A seismic facies analysis method and system based on an improved self-organizing neural network

Abstract

The application provides a seismic facies analysis method and system based on an improved self-organizing neural network, which comprises the following steps: S1, initializing the self-organizing neural network, wherein the self-organizing neural network is a two-layer self-organizing neural network with K neurons, and K is the number of seismic facies types; S2, calculating the similarity between input data and each neuron one by one and determining the winning neuron according to the similarity; S3, updating the related parameters of the neuron; and S4, every M time steps, adding or inserting a new neuron according to a first preset rule, and deleting or discarding a neuron according to a second preset rule. According to the lateral variability of the seismic signal in a certain target layer, the improved self-organizing neural network is used to classify the seismic trace shape, and the classification result forms discrete'seismic facies'. Further, the rock physical parameters or the plane distribution rule of the sedimentary facies are described by using the seismic data.

Description

TECHNICAL FIELD

[0001] The present application relates to oil and gas and coal bed gas seismic exploration and development, and particularly relates to a seismic facies analysis method and system based on an improved self-organizing neural network. BACKGROUND

[0002] In the exploration and development of underground sedimentary mineral resources such as oil and coal, sedimentary facies research is of great significance. However, since the target layer is deeply buried underground, the research methods and means used are quite different from those used in outcrop sedimentary facies research.

[0003] In underground facies analysis, only through rock data can the sedimentary facies markers of interest be observed, and drilling coring is generally not continuous, and the full-well coring rate of a single well is often only a few percent to tens of percent, which greatly hinders sedimentary facies research. Although well logging facies analysis can make continuous sedimentary facies interpretation for the full well, it has strong multi-solution, and therefore, in addition to the above two types of data, more information needs to be obtained from other data to improve the accuracy of sedimentary facies interpretation.

[0004] More importantly, even if the single-well facies analysis data are sufficient, the information obtained by using traditional research methods is still only a part of the information, and important information such as stratigraphic stacking patterns and sedimentary body shapes is not utilized. Further, even if the interpretation is completely correct, it is still only a "one-hole view". In order to further understand the planar distribution characteristics of sedimentary facies, a large number of sufficiently dense drill holes are required, which is difficult to meet at the exploration stage. Therefore, a new method that can better understand the planar variation characteristics of sedimentary facies using a small number of drill holes is urgently needed.

[0005] Seismic facies analysis is exactly what is needed to meet the above urgent needs. Seismic facies is the appearance of reflected waves on a seismic reflection time section. Seismic facies analysis is the interpretation and inference of sedimentary facies based on seismic facies characteristics. In oil exploration and the exploration of some coalfields and salt mines, seismic exploration data are essential and important basic data. These data can be obtained at the early stage of exploration, and generally cover the entire basin, and contain a wealth of stratigraphic, structural and sedimentary facies information, and therefore are extremely valuable basic data for underground geological analysis. Seismic facies analysis, as an important part of seismic stratigraphy, was born around 1977 and quickly spread around the world. Over the past decade, it has been continuously developed and improved in extensive practice, and has become an indispensable sharp weapon for underground facies analysis.

[0006] The method of seismic facies analysis is to identify the unique characteristics of seismic reflection wave groups within each sequence and their shape combinations, and to assign them certain geological meanings, and then to interpret the sedimentary facies, which is called seismic facies analysis.

[0007] There are two methods for seismic facies analysis and identification. The first method is to observe seismic reflection characteristics by artificial means and compare them with established standard seismic facies characteristics to determine to which seismic facies they belong. This method is generally applied to local seismic data interpretation and analysis, and the interpretation and identification accuracy is relatively low. The other method is to analyze and calculate seismic data or seismic attribute data by using seismic data processing technology, computer technology and certain mathematical methods to extract seismic facies that can reflect sedimentary facies changes. This is an efficient, advanced and quantitative seismic facies identification method.

[0008] Seismic waveform is the basic property of seismic data, which contains all qualitative and quantitative information such as reflection pattern, phase, frequency and amplitude, and is the overall characteristics of seismic information. The dynamic changes thereof contain rich internal information and can truly reflect the characteristics of underground structure. Waveform classification method is the most commonly used method for seismic facies analysis. Through classification of seismic signal waveforms, the division of seismic facies can be realized.

[0009] Traditional waveform classification methods include K-means and self-organizing neural network methods. They both use unsupervised learning method to complete data classification in a self-organizing manner. However, the number of data categories needs to be given in advance. A general self-organizing artificial neural network is composed of an input layer and a competitive layer (output layer). The number of input layer neurons is N, and the competitive layer is a one-dimensional or two-dimensional plane array composed of M neurons. The network is fully connected, that is, each input node is connected to all output nodes (however, only adjacent neurons are connected between neurons). It can map an arbitrary-dimensional input mode into a one-dimensional or two-dimensional graph in the output layer while keeping the topological structure unchanged. Through repeated learning of the input mode, the weight vector space can be consistent with the probability distribution of the input mode, that is, the probability is maintained. Each neuron in the competitive layer of the network competes for the response opportunity of the input mode, and the weights related to the winning neuron are adjusted in the direction more conducive to its competition, that is, the winning neuron is taken as the center, the near neighbors show excitatory lateral feedback, and the far neighbors show inhibitory lateral feedback. In actual implementation, the training algorithm of the general self-organizing artificial neural network generally uses Kohonen algorithm. The self-organizing artificial neural network based on Kohonen algorithm has some limitations, such as: the network structure is fixed and cannot be dynamically changed; generally, there are only two layers, input layer and competitive layer (output layer), and the number of neurons in the competitive layer (output layer) is specified in advance, that is, the number of categories. SUMMARY

[0010] The features and advantages of the present application are set forth in the description that follows, and in part will be apparent from that description, or can be learned by practice of the present application.

[0011] To overcome the problems of the prior art, the present application provides a seismic facies analysis method based on an improved self-organizing neural network, comprising the steps of:

[0012] S1, initializing a self-organizing neural network, the self-organizing neural network being a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies categories;

[0013] S2, calculating the similarity of input data with each neuron one by one and determining the winning neuron therefrom;

[0014] S3, updating the relevant parameters of the neurons;

[0015] S4, every M time steps, adding or inserting a new neuron according to a first preset rule and deleting or discarding a neuron according to a second preset rule.

[0016] Preferably, in the step S2, the similarity d j of input data with each neuron is calculated as follows:

[0017]

[0018] wherein, is the weight vector of each neuron, and x is the current input data.

[0019] Preferably, the neuron most similar to the input data is the winning neuron in the step S2.

[0020] Preferably, the relevant parameters of the neurons include the winning frequency of the neurons and the weights of the neurons.

[0021] Preferably, the first preset rule is as follows:

[0022] The neuron v q with the largest winning frequency is counted.

[0023] The neuron v q farthest from the neuron v q in the neighborhood of the neuron v f is obtained by a neighborhood function, and the difference e between the two weight vectors is calculated for the neurons in this range.

[0024]

[0025] wherein j ∈ [q, f], and for given parameters a and b, when a < e < b, a neuron v r is inserted between j and j+1.

[0026] Preferably, the step S4 comprises:

[0027] The inserted neuron v r is assigned a weight value as follows:

[0028] The winning frequency adjustment is as follows:

[0029] Preferably, the second preset rule is:

[0030] The neuron with the minimum winning frequency is counted, and when the winning frequency C p of the neuron with the minimum winning frequency is less than a preset threshold, the neuron with the minimum winning frequency is discarded.

[0031] The present application provides a seismic facies analysis system based on an improved self-organizing neural network, comprising:

[0032] An initialization module is configured to initialize the self-organizing neural network, which is a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies categories;

[0033] A calculation module is configured to calculate the similarity between input data and each neuron one by one and determine the winning neuron accordingly;

[0034] A parameter updating module is configured to update the relevant parameters of the neurons;

[0035] A neuron updating module is configured to add or insert a new neuron every M time steps according to a first preset rule and delete or discard a neuron according to a second preset rule.

[0036] Preferably, the first preset rule is:

[0037] The neuron v q with the maximum winning frequency is counted;

[0038] The neuron v q that is the farthest from v q within the neighborhood of v f is obtained, and the difference e between the two weight vectors is calculated for the neurons within this range;

[0039]

[0040] wherein j∈[q, f], and for given parameters a and b, when a<e<b, a neuron v r is inserted between j and j+1.

[0041] The second preset rule is:

[0042] The neuron with the minimum winning frequency is counted, and when the winning frequency C p of the neuron with the minimum winning frequency is less than a preset threshold, the neuron with the minimum winning frequency is discarded.p When the winning frequency is less than a preset threshold, the neuron with the minimum winning frequency is discarded.

[0043] The application provides a computer readable storage medium, which stores at least one program executable by a computer, and the at least one program causes the computer to execute steps in a self-organizing neural network seismic facies analysis method provided by an embodiment of the application when the at least one program is executed by the computer.

[0044] The skilled in the art will better understand the features and contents of the technical solutions by reading the specification. BRIEF DESCRIPTION OF DRAWINGS

[0045] The advantages and implementation manners of the application will be more obvious by specifically describing the application below by referring to the drawings, and the contents shown in the drawings are only used for explaining the application and do not constitute any limitation on the application, and in the drawings:

[0046] Figure 1 The figure is a flowchart of the self-organizing neural network seismic facies analysis method of the embodiment of the application.

[0047] Figure 2 The figure is a structure diagram of the self-organizing neural network seismic facies analysis system of the embodiment of the application. DETAILED DESCRIPTION

[0048] The application is specifically described below by combining with specific embodiments, and it is necessary to point out here that the following embodiments are only used for further explaining the application and cannot be understood as limiting the protection scope of the application, and some non-essential improvements and adjustments of the application made by the skilled in the art according to the content of the application still belong to the protection scope of the application.

[0049] Under the condition of a given N-channel seismic data set D and the number K of seismic facies to be distinguished, the seismic data is organized into K (K≤N) sub-zones S k The data set of each zone after sub-zoning is represented, wherein one sub-zone is called a cluster, and each cluster represents one seismic facies.

[0050] First, for a given post-stack seismic data volume (in the time domain or in the depth domain), data to be used for classification is selected. There are two ways of selection: the first way is to select a data block with a center point as the center of a time window and a time window size as the length of the time window as the classification data by giving the center point and the time window size. The second way is to select a data block with an interpreted horizon (the horizon needs to be interpreted in advance) as the center of a time window and a time window size as the length of the time window as the classification data by giving the interpreted horizon and the time window size. Thus, a seismic data set D is generated.

[0051] Suppose we want to classify P-dimensional data X on N data points into K classes. The data can be represented as where i = 1, 2, …, N denotes the data point index, denotes the data on each data point (length of seismic trace). The standard Euclidean distance is usually used as the similarity measure between two data: Of course, other distances can also be used, such as Manhattan distance or Minkowski distance:

[0052] The number of neurons and the topology of a general self-organizing artificial neural network are defined before initialization, i.e. fixed. The number of neurons is the number of classes, i.e. the number of classes is determined in advance. When the number of classes is not known in advance, different classification number distributions need to be tested, and then the classification results of different classification numbers are compared to determine the optimal classification number. The improved self-organizing artificial neural network of the present application refers to the number of neurons which is not fixed but changes with the training time step, and correspondingly the topology between neurons also changes.

[0053] As shown in Figure 1 , the present application provides a seismic facies analysis method based on an improved self-organizing neural network, and the steps are as follows:

[0054] S1, initializing a self-organizing neural network, the self-organizing neural network being a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies types;

[0055] In this embodiment, a two-layer self-organizing neural network with K neurons is selected as the initial self-organizing neural network, which is not different from a general self-organizing artificial neural network; at the same time, the weight vector of each neuron is initialized j = 1: K denotes the neuron index, and the initialization of the neuron weight vector is to assign it as a random vector, and the size of each component is between 0 and 1; the network training time step t = 0 is initialized; the winning frequency C j of all neurons is initialized to 0; a learning rate function η(t) and a neighborhood function h j (t) are selected; the learning rate is a function, which is a function decreasing with iteration time, and can also be a constant. Commonly used function forms include: linear function inverse time function power function where η0 is the initial learning rate, which is usually a positive number less than 1, for example 0.1, η Tis the final learning rate, usually smaller than η0 and close to 0, for example 0.001, t is the current iteration time (number), T is the total iteration time (number), usually T takes the total sample number N, C is a constant, usually takes a positive integer larger than 1 and smaller than T.

[0056] The neighborhood function is a function decreasing with time. It can have several choices. One common choice is a constant, i.e. the best matching unit is adjusted by a fixed number of neurons in its neighborhood. Another choice is a Gaussian function, that is, it contains a relatively large neighborhood range at the beginning of the iteration, and becomes zero at the end of the iteration. When the M neurons are a two-dimensional planar array, each neuron is represented by a pair of positions: Suppose the best matching unit is labeled as Any neuron in its neighborhood is labeled as The Gaussian neighborhood function is generally expressed as:

[0057] where: σ(t) is the neighborhood radius, which changes with the iteration time, it can be expressed as or where σ0 is the initial neighborhood radius, σ f is the final neighborhood radius, t is the current iteration time (number), T is the total iteration time (number).

[0058] S2, calculate the similarity of the input data with each neuron and determine the winning neuron accordingly;

[0059] In specific implementation, a trace is randomly selected from the seismic data set D as input data 3, for the input data Calculate the similarity d j of the input data with each neuron:

[0060]

[0061] where, is the weight vector of each neuron, is the current input data

[0062] The winning neuron is determined according to the competition principle, specifically:

[0063] Suppose the winning neuron is labeled as BMU, and its corresponding weight vector is Then the winning neuron satisfies: That is, the neuron most similar to the input data is the winning neuron.

[0064] S3, updating the related parameters of the neuron; the related parameters of the neuron include the winning frequency of the neuron and the weight of the neuron.

[0065] In this embodiment, step S3 specifically includes:

[0066] (1) updating the winning frequency of the neuron: assuming that the winning frequency of a certain neuron is C j , the winning frequency of the neuron is updated according to the following rules:

[0067]

[0068] wherein N BMU represents the serial number of the neuron adjacent to the winning neuron;

[0069] (2) updating the weight of the neuron: the weight of the neuron is updated according to the following rules:

[0070]

[0071] wherein η(t) is a learning rate function, h j (t) is a neighborhood function of the neuron.

[0072] (3) updating the learning rate function and the neighborhood function, i.e. calculating η(t+1) and h j (t+1).

[0073] S4, every M time steps, a new neuron is added or inserted according to a first preset rule, and a neuron is deleted or discarded according to a second preset rule.

[0074] The first preset rule includes: every M (here, M is a value of the iteration number t) time steps, the neuron v q with the largest winning frequency is counted, and the neuron v q farthest from v q in the neighborhood and obtained by the neighborhood function is v f , then for the neurons in this range, the difference e between the two weight vectors is calculated:

[0075]

[0076] wherein j∈[q, f], for given parameters a and b, when a<e<b, for example when 0.5<e<3, a neuron v r is inserted between j and j+1, and the weight of the newly inserted neuron is assigned as: the winning frequency is adjusted as: M can be given as a constant at initialization, for example M=1000.

[0077] The second preset rule comprises: every M time steps, the neuron with the minimum winning frequency is counted p When the winning frequency C p of the neuron is less than a threshold, for example, C p <5, the neuron is discarded.

[0078] S5, repeating the above S2-S4 steps until the total iteration time T.

[0079] After the iteration is completed, the number of output layer neurons represents the number of data classifications, and the weight of each neuron represents the corresponding standard of the data of the class.

[0080] After step S5, the waveform classification result can be used for sedimentary facies analysis. The waveform classification result forms discrete "seismic facies" on a plane, and the seismic facies map is calibrated with petrophysical parameters or sedimentary facies, so that numerical values are added to the interpretation result, so that the qualitative seismic facies map becomes a quantitative map showing the spatial variation of the selected petrophysical parameter or sedimentary facies.

[0081] The present application provides a method for classifying seismic waveforms using an improved self-organizing neural network, which compares actual seismic data in a layer by trace (i.e., step S2), carefully depicts the lateral variation of seismic signals, and thus obtains the planar distribution rule of seismic anomalies.

[0082] The present application provides a computer-readable storage medium storing at least one computer-executable program, which, when executed by a computer, causes the computer to perform the steps of the above method.

[0083] As Figure 2 shown, a seismic facies analysis system based on an improved self-organizing neural network comprises an initialization module 10, a calculation module 20, a parameter updating module 30, and a neuron updating module 40.

[0084] The initialization module 10 is used to initialize the self-organizing neural network, which is a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies types.

[0085] In this embodiment, a two-layer self-organizing neural network with K neurons is selected as the initial self-organizing neural network, which is not different from a general self-organizing artificial neural network; and the weight vector of each neuron is initialized, wherein j=1:K represents the neuron number, and the weight vector The initialization of the input data is to assign it as a random vector, and the size of each component is between [0, 1]; initialize the network training time step t = 0; initialize the winning frequency C of all neurons j = 0; select learning rate function η(t) and neighborhood function h j (t); the learning rate is a function, which is a function of iteration time, and can also be a constant. Commonly used functions are: linear function inverse time function power function Where η0 is the initial learning rate, usually a positive number less than 1, such as 0.1, η T is the final learning rate, usually less than η0 and close to 0, such as 0.001, t is the current iteration time (number), T is the total iteration time (number), usually T takes the total sample number N, C is a constant, usually takes a positive integer greater than 1 and less than T.

[0086] The neighborhood function is a function that decreases with time. There are many choices. A common choice is a constant, that is, the best matching unit is adjusted by a fixed number of neurons. Another is the Gaussian function, that is, in the early stage of iteration, it contains a relatively large field range, and at the end of iteration it becomes zero. When M neurons are two-dimensional plane arrays, each neuron is represented by a pair of positions: Assume that the best matching unit is marked as Any neuron in its field is marked as The Gaussian field function is generally expressed as:

[0087] Where: σ(t) is the field radius, which changes with iteration time, which can be expressed as Or Where σ0 is the initial field radius, σ f is the final neighborhood radius, t is the current iteration time (number), T is the total iteration time (number).

[0088] The calculation module 20 is connected with the initialization module 10, which is used to calculate the similarity between the input data and each neuron one by one and determine the winning neuron accordingly;

[0089] In specific implementation, a trace is randomly selected from the seismic data set D as input data 3, for the input data Calculate the similarity d j between the input data and each neuron:

[0090]

[0091] Where, It is the weight vector of each neuron. For the current input data

[0092] The winning neuron is determined according to the principle of competition, specifically:

[0093] Assuming the winning neuron is labeled BMU, its corresponding weight vector is... The winning neuron then satisfies: In other words, the neuron most similar to the input data is the winning neuron.

[0094] The parameter update module 30 is connected to the calculation module 20 and is used to update the relevant parameters of the neuron. In this embodiment, the parameter update module 30 is specifically used for:

[0095] (1) Update the win frequency of a neuron: Assume that the win frequency of a certain neuron is C. j Then, the win frequency of the neuron is updated according to the following rules:

[0096]

[0097] Where N BMU Indicates the sequence number of the neuron adjacent to the winning neuron;

[0098] (2) Update the neuron weights: Update the neuron weights according to the following rules:

[0099]

[0100] Where η(t) is the learning rate function, h j (t) is the neighborhood function of the neuron.

[0101] (3) Update the learning rate function and the neighborhood function, i.e., calculate η(t+1) and h j (t+1);

[0102] The neuron update module 40 is connected to the parameter update module 30. It is used to add or insert a new neuron every M time steps according to a first preset rule, and delete or discard a neuron according to a second preset rule. The first preset rule includes: every M time steps (where M is a value for the iteration number t), the neuron v with the highest winning frequency is identified. q v is obtained from the neighborhood function. q Within the neighborhood and with v q The farthest neuron v f Then, for neurons within this range, the difference e between the two weight vectors is calculated accordingly:

[0103]

[0104] where j∈[q, f], for given parameters a and b, when a<e<b, for example when 0.5<e<3, a neuron v is inserted between j and j+1 r , and the weight of the newly inserted neuron is assigned as: The winning frequency adjustment is: M can be given as a constant at initialization, for example M=1000.

[0105] The second preset rule includes: every M time steps, the neuron v with the minimum winning frequency is counted p , when the winning frequency C p of the neuron is less than a certain threshold, for example C p <5, the neuron is discarded.

[0106] After the iteration is completed, the number of output layer neurons represents the number of data classifications, and the weight of each neuron corresponds to the standard of the corresponding class data.

[0107] Another embodiment of the present application provides a seismic facies analysis system based on the improved self-organizing neural network, in addition to the initialization module 10, the calculation module 20, the parameter updating module 30, and the neuron updating module 40, further comprising a sedimentary facies analysis module for performing sedimentary facies analysis by using the waveform classification result. The waveform classification result forms discrete "seismic facies" on a plane, and the seismic facies map is calibrated by rock physical parameters or sedimentary facies, so that numerical values are added to the interpretation result, and the qualitative seismic facies map becomes a quantitative map showing the spatial variation of the selected rock physical parameter or sedimentary facies.

[0108] The present application provides a seismic facies analysis method and system based on the improved self-organizing neural network. First, according to the lateral variability of the seismic signal in a certain target layer (i.e. the similarity between the input data and each neuron), the waveform of the seismic trace is classified by using the improved self-organizing neural network, and the classification result forms discrete "seismic facies". Then, the seismic facies is calibrated by rock physical parameters or sedimentary facies, so as to realize the description of the planar distribution of the rock physical parameters or sedimentary facies by using the seismic data.

[0109] The preferred embodiments of the present application have been described above with the aid of numerous drawings. Various modifications of the preferred embodiments of the application will be apparent to those skilled in the art and can be made without departing from the scope and spirit of the application. For example, features illustrated and described in relation to one embodiment can be incorporated into another embodiment. The description and drawings are accordingly intended to be illustrative and not restrictive.

Claims

1. A method of seismic facies analysis based on an improved self-organizing neural network, characterized in that, The method comprises the steps of: S1, initializing a self-organizing neural network, the self-organizing neural network being a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies categories; S2, calculating the similarity between input data and each neuron one by one and determining the winning neuron according to the similarity; S3, updating the related parameters of the neuron, the related parameters of the neuron including the winning frequency of the neuron and the weight of the neuron; S4, inserting a new neuron every M time steps according to a first preset rule and discarding a neuron according to a second preset rule; The S3 comprises: updating the winning frequency of the neuron, the rule being that: wherein, represents the winning frequency of the neuron, t represents the current iteration number, j represents the neuron number, represents the neuron number adjacent to the winning neuron; updating the weight of the neuron, the rule being that: wherein, represents a weight vector of each neuron, is a learning rate function, is a neighborhood function of the neuron; Updating a learning rate function and a neighborhood function ; The first preset rule is that: counting the neuron with the highest winning frequency ; from the neighborhood function within the neighborhood and the farthest neuron and for each neuron within this range, the difference e of the two weight vectors is calculated: ， wherein , for a given parameter and , when , insert a neuron between and ;​ The step S4 comprises: weighting of the inserted neuron is: ; The win frequency adjustment is: .

2. The seismic facies analysis method based on the improved self-organizing neural network according to claim 1, characterized in that, In the step S2, the similarity of the input data with each neuron is calculated is: wherein, is a weight vector for each neuron, is the current input data.

3. The seismic facies analysis method based on the improved self-organizing neural network according to claim 1, characterized in that, The neuron most similar to the input data in the step S2 is the winning neuron.

4. The seismic facies analysis method based on the improved self-organizing neural network according to claim 1, characterized in that, The second preset rule is that: The neuron with the lowest winning frequency is identified. The winning frequency of the neuron with the lowest winning frequency is... If the frequency is less than a preset threshold, the neuron with the lowest winning frequency is discarded.

5. A system for facies analysis based on improved self-organizing neural network according to any one of claims 1 to 4, characterized in that, The method comprises: an initialization module configured to initialize a self-organizing neural network, the self-organizing neural network being a two-layer self-organizing neural network with K neurons, wherein K is the number of seismic facies categories; a calculation module configured to calculate the similarity between input data and each neuron one by one and determine the winning neuron according to the similarity; a parameter updating module configured to update the related parameters of the neuron; a neuron updating module configured to insert a new neuron every M time steps according to a first preset rule and discard a neuron according to a second preset rule; The first preset rule is that: counting the neuron with the highest winning frequency ; from the neighborhood function within the neighborhood and the farthest neuron and for each neuron within this range, the difference e of the two weight vectors is calculated: ， wherein , for a given parameter and , when , insert a neuron between and . ; The second preset rule is that: The neuron with the lowest winning frequency is identified. The winning frequency of the neuron with the lowest winning frequency is... If the frequency is less than a preset threshold, the neuron with the lowest winning frequency is discarded.

6. A computer-readable storage medium storing at least one program computer-executable, the at least one program comprising instructions for causing a computer to perform the method of any one of claims 1-5. The at least one program, when executed by the computer, causes the computer to perform the steps in the method of any one of claims 1-4.

Images