Cloud-based neuroscience data processing methods

Abstract

This invention relates to the field of cloud-edge collaboration technology, and more particularly to a neuroscience data processing method based on a cloud platform. The method includes: generating neuronal firing timestamp sequences and behavioral timestamp sequences through a target edge computing unit for synchronization analysis; determining whether the edge device meets performance requirements based on the synchronization; generating a historical sequence of synchronization timestamps through a cloud platform computing unit; performing cluster analysis on the historical sequence of synchronization timestamps to generate a historical pattern of synchronization decay trends; determining the synchronization decay trend pattern of the target user; further determining the historical pattern corresponding to the synchronization decay trend and obtaining sample training weights; training to generate a latest decoder; and determining the execution of the latest decoder and its refresh cycle based on the stability index of the latest decoder. This invention improves the training and refresh efficiency of the user's edge device decoder.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge collaboration technology, and in particular to a neuroscience data processing method based on a cloud platform. Background Technology

[0002] With the convergence of neural engineering and cloud computing technologies, invasive brain-computer interface (BCI) systems have gradually moved from laboratory research to practical applications. These systems acquire cortical neural signals through implanted electrode arrays and decode them into motor commands, enabling users to directly control external devices such as robotic arms to perform daily tasks. Current technologies have achieved preliminary real-time acquisition and motion decoding of neural signals and established a cloud-based model training framework, which can support remote maintenance and parameter updates of BCI systems to a certain extent.

[0003] However, existing technologies generally treat the time-varying nature of neural signals as a single-dimensional noise interference, adopt a uniform full-model retraining strategy to deal with all types of decoding performance degradation, do not consider the distinguishable features presented on different degradation trajectories, and cannot utilize the degradation patterns accumulated by the population to transfer and optimize individual diagnosis and intervention strategies.

[0004] Chinese Patent Publication No. CN107212883A discloses a robotic arm writing device and control method based on EEG control, comprising: the device receiving information collected by the EEG acquisition device and processing the collected information; the specific data processing process is as follows: data filtering, wavelet filtering, feature extraction, classification, classification result conversion, and the robotic arm performing the writing action; the robotic arm receiving control instructions from the signal processing device and executing corresponding instruction actions according to the above control instructions.

[0005] Therefore, it is evident that the existing technology has the following problems: Existing technologies do not consider monitoring the performance degradation trend of target users and parallel users' edge devices through cloud platforms, nor do they consider analyzing the group performance degradation pattern and transferring learning by combining edge and cloud platforms, which leads to low efficiency in training and refreshing user edge device decoders. Summary of the Invention

[0006] To address this, the present invention provides a cloud-based neuroscience data processing method to overcome the problems in existing technologies that fail to consider monitoring the performance degradation trend of target users and parallel users' edge devices through the cloud platform, and fail to consider analyzing the group performance degradation pattern and transfer learning through the combination of edge and cloud platforms, resulting in low training and refresh efficiency of user edge device decoders.

[0007] To achieve the above objectives, the present invention provides a neuroscience data processing method based on a cloud platform, comprising: The target edge computing unit acquires neural signal data and edge device behavior data when the target user uses the edge device, in order to generate neuron firing timestamp sequences and behavior timestamp sequences, wherein the behavior data includes the three-dimensional coordinates and velocity of the robotic arm end effector; The synchronization degree of the target user is determined by analyzing the neuron firing timestamp sequence and behavior timestamp sequence based on the target edge computing unit, and then sent to the cloud platform computing unit. Based on the synchronization degree, determine whether the edge device meets the performance requirements, so as to obtain several synchronization degrees of several parallel users through the cloud platform computing unit and generate a historical sequence of synchronization degree timestamps for each parallel user. Cluster analysis is performed on the historical sequences of the synchronization timestamps to generate historical patterns of synchronization decay trends. The cloud platform computing unit obtains several historical synchronization degrees of the target user to generate a historical sequence of synchronization degree timestamps of the target user and analyzes the time warping morphology characteristics to determine the synchronization degree decay trend pattern of the target user. The time warping morphology characteristics include decay slope, abrupt change point position and fluctuation amplitude. By comparing the synchronization decay trend pattern of the target user with the historical synchronization decay trend pattern, the historical pattern corresponding to the synchronization decay trend of the target user is determined. Based on the historical pattern corresponding to the synchronicity decay trend, the cloud platform computing unit obtains the sample training weights of the corresponding parallel users, uses the neuron firing timestamp sequence to train and generate the latest decoder based on the sample training weights, and feeds it back to the target edge computing unit. During the target user's use of the edge device, the original decoder is used to execute and obtain the execution speed, while the latest decoder is run and the predicted speed is obtained. The execution speed and the predicted speed are compared to determine the stability index of the latest decoder. Based on the stability metrics of the latest decoder, the latest decoder fed back by the execution cloud platform computing unit is determined, and the refresh cycle of the decoder is determined by the cloud platform computing unit based on the synchronization decay trend pattern of the target user.

[0008] Furthermore, the process of performing synchronization analysis on neuronal firing timestamp sequences and behavioral timestamp sequences based on target edge computing units includes: The timing error is calculated based on the neuron firing timestamp sequence and the behavior timestamp sequence; The synchronization degree of the edge device is determined by comparing the timing error with a preset timing error threshold.

[0009] Furthermore, the process of determining whether edge devices meet performance requirements includes: Based on the comparison between the synchronization degree and the preset synchronization degree threshold, if the proportion of the synchronization degree greater than the preset synchronization degree threshold is greater than the preset proportion threshold, then it is determined that the edge device does not meet the performance requirements.

[0010] Furthermore, the process of determining the historical pattern of synchronicity decay trends includes: When it is determined that the edge device does not meet the performance requirements, several cluster centers are determined by performing cluster analysis on the synchronization timestamp historical sequence, and the corresponding synchronization decay trend historical pattern is determined based on each cluster center.

[0011] Furthermore, the process of determining the historical patterns corresponding to the synchronization decay trend of the target user includes: Calculate the distance between the target user's synchronization decay trend pattern and the historical synchronization decay trend pattern; Based on the minimum value of the distance, the historical pattern corresponding to the synchronization decay trend of the target user is determined.

[0012] Furthermore, the process of obtaining the sample training weights for corresponding parallel users through the cloud platform computing unit includes: Based on the historical patterns corresponding to the aforementioned synchronization decay trend, the synchronization of several parallel users after training is obtained. The training weights are obtained by comparing the synchronization degree after training with a preset synchronization degree threshold. If the proportion of samples with a synchronization degree less than the preset synchronization degree threshold is less than or equal to the preset proportion threshold, the corresponding sample training weights are obtained.

[0013] Furthermore, the process of determining the stability metrics of the latest decoder includes: Calculate the difference between several execution speeds and the predicted speeds per unit time to generate a difference sequence; The mean and standard deviation of the differences are calculated based on the difference sequence to determine the stability index of the latest decoder.

[0014] Furthermore, the process of determining the latest decoder fed back by the cloud platform computing unit includes: The stability index of the latest decoder is compared with a preset stability index threshold. If the stability index is less than the preset stability index threshold, the latest decoder fed back by the cloud platform computing unit is determined to be executed.

[0015] Furthermore, the process of determining the activation of the cloud platform computing unit to determine the decoder refresh cycle includes: The stability index of the latest decoder is compared with a preset stability index threshold. If the stability index is less than the preset stability index threshold, the cloud platform computing unit is activated to determine the refresh cycle of the decoder.

[0016] Furthermore, the process of determining the decoder's refresh cycle includes: The refresh cycle of the decoder is determined based on the slope of the target user's synchronization decay and the slope of the historical pattern corresponding to the synchronization decay trend.

[0017] Compared with existing technologies, the advantages of this invention lie in its ability to integrate neural signal acquisition, performance monitoring, group clustering, pattern matching, encoder training and verification, and adaptive refresh into a complete closed loop by constructing a collaborative processing architecture that includes edge computing units and cloud platform computing units. This enables a shift from passive, reactive maintenance to proactive, predictive management of the target user's edge devices. The edge computing unit is responsible for local real-time signal processing and decoding inference; the cloud platform computing unit is responsible for the aggregation and analysis of cross-user group data and deep model training, allowing individuals to benefit from the degradation patterns and intervention experience accumulated by parallel users. Through continuous tracking of synchronization indicators and automatic identification of decay trend patterns, performance degradation can be accurately located and targeted interventions implemented. The dynamic calculation of the decoder refresh cycle can proactively initiate the evaluation process before performance has significantly degraded, transforming the maintenance mode from remediation after problems occur to prevention before degradation occurs, thus improving the training and refresh efficiency of the user's edge device decoder.

[0018] Furthermore, this invention quantifies the temporal coupling relationship between neural signals and motion output into a continuously trackable synchronization index by aligning the neuronal firing timestamp sequence with the behavioral timestamp sequence and calculating the timing error. This provides a unified quantitative basis for subsequent performance assessment and trend analysis. By comparing the current synchronization with a dynamic threshold constructed based on the user's own stable period and using the percentage of consecutive days exceeding the limit as the judgment criterion, a quantitative assessment of the performance requirements of edge devices is achieved.

[0019] Furthermore, this invention performs cluster analysis on synchronization trajectories in parallel user groups where trigger performance does not meet requirements, obtaining degradation patterns with different morphological characteristics, thus enabling the generated historical decay trend patterns to have generalization capabilities. The time-warped morphological features corresponding to each pattern reflect the dynamic evolution law of the degradation process, providing a quantifiable comparison template for subsequent individual matching. By performing the same morphological feature extraction on the recent synchronization sequences of target users as on historical patterns, a feature basis is provided for subsequent accurate matching with group patterns, ensuring feature comparability.

[0020] Furthermore, this invention achieves accurate attribution of decoding performance degradation by calculating the distance between the target user's morphological feature vector and each historical pattern, and determining the corresponding pattern by minimizing the distance. By selecting successfully intervened model parameters from parallel users based on the attribution results as training weights for transfer learning, personalized and precise customization of intervention strategies is achieved. Cloud-based training ensures model capacity and accuracy; after training, the model is pushed to edge devices, allowing edge devices to achieve continuously optimized decoding performance without bearing the training load.

[0021] Furthermore, this invention implements a verification mechanism on the edge device side, allowing the old and new decoders to infer in parallel while only the original decoder actually controls the device. Comparative data is continuously collected for three days without interfering with the target user's normal use, fully verifying the potential abnormal output risk of the latest decoder for individualized neural activity patterns. Analysis of the difference between execution speed and prediction speed sequence quantitatively assesses the stability of the latest decoder's output.

[0022] Furthermore, this invention determines whether to switch to a new decoder by comparing stability indicators with preset thresholds, avoiding operational risks caused by direct deployment. The preset stability indicator thresholds ensure that the new encoder must reach a stable level before it can be enabled, guaranteeing the continuity of the user experience for the target users. The switching process is completed before the device is powered on the next day, ensuring that the target users are unaware of the change.

[0023] Furthermore, this invention dynamically calculates the decoder refresh cycle based on individual degradation slopes and group historical experience, enabling proactive initiation of the evaluation process before performance significantly degrades. This achieves an upgrade in maintenance mode from passive response to proactive prediction. After the refresh cycle is sent to the edge device, the device automatically triggers a complete cloud-based evaluation process when the cycle arrives. This allows for early detection and intervention of potential degradation trends, preventing passive response only after performance has accumulated to a perceptible level, and improving the control stability of the edge device. Attached Figure Description

[0024] Figure 1 This is a flowchart of a cloud-based neuroscience data processing method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the synchronization analysis of neuronal firing timestamp sequences and behavioral timestamp sequences according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating how to determine whether an edge device meets performance requirements according to an embodiment of the present invention. Figure 4 This is a flowchart illustrating the historical patterns corresponding to the synchronization decay trend of a target user in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0026] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0027] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] It is understood that this invention is primarily applied to clinical rehabilitation scenarios involving invasive brain-computer interfaces, enabling long-term adaptive and stable control of external devices used by implanted users (i.e., target users). In this scenario, the target user is a user with an implanted cortical electrode array who uses an edge device, requiring them to control the external device through neural commands to complete daily activities. The edge device refers to the external execution device directly used by the target user, such as a robotic arm, used to receive decoded commands and perform grasping and moving operations. The edge computing unit is deployed locally on the edge device, such as an embedded computer installed on the robotic arm, responsible for real-time acquisition of the target user's neural signals, execution of front-end signal processing, running a lightweight decoder for motion command inference, and synchronous acquisition of motion state data from the edge device. The cloud platform computing unit is deployed on a cloud server cluster, responsible for aggregating anonymized neural and behavioral data from multiple centers and users, performing population clustering analysis to generate degradation patterns, training a high-precision decoder based on transfer learning, and remotely distributing the data. The decoder is a machine learning model based on recurrent neural networks or temporal convolutional networks, with input being a population neuron firing rate vector and output being the three-dimensional spatial motion commands of the edge device, achieving accurate mapping from neural signals to motion intentions through continuous optimization.

[0029] Please see Figure 1 The diagram shows a flowchart of a cloud-based neuroscience data processing method according to an embodiment of the present invention. The cloud-based neuroscience data processing method according to an embodiment of the present invention includes: Step S1: Obtain neural signal data and behavioral data of the edge device when the target user uses the edge device through the target edge computing unit to generate neuron firing timestamp sequence and behavioral timestamp sequence, wherein the behavioral data includes the three-dimensional coordinates and speed of the robotic arm end; Step S2: Based on the target edge computing unit, perform synchronization analysis on the neuron firing timestamp sequence and behavior timestamp sequence to determine the synchronization of the target user, and send it to the cloud platform computing unit; Step S3: Based on the synchronization degree, determine whether the edge device meets the performance requirements, so as to obtain several synchronization degrees of several parallel users through the cloud platform computing unit and generate a synchronization degree timestamp historical sequence for each parallel user. Step S4: Generate a historical pattern of synchronization decay trend by performing cluster analysis on the historical sequences of each synchronization time stamp. Step S5: Obtain several historical synchronization degrees of the target user through the cloud platform computing unit to generate a historical sequence of synchronization degree timestamps of the target user and analyze the time warping morphology characteristics to determine the synchronization degree decay trend pattern of the target user. The time warping morphology characteristics include decay slope, abrupt change point position and fluctuation amplitude. Step S6: Based on the comparison between the synchronization decay trend pattern of the target user and the historical synchronization decay trend pattern, determine the historical pattern corresponding to the synchronization decay trend of the target user. Step S7: Based on the historical pattern corresponding to the synchronization decay trend, obtain the sample training weights of the corresponding parallel users through the cloud platform computing unit, use the neuron firing timestamp sequence to train and generate the latest decoder based on the sample training weights, and feed it back to the target edge computing unit. Step S8: During the process of the target user using the edge device, the original decoder is used to execute and the execution speed is obtained. At the same time, the latest decoder is run and the predicted speed is obtained. The execution speed and the predicted speed are compared to determine the stability index of the latest decoder. Step S9: Based on the stability index of the latest decoder, determine the latest decoder fed back by the execution cloud platform computing unit, and determine the refresh cycle of the decoder by the cloud platform computing unit based on the synchronization decay trend pattern of the target user.

[0030] In this embodiment, in step S1, the target edge computing unit is installed in the edge device used by the target user. After the edge device is started, the edge computing unit continuously acquires the target user's raw broadband neural signals at a sampling rate of 30kHz. This signal contains complete waveform information of neuronal action potentials. Simultaneously, it synchronously acquires the end-effector three-dimensional coordinates and execution speed fed back by the robotic arm joint decoder at a frequency of 100Hz. Internally, the edge computing unit uses any of the methods in the prior art to perform bandpass filtering of 300Hz to 6000Hz on the raw broadband signal to extract peak frequency bands. Each time a peak is detected, the firing time is recorded with millisecond precision, generating a neuronal firing timestamp sequence. Simultaneously, the robotic arm motion data is aligned with a unified timestamp to generate a behavior timestamp sequence containing the end-effector position and speed at each moment.

[0031] Please see Figure 2 As shown, it is a flowchart of the synchronization analysis of neuronal firing timestamp sequences and behavioral timestamp sequences in an embodiment of the present invention.

[0032] Specifically, in step S2, the process of performing synchronization analysis on the neuronal firing timestamp sequence and behavioral timestamp sequence based on the target edge computing unit includes: Step S21: Calculate the timing error based on the neuron firing timestamp sequence and the behavior timestamp sequence; Step S22: The timing error is compared with a preset timing error threshold to determine the synchronization degree of the edge device.

[0033] In this embodiment, the target edge computing unit aligns the neuron firing timestamp sequence and the behavior timestamp sequence, calculates the time difference between neuron firing and the corresponding start of behavior as the timing error, and calculates the average timing error per hour. This average is then compared to a preset timing error threshold, and the ratio is used as the neuro-behavioral synchronization index (synchronization = average / preset timing error threshold). The neuro-behavioral synchronization index is acquired on a 24-hour cycle, and the edge computing unit sends the cycle's neuro-behavioral synchronization index to the cloud platform computing unit. The preset timing error threshold is determined by the 90th percentile of the performance requirements met by the edge devices of parallel users. It can be understood that a lower synchronization value indicates a tighter coupling between the neural firing pattern and the motion output, resulting in higher decoding accuracy.

[0034] Please see Figure 3 As shown, it is a flowchart for determining whether an edge device meets the performance requirements according to an embodiment of the present invention.

[0035] Specifically, in step S3, the process of determining whether the edge device meets the performance requirements includes: Step S31: Based on the comparison between the synchronization degree and the preset synchronization degree threshold, if the proportion of the synchronization degree greater than the preset synchronization degree threshold is greater than the preset proportion threshold, then it is determined that the edge device does not meet the performance requirements.

[0036] In this embodiment, after receiving the target user's neural-behavioral synchronization index for the current period, the cloud platform computing unit generates a synchronization timestamp sequence based on the timestamp of each index. The cloud platform pre-stores all synchronization records of the target user, obtains the synchronization of those periods marked as stable by the target user, calculates their average, and uses 1.5 times the average as the synchronization threshold. The cloud platform obtains the synchronization over three consecutive days. If the percentage of synchronization exceeding the synchronization threshold for three consecutive days is greater than 60%, the current edge device is determined to not meet the performance requirements. If the percentage of synchronization exceeding the synchronization threshold for three consecutive days is less than or equal to 60%, the current edge device is determined to meet the performance requirements, and the target user's synchronization index is continuously monitored.

[0037] Specifically, in step S4, the process of generating a historical pattern of synchronization decay trend includes: Step S41: When it is determined that the edge device does not meet the performance requirements, several cluster centers are determined by performing cluster analysis on the synchronization time stamp history sequence, and the corresponding synchronization decay trend history pattern is determined based on each cluster center.

[0038] In this embodiment, when a target user triggers a condition that is not met, the cloud platform computing unit obtains synchronization data of all parallel users with the same implanted brain regions and complete synchronization records (without involving the reading of user privacy data) that have read permissions. It extracts the synchronization timestamp sequences of the parallel users from the past 12 months where edge devices do not meet performance requirements, and performs cluster analysis on these sequences, setting the minimum number of samples within each cluster to 5. The synchronization timestamp sequences corresponding to the cluster centers from the past 6 months are analyzed for time warping morphology characteristics, including decay slope, mutation point location, and fluctuation amplitude. The multidimensional feature vector formed by the time warping morphology characteristics corresponding to each cluster center is determined as the historical pattern of synchronization decay trend. The process of determining the time warping morphology characteristics includes: identifying mutation points in the sequence through neighborhood comparison; defining mutation points as the time positions where the synchronization change within 3 days exceeds 3 times the standard deviation of historical fluctuation; performing linear regression on the stable segments between mutation points to calculate the decay slope, with a positive slope indicating that synchronization deteriorates over time; and calculating the fluctuation amplitude of the sequence within a 6-month window, i.e., the difference between the maximum and minimum values ​​divided by the mean. The multidimensional feature vector formed by the above features fully describes the synchronicity decay pattern of historical patterns. If there are multiple mutation points in the sequence, the decay segment after the most recent mutation point is selected for analysis; if there are no mutation points, the entire 6-month window is used.

[0039] In this embodiment, in step S5, the cloud platform computing unit retrieves the complete synchronization time stamp history sequence of the target user since implantation, focusing on the data window of the most recent 6 months. After performing the same dynamic time warping morphological feature extraction as in S4 on this window sequence, a multi-dimensional feature vector is constructed, serving as the synchronization decay trend pattern of the target user. This vector fully describes the current synchronization decay pattern of the target user. This feature vector is temporarily stored in the cloud as input for subsequent pattern matching.

[0040] Please see Figure 4 As shown, it is a flowchart of the historical pattern corresponding to the synchronization decay trend of the target user in an embodiment of the present invention.

[0041] Specifically, in step S6, the process of determining the historical pattern corresponding to the synchronization decay trend of the target user includes: Step S61: Calculate the distance between the synchronization decay trend pattern of the target user and the historical synchronization decay trend pattern; Step S62: Determine the historical pattern corresponding to the synchronization decay trend of the target user based on the minimum value of the distance.

[0042] In this embodiment, the cloud platform computing unit compares the synchronization decay trend pattern of the target user generated in step S5 with the historical synchronization decay trend pattern generated in step S4, and calculates the Mahalanobis distance between the target user's synchronization decay trend pattern and each historical synchronization decay trend pattern. It is understood that Mahalanobis distance is used as a similarity measure because it can consider the correlation between various feature dimensions and avoid the influence of differences in dimensions on the results. The historical pattern corresponding to the minimum distance is taken as the historical pattern corresponding to the synchronization decay trend of the target user.

[0043] Specifically, in step S7, the process of obtaining the sample training weights of the corresponding parallel users through the cloud platform computing unit includes: Step S71: Obtain the synchronization degree of several parallel users after training based on the historical pattern corresponding to the synchronization degree decay trend. Step S72: Based on the comparison between the synchronization degree after training and the preset synchronization degree threshold, if the proportion of the synchronization degree less than the preset synchronization degree threshold is less than or equal to the preset proportion threshold, then the corresponding sample training weight is obtained.

[0044] In this embodiment, based on the historical pattern corresponding to the synchronization decay trend of the target user determined in step S6, the cloud platform computing unit identifies several parallel users matching this pattern, obtains the synchronization of these parallel users for each of the three consecutive days after retraining, calculates the percentage of each parallel user whose synchronization exceeds the synchronization threshold, obtains the historical records of parallel users whose percentage is less than or equal to 60%, and extracts the decoder parameters used by these parallel users as sample training weights based on the historical records. Parallel users whose synchronization exceeds the synchronization threshold by more than 60% are not subject to decoder parameter acquisition. It can be understood that decoder parameters refer to all trainable weight matrices and bias terms constituting a recurrent neural network or temporal convolutional network, specifically including the connection weights from the input layer to the hidden layer, the recurrent connection weights of the hidden layer, the connection weights from the hidden layer to the output layer, and the bias vectors of each layer. These parameters completely define the mapping function from the neuron firing rate vector to the three-dimensional spatial velocity command. In transfer learning, decoder parameters obtained from parallel users are used as training weights for samples. This is because parallel users with the same historical decay trend share similar neural signal degradation mechanisms, and the feature extraction patterns formed by their decoders during long-term adaptation are valuable for the current target user. Using these parameters as training weights for transfer learning is equivalent to allowing the target user's current encoder to continue learning from a point where it has already learned to handle similar degradation patterns, rather than starting from random initialization. This significantly reduces the amount of training data required and accelerates convergence. The cloud platform uses these training weights as training weights for transfer learning, loading the target user's neuronal firing timestamp sequences and behavioral timestamp sequences from the past two weeks as supervised training data. Training is performed on the cloud platform's GPU cluster, and after generating the latest decoder, it is pushed to the target edge computing unit and marked as awaiting activation. It is understandable that using training data to generate the decoder is a conventional technique, and will not be elaborated upon here.

[0045] Specifically, in step S8, the process of determining the stability metrics of the latest decoder includes: Step S81: Calculate the difference between several execution speeds and the predicted speeds per unit time to generate a difference sequence; Step S82: Calculate the mean difference and standard deviation of the difference based on the difference sequence to determine the stability index of the latest decoder.

[0046] In this embodiment, the target edge computing unit enters the verification phase after receiving the latest decoder from the cloud platform. During normal use by the target user the following day, the edge device simultaneously loads the original decoder and the latest decoder in the background, inputting the same real-time neural signal data into the two decoders for parallel inference to obtain the execution speed of the original decoder and the prediction speed corresponding to the latest decoder. The output of the original decoder is always sent to the robotic arm controller to ensure that the target user's operation is not affected; the output of the latest decoder is only stored in memory for subsequent comparison. The edge computing unit continuously calculates the differences between several execution speeds and prediction speeds over three days of use, generating a difference sequence. Based on the difference sequence, the mean and standard deviation of the differences are calculated, and the mean and standard deviation are determined as stability indicators of the latest decoder.

[0047] Specifically, in step S9, the process of determining the latest decoder fed back by the cloud platform computing unit includes: Step S91: Compare the stability index of the latest decoder with the preset stability index threshold. If the stability index is less than the preset stability index threshold, then determine the latest decoder fed back by the cloud platform computing unit.

[0048] Specifically, in step S9, the process of determining to start the cloud platform computing unit to determine the decoder refresh cycle includes: Step S92: Compare the stability index of the latest decoder with the preset stability index threshold. If the stability index is less than the preset stability index threshold, then start the cloud platform computing unit to determine the refresh cycle of the decoder.

[0049] Specifically, in step S9, the process of determining the decoder's refresh cycle includes: Step S93: Determine the refresh cycle of the decoder based on the target user's synchronization attenuation slope and the attenuation slope of the historical mode corresponding to the synchronization attenuation trend.

[0050] In this embodiment, the stability index of the latest decoder is compared with a preset stability index threshold. If the stability index is less than the preset stability index threshold, the latest decoder is determined to be stable. Specifically, the mean difference is compared with a preset mean difference threshold, and the standard deviation of the difference is compared with a preset standard deviation threshold. If both the mean and standard deviation are less than their respective thresholds, the stability index of the latest decoder is determined to meet the requirements, indicating that the average deviation between the latest decoder output and the current actual control command is small and the fluctuation is controllable, thus meeting the switching conditions. Otherwise, the stability index is determined to not meet the requirements, the latest decoder output deviation is too large or the fluctuation is drastic, and the switch is not allowed and retraining is required. The preset mean difference threshold is set to 10% of the average execution speed of the original decoder during the historical period when its performance met the requirements, and the preset standard deviation threshold is set to 1.5 times the standard deviation of the original decoder's execution speed during the same historical period. The edge device performs the latest decoder switching operation before the target user powers on the next day. The switching process does not require restarting the device, and the target user is unaware of the switch when powering on the device. While ensuring the latest decoder possesses stability (i.e., the stability index is less than a preset stability index threshold), the cloud platform computing unit calculates the ratio of the average decay slope of the synchronization decay trend corresponding to the historical pattern to the decay slope of the target user's synchronization decay trend pattern, and calculates the average interval time from the first intervention to the second triggered intervention for parallel users corresponding to the historical pattern of the synchronization decay trend. The product of the average interval time and the ratio is determined as the decoder's refresh cycle. The final calculated refresh cycle, in days, is sent to the target edge computing unit via an encrypted channel. The edge device stores this parameter in its local configuration and automatically triggers a complete decoder retraining and stability verification process when the cycle arrives.

[0051] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A cloud platform-based method for processing neuroscientific data, characterized in that, include: The target edge computing unit acquires neural signal data and edge device behavior data when the target user uses the edge device, in order to generate neuron firing timestamp sequences and behavior timestamp sequences. The behavior data includes the three-dimensional coordinates and velocity of the robotic arm end effector. The synchronization degree of the target user is determined by analyzing the neuron firing timestamp sequence and behavior timestamp sequence based on the target edge computing unit, and then sent to the cloud platform computing unit. Based on the synchronization degree, determine whether the edge device meets the performance requirements, so as to obtain several synchronization degrees of several parallel users through the cloud platform computing unit and generate a historical sequence of synchronization degree timestamps for each parallel user. Cluster analysis is performed on the historical sequences of the synchronization timestamps to generate historical patterns of synchronization decay trends. The cloud platform computing unit obtains several historical synchronization degrees of the target user to generate a historical sequence of synchronization degree timestamps of the target user and analyzes the time warping morphology characteristics to determine the synchronization degree decay trend pattern of the target user. The time warping morphology characteristics include decay slope, abrupt change point position and fluctuation amplitude. By comparing the synchronization decay trend pattern of the target user with the historical synchronization decay trend pattern, the historical pattern corresponding to the synchronization decay trend of the target user is determined. Based on the historical pattern corresponding to the synchronicity decay trend, the cloud platform computing unit obtains the sample training weights of the corresponding parallel users, uses the neuron firing timestamp sequence to train and generate the latest decoder based on the sample training weights, and feeds it back to the target edge computing unit. During the target user's use of the edge device, the original decoder is used to execute and obtain the execution speed, while the latest decoder is run and the predicted speed is obtained. The execution speed and the predicted speed are compared to determine the stability index of the latest decoder. Based on the stability metrics of the latest decoder, the latest decoder fed back by the execution cloud platform computing unit is determined, and the refresh cycle of the decoder is determined by the cloud platform computing unit based on the synchronization decay trend pattern of the target user. 2.The cloud platform-based neuroscience data processing method of claim 1, wherein, The process of performing synchronization analysis on neuronal firing timestamp sequences and behavioral timestamp sequences based on target edge computing units includes: The timing error is calculated based on the neuron firing timestamp sequence and the behavior timestamp sequence; The synchronization degree of the edge device is determined by comparing the timing error with a preset timing error threshold. 3.The cloud platform-based neuroscience data processing method of claim 2, wherein, The process of determining whether an edge device meets performance requirements includes: Based on the comparison between the synchronization degree and the preset synchronization degree threshold, if the proportion of the synchronization degree greater than the preset synchronization degree threshold is greater than the preset proportion threshold, then it is determined that the edge device does not meet the performance requirements. 4.The cloud platform-based neuroscience data processing method of claim 3, wherein, The process of determining historical patterns of synchronicity decay trends includes: When it is determined that the edge device does not meet the performance requirements, several cluster centers are determined by performing cluster analysis on the synchronization timestamp historical sequence, and the corresponding synchronization decay trend historical pattern is determined based on each cluster center. 5.The cloud platform-based neuroscience data processing method of claim 4, wherein, The process of determining the historical patterns corresponding to the synchronization decay trend of a target user includes: Calculate the distance between the target user's synchronization decay trend pattern and the historical synchronization decay trend pattern; Based on the minimum value of the distance, the historical pattern corresponding to the synchronization decay trend of the target user is determined. 6.The cloud platform-based neuroscience data processing method of claim 5, wherein, The process of obtaining the sample training weights for corresponding parallel users through the cloud platform computing unit includes: Based on the historical patterns corresponding to the aforementioned synchronization decay trend, the synchronization of several parallel users after training is obtained. The training weights are obtained by comparing the synchronization degree after training with a preset synchronization degree threshold. If the proportion of samples with a synchronization degree less than the preset synchronization degree threshold is less than or equal to the preset proportion threshold, the corresponding sample training weights are obtained.

7. The neuroscience data processing method based on a cloud platform according to claim 6, characterized in that, The process of determining the stability metrics of the latest decoder includes: Calculate the difference between several execution speeds and the predicted speeds per unit time to generate a difference sequence; The mean and standard deviation of the differences are calculated based on the difference sequence to determine the stability index of the latest decoder.

8. The neuroscience data processing method based on a cloud platform according to claim 7, characterized in that, The process of determining the latest decoder fed back by the cloud platform computing unit includes: The stability index of the latest decoder is compared with a preset stability index threshold. If the stability index is less than the preset stability index threshold, the latest decoder fed back by the cloud platform computing unit is determined to be executed.

9. The neuroscience data processing method based on a cloud platform according to claim 8, characterized in that, The process of determining the start-up of the cloud platform computing unit to determine the decoder refresh cycle includes: The stability index of the latest decoder is compared with a preset stability index threshold. If the stability index is less than the preset stability index threshold, the cloud platform computing unit is activated to determine the refresh cycle of the decoder.

10. The neuroscience data processing method based on a cloud platform according to claim 9, characterized in that, The process of determining the decoder's refresh cycle includes: The refresh cycle of the decoder is determined based on the slope of the target user's synchronization decay and the slope of the historical pattern corresponding to the synchronization decay trend.

Images