Method and device for extracting features of neural pulse signals
By calculating the pulse spacing sequence and defining standard pulse sequences, the characteristics of neural pulse signals are extracted, and the problem of processing a single neural pulse sequence in the prior art is solved, and high-precision neural pulse signal recognition and information utilization are achieved.
Patent Information
- Application Number
- CN202110381484.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-04-09
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-04-09
AI Technical Summary
The prior art is difficult to directly process and calculate a single neural pulse sequence, resulting in limited application of neural pulse signals in machine learning and neural networks, and traditional metrics cannot process a single pulse sequence.
By calculating the pulse spacing sequence, dividing the time window, defining the standard pulse sequence and extracting statistical, pulse metrics and vector measurement characteristics, the neural pulse signal is directly calculated, which is suitable for neural pulse signal processing of single or multiple neurons.
It realizes high-precision recognition of neural pulse signals, more information utilization is made, and the recognition accuracy is improved. It is suitable for various machine learning algorithms and artificial neural networks.
Smart Images

Figure CN113255436B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the technical field of machine learning of neural pulse signals, and in particular to a method and device for extracting features of neural pulse signals. Background Art
[0002] Modern brain science research shows that when the human body is stimulated by external stimuli, action potentials will be generated in the stimulated parts. When the action potential exceeds a certain threshold, cell excitation or nerve impulses will occur. Action potentials are also called nerve pulses. The transmission of signals in the human body is carried out in the form of nerve pulses. Specifically, different information is encoded by the time when the nerve pulse occurs. Therefore, the action potential sequence, or nerve pulse sequence, can be regarded as a sequence composed of a series of time values. Many features of the sequence can characterize the transmitted information, such as the time of the first pulse, the pulse interval, the number of pulses in a period of time, etc. The research on the decoding mechanism of nerve pulse signals and the development of related technologies are conducive to promoting the development of scientific fields and high-tech industries such as life sciences, artificial intelligence technology, dexterous operation robots, and medical prostheses.
[0003] At present, the relevant technical research in this field is mainly divided into two aspects. On the one hand, in order to simulate the human nervous system, a variety of artificial neural networks (ANN, Artificial Neural Network) and machine learning models have been developed to achieve regression model fitting, type differentiation, outlier detection, data analysis and other purposes. However, these technologies can only process digital signals and cannot be directly applied to the calculation of neural pulse signals. On the other hand, many researchers directly turned to the development of specialized neural pulse network models (SNN, Spiking Neural Network) or neural signal processing chips that can realize related calculations. However, these specialized neural pulse networks or chips often require high usage and development costs and are difficult to be widely used. Although some studies directly use the statistical characteristics of neural pulses, such as the number of pulses, as features, extract them and input them into traditional machine learning or neural network algorithms for training and modeling, due to insufficient information utilization, the actual effect is often not good, which is manifested in low prediction accuracy or recognition accuracy.
[0004] In addition, in recent years, many scholars have proposed a variety of metrics for measuring the information contained in neural pulse sequences, such as Van Rossum Distance, Victor-Purpura Distance, etc., which are suitable for evaluating the similarity or difference between neural pulse sequences, and their effectiveness has been confirmed by most studies. In the existing traditional measurement field, there are also a variety of metrics for digital vectors, such as Euclidean Distance, Cityblock Distance, etc., which are suitable for calculating the distance between vectors with the same dimension and are widely used in the field of result evaluation and signal analysis. However, both of these metrics require that the extracted features must be based on two pulse sequences (or vectors) and cannot process a single pulse sequence. Summary of the invention
[0005] The present disclosure provides a method and device for extracting features of neural pulse signals. The technical purpose is to directly calculate the neural pulse signals emitted by a single or multiple neurons by introducing pulse spacing sequences, standard pulse sequences and time windows, so as to solve the problem that traditional machine learning or neural network models are difficult to apply to neural pulse signal processing.
[0006] The above technical objectives of the present disclosure are achieved through the following technical solutions:
[0007] A method for extracting features of a neural pulse signal, comprising:
[0008] The pulse spacing sequence DT is calculated by the neural pulse sequence ST, including: the neural pulse sequence generated by each neuron is ST = {T(i)} = {T(1), T(2), ..., T(N)}, i∈(1, N), N≥2; T(i) represents the time when the i-th pulse in the neural pulse sequence appears; then the pulse spacing sequence is:
[0009] DT={D(1),D(2),...,D(N-1)}={T(2)-T(1),T(3)-T(2),...,T(N)-T(N-1)};
[0010] The neural pulse sequence and the pulse interval sequence are divided into different subsequences according to the time window length W. <W≤T(N);
[0011] According to the time window length W and N RM Get the standard pulse sequence for each subsequence; where N RM Indicates the maximum number of pulses in all subsequences, N RM ≥2;
[0012] Statistical features, pulse metric features and vector metric features of each subsequence are extracted according to the standard pulse sequence.
[0013] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the program.
[0014] The beneficial effects of the present disclosure are as follows: the feature extraction method and device for neural pulse signals described in the present application can directly calculate the neural pulse signals without applying a specific function to convert the pulse signals into digital signals and then inputting them into a machine learning or neural network model; there is no restriction on the use of classification models, and it is applicable to various machine learning algorithms and artificial neural networks with classification and recognition functions; compared with the existing technology of directly inputting the statistical features of pulse sequences (such as the number of pulses or the time average, etc.) into the machine learning model to realize type recognition, the method proposed in the present application can extract more effective features, make more effective use of information, and achieve higher recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 is a flow chart of the feature extraction method of the present invention;
[0016] Figure 2 is a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0017] The technical solution of the present disclosure will be described in detail below with reference to the accompanying drawings.
[0018] Figure 1 is a flow chart of the feature extraction method of the present invention, as shown in Figure 1 As shown, S1: Calculate the pulse spacing sequence DT through the neural pulse sequence ST, including: the neural pulse sequence generated by each neuron is ST = {T(i)} = {T(1), T(2), ..., T(N)}, i∈(1, N), N≥2; T(i) represents the time when the i-th pulse in the neural pulse sequence appears; then the pulse spacing sequence is:
[0019] DT={D(1),D(2),...,D(N-1)}={T(2)-T(1),T(3)-T(2),...,T(N)-T(N-1)}.
[0020] S2: Divide both the neural pulse sequence and the pulse interval sequence into different subsequences according to the time window length W, where 0 < W ≤ T(N). It is possible that in the same time window length W, the number of pulses in the corresponding different subsequences is different, because the occurrence time of each pulse in the neural pulse sequence is not fixed. Some subsequences may have more pulses, while some may have fewer pulses.
[0021] S3: According to the time window length W and N RM obtain the standard pulse sequence of each subsequence; where N RM represents the maximum number of pulses in all subsequences, and N RM ≥ 2.
[0022] Specifically, use j to represent the serial number of the subsequence of the neural pulse sequence ST. Then the standard pulse sequence of the subsequence of the neural pulse sequence ST is:
[0023] where 1 ≤ j ≤ N bin , N bin represents the total number of subsequences of the neural pulse sequence ST, and N bin = ceil(T(N) / W), and the ceil() function represents the smallest integer not less than its argument; when j = 1, the standard pulse sequence of the first subsequence of the neural pulse sequence ST is
[0024] Use k to represent the serial number of the subsequence of the pulse interval sequence DT. Then the standard pulse sequence RDT of the pulse interval sequence DT k = RST j , and when k = 1, the standard pulse sequence of the first subsequence of the pulse interval sequence DT is
[0025] That is to say, the number of pulses in the standard pulse sequences corresponding to all subsequences is the same. For example, if N RM = 3, that is, the maximum number of pulses in all subsequences is 3, and the time window length W = 100 ms, then the uniform interval of the standard pulse sequence Then the standard pulse sequence corresponding to the first subsequence of the neural pulse sequence is RST1 = {0, 50, 100}, the standard pulse sequence corresponding to the second subsequence is RST2 = {100, 150, 200}, the standard pulse sequence corresponding to the third subsequence is RST3 = {200, 250, 300}, and so on until the standard pulse sequences of all subsequences are obtained.
[0026] The number of subsequences of the pulse spacing sequence is generally determined according to the total pulse time of the neural pulse sequence and the time window length W.
[0027] In addition, when extracting the vector measurement feature, the two measured vectors should have the same dimension. If the number of pulses in the subsequence is less than the number of pulses in the standard pulse sequence, the method of end zero padding is used to extract the feature, including:
[0028] If N j represents the total number of pulses in the jth subsequence of the neural pulse sequence ST, N j ≥1, then in, Indicates the zero-filling sequence at the end, N 0j Indicates the number of zero padding, N 0j =N RM -N j ; The dist() function has the same definition as the vector metric feature.
[0029] If N k represents the total number of pulses in the kth subsequence in the pulse spacing sequence DT, N k ≥1, then in, Indicates the zero-filling sequence at the end, N 0k Indicates the number of zero padding, N 0k =N RM -N k .
[0030] S4: Extracting statistical features, pulse metric features and vector metric features of each subsequence according to the standard pulse sequence.
[0031] As a specific embodiment, when the present invention is applied to the pulse signal calculation of multiple neurons that perceive the same external stimulus, taking three neurons as an example, Figure 2 As shown, for the neural pulse sequence generated by each neuron, its pulse spacing sequence is first calculated, and the neural pulse sequence and the pulse spacing sequence are divided into several subsequences according to the length of the time window; then the standard pulse sequence is defined, and the statistical characteristics, pulse measurement characteristics and vector measurement characteristics of the subsequences in each time length window are extracted; all the extracted features are used as the input of the machine learning or neural network classification model, and the type to be identified is used as the output of the classification model. After training, multi-type intelligent recognition based on neural pulse signals can be realized.
[0032] If there are 3 neurons, then there will be 3 neural pulse sequences and corresponding 3 pulse interval sequences, a total of 6 sequences. Then these 6 sequences are divided into several subsequences according to the length of the time window, and then the standard pulse sequence of each subsequence is defined. Then, the features of the neural pulse sequence are extracted based on the standard pulse sequence. Finally, when the extracted features are input into the classification model, there is no need to establish multiple classification models. Instead, one model is established and all the above features are input into the model. After training the model, multi-type intelligent recognition based on neural pulse signals can be realized.
[0033] As a specific embodiment, the pulse metric features include ISI Distance, Spike Distance, Spike Synchronization, Van Rossum Distance, and Victor-Purpura Distance. The vector metric features include Euclidean Distance, Cityblock Distance, Cosine Distance, Correlation Distance, and Spearman Distance.
[0034] The above are exemplary embodiments of the present invention. The protection scope of the present invention is defined by the claims and their equivalents.
Claims
1. A feature extraction method for a neural pulse signal, characterized in that: include: The pulse spacing sequence DT is calculated by the neural pulse sequence ST, including: the neural pulse sequence generated by each neuron is ST = {T(i)} = {T(1), T(2), ..., T(N)}, i∈(1, N), N≥2; T(i) represents the time when the i-th pulse in the neural pulse sequence appears; then the pulse spacing sequence is: DT={D(1),D(2),...,D(N-1)}={T(2)-T(1),T(3)-T(2),...,T(N)-T(N-1)}; The neural pulse sequence and the pulse interval sequence are divided into different subsequences according to the time window length W. <W≤T(N); According to the time window length W and N RM Get the standard pulse sequence for each subsequence; where N RM Indicates the maximum number of pulses in all subsequences, N RM ≥2; Extracting statistical features, pulse metric features and vector metric features of each subsequence according to the standard pulse sequence; Wherein, j is used to represent the serial number of the subsequence of the neural pulse sequence ST, and the standard pulse sequence of the subsequence of the neural pulse sequence ST is: Among them, 1≤j≤N bin , N bin represents the total number of subsequences of the neural pulse sequence ST, N bin = ceil(T(N) / W), where the ceil() function represents the smallest integer not less than its argument; when j = 1, the standard pulse sequence of the first subsequence of the neural pulse sequence ST is Using k to represent the sequence number of the subsequence of the pulse spacing sequence DT, the standard pulse sequence RDT of the pulse spacing sequence DT is k =RST j , when k=1, the standard pulse sequence of the first subsequence of the pulse spacing sequence DT is 2. The feature extraction method of a neural pulse signal according to claim 1, characterized in that: When extracting the vector measurement feature, the two measured vectors should have the same dimension. If the number of pulses in the subsequence is less than the number of pulses in the standard pulse sequence, the feature extraction is performed by using the end zero padding method, including: If N j represents the total number of pulses in the jth subsequence of the neural pulse sequence ST, N j ≥1, then in, Indicates the zero-filling sequence at the end, N 0j Indicates the number of zero padding, N 0j =N RM -N j ; The dist() function has the same definition as the vector metric feature; If N k represents the total number of pulses in the kth subsequence in the pulse spacing sequence DT, N k ≥1, then in, Indicates the zero-filling sequence at the end, N 0k Indicates the number of zero padding, N 0k =N RM -N k .
3. The feature extraction method of neural pulse signal according to claim 2, characterized in that: The pulse metric features include inter-pulse distance, pulse distance, pulse synchronization, Van Rossum distance, and Victor Pulp distance; The vector metric features include Euclidean distance, city block distance, cosine distance, correlation distance and Spearman distance.
4. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 3 is implemented.
