0024] Using a multi-electrode array chip, cells were seeded and an in vitro cultured biological neural network was constructed. The chip surface was cleaned by treating it with freshly prepared 1% Terg-a-zyme solution at room temperature for 2 hours. After that, it was rinsed three times with deionized water and dried thoroughly. The dried chip was sterilized in 70% ethanol for 12 to 16 hours, rinsed three times with sterile deionized water, and then placed in a 90 mm sterile culture dish. A high humidity chamber was constructed by using a 30 mm moistening dish. 0.6 mL of complete culture medium was added to each chip for pre-culture and then placed in a 37°C, 5% CO2 incubator for two days to improve surface compatibility.

[0025] After pre-culture, the chip surface was subjected to double coating treatment. First, it was incubated with 0.1 mg / mL poly-D-lysine (PDL, Sigma-Aldrich, #P6407) at 37°C for 3 hours, rinsed three times with sterile water and dried. Then, it was incubated with 0.02 mg / mL laminin at 37°C for 1 hour as the second coating layer. After incubation, no washing was required, and cells were directly seeded to obtain the biochip.

[0026] In some embodiments of the present invention, during cell seeding, 50 μL of neuronal cell suspension is dropped onto the central electrode area of ​​the chip. The suspension concentration is set between 12,000 and 16,000 cells / μL depending on the cell type. After seeding, the chip is covered with a sterile chip cap and incubated at 37°C for 2 hours to promote cell adhesion. Subsequently, 0.6 mL of intact culture medium is slowly added along the side of the chip. The culture environment is controlled at 37°C, 5% CO2, and relative humidity greater than 95%.

[0027] After cell adhesion is complete, the cell culture enters the routine culture stage: starting from day 1, the culture medium is changed three times a week, with 50% of the original volume replaced each time. Neural electrical activity is recorded 4 to 24 hours after changing the culture medium, and the time interval between replacement and recording is kept as consistent as possible. As the culture time progresses, neurons gradually grow, differentiate, and form a stable synaptic connection network. An in vitro neural network with bioelectrical activity is formed on the chip, providing a basis for subsequent electrical stimulation, training, and control experiments.

[0028] In some embodiments of the present invention, in the step of selecting the electrode region of the biochip based on the firing rate obtained from channel scanning of the biochip:

[0029] The overall activity of the biochip is scanned and recorded, and the spontaneous discharge rate of each channel is calculated;

[0030] Based on the discharge activity, the channels with the highest discharge rate and the first preset number of channels are selected as candidate recording channels, and a two-dimensional spatial discharge heat map is constructed.

[0031] A two-dimensional discharge density search method based on sliding windows is adopted to divide the two-dimensional spatial discharge heat map into non-overlapping windows of fixed size, and to calculate the average discharge rate of the channel by traversing each window.

[0032] All windows are sorted in descending order of average discharge rate. The window region in the first interval is selected as the candidate stimulation region, and two candidate stimulation regions with a distance greater than the first distance are selected as electrode region one and electrode region two, respectively.

[0033] In some embodiments of the present invention, the step of selecting the electrode regions of the biochip based on the firing rate obtained by scanning the channels of the biochip further includes selecting electrode region three and electrode region four from the candidate stimulation regions, wherein electrode region three is used to apply stimulation based on a specific category of target body; and electrode region four is used to apply stimulation based on multiple target bodies.

[0034] In some embodiments of the present invention, the step of selecting the electrode region of the biochip based on the firing rate obtained by scanning the channels of the biochip further includes selecting electrode region five from the candidate stimulation regions, the electrode region five being used to apply stimulation after the body action is successful.

[0035] In some embodiments of the present invention, electrode region three is used to represent a target: if the target is of a special category, a 100mV pulse is synchronously applied to the third electrode; otherwise, it is not activated, thus guiding priority response to critical targets. Electrode region four is used to encode local environmental complexity: when multiple targets exist simultaneously in the scene, a 200mV pulse is applied to electrode region four; if there is only one target, it is not activated. This encoding can guide BNNs to enhance processing priority or increase response intensity in multi-target situations, helping the system maintain a stable control strategy in complex environments. Electrode region five carries historical hit feedback: if the previous action was successful, a 200mV reward pulse is applied to electrode region five 1 second after the action is completed, simulating short-term memory modulation.

[0036] In some embodiments of the present invention, the method further includes training the biochip, the steps of which include:

[0037] Symmetrical bidirectional pulses with the same preset width and multiple preset pulse amplitudes are applied to electrode region one and electrode region two respectively to form multiple stimulation modes. Each mode is repeated at equal time intervals. Multiple stimulation modes are performed in sequence to form one round of training.

[0038] In each round of training, a perturbation stimulus mechanism is introduced. A preset number of random repetitive pulses are added between each training mode. The pulse amplitude is within a preset range, the interval is the same, and the stimulus form is a symmetrical biphasic pulse.

[0039] A second aspect of the present invention also provides an intelligent control system based on an in vitro cultured biological neural network. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system performs the steps of the method described above.

[0040] A third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aforementioned intelligent control method based on in vitro cultured biological neural networks.

[0041] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.

[0042] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0043] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.

[0044] Figure 1 This is a schematic diagram of one implementation of the intelligent control method based on in vitro cultured biological neural networks according to this scheme;

[0045] Figure 2 This is a schematic diagram of the overall architecture of the intelligent control method based on in vitro cultured biological neuronal networks in this scheme;

[0046] Figure 3 This is a schematic diagram of the closed-loop control in this scheme;

[0047] Figure 4 This is a schematic diagram of Spike's distribution rate activity.

[0048] Figure 5 This is a schematic diagram illustrating the training of the biochip in this scheme;

[0049] Figure 6 This is a statistical diagram illustrating the discharge activity of the biochip in this scheme;

[0050] Figure 7 This is a histogram diagram illustrating the electrode channel distribution in this scheme;

[0051] Figure 8 This is a schematic diagram of the electrode channel correlation thermogram for this scheme;

[0052] Figure 9 This is a comparison chart of the average correlation coefficients of this scheme;

[0053] Figure 10 This is a comparison chart of the overall efficiency of this solution;

[0054] Figure 11 This is a comparison chart of the modular coefficients of this scheme;

[0055] Figure 12 The diagram shows the accuracy of random stimulus classification in this scheme. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0057] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0058] Existing technologies generally suffer from problems such as insufficient utilization of stimulus information, inadequate mining of spatiotemporal features, and limited task complexity, and have not yet fully demonstrated their potential in dynamic perception and adaptive learning.

[0059] like Figure 1-3 As shown, this invention proposes an intelligent control method based on in vitro cultured biological neural networks. The steps of this method include:

[0060] Step S100: Extract neurons from the organism and culture them in vitro, and construct a biochip using a microelectrode array chip;

[0061] Step S200: Scan the biochip and select the electrode region of the biochip based on the emission rate obtained from the channel scanning of the biochip. The electrode region of the biochip includes electrode region one and electrode region two.

[0062] Step S300: Obtain the real-time task screen to be judged, determine the relative distance between the position of the body and the position of the target body in the real-time task screen to be judged, select electrode region one or electrode region two to apply stimulation to the biochip and the stimulation intensity based on the relative distance.

[0063] The real-time task screen is a captured image of a dynamic task scene, which can be a two-dimensional or three-dimensional image.

[0064] Step S400: Obtain the feedback of the biochip to the stimulus and construct it as an initial feedback vector. Calculate the significance parameter for each channel in the initial feedback vector. Select each channel based on the significance parameter of each channel and construct a channel vector based on the selected channels.

[0065] In the specific implementation process, to ensure the consistency of input feature dimensions, when the number of effective channels is less than 100, empty channels (zero vectors) are used to fill the gaps; if it exceeds 100, the top 100 channels are retained according to their significance scores to construct channel vectors. The final constructed channel number sorting vector not only reflects the firing response intensity of neurons to specific stimuli, but also covers their selective characteristics in spatial distribution, exhibiting good distinguishability and can be used for subsequent classification decoding and behavior control strategy implementation.

[0066] Step S500: Input the channel vector into the pre-trained classification model, and output the action through the classification model.

[0067] Specifically, the Bagging Tree algorithm is used as the classification model. This method is particularly suitable for neural data classification tasks with limited samples and high feature dimensionality. Through three rounds of training, such as... Figure 8 , 9 As shown in Figures 10, 11, and 12, the system establishes a label mapping relationship between each stimulus pattern and its corresponding behavioral instruction. After training, the classification model is solidified and invoked in the real-time recognition stage, inputting the neural features corresponding to each stimulus response into the classifier and outputting the corresponding action.

[0068] Figure 8 and 9 The heatmaps and average correlation coefficient changes between electrode channels are presented for four training phases (n = 8). Overall, the level of coordinated firing between neurons shows a continuous upward trend during training. (Before training...) Figure 8 As shown in (a), the correlation between channels is already at a high level, reflecting a certain degree of synchronization and activity redundancy in the biological neural network. With the progress of external stimulus training... Figure 8 As shown in (b), (c), and (d), the average correlation coefficient increased after each round of training, indicating a continuous strengthening of the co-firing relationship between neurons. This trend suggests that during the adaptation to repetitive stimuli, the network's internal connections are gradually optimized and reorganized, forming a tighter and more efficient co-firing pattern. The continuous enhancement of firing correlation not only reflects the remodeling of neural activity patterns but also suggests that training promotes the improvement of the network's functional integration ability. These results reveal the activity plasticity of in vitro biological neural networks under external intervention and the potential for training-driven functional remodeling.

[0069] Figure 10The changes in global efficiency (n = 8) of the neural network before training and after three rounds of training are shown. Experimental results show that global efficiency improves to varying degrees after each training round, exhibiting a continuous upward trend. As an important topological indicator for measuring the information integration ability of a neural network, global efficiency reflects the inverse value of the average shortest path between nodes in the network; a higher value indicates more efficient information propagation throughout the network. This result indicates that under the influence of repeated stimulus training, the dispersed neurons within the network gradually establish a tighter connection structure, thereby enhancing the communication efficiency and collaborative processing ability between regions. The gradual improvement in global efficiency also reflects the continuous optimization of the neural network's overall information integration ability and dynamic response performance under the guidance of external stimuli, demonstrating good functional plasticity.

[0070] Figure 11 This study demonstrates the changing trend of the modularity coefficient of the neural network before training and after three rounds of training (n = 8). Experimental results show that the modularity coefficient gradually decreases with increasing training rounds, indicating that the network gradually transitions from an initial state with distinct functional partitions and obvious local module collaboration to a more integrated and globally collaborative structure. The modularity coefficient reflects the degree of dense intra-module connections and sparse inter-module connections after the network is divided into several modules (clusters). High modularity means that the subgroups in the network are relatively independent, with information exchange mostly occurring locally; while low modularity indicates that the network tends towards overall collaboration and information integration. The gradual decrease in modularity indicates that, under the influence of repeated training stimuli, the neural network breaks down the original local information silos, promotes collaborative interaction between different functional areas, and thus supports a more efficient and unified information processing mechanism. This further verifies the process of functional structure reorganization and integration capability improvement of the network during training, demonstrating strong plasticity.

[0071] Figure 12This study demonstrates the changes in classification accuracy of the neural network across four training phases, evaluated using 10-fold cross-validation. The results show that as training progresses, the network's classification accuracy exhibits a significant upward trend: from 66.5% in the untrained phase to 84.5% after the third training phase, reflecting the continuously enhanced ability of the neural network to recognize and respond to external stimulus patterns. This trend indicates that external stimulus training, especially under perturbation stimuli, effectively promotes the plasticity and functional reconstruction of the neural network. At the structural level, the aforementioned analysis reveals that training leads to more integrated connection patterns in the network (increased global efficiency, decreased modularity coefficient) and improved functional synergy between channels (enhanced correlation). These changes provide a more favorable foundation for information transmission and computational power support for the neural network. At the functional level, the improvement in classification accuracy indicates that the network has not only undergone structural optimization but also demonstrated stronger information representation capabilities and output discriminability at the task level. The firing patterns of neurons under different stimulus conditions tend to stabilize and become distinguishable, and the network as a whole exhibits an evolution from "spontaneous response" to "task-driven response."

[0072] Joint analysis: The combined results of the discharge correlation heatmap, global efficiency, modularity coefficient, and electrode density distribution show that the in vitro cultured biological neural network exhibits significant structural remodeling and functional plasticity during training. Initially, the network displays strong cooperative synchronization characteristics, with high redundancy in inter-channel discharge activity, significant modularity, and an overall localized, partitioned, primary integration state. After the first training cycle, the correlation slightly increases, the modularity coefficient slightly decreases, and the global efficiency begins to rise. Simultaneously, the degree values ​​of some electrodes increase, indicating an initial trend of decorrelation between neuronal responses and a reshaping of information transmission paths. This stage can be considered the initial stage of network functional decoupling and reconstruction. As the number of training rounds increases, the correlation coefficient rises again, the global efficiency continues to increase, the modularity coefficient further decreases, and the overall electrode density improves, indicating that the network has undergone a dynamic adjustment process from "desynchronization to cooperative reconstruction." More functional connections are established between neurons, breaking the original modular isolation pattern and giving the network stronger information integration capabilities and global coordination characteristics.

[0073] This approach not only demonstrates the technical feasibility of using biological neural networks for complex behavior recognition and control tasks, but also reflects the advantages of deep integration of "artificial systems + biological networks". BNNs, as a physical reservoir, possess high parallelism, low power consumption, and self-organizing structural characteristics. They can achieve effective recognition and control with small sample sizes without deep training of network weights, showcasing the potential application value of natural neural systems in intelligent control. Simultaneously, the modulation stimulation mechanism introduced by the system further enhances the learning ability and task adaptability of BNNs, enabling them to achieve plasticity enhancement and experience accumulation through key behavioral feedback, thus initially possessing adaptive regulation capabilities based on biological feedback. In the future, this method is expected to be extended to more complex task environments, such as multi-objective coordinated control, biomimetic robot navigation, and human-computer interaction systems, driving the development of brain-like intelligence towards practical application.

[0074] This scheme, employing the aforementioned approach, induces in vitro biological neural networks (BNNs) to generate specific responses to external inputs, achieving dynamic control of the neural state space and providing stable, distinguishable high-dimensional state vectors for intelligent control systems. In this scheme, BNNs are modeled as a physical implementation of a reservoir computing (RC) system. Reservoir computing is a structure oriented towards time-series data processing. Its core mechanism lies in mapping inputs to a high-dimensional state space through a reservoir with nonlinear dynamic characteristics, while the output is extracted by a linearly trainable readout layer. Traditional RC systems often rely on recurrent neural networks (RNNs) as reservoirs, while in vitro BNNs, due to their inherent biophysical properties, become an ideal carrier for constructing physical reservoirs. In vitro BNNs possess two key characteristics: nonlinear computational capability and network plasticity. Neurons integrate spatial and temporal information through complex synaptic networks, exhibiting highly nonlinear signal processing capabilities. At the same time, the neural state undergoes temporary changes after receiving input and gradually returns to the ground state without continuous stimulation, demonstrating typical short-term memory characteristics. Furthermore, the saliency parameters of each channel can be calculated through feedback from the biochip, and the channel vector can be selected to locate representative neurons and ensure control precision.

[0075] In some embodiments of the present invention, in the step of obtaining the feedback of the biochip to the stimulus and constructing it as an initial feedback vector, and calculating the significance parameter for each channel in the initial feedback vector:

[0076] The channel number of each Spike response that occurs within the first preset time period after the biochip is stimulated is used to construct the initial feedback vector;

[0077] For each channel in the initial feedback vector, a significance parameter is calculated based on the number of Spike responses for that channel in the initial feedback vector and the number of Spike responses for that channel during a second preset time period before it is stimulated.

[0078] In the specific implementation process, within a time window of 100–200 ms after the application of stimulation to the biochip, the firing data of BNNs are collected in real time through the MEA platform. Subsequently, feature extraction is performed on the collected neural activity using a dynamic channel screening method based on a combination of response saliency and spatial sparsity. This method first counts the baseline firing count of each channel within 100 ms before stimulation and the response firing count within 100–200 ms after stimulation, and then calculates its response saliency parameter.

[0079] In some embodiments of the present invention, in the step of calculating the significance parameter based on the number of Spike responses of the channel in the initial feedback vector and the number of Spike responses of the channel during a second preset time period before stimulation, the significance parameter is calculated using the following formula:

[0080]

[0081] in, Indicates channel The number of Spike responses in the initial feedback vector; Indicates channel The number of Spike responses in the second preset time period before being stimulated. Indicates channel The significance parameter.

[0082] In some embodiments of the present invention, in the step of selecting each channel based on the significance parameter of each channel and constructing a channel vector based on the selected channel, the significance parameter is compared with a preset threshold to determine the selected channel, and the channel number of the selected channel is used as the value of each dimension in the channel vector to obtain the channel vector.

[0083] All channels with significance scores exceeding a set threshold (e.g., 1.0) are retained to form a candidate set, and spatial sparsity constraints are introduced: the MEA chip is divided into an equidistant spatial grid, or the K-means clustering algorithm is used to partition the channels, and the channel with the highest response score is selected in each sub-region to ensure the balance of the final channel set in spatial distribution.

[0084] In some embodiments of the present invention, the steps of extracting and culturing neurons from an organism in vitro, and constructing a biochip using a microelectrode array chip include:

[0085] After peeling off the fetal rat cortical tissue and cutting it into small pieces, the tissue was washed in DMEM culture medium and then transferred to a digestion solution containing 0.125% trypsin and deoxyribonuclease I. The tissue was digested at 37°C for 20 minutes.

[0086] After complete dissociation by pipetting, a single-cell suspension was obtained, and the cells were separated by centrifugation at 80×g for 5 minutes. The supernatant was then removed.

[0087] The obtained cells were resuspended in a complete medium containing Neurobasal-Plus medium, B27-Plus, GlutaMax, and penicillin-streptomycin antibiotics;

[0088] Specifically, fetal cortical tissue from pregnant Beijing Spaford rats at 15.5 days of gestation was used. To ensure cell viability and tissue integrity, the entire tissue acquisition process was performed on ice. After being peeled, the cortical tissue was cut into small pieces and gently washed twice in DMEM culture medium from Zhongke Maichen. Then, it was transferred to a digestion solution containing 0.125% trypsin (#CC017) and deoxyribonuclease I (#D8071) and digested at 37°C for 20 minutes. After thorough dissociation by pipetting, a single-cell suspension was obtained and centrifuged at 80×g for 5 minutes to separate the cells, and the supernatant was removed. The obtained cells were resuspended in complete medium containing Neurobasal-Plus medium (Gibco, #A3582901), B27-Plus (#A3582801), GlutaMax (#35050061), and penicillin-streptomycin antibiotics (#CC004).

[0089] Using a multi-electrode array chip, cells were seeded and an in vitro cultured biological neural network was constructed. The chip surface was cleaned by treating it with freshly prepared 1% Terg-a-zyme solution at room temperature for 2 hours. After that, it was rinsed three times with deionized water and dried thoroughly. The dried chip was sterilized in 70% ethanol for 12 to 16 hours, rinsed three times with sterile deionized water, and then placed in a 90 mm sterile culture dish. A high humidity chamber was constructed by using a 30 mm moistening dish. 0.6 mL of complete culture medium was added to each chip for pre-culture and then placed in a 37°C, 5% CO2 incubator for two days to improve surface compatibility.

[0090] After pre-culture, the chip surface was subjected to double coating treatment. First, it was incubated with 0.1 mg / mL poly-D-lysine (PDL, Sigma-Aldrich, #P6407) at 37°C for 3 hours, rinsed three times with sterile water and dried. Then, it was incubated with 0.02 mg / mL laminin at 37°C for 1 hour as the second coating layer. After incubation, no washing was required, and cells were directly seeded to obtain the biochip.

[0091] In some embodiments of the present invention, during cell seeding, 50 μL of neuronal cell suspension is dropped onto the central electrode area of ​​the chip. The suspension concentration is set between 12,000 and 16,000 cells / μL depending on the cell type. After seeding, the chip is covered with a sterile chip cap and incubated at 37°C for 2 hours to promote cell adhesion. Subsequently, 0.6 mL of intact culture medium is slowly added along the side of the chip. The culture environment is controlled at 37°C, 5% CO2, and relative humidity greater than 95%.

[0092] After cell adhesion is complete, the cell culture enters the routine culture stage: starting from day 1, the culture medium is changed three times a week, with 50% of the original volume replaced each time. Neural electrical activity is recorded 4 to 24 hours after changing the culture medium, and the time interval between replacement and recording is kept as consistent as possible. As the culture time progresses, neurons gradually grow, differentiate, and form a stable synaptic connection network. An in vitro neural network with bioelectrical activity is formed on the chip, providing a basis for subsequent electrical stimulation, training, and control experiments.

[0093] Specifically, using MaxOne multi-electrode array (MEA) chips from MaxWell Biosystems, neuronal cells were seeded to construct in vitro cultured biological neural networks. The chip surface was cleaned by treating with freshly prepared 1% Terg-a-zyme solution at room temperature for 2 hours, followed by rinsing three times with deionized water and thorough drying. The dried chips were then sterilized in 70% ethanol for 12 to 16 hours, rinsed three times with sterile deionized water, and placed in 90 mm sterile culture dishes, along with 30 mm moistening dishes to construct a high-humidity chamber. Each chip was pre-cultured with 0.6 mL of complete culture medium and placed in a 37°C, 5% CO2 incubator for two days to improve surface compatibility. After pre-culture, the chip surface underwent a double coating treatment. First, cells were incubated with 0.1 mg / mL poly-D-lysine (PDL, Sigma-Aldrich, #P6407) at 37°C for 3 hours, then rinsed three times with sterile water and dried. Next, cells were incubated with 0.02 mg / mL laminin (Sigma-Aldrich, #L2020) at 37°C for 1 hour as a second coating layer. No washing was performed after incubation, and cells were directly seeded. Cell seeding was performed using the "DotPlating" method, with 50 μL of neuronal cell suspension added dropwise to the central electrode area of ​​the chip. The suspension concentration was set between 12,000 and 16,000 cells / μL depending on the cell type. After seeding, the chip was covered with a sterile cap and incubated at 37°C for 2 hours to promote cell adhesion. Then, 0.6 mL of intact culture medium was slowly added along the side of the chip. The culture environment was controlled at 37°C, 5% CO2, and a relative humidity greater than 95%. After cell adhesion, the cells entered the routine culture phase: starting from day 1, the culture medium was changed three times a week, replacing 50% of the original volume each time. Neural electrical activity was recorded 4 to 24 hours after medium change, with the time interval between medium change and recording kept as consistent as possible. As the culture progressed, neurons gradually grew, differentiated, and formed a stable synaptic network, creating an in vitro neural network with bioelectrical activity on the chip, providing a foundation for subsequent electrical stimulation, training, and control experiments.

[0094] In some embodiments of the present invention, in the step of selecting the electrode region of the biochip based on the firing rate obtained from channel scanning of the biochip:

[0095] The overall activity of the biochip is scanned and recorded, and the spontaneous discharge rate of each channel is calculated;

[0096] Based on the discharge activity, the channels with the highest discharge rate and the first preset number of channels are selected as candidate recording channels, and a two-dimensional spatial discharge heat map is constructed.

[0097] Specifically, the MaxLab Live system was first used to scan and record the overall activity of the biochip, and the spontaneous discharge rate of each electrode was calculated. Based on the discharge activity, the top 1024 channels with the highest discharge rates were selected as candidate recording channels, and a two-dimensional spatial discharge thermogram was constructed.

[0098] A two-dimensional discharge density search method based on sliding windows is adopted to divide the two-dimensional spatial discharge heat map into non-overlapping windows of fixed size, and to calculate the average discharge rate of the channel by traversing each window.

[0099] Specifically, to achieve systematic analysis of the discharge structure, the system introduces a two-dimensional discharge density search method based on a sliding window. The entire electrode array is divided into fixed-size 3×3 non-overlapping windows, and the average discharge rate of the electrodes within each window is calculated to evaluate the partial discharge density.

[0100] All windows are sorted in descending order of average discharge rate. The window region in the first interval is selected as the candidate stimulation region, and two candidate stimulation regions with a distance greater than the first distance are selected as electrode region one and electrode region two, respectively.

[0101] The first distance is 1.0 mm.

[0102] like Figure 4 As shown, further, in order to construct richer input encoding, 2 to 5 local high-density regions are selected from the candidate regions as stimulus encoding regions according to task requirements. The specific number can be dynamically set according to the task complexity.

[0103] In some embodiments of the present invention, the step of selecting the electrode regions of the biochip based on the firing rate obtained by scanning the channels of the biochip further includes selecting electrode region three and electrode region four from the candidate stimulation regions, wherein electrode region three is used to apply stimulation based on a specific category of target body; and electrode region four is used to apply stimulation based on multiple target bodies.

[0104] In some embodiments of the present invention, the step of selecting the electrode region of the biochip based on the firing rate obtained by scanning the channels of the biochip further includes selecting electrode region five from the candidate stimulation regions, the electrode region five being used to apply stimulation after the body action is successful.

[0105] In some embodiments of the present invention, based on the relative position between the target and the body in the scene, six stimulation modes (with regional and amplitude settings) are mapped to act on BNNs and induce them to generate specific discharge activities; then, discharge data are collected within a preset response window, and the sorted active channel numbers are extracted as feature vectors; the feature vectors are input into a trained Bagging Tree model for classification and recognition, and the output category labels are mapped and converted into behavior control instructions to guide the body to complete actions.

[0106] The control process constitutes a complete closed-loop neural signal pathway: input encoding → neural response acquisition → feature extraction → pattern recognition → behavioral output. From real-time changes in environmental input to the internal dynamic response of the neural network, and then to the executable control at the system behavior level, a closed-loop flow of "perception-computation-execution" is formed between the brain-like system and the external world.

[0107] like Figure 5 and 6 As shown, in some embodiments of the present invention, the method further includes training the biochip, the steps of which include:

[0108] Symmetrical bidirectional pulses with the same preset width and multiple preset pulse amplitudes are applied to electrode region one and electrode region two respectively to form multiple stimulation modes. Each mode is repeated at equal time intervals. Multiple stimulation modes are performed in sequence to form one round of training.

[0109] In each round of training, a perturbation stimulus mechanism is introduced. A preset number of random repetitive pulses are added between each training mode. The pulse amplitude is within a preset range, the interval is the same, and the stimulus form is a symmetrical biphasic pulse.

[0110] Figure 6 The study demonstrates the differences in spontaneous firing activity of neural networks under two training conditions. In the absence of added perturbation stimuli, the network's spontaneous activity exhibits numerous synchronized network bursts (SNBs), where multiple electrode channels fire densely within a similar timeframe, demonstrating a highly synchronized activation pattern, such as... Figure 6 As shown in (a). However, after the introduction of perturbation stimuli, the SNB phenomenon significantly decreased, the firing rhythm became more dispersed, and neuronal activity tended to become asynchronous, as shown in (a). Figure 6 As shown in (b), perturbation training helps to break up the synchronous discharge circuit formed by the original synaptic connections, making the network more dynamically stable, reducing structural redundancy, and providing a more stable dynamic basis for subsequent task learning.

[0111] Figure 7 The histograms of electrode channel degree distribution are shown for four training phases. In the untrained phase, such as... Figure 7As shown in (a), most electrodes exhibit low connectivity, indicating that functional connections between neurons are not yet fully established, resulting in a sparse network structure. With increasing training rounds... Figure 7 As shown in (b), (c), and (d), the overall electrode intensity shows an upward trend, with the number of high-order nodes gradually increasing, reflecting the continuous enhancement of network connection density and complexity. This change indicates that external electrical stimulation training promotes the establishment of new functional connections between neurons, and the network structure gradually evolves from loose to integrated. The increase in electrode intensity suggests that more neurons participate in collaborative activities, further supporting the structural plasticity and functional reconstruction capabilities of neural networks under training-driven conditions.

[0112] In summary, this invention constructs a closed-loop control system based on in vitro cultured biological neural networks (BNNs), such as... Figure 2 As shown, the system consists of four main functional modules: First, the in vitro biological neural network module: the core computing unit of the system is BNNs cultured on the MEA, which originate from neurons in the cerebral cortex of pregnant mouse embryos. During the culture process, they spontaneously form complex connection structures, possessing natural nonlinear dynamic characteristics and network plasticity. Second, the MEA platform module: enabling high-density signal acquisition and precise electrical stimulation of the BNNs, serving as a bridge connecting the neural network and the control task. Third, the recognition algorithm module: using a Bagging Tree ensemble learning model to classify and recognize the aforementioned feature vectors, mapping different neural response patterns to corresponding discrete control commands. Fourth, the scene module: this module, on the one hand, encodes the relative position information of the target and the host in the environment and generates corresponding electrical stimulation patterns, transmitting them to the MEA platform for stimulating the BNNs; on the other hand, it uses the control commands output by the recognition algorithm module to drive the host to complete movement operations, thereby achieving real-time closed-loop control of neural network response and behavior output.

[0113] In the experiment, this scheme sets the scenario as a simplified flight shooting game. The game interface is divided horizontally into four parallel tracks, and both the main aircraft (body) and the enemy aircraft (target) are restricted to moving within these tracks. Enemy aircraft are randomly generated in the upper area and move forward, without attack capabilities; the main aircraft is located in the lower area and can move left or right in units of the track based on the recognition results. The control range includes seven possible behaviors: moving left by 3, 2, or 1 space, remaining stationary, and moving right by 1, 2, or 3 spaces. This setting simplifies the complex control problem into finite state control, facilitating the effective mapping of neural network signals.

[0114] The design of the input stimuli is determined by the enemy aircraft's position relative to the host aircraft. When the enemy aircraft is positioned to the left of the host aircraft, electrical stimulation is applied to "Region 1" of the BNNs; if the enemy aircraft is positioned to the right, "Region 2" is stimulated. Furthermore, the stimulation amplitude is further set according to the horizontal distance between the host and enemy aircraft: distances of 3 units, 2 units, and 1 unit correspond to symmetrical biphasic electrical stimulation of 500 mV, 300 mV, and 100 mV, respectively; if the enemy aircraft coincides with the host aircraft (distance of 0 units), no stimulation is applied, indicating that the target is aligned and no adjustment is needed. This constructs six input stimulation patterns, comprehensively covering the six types of movement commands required by the host aircraft.

[0115] This invention also provides an intelligent control system based on in vitro cultured biological neural networks. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.

[0116] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned intelligent control method based on in vitro cultured biological neural networks. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0117] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0118] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0119] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart control method based on in vitro cultured biological neural networks, characterized in that, The steps of this method include: Neurons from organisms are extracted and cultured in vitro, and biochips are constructed using microelectrode array chips; The biochip is scanned, and the electrode regions of the biochip are selected based on the firing rate obtained from the channel scanning of the biochip. The electrode regions of the biochip include electrode region one and electrode region two. The real-time task image to be judged is acquired, and the relative distance between the position of the body and the position of the target body in the real-time task image to be judged is determined. Based on the relative distance, electrode region one or electrode region two for applying stimulation to the biochip, as well as the stimulation intensity, are selected. The feedback of the biochip to the stimulus is obtained and constructed as an initial feedback vector. For each channel in the initial feedback vector, a significance parameter is calculated. The channel number of each Spike response that occurs within a first preset time period after the biochip is stimulated is constructed as the initial feedback vector. For each channel in the initial feedback vector, a significance parameter is calculated based on the number of Spike responses of that channel in the initial feedback vector and the number of Spike responses of that channel in a second preset time period before the channel is stimulated. Each channel is selected based on its significance parameter, and a channel vector is constructed based on the selected channels. The channel vector is input into a pre-trained classification model, and the classification model outputs the action.

2. The intelligent control method based on in vitro cultured biological neural networks according to claim 1, characterized in that, In the step of calculating the significance parameter based on the number of Spike responses of the channel in the initial feedback vector and the number of Spike responses of the channel in the second preset time period before stimulation, the significance parameter is calculated using the following formula: in, Indicates channel The number of Spike responses in the initial feedback vector; Indicates channel The number of Spike responses in the second preset time period before being stimulated. Indicates channel The significance parameter.

3. The intelligent control method based on in vitro cultured biological neural networks according to claim 1 or 2, characterized in that, In the step of selecting each channel based on the significance parameter of each channel and constructing a channel vector based on the selected channel, the significance parameter is compared with a preset threshold to determine the selected channel, and the channel number of the selected channel is used as the value of each dimension in the channel vector to obtain the channel vector.

4. The intelligent control method based on in vitro cultured biological neural networks according to claim 1, characterized in that, The steps of extracting and culturing neurons from an organism in vitro, and constructing a biochip using a microelectrode array chip include: After peeling off the fetal rat cortical tissue and cutting it into small pieces, the tissue was washed in DMEM culture medium and then transferred to a digestion solution containing 0.125% trypsin and deoxyribonuclease I. The tissue was digested at 37°C for 20 minutes. After complete dissociation by pipetting, a single-cell suspension was obtained, and the cells were separated by centrifugation at 80×g for 5 minutes. The supernatant was then removed. The obtained cells were resuspended in a complete medium containing Neurobasal-Plus medium, B27-Plus, GlutaMax, and penicillin-streptomycin antibiotics; Using a multi-electrode array chip, cells were seeded and an in vitro cultured biological neural network was constructed. The chip surface was cleaned by treating it with freshly prepared 1% Terg-a-zyme solution at room temperature for 2 hours. After that, it was rinsed three times with deionized water and dried thoroughly. The dried chip was sterilized in 70% ethanol for 12 to 16 hours, rinsed three times with sterile deionized water, and then placed in a 90 mm sterile culture dish. A high humidity chamber was constructed by using a 30 mm moistening dish. 0.6 mL of complete culture medium was added to each chip for pre-culture and then placed in a 37°C, 5% CO2 incubator for two days to improve surface compatibility. After pre-culture, the chip surface was subjected to double coating treatment. First, it was incubated with 0.1 mg / mL poly-D-lysine at 37°C for 3 hours, rinsed three times with sterile water and dried. Then, it was incubated with 0.02 mg / mL laminin at 37°C for 1 hour as the second coating layer. After incubation, no washing was required, and cells were directly seeded to obtain the biochip.

5. The intelligent control method based on in vitro cultured biological neural networks according to claim 1, characterized in that, In the step of selecting the electrode region of the biochip based on the firing rate obtained from channel scanning of the biochip: The overall activity of the biochip is scanned and recorded, and the spontaneous discharge rate of each channel is calculated; Based on the discharge activity, the channels with the highest discharge rate and the first preset number of channels are selected as candidate recording channels, and a two-dimensional spatial discharge heat map is constructed. A two-dimensional discharge density search method based on sliding windows is adopted to divide the two-dimensional spatial discharge heat map into non-overlapping windows of fixed size, and to calculate the average discharge rate of the channel by traversing each window. All windows are sorted in descending order of average discharge rate. The window region in the first interval is selected as the candidate stimulation region, and two candidate stimulation regions with a distance greater than the first distance are selected as electrode region one and electrode region two, respectively.

6. The intelligent control method based on in vitro cultured biological neural networks according to claim 5, characterized in that, The step of selecting electrode regions of a biochip based on the firing rate obtained from channel scanning of the biochip further includes selecting electrode region three and electrode region four from the candidate stimulation regions, wherein electrode region three is used to apply stimulation based on a specific category of target body; and electrode region four is used to apply stimulation based on multiple target bodies.

7. The intelligent control method based on in vitro cultured biological neural networks according to claim 5, characterized in that, The step of selecting the electrode region of the biochip based on the firing rate obtained from the channel scanning of the biochip further includes selecting electrode region five from the candidate stimulation regions, the electrode region five being used to apply stimulation after the body action is successful.

8. The intelligent control method based on in vitro cultured biological neural networks according to claim 1, characterized in that, The method further includes training the biochip, the steps of which include: Symmetrical bidirectional pulses with the same preset width and multiple preset pulse amplitudes are applied to electrode region one and electrode region two respectively to form multiple stimulation modes. Each mode is repeated at equal time intervals. Multiple stimulation modes are performed in sequence to form one round of training. In each round of training, a perturbation stimulus mechanism is introduced. A preset number of random repetitive pulses are added between each training mode. The pulse amplitude is within a preset range, the interval is the same, and the stimulus form is a symmetrical biphasic pulse.

9. An intelligent control system based on in vitro cultured biological neural networks, characterized in that, The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method as described in any one of claims 1 to 8.

Images