A gravity matching method based on probabilistic neural network

Abstract

The invention discloses a gravity matching method based on a probabilistic neural network. The probabilistic neural network is used to match sampling points of an inertial navigation indication trajectory. The method has the advantages of fast training speed, simple learning process and minimum expected risk of misclassification. The method can obtain a Bayesian optimal estimation result, i.e., a matching point with a higher posterior probability, a high matching rate and a good matching effect, and can overcome the influence of gravity measurement errors. For points whose initial matching results are unreliable, after considering the position correlation of the inertial navigation sampling points, a particle filtering method is used to obtain the secondary matching results. Since the particle filtering method can better handle nonlinear non-Gaussian problems, the matching accuracy can be better improved, and the matching results can be made more accurate and reliable.

Description

Technical Field

[0001] The invention belongs to the technical field of inertial navigation, and in particular relates to a gravity matching method based on a probabilistic neural network. Background Art

[0002] In underwater navigation, the positioning error of the inertial navigation system will accumulate over time. Therefore, gravity matching, as an auxiliary navigation method, is often used together with inertial navigation to form a combined navigation system to obtain more accurate navigation and positioning results.

[0003] In 2007, Cheng Li and Cai Tijing proposed a gravity matching algorithm based on pattern recognition neural network. That is, using probabilistic neural network (PNN) to determine the search center point in the navigation trajectory of the inertial navigation system during gravity matching, the coordinate point closest to the gravity feature is found on the gravity background map, and the search range is planned to be a number of grid points. According to the relative position coordinate relationship of the sampling points of the inertial navigation trajectory, a certain dimension of gravity measurement value sequence is extracted from the gravity background map, and defined together with the position coordinates of the corresponding grid points into multiple pattern classes to create a pattern recognition probabilistic neural network. Input a real-time gravity measurement sequence of a certain dimension, identify it as a certain pattern class according to the correlation degree of the gravity feature, and compare the output pattern class label to determine the carrier position. [4] The algorithm has a high matching rate and good matching effect, can overcome the influence of gravity measurement errors, and has strong real-time performance.

[0004] In 2016, Yu Li et al. proposed to introduce a particle filter based on Bayesian estimation, adding the phase correlation between adjacent points of the inertial navigation indication position into the matching algorithm to form a vector matching algorithm. [5] In areas with large gravity anomalies, particle filtering is used. The vector matching algorithm takes into account the position relationship between the inertial navigation indicator and its adjacent points during matching, and corrects the matching results based on the position correlation between adjacent points. [6] The vector matching algorithm is robust to the changes in gravity anomalies in the matching area, and the matching results are more accurate and reliable. Summary of the invention

[0005] In view of this, an object of the present invention is to provide a gravity matching method based on a probabilistic neural network, which can achieve higher matching accuracy.

[0006] A gravity matching method based on probabilistic neural network includes the following steps:

[0007] Step 1: Take each sampling point in the inertial navigation indication trajectory as the starting point, take L sampling points forward along the timeline of the inertial navigation indication trajectory, and obtain a trajectory. Set the starting point as point C; take point C as the search center, and in the grid map divided by the gravity background map, take M grids in the longitude direction and N grids in the latitude direction to form a search area with point C as the center point. There are S grid points in the area:

[0008] S=(2M+1)×(2N+1) (2-1)

[0009] Wherein, L is at least 3; M and N are at least 2 respectively;

[0010] Step 2: Determine the relative position of the sampling points of the inertial navigation indication trajectory: The longitude and latitude coordinates of the center point C are expressed as On the trajectory starting from C, each sampling point B i The latitude and longitude distance to the starting point is recorded as:

[0011] Δλ(CB i )=|λ C -λ Bi |,i=1,2...L (2-2)

[0012]

[0013] in For any point B i The latitude and longitude coordinates of;

[0014] Step 3: In the search area, the coordinate point of the kth grid point in the gravity background map is expressed as (i k ,j k ), which is defined as:

[0015] i k =i-M+R k (2-4)

[0016] j k =j+N+R k ×(2N+1)-k+1 (2-5)

[0017] Where k = 1, 2, ..., S; R k The quotient of k divided by (2N+1) is an integer. The longitude and latitude coordinates of the point are determined according to the coordinate correspondence of the gravity background map. According to the trajectory indicated by the inertial navigation, take the kth grid point as the starting point, take L sampling points forward, and get a trajectory B with point k as the starting point k According to the relative position of the trajectory sampling points indicated by the inertial navigation in step 2, trajectory B kThe corresponding latitude and longitude positions of the L points are:

[0018] λ(k)=Δλ(CB i )+λ(i k ) (2-6)

[0019]

[0020] Then the trajectory B k Defined as mode k, the gravity measurement values ​​of L points in mode k are determined according to the gravity background map, and the gravity measurement value sequence is expressed as g mk , the gravity measurement sequence corresponding to all S modes in the search area forms the matrix P = [g m1 g m2 ... g mS ];

[0021] Step 4: Construct a probabilistic neural network, specifically:

[0022] Construct a 4-layer probabilistic neural network with input layer, pattern layer, accumulation layer and output layer; the input layer has L nodes; the number of nodes in the pattern layer and accumulation layer are both S;

[0023] The output layer performs competitive operation on the correlation data of the accumulation layer, takes the pattern with the highest probability of approaching the gravity measurement value of the center point as the result, and outputs it through the output layer. The grid point corresponding to the pattern is the matching point of the center point.

[0024] Step 5: Input the gravity measurement value sequence of dimension L corresponding to each inertial navigation sampling point to be matched, and the matrix P consisting of the gravity measurement value sequence composed of the corresponding S patterns, into the probabilistic neural network for gravity matching to obtain the corresponding matching points; determine the coordinate information of the underwater vehicle according to the matching points.

[0025] Furthermore, it is checked whether the credibility of each matching point meets the requirements. If not, the matching is performed again.

[0026] Preferably, the method for checking the credibility of each matching point is:

[0027] Suppose the sampling point at sampling time h in the matching point trajectory is P h ', according to the inertial navigation indicated trajectory, the four matching points before the sampling time h are obtained and form a point set {P h-4 ',P h-3 ',P h-2 ',P h-1 ',P h '};

[0028] The sampling point of the inertial navigation system at sampling time k during navigation is set as P h, take the 4 sampling points before the sampling time h, and the point set is {P h ,P h-1 ,P h-2 ,P h-3 ,P h-4};

[0029] When it satisfies formula (2-12), the matching point P h 'Satisfy credibility requirements;

[0030] n×dist(P h ,P h-n )-σ≤dist(P h ',P h-n ')≤n×dist(P h ,P h-n )+σ,n=1,2,3,4 (2-12)

[0031] Among them, σ represents the accuracy of the longitude and latitude of the gravity background map;

[0032]

[0033]

[0034] (m h ,n h ) and (m h-n , n h-n ) are the sampling points P h and P h-n The longitude and latitude information, (m′ h ,n′ h ) and (m′ h-n , n′ h-n ) are the matching points P h ' and P′ h-n The longitude and latitude information.

[0035] Preferably, the rematching method is:

[0036] When P h ' is an untrustworthy point, according to the particle filtering algorithm, P h 'The four previous matching points are used as the center point, the particle swarm is initialized, the longitude and latitude information of each particle is obtained, and then the gravity measurement value of each particle is determined. h-n The weight of the particle is determined by the difference in the corresponding gravity measurement values. The smaller the difference, the greater the weight. Then, multiple new particles are scattered around the particle with the largest weight to obtain a new particle group, and the geometric center of the new particle group is calculated and set as Then the geometric centers of four particle groups are obtained;

[0037] P h 'The geometric center of the particle swarm corresponding to the previous four sampling points is calculated by weighted least squares method to obtain the final result as the final matching point.

[0038] Preferably, the activation function of the pattern layer is a Gaussian function, and the node bias is set to σ z =δ z ξ;

[0039] ξ is the node bias coefficient; δ z It is the average roughness of the gravity roughness in the longitude and latitude directions of the search area.

[0040] Preferably, the average roughness δ z The calculation methods include:

[0041] Calculate the gravity roughness of the set area: Based on the inertial navigation indication trajectory, calculate the grids crossed by the longitude and latitude directions on the trajectory within (L-1)Δt time, thereby forming a crossing area; expand M grids on both sides of the longitude direction of the crossing area, and expand N grids on both sides of the latitude direction, then the number of grids crossed in the longitude direction of the expanded area is X=2M+Lx, and the number of grids crossed in the latitude direction is Y=2N+Ly, where Lx and Ly are the number of grids crossed in the longitude and latitude directions within (L-1)Δt time, and H is used in this area. ij Represents the gravity measurement value at the coordinate point (i, j) in the search area. The roughness calculation in the longitude and latitude directions corresponds to equations (2-14) and (2-15) respectively:

[0042]

[0043]

[0044] Mean roughness definition:

[0045]

[0046] Preferably, the connection weight matrix between the accumulation layer nodes and the pattern layer is a pattern class matrix.

[0047] The present invention has the following beneficial effects:

[0048] The invention discloses a gravity matching method based on a probabilistic neural network, which uses a probabilistic neural network to match sampling points of an inertial navigation indication trajectory. The method has fast training speed, simple learning process, and minimum expected risk of misclassification. The method can obtain a Bayesian optimal estimation result, that is, a matching point with a higher posterior probability, a high matching rate, and a good matching effect, and can overcome the influence of gravity measurement errors.

[0049] For points whose initial matching results are unreliable, after considering the position correlation of the inertial navigation sampling points, the particle filtering method is used for secondary matching results. Since the particle filtering method can better handle nonlinear non-Gaussian problems, it can better improve the matching accuracy and make the matching results more accurate and reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The grid lines of the gravity map of the present invention;

[0051] Figure 2 It is a schematic diagram of the probabilistic neural network structure;

[0052] Figure 3 is the sampling scheme of the vector matching algorithm;

[0053] Figure 4 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0054] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0055] like Figure 4 As shown, a gravity matching method based on a probabilistic neural network of the present invention specifically includes the following steps:

[0056] Step 1: Determine the search area: When the heading and speed of the underwater vehicle remain unchanged, collect a gravity measurement data point every interval Δt. Assume that the current inertial navigation indication trajectory has a total of L_sum sampling points. Assume that at a certain sampling time t s When , the sampling point is C, and each sampling point is used as the center point to be matched; taking each center point as the starting point, take L sampling points forward along the inertial navigation indication trajectory timeline to obtain a trajectory with point C as the starting point. The gravity measurement data corresponding to the L sampling points are arranged in order into a vector All sampling points will form a gravity measurement value with a vector dimension of L_sum. According to its inertial navigation data, the coordinates of the center C on the gravity map grid map can be determined. Take point C as the search center, in the grid map divided by the gravity background map, take M grids in the longitude direction and N grids in the latitude direction, forming a search area with point C as the center point. There are S grid points in the area:

[0057] S=(2M+1)×(2N+1) (2-1)

[0058] In this embodiment, L is at least 3, and M and N are 2 respectively;

[0059] Step 2: Determine the relative position of the sampling points of the inertial navigation indication trajectory: In the inertial navigation data, the longitude and latitude coordinates of the center point C are expressed as On the trajectory starting from C, each sampling point B i The latitude and longitude distance to the starting point is recorded as:

[0060] Δλ(CB i )=|λ C -λ Bi |,i=1,2...L (2-2)

[0061]

[0062] in For any point B i The latitude and longitude coordinates of .

[0063] Step 3: Define k-mode: In the search area, the coordinate point of the k-th grid point in the gravity background map is expressed as (i k ,j k ), which is defined as:

[0064] i k =i-M+R k (2-4)

[0065] j k =j+N+R k ×(2N+1)-k+1 (2-5)

[0066] Where k = 1, 2, ..., S; R k The quotient of k divided by (2N+1) is an integer. According to the coordinate correspondence of the gravity background map, the longitude and latitude coordinates of the point can be determined as According to the trajectory indicated by the inertial navigation, take the kth grid point as the starting point, take L sampling points forward, and get a trajectory B with point k as the starting point k According to the relative position of the trajectory sampling points indicated by the inertial navigation in step 2, trajectory B k The corresponding latitude and longitude positions of the L points are:

[0067] λ(k)=Δλ(CB i )+λ(i k ) (2-6)

[0068]

[0069] Then the trajectory B k Defined as mode k, the gravity measurement values ​​of L points in mode k are determined according to the gravity background map, and the gravity measurement value sequence is g mk , the gravity measurement sequence corresponding to all S modes in the search area forms the matrix P = [g m1 g m2 ... g mS ].

[0070] Step 4: Establish a probabilistic neural network:

[0071] according to Figure 2 Construct a 4-layer feedforward neural network, including input layer, pattern layer, accumulation layer and output layer.

[0072] The input layer has L nodes (the feature dimension of each sample), which pass the correlation between the input vector and the sample vector to the pattern layer;

[0073] The number of nodes in the pattern layer and the accumulation layer is S. Each node in the pattern layer is the corresponding column vector of the training sample matrix P. The activation function is the Gaussian function, and the node bias calculation is σ z =δ z ξ, the gravitational roughness Q in the longitude and latitude directions of the search area λ , Q Φ The node bias coefficient ξ takes different values ​​in search areas with different regional feature information because of different gravity anomaly change rates;

[0074] All nodes in the accumulation layer have no bias numbers, and the connection weight matrix with the pattern layer is the pattern class matrix T2;

[0075] The output layer performs competitive operations on the correlation data of the accumulation layer, and takes the pattern with the highest probability of approaching the gravity measurement value of the center point as the result output. The grid point corresponding to the pattern is the matching point of the center point, and the coordinate information of the underwater vehicle is further determined based on the matching point.

[0076] Specific parameter settings of probabilistic neural network:

[0077] (1) When the output matrix of the accumulation layer is H:

[0078]

[0079] Among them, h ij Represents the jth input sample P j The output of a node in the pattern layer belongs to the i-th pattern class in the accumulation layer. Let the weight matrix from the accumulation layer to the output layer be W = [w1 w2 ... w S ], that is, w i Represents the connection weight from the ith node of the accumulation layer to the output layer node, then the target matrix can be calculated as T2=WH.

[0080] (2) Pattern indicator matrix T: Set the pattern indicator matrix to T = [1 2 ... S]. The matrix P and the matrix T can be trained through a neural network to obtain the target matrix, which is an S-dimensional square matrix. When a certain pattern k_result is determined to be the most likely recognition result, the value in the k_result row and the k_result column is 1, and the rest of the positions are 0.

[0081] (3) Calculate the gravity roughness of the set area: Based on the inertial navigation indication trajectory, calculate the grids that the trajectory crosses in the longitude and latitude directions within (L-1)Δt time on the trajectory, thereby forming a crossing area; expand M grids on both sides of the longitude direction of the crossing area, and expand N grids on both sides of the latitude direction. The number of grids crossed in the longitude direction of the expanded area is X=2M+Lx, and the number of grids crossed in the latitude direction is Y=2N+Ly, where Lx and Ly are the number of grids crossed in the longitude and latitude directions within (L-1)Δt time. H is used in this area. ij Represents the gravity measurement value at the coordinate point (i, j) in the search area. The roughness calculation in the longitude and latitude directions corresponds to equations (2-14) and (2-15) respectively:

[0082]

[0083]

[0084] The average roughness is:

[0085]

[0086] Step 5, Probabilistic neural network to achieve gravity matching: The gravity measurement value sequence of dimension L corresponding to each inertial navigation sampling point to be matched, and the matrix P composed of the gravity measurement value sequence composed of the corresponding S patterns, are input into the neural network for gravity matching. In the present invention, the last point of the INS trajectory is taken as the starting point, and the center point is moved forward in sequence to update, and the previous steps are repeated until the matching is terminated when there are L-1 unmatched points left, and the initial matching points of all inertial navigation trajectory sampling points can be obtained.

[0087] Step 6: Check the credibility of the initial matching points:

[0088] The positional relationship between the inertial navigation system sampling point and the initial matching point can be used to obtain vector elements. The initial matching point can be verified by vectors, specifically:

[0089] The sampling scheme of the vector matching algorithm is as follows Figure 3 As shown: Let the sampling point at sampling time h in the matching point trajectory be P h ', according to the inertial navigation indicated trajectory, the four matching points before the sampling time h are obtained and form a point set {P h-4 ',Ph-3 ',P h-2 ',P h-1 ',P h '};

[0090] The sampling point of the inertial navigation system at sampling time k during navigation is set as P h , take the 4 sampling points before the sampling time h, and the point set is {P h ,P h-1 ,P h-2 ,P h-3 ,P h-4};

[0091] When it satisfies formula (2-12), the matching point P h 'Satisfy credibility requirements;

[0092] n×dist(P h ,P h-n )-σ≤dist(P h ',P h-n ')≤n×dist(P h ,P h-n )+σ,n=1,2,3,4 (2-12)

[0093] Among them, σ represents the accuracy of the longitude and latitude of the gravity background map;

[0094]

[0095]

[0096] (m h ,n h ) and (m h-n , n h-n ) are the sampling points P h and P h-n The longitude and latitude information, (m′ h ,n′ h ) and (m′ h-n , n′ h-n ) are the matching points P h ' and P′ h-n The longitude and latitude information.

[0097] Step 7: Untrusted matching points need to be matched again:

[0098] When P h ' is an untrustworthy point and needs to be re-matched. According to the particle filter algorithm, P h 'The previous four matching points P h-n' is the center point, the particle swarm is initialized, the longitude and latitude information of each particle is obtained, and then the gravity measurement value of each particle is determined. h-n The difference of the corresponding gravity measurement values ​​determines the weight of the particle. The smaller the difference, the greater the weight. The particle with the largest weight needs to be resampled, that is, some new particles are scattered around the particle with large weight to obtain a new particle group, and the geometric center of the new particle group is calculated, which is set as The local optimal solution of the geometric center of the four particle swarms can be obtained

[0099] P h 'The geometric center of the particle swarm corresponding to the previous four sampling points is calculated by weighted least squares method to obtain the final result As the final matching point, its latitude and longitude information is output; the weighted model is as follows:

[0100]

[0101] Normalize these weights:

[0102]

[0103] The final weighted P k Estimated value:

[0104]

[0105] Step 8: Loop through steps 6 and 7 until the last point is matched.

[0106] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A gravity matching method based on probabilistic neural network, characterized in that: The steps include: Step 1: Take each sampling point in the inertial navigation indication trajectory as the starting point, take L sampling points forward along the time line of the inertial navigation indication trajectory, and obtain a trajectory, and set the starting point as point C; With point C as the search center, in the grid map divided by the gravity background map, take M grids in the front and back directions in the longitude direction, and take N grids in the top and bottom directions in the latitude direction to form a search area with point C as the center point. There are S grid points in the area: S=(2M+1)×(2N+1) (2-1) Wherein, L is at least 3; M and N are at least 2 respectively; Step 2: Determine the relative position of the sampling points of the inertial navigation indication trajectory: The longitude and latitude coordinates of the center point C are expressed as On the trajectory starting from C, each sampling point B i The latitude and longitude distance to the starting point is recorded as: Dλ(CB i )=|λ C -l Bi |,i=1,2...L (2-2) in For any point B i The latitude and longitude coordinates of; Step 3: In the search area, the coordinate point of the kth grid point in the gravity background map is expressed as (i k ,j k ), which is defined as: i k =i-M+R k (2-4) <h2 style=";text-align:left;direction:ltr">j<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> (j+N+R)<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> ×(2N+1)-k+1 (2-5) Where k = 1, 2, ..., S; R k The quotient of k divided by (2N+1) is an integer. The longitude and latitude coordinates of the point are determined according to the coordinate correspondence of the gravity background map. According to the trajectory indicated by the inertial navigation, take the kth grid point as the starting point, take L sampling points forward, and get a trajectory B with point k as the starting point k , according to the relative position of the trajectory sampling points indicated by the inertial navigation in step 2, trajectory B k The corresponding latitude and longitude positions of the L points are: λ(k)=Δλ(CB i )+λ(i k ) (2-6) Then the trajectory B k Defined as mode k, the gravity measurement values ​​of L points in mode k are determined according to the gravity background map, and the gravity measurement value sequence is expressed as g mk , the gravity measurement sequence corresponding to all S modes in the search area forms the matrix P = [g m1 g m2 ... mS ]; Step 4: Construct a probabilistic neural network, specifically: Construct a 4-layer probabilistic neural network with input layer, pattern layer, accumulation layer and output layer; the input layer has L nodes; the number of nodes in the pattern layer and accumulation layer are both S; The output layer performs competitive operation on the correlation data of the accumulation layer, takes the pattern with the highest probability of approaching the gravity measurement value of the center point as the result, and outputs it through the output layer. The grid point corresponding to the pattern is the matching point of the center point. Step 5: Input the gravity measurement value sequence of dimension L corresponding to each inertial navigation sampling point to be matched, and the matrix P consisting of the gravity measurement value sequence composed of the corresponding S patterns, into the probabilistic neural network for gravity matching to obtain the corresponding matching points; determine the coordinate information of the underwater vehicle according to the matching points.

2. A gravity matching method based on probabilistic neural network as claimed in claim 1, characterized in that: Check whether the credibility of each matching point meets the requirements. If not, re-match.

3. A gravity matching method based on probabilistic neural network as claimed in claim 2, characterized in that: The method to check the credibility of each matching point is: Suppose the sampling point at sampling time h in the matching point trajectory is P h ', according to the inertial navigation indicated trajectory, the four matching points before the sampling time h are obtained and form a point set {P h-4 ',P h-3 ',P h-2 ',P h-1 ',P h '}; The sampling point of the inertial navigation system at sampling time k during navigation is set as P h , take the 4 sampling points before the sampling time h, and the point set is {P h ,P h-1 ,P h-2 ,P h-3 ,P h-4 }; When it satisfies formula (2-12), the matching point P h 'Satisfy credibility requirements; n×dist(P h ,P h-n )-σ≤dist(P h ',P h-n ')≤n×dist(P h ,P h-n )+σ,n=1,2,3,4 (2-12) Among them, σ represents the accuracy of the longitude and latitude of the gravity background map; (m h ,n h ) and (m h-n , n h-n ) are the sampling points P h and P h-n The longitude and latitude information, (m′ h ,n′ h ) and (m′ h-n , n′ h-n ) are the matching points P h ' and P′ h-n The longitude and latitude information.

4. A gravity matching method based on probabilistic neural network as claimed in claim 2, characterized in that: The rematching method is: When P h ' is an untrustworthy point, according to the particle filtering algorithm, P h 'The four previous matching points are used as the center point, the particle swarm is initialized, the longitude and latitude information of each particle is obtained, and then the gravity measurement value of each particle is determined. h-n The weight of the particle is determined by the difference in the corresponding gravity measurement values. The smaller the difference, the greater the weight. Then, multiple new particles are scattered around the particle with the largest weight to obtain a new particle group, and the geometric center of the new particle group is calculated and set as Then the geometric centers of four particle groups are obtained; P h 'The geometric center of the particle swarm corresponding to the previous four sampling points is calculated by weighted least squares method to obtain the final result as the final matching point.

5. A gravity matching method based on probabilistic neural network as claimed in claim 1, characterized in that: The activation function of the pattern layer is a Gaussian function, and the node bias is set to σ z =δ z ξ; ξ is the node bias coefficient; δ z It is the average roughness of the gravity roughness in the longitude and latitude directions of the search area.

6. A gravity matching method based on probabilistic neural network as claimed in claim 5, characterized in that: Average roughness δ z The calculation methods include: Calculate the gravity roughness of the set area: Based on the inertial navigation indication trajectory, calculate the grids crossed by the longitude and latitude directions on the trajectory within (L-1)Δt time, thereby forming a crossing area; expand M grids on both sides of the longitude direction of the crossing area, and expand N grids on both sides of the latitude direction, then the number of grids crossed in the longitude direction of the expanded area is X=2M+Lx, and the number of grids crossed in the latitude direction is Y=2N+Ly, where Lx and Ly are the number of grids crossed in the longitude and latitude directions within (L-1)Δt time, and H is used in this area. ij Represents the gravity measurement value at the coordinate point (i, j) in the search area. The roughness calculation in the longitude and latitude directions corresponds to equations (2-9) and (2-10) respectively: Mean roughness definition:

7. A gravity matching method based on probabilistic neural network as claimed in claim 1, characterized in that: The connection weight matrix between the accumulation layer nodes and the pattern layer is the pattern class matrix.

Images