Medical data-oriented deep convolutional fuzzy neural network and training method thereof
By combining deep convolutional fuzzy neural networks with fuzzy logic and deep learning methods, the problems of computational cost and interpretability of high-dimensional medical data are solved, improving data processing capabilities and training efficiency, and enabling better analysis of patient medical data.
Patent Information
- Application Number
- CN202310431951.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-21
- Publication Date
- 2025-11-25
- Estimated Expiration
- 2043-04-21
AI Technical Summary
Existing fuzzy inference systems are computationally expensive and struggle to express data relationships when processing high-dimensional medical data. Neural networks lack interpretability when processing medical data, which affects the reliability of prediction results.
A deep convolutional fuzzy neural network is adopted, which combines fuzzy logic antecedent, deep convolutional computation and fuzzy result representation. Medical data is processed through membership functions, convolutional layers and fully connected layers. The network is trained using a hybrid learning method, including grid partitioning initialization and gradient descent algorithm.
It has improved the processing capabilities of high-dimensional medical data while maintaining interpretability, enhanced training efficiency and data processing capabilities, and enabled better mining of correlation information in patients' physiological information.
Smart Images
Figure CN116384450B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and in particular to a deep convolutional fuzzy neural network for medical data and its training method. Background Technology
[0002] Fuzzy inference systems are limited by their data pathways and complex structures. When faced with large-scale, high-dimensional data, they often require extremely high computational costs. At the same time, it is difficult to express the relationships between data, resulting in the prediction accuracy of fuzzy inference systems being lower than that of some algorithmic models with network structures.
[0003] While neural network algorithms can effectively process high-dimensional data, the feedforward propagation network path, constructed through the topological relationships of a large number of neurons, results in the weight coefficients of the neurons in the hidden layers not being directly related to the results or representing any real-world significance directly related to the task being processed. This is what is commonly referred to as the "black box" nature, which leads to a lack of interpretability in neural networks and affects the credibility and acceptability of the results.
[0004] Compared to other types of data, medical data often has more potential correlations among its various attributes and characteristics. There are some logical connections between different physiological indicators or physical characteristics of patients, which makes some ordinary reasoning models unable to handle medical data well. Summary of the Invention
[0005] In view of this, the present invention aims to propose a deep convolutional fuzzy neural network for medical data, which incorporates the concept of deep neural networks into the fuzzy inference system to achieve both interpretability and the ability to process high-dimensional data.
[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:
[0007] A deep convolutional fuzzy neural network for medical data is characterized by including a medical data interpretability prediction model (IP-DCFNN) based on a deep convolutional fuzzy neural network.
[0008] The IP-DCFNN consists of three parts:
[0009] Fuzzy logic preamble: The fuzzy logic preamble extracts input data, and the input data is transformed from numerical values into a set of membership degree values for fuzzy linguistic scalars through the operation of the membership function in the fuzzy logic preamble;
[0010] The depthwise convolution calculation part extracts hidden features from the input rule weights and converts the hidden layer weights into high-dimensional information representations.
[0011] Fuzzy result representation part: The fuzzy result representation part is used to handle the defuzzification process in fuzzy inference.
[0012] Furthermore, the membership function Used to map numerical input data to the range of 0 and 1;
[0013] The input data x i For [x1, ..., x] n Within the interval, the membership function Input data x i Mapped to the fuzzy set [low, high], where k = 1, ..., K and represents the selection of fuzzy linguistic variables, the parameters of the membership function are μ and σ;
[0014] The fuzzy logic antecedent is formed using the Takagi-Sugeno fuzzy model.
[0015]
[0016]
[0017] The fuzzy rules described in the text;
[0018] The fuzzy variables generated from different dimensions of the input data are arranged and combined to construct the antecedent of the rule. By multiplying the membership degree corresponding to each fuzzy variable in the rule antecedent, the initial excitation intensity of the rule can be obtained. The initial excitation intensity represents the functional weight of the rule.
[0019] The initial excitation intensity of rule j is calculated from the membership degrees of all fuzzy variables in the rule's antecedents, using the following formula: That is, the weight of the fuzzy rule is obtained by performing an "AND" fuzzy logic operation on each fuzzy variable.
[0020] Furthermore, the deep convolution part takes the initial excitation intensity generated by the fuzzy logic preamble as the initial input and outputs the final excitation intensity, which has the same dimension as the initial excitation intensity.
[0021] The depthwise convolution calculation part mainly consists of convolutional layers and fully connected layers;
[0022] The convolutional layer includes a one-dimensional convolutional layer, a ReLU activation layer, and a max pooling layer;
[0023] The fully connected layer includes a linear layer and a ReLU activation layer;
[0024] The initial input passed through The convolutional processing is extended to a higher dimension by the one-dimensional convolutional kernel and flattened back to the initial dimension in the fully connected layer, wherein, This represents the i-th data unit in the l-th convolutional layer. Let b represent the weight of the l-th convolutional kernel, where b is the weight of the a-th kernel. l Indicates the bias of this layer;
[0025] The network layers following the convolutional layer are fully connected, meaning that each node in layer l is connected to all nodes in layer (l-1). The parameter w represents the connection weight, and b represents the node bias. The formula for the feedforward propagation of the neural network is:
[0026] In the network layers following the convolutional layer, the activation function is defined as σ(x) = 1 / (1+e^x). -x One-dimensional convolution is used to process the input excitation intensity.
[0027] Furthermore, the blurred result representation part represents the final excitation intensity [W′1, W′2, ..., W′] output in the depthwise convolution calculation part. N The normalization algorithm is used to calculate the normalized weights and represent the activation level of each rule;
[0028] The formula for the normalization algorithm is: in, This represents the normalized excitation intensity of rule j and its contribution to the total weight.
[0029] The consequent of a fuzzy rule is a linear combination of the input data, expressed as: and That is, the consequent parameter of the j-th rule;
[0030] The total output Z is defined as the weighted sum of the outputs of all rules, and the specific formula is as follows: The weight of each rule refers to the normalized excitation intensity of that rule.
[0031] For subsequent parameters The calculation uses a least squares estimator (LSE) to obtain the optimization result, letting... And θ = [θ1, θ2, ..., θ N ] T Let F = [f1, f2, ..., f N Let the estimation error be e, then the corrected formula for Z is Z = Fθ + e, and the mean square error function of the result is... In the definition, m represents the length of the dataset, and y... t This represents the actual label of the t-th data point in the dataset. This represents the t-th input data from the dataset;
[0032] The optimization objective is to minimize the squared error when... When Z = Fθ + e is substituted, the data can be fitted using LSE and the consequent parameters in F can be obtained.
[0033] Furthermore, the fuzzy result represents a portion of the output prediction result. The loss value is calculated using the mean squared error (MSE), and the parameters of IP-FDCNN are updated using the backpropagation algorithm (GD). The loss function is:
[0034] Compared with the prior art, the present invention has the following advantages:
[0035] The deep convolutional fuzzy neural network for medical data described in this invention achieves both interpretability and the ability to process high-dimensional data by incorporating the concept of a deep neural network into a fuzzy inference system. Specifically, this invention represents the fuzzification and defuzzification parts of the fuzzy inference system using a network architecture, and inserts a deep convolutional computation module between fuzzification and defuzzification. Through the high-dimensional data mapping and processing capabilities of the deep convolutional network in this module, the data processing capability of the entire model system is improved. Although the hidden weights of the deep convolutional part cannot be well interpreted due to the topological limitations of the neural network, the fuzzification and defuzzification parts can extract a relatively rich rule base and membership information, thereby improving the system's data processing capability without sacrificing interpretability.
[0036] Another objective of this invention is to propose a training method for deep convolutional fuzzy neural networks for medical data. By proposing a fuzzy membership parameter initialization algorithm based on grid partitioning, the parameters of the fuzzy membership function in the model can be adaptively initialized according to the data distribution, which greatly improves the training efficiency.
[0037] To achieve the above objectives, the technical solution of the present invention is implemented as follows:
[0038] A training method for a deep convolutional fuzzy neural network for medical data includes two stages: initialization of model parameters and iterative update of parameters. The training method for the deep convolutional fuzzy neural network for medical data adopts a hybrid learning approach, in which specific parameters are learned and updated through different algorithms during feedforward and backpropagation.
[0039] Furthermore, the training method for deep convolutional fuzzy neural networks for medical data employs a hybrid learning approach as follows:
[0040] The antecedent parameters are located in the antecedent part of the fuzzy logic. Their initialization algorithm is the initialization algorithm of grid partitioning, their update timing is backpropagation, and their update algorithm is the gradient descent algorithm.
[0041] The consequent parameters are located in the fuzzy result representation part. Their initialization algorithm is the zero initialization algorithm, their update timing is feedforward propagation, and their update algorithm is the least squares estimation algorithm.
[0042] The node weights are located in the depthwise convolution calculation part. Their initialization algorithm is the He-normalization algorithm, their update timing is backpropagation, and their update algorithm is gradient descent.
[0043] The node bias is located in the depthwise convolution computation part. Its initialization algorithm is the He-normalization algorithm, its update timing is backpropagation, and its update algorithm is the gradient descent algorithm.
[0044] Furthermore, the antecedent parameter of the fuzzy logic antecedent part is represented as: In the membership functions of the fuzzy function {μ, σ}, μ represents the fuzzy center and σ represents the fuzzy width;
[0045] The antecedent parameters are initialized using a grid partition-based initialization algorithm.
[0046] The grid partitioning is a method for dividing data space, which divides the input data space into grid subspaces parallel to the axes based on the membership function of each feature;
[0047] The initialization algorithm based on grid partitioning initializes the parameters of each membership function according to the predefined partition grid divided by the grid. For data samples falling in the grid, the fuzzy center μ is initialized to the median of the data in this grid, and the fuzzy width σ is initialized to a linear approximation of the data size in the grid, with its coefficients coming from linear fitting.
[0048] The parameters for the depthwise convolution computation are the weights and biases of the neurons in the convolutional and fully connected layers. The He-uniform algorithm is used to initialize these parameters, and the specific steps are as follows:
[0049] A. Samples are drawn to form a uniform distribution of [-limit, limit];
[0050] B. Limit in In the calculation, This represents the number of input neurons in layer l;
[0051] The consequent parameters of the fuzzy result representation are the parameters in the linear expression of the output of each rule, specifically the Rule. j :
[0052] The results section of the rules is shown below;
[0053] Before training the model, apply simple zero initialization to all consequent parameters (i.e., assign all consequent parameters in the model to 0 at the start of training).
[0054] Furthermore, the IP-DCFNN model is trained using a hybrid learning method;
[0055] The model updates the consequent parameters using the least squares estimation algorithm (LSE) during the feedforward process and updates the antecedent and hidden layer parameters using the gradient descent algorithm (GD) during the backpropagation process.
[0056] During backpropagation, the gradient is as follows: The calculation is shown, where C represents... The loss value calculated in the figure, #(n) represents the number of nodes in the l-th layer that are affected by the nodes in the (l+1)-th layer;
[0057] The weights of nodes in the network layer are then updated in reverse order using the calculated gradient values.
[0058] Compared with the prior art, the present invention has the following advantages:
[0059] The training method for deep convolutional fuzzy neural networks for medical data described in this invention uses a fuzzy membership parameter initialization algorithm based on grid partitioning. This algorithm can adaptively initialize the parameters of the fuzzy membership function in the model according to the data distribution, which greatly improves the training efficiency. Attached Figure Description
[0060] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0061] Figure 1 This is a schematic diagram of the structure of the deep convolutional fuzzy neural network for medical data and its training method as described in an embodiment of the present invention;
[0062] Figure 2 This refers to the membership function of the fuzzy logic preamble of the deep convolutional fuzzy neural network for medical data and its training method described in this embodiment of the invention.
[0063] Figure 3 This is a simplified structure of the deep convolution calculation part of the deep convolutional fuzzy neural network for medical data and its training method as described in the embodiments of the present invention;
[0064] Figure 4This refers to the Conv1d layer of the deep convolution calculation part of the deep convolutional fuzzy neural network for medical data and its training method described in this embodiment of the invention.
[0065] Figure 5 This is a schematic diagram of grid partitioning for a deep convolutional fuzzy neural network for medical data and its training method as described in an embodiment of the present invention. Detailed Implementation
[0066] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0067] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "back," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0068] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0069] This embodiment relates to a deep convolutional fuzzy neural network for medical data and its training method, including a medical data interpretability prediction model (IP-DCFNN) based on a deep convolutional fuzzy neural network.
[0070] like Figure 1 As shown, IP-DCFNN consists of three parts: a fuzzy logic preamble, which extracts the input data and transforms it from numerical values into a set of membership values for fuzzy linguistic scalars through membership function operations; a deep convolution computation part, which extracts hidden features from the input rule weights and converts them into high-dimensional information representations; and a fuzzy result representation part, which handles the defuzzification process in fuzzy inference. Specifically, the fuzzy logic preamble and fuzzy result representation part serve as tools for fuzzification and defuzzification of fuzzy rules in the model, respectively, while the deep convolution computation part processes the hidden weights using high-dimensional representations. Although the deep convolution process largely adopts a neural network architecture, and the hidden weights are not interpretable, interpretable information can be obtained from the other two parts, such as fuzzy membership degrees, fuzzy rules, and the trigger strength of each rule. Therefore, this model can process medical data in an interpretable way and avoid a decrease in prediction accuracy.
[0071] The fuzzy logic antecedent part is used to form the basic framework of fuzzy if-then rules and represent the fuzzy antecedents in the rules. In this part, the input data is extracted by fuzzy logic, and the input data is transformed from numerical values into a set of membership degree values for fuzzy linguistic scalars through membership function operations.
[0072] like Figure 2 As shown, the membership function μ * (x) is used to map numerical input data to a value within the interval of 0 and 1, representing the membership degree in the defined fuzzy language set. For membership functions, the Gaussian membership function has stronger expressive power and better performance compared to other function forms; therefore, in the model proposed in this embodiment, the Gaussian membership function is used to calculate the membership degree. Specifically, for [x1, ..., x...], ... n Input data x in ] i Using membership functions It is mapped to a fuzzy set [low, high], where k = 1, ..., K (K represents the size of the fuzzy set, which is 2 here) and represents the selection of fuzzy linguistic variables; the parameters of the membership function are μ and σ, which greatly affect the approximation ability of the fuzzy system. In the IP-DCFNN model, this embodiment uses a grid partitioning-based initialization algorithm to initialize these parameters and updates them in each iteration during model training using the gradient descent algorithm.
[0073] The antecedent part of the fuzzy logic is formed using the Takagi-Sugeno fuzzy model.
[0074]
[0075] The fuzzy rules (TS rules) described in [the document]. j (j = 1, ..., N and N = K) n The IF statement describes the content of the j-th fuzzy rule. The part following "IF" is the antecedent of the rule, represented by fuzzy linguistic variables for each input data point, where x... i (i = 1, 2, ..., n) are the input data. These are fuzzy variables in a fuzzy set. The part after "THEN" is the consequent of the rule, which represents the result of the rule through a nonfuzzy linear combination of the input variables, where the output f... j This represents the final excitation intensity of rule j.
[0076] The fuzzy logic antecedent can adaptively generate N rules (specifically referring to the antecedent for rule generation), where N = K. n(where n is the dimension of the input data and K is the size of the fuzzy set). From this, we can see that the basic method of adaptive generation is to construct the antecedent of the rule by arranging and combining the fuzzy variables generated from different dimensions of the input data. By multiplying the membership degree corresponding to each fuzzy variable in the rule antecedent, we can obtain the initial excitation intensity of the rule, which represents the functional weight of the rule. As in... In this context, the initial excitation intensity of the j-th rule is calculated from the membership degree of all fuzzy variables in the rule's antecedents. That is, the weight of the fuzzy rule is obtained by performing an AND fuzzy logic operation on each fuzzy variable.
[0077] The deep convolutional computation part is responsible for extracting hidden features from the input rule weights and converting the hidden layer weights into a high-dimensional information representation to characterize the hidden relationships between different rules. The complex structure of the neural network can represent non-linear relationships, greatly improving the processing power of the entire model; specifically, the simplified structure of this part is as follows: Figure 3 As shown, the deep convolution part takes the initial excitation intensity generated by the fuzzy logic antecedent part as the initial input and outputs the final excitation intensity, which has the same dimension as the initial excitation intensity. The deep convolution calculation part mainly consists of convolutional layers and fully connected layers. The data will be expanded to high dimensions by one-dimensional convolutional kernels in the convolutional layers and flattened to the initial dimension in the fully connected layers.
[0078] Input data after Convolution processing. This represents the i-th data unit in the l-th convolutional layer. Let b represent the weight of the l-th convolutional kernel, where b is the weight of the a-th kernel. l This represents the bias of that layer. In the network, the activation function is defined as σ(x) = 1 / (1+e^x). -x ).like Figure 4 As shown, this embodiment uses one-dimensional convolution to process the input excitation intensity, thereby uncovering the relationship between the input weights. The network layers connected in series after the convolutional layers are fully connected, meaning that each node in the l-th layer is connected to all nodes in the (l-1)-th layer. The parameter w represents the connection weight, and b represents the node bias. The formula for feedforward propagation is... This layer helps reduce the number of dimensions that hide the weights to the same size as the input for this part, and prepares for the next part of deblurring.
[0079] The fuzzy result representation part is used to handle the defuzzification process in fuzzy inference, which helps to convert the intermediate quantities of fuzzy logic operations into explicit results in the final output of the model; for the final excitation intensity [W′1, W′2, ..., W′] of the output of the previous part. N This section uses a normalization algorithm to calculate normalized weights and represent the activation level of each rule. The normalization formula is as follows: As shown, This represents the normalized excitation intensity of rule j and signifies its contribution to the total weight. From the normalized excitation intensity, the importance of each rule in the rule base can be interpreted.
[0080] In the TS rules described above, the consequent of the fuzzy rule is a linear combination of the input data, expressed as: and This refers to the consequent parameter of the j-th rule; the total output Z is defined as the weighted sum of the outputs of all rules. Specifically, the weight of each rule refers to the weight obtained through... The normalized excitation intensity; thus, in the fuzzy result representation part, the rule result will be formed and the intermediate quantities of fuzzy logic will be converted into numerical outputs, enabling the model to provide prediction results; for subsequent parameters In this embodiment, the least squares estimator (LSE) is used to obtain the optimization result; let And θ = [θ1, θ2, ..., θ N ] T Let F = [f1, f2, ..., f N ], and assuming the estimation error is e, then the corrected formula for z is Z = Fθ + e; the mean square error function of the result is in In the definition, m represents the length of the dataset, and y represents the length of the dataset. t This represents the actual label of the t-th data point in the dataset. Let represent the t-th input data from the dataset; therefore, the optimization objective is to minimize the squared error; when Substituting Z = Fθ + e, we can use LSE to fit the data and obtain the consequent parameters in F; the fuzzy result represents part of the output prediction result. This embodiment uses mean squared error (MSE) to calculate the loss value and uses the backpropagation algorithm (GD) to update the parameters of IP-FDCNN. The loss function is as follows:
[0081] IP-DCFNN employs a hybrid learning approach. The training process of the IP-DCFNN model can be divided into two stages: model parameter initialization and iterative parameter update. During training, some parameters play a crucial role in the final performance of the model. These parameters are shown in Table 1.
[0082]
[0083] Table 1
[0084] These parameters are distributed across the three components of the IP-DCFNN; the consequent parameters are updated by the least squares estimation algorithm (LSE) during the feedforward propagation, while the other parameters are updated by the gradient descent (GD) algorithm during the backpropagation; that is, these parameters are learned and updated through different algorithms during the feedforward and backpropagation processes.
[0085] The parameter initialization algorithm is shown in Table 1. The antecedent parameters of the fuzzy logic antecedent part are expressed as follows: The membership functions {μ, σ} represent the fuzzy center and σ the fuzzy width. The antecedent parameters are initialized using a grid partition-based initialization algorithm. Specifically, grid partitioning is a method for dividing the data space, which divides the input data space into grid subspaces parallel to the axes based on the membership function of each feature. Figure 5 As shown, taking two dimensions as an example, in each dimension, the fuzzy function corresponding to the fuzzy set divides this dimension in a direction parallel to the axis. The division of the data into multiple dimensions divides the entire dataset into small grids. Each grid partition can represent a combination of fuzzy variables. The initialization algorithm based on grid partitioning initializes the parameters of each membership function according to the predefined partition grids divided by the grid. For data samples falling in the grid, the fuzzy center μ is initialized to the median of the data in this grid, and the fuzzy width σ is initialized to a linear approximation of the data size in the grid, with its coefficients coming from linear fitting.
[0086] Based on the foregoing, the initialization algorithm steps for the antecedent parameters are as follows:
[0087] 1) Divide the input space into different dimensions according to attribute features;
[0088] 2) For each data dimension, sort the data and divide the current single-dimensional feature data equally according to the number of fuzzy variables in the fuzzy set of the corresponding feature;
[0089] 3) For each segmented data unit, calculate the membership function parameters (u, θ) of the corresponding fuzzy variable:
[0090] Where u is the median of the feature data values in the current data unit, and θ is ρ times the difference between the maximum and minimum feature data values in the current data unit, where ρ comes from the linear fitting coefficient between the membership function calculation result and the feature data distribution range.
[0091] The parameters for the depthwise convolution computation are the weights and biases of the neurons in the convolutional and fully connected layers. The He-uniform algorithm is used to initialize these parameters, and the specific steps are as follows:
[0092] A. Samples are drawn to form a uniform distribution of [-limit, limit];
[0093] B. Limit in Calculation in the middle;
[0094] The number of input neurons in layer l is represented; the result parameters (posterior parameters) of the fuzzy result representation part are the parameters in the linear expression of the output of each rule, specifically:
[0095]
[0096] The results section of the rules is shown; unlike other parameters, the result parameters are generated during the forward propagation process and depend on the single iteration during the forward propagation process, but are independent of the historical iterations, that is, the consequent parameters in two adjacent iterations are not directly related; before training the model, all consequent parameters are given a simple zero initialization (i.e., all consequent parameters in the model are assigned a value of 0 at the beginning of training).
[0097] The IP-DCFNN model is trained using a hybrid learning approach. During the feedforward process, the consequent parameters are updated using the least squares estimation algorithm (LSE), and during backpropagation, the antecedent and hidden layer parameters are updated using the gradient descent algorithm (GD). During backpropagation, the gradient is calculated as follows: The calculation is shown, where C represents... The loss value calculated in the algorithm, #(n) represents the number of nodes in the l-th layer that are affected by the nodes in the l+1-th layer; the weights of the nodes in the network layer are then updated layer by layer in reverse by using the calculated gradient values.
[0098] The model training strategy is as follows:
[0099] 1) Initialize the model parameters using the initialization algorithm;
[0100] 2) The input data is transformed into membership values relative to the fuzzy variables using a membership function;
[0101] 3) By permuting and combining fuzzy variables, all rules in the rule base and their initial excitation intensities are generated;
[0102] 4) The initial excitation intensity is processed by the convolution kernel in the convolution calculation unit to obtain the final excitation intensity;
[0103] 5) Normalize the excitation intensity and calculate the fuzzy consequent parameters using the least squares estimation method;
[0104] 6) Calculate the final result degree value using the input linear computation function in the fuzzy consequent;
[0105] 7) Calculate the loss value using mean squared error, and calculate the gradient values of the fuzzy logic antecedent and the depthwise convolution calculation part;
[0106] 8) Update the fuzzy antecedent parameters and hidden layer parameters (node weights and node biases) using the gradient descent algorithm.
[0107] The deep convolutional fuzzy neural network and its training method for medical data described in this embodiment incorporate the concept of deep neural networks into the fuzzy inference system to achieve both interpretability and the ability to process high-dimensional data. Specifically, the fuzzification and defuzzification parts of the fuzzy inference system are represented using a network architecture, and a deep convolution calculation module is inserted between fuzzification and defuzzification. Through the high-dimensional data mapping and processing capabilities of the deep convolutional network in this module, the data processing capability of the entire model system is improved. Although the hidden weights of the deep convolution part cannot be well interpreted due to the topological limitations of the neural network, the fuzzification and defuzzification parts can extract relatively rich rule bases and membership information, thereby improving the system's data processing capability without sacrificing the system's interpretability.
[0108] Meanwhile, by proposing a fuzzy membership parameter initialization algorithm based on grid partitioning, the parameters of the fuzzy membership function in the model can be adaptively initialized according to the data distribution. This avoids the situation where parameter initialization for the fuzzy inference system part will greatly affect the training effect and final capability, thus greatly improving training efficiency.
[0109] This model combines fuzzy inference networks with deep convolutional computation units to perform targeted processing of patient medical data. The convolutional units in the model can further mine the correlation information and potential feature expressions in the patient's physiological information through high-dimensional mapping calculation of fuzzy rule weights. Therefore, this model is more conducive to inference analysis of tasks related to patient medical data.
[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A deep convolutional fuzzy neural network method for medical data, characterized in that: This includes the IP-DCFNN, a medical data interpretability prediction model based on deep convolutional fuzzy neural networks; The IP-DCFNN consists of three parts: Fuzzy logic preamble: The fuzzy logic preamble extracts input data, and the input data is transformed from numerical values into a set of membership degree values for fuzzy linguistic scalars through the operation of the membership function in the fuzzy logic preamble; The depthwise convolution calculation part extracts hidden features from the input rule weights and converts the hidden layer weights into high-dimensional information representations. Fuzzy result representation part: The fuzzy result representation part is used to handle the defuzzification process in fuzzy inference; The deep convolution part takes the initial excitation intensity generated by the fuzzy logic preamble as the initial input and outputs the final excitation intensity, which has the same dimension as the initial excitation intensity. The depthwise convolution calculation part mainly consists of convolutional layers and fully connected layers; The convolutional layer includes a one-dimensional convolutional layer, a ReLU activation layer, and a max pooling layer; The fully connected layer includes a linear layer and a ReLU activation layer; The initial input passed through The convolutional processing is extended to high dimension by a one-dimensional convolutional kernel and flattened back to the initial dimension in the fully connected layer, where represents the . The middle convolutional layer One data unit, Indicates the first The first layer of convolution kernel Weight, Indicates the bias of this layer; The network layer following the convolutional layer is fully connected, indicating that the first... Each node in the layer is connected to the first All nodes of the layer, parameters Represents connection weights. The bias of the node is represented by the formula for the feedforward propagation of the neural network. ; In the network layers connected in series after the convolutional layer, the activation function is defined as One-dimensional convolution is used to process the input excitation intensity; The blurred result representation part represents the final excitation intensity output in the depthwise convolution calculation part. The normalization algorithm is used to calculate the normalized weights and represent the activation level of each rule; The formula for the normalization algorithm is as follows: ,in, Representation rules The normalized excitation intensity represents the contribution of the rule to the total weight; The consequent of a fuzzy rule is a linear combination of the input data, expressed as: ,and That is, the first The consequent parameter of the rule; Total output Defined as the weighted sum of all rule outputs, the specific formula is as follows: The weight of each rule refers to the normalized excitation intensity of that rule. For subsequent parameters The calculation uses a least squares estimator to obtain the optimization result, let and ,make And assuming the estimation error is , then we can obtain The corrected formula is The mean square error function of the result is in The definition in the middle, where This indicates the length of the dataset, while Represents the first of the datasets The actual label of each data point Represents the first from the dataset One input data; The optimization objective is to minimize the squared error when... Substitution At that time, LSE can be used to fit the data and obtain The consequent parameter in the text.
2. The deep convolutional fuzzy neural network method for medical data according to claim 1, characterized in that: The membership function Used to map numerical input data to the range of 0 and 1; The input data for Within the interval, the membership function Input data Mapped to the fuzzy set [low, high], where And it represents the selection of fuzzy linguistic variables, with the parameters of the membership function being... and ; The fuzzy logic antecedent is formed using the Takagi-Sugeno fuzzy model. ; ； The fuzzy rules described in the text; The fuzzy variables generated from different dimensions of the input data are arranged and combined to construct the antecedent of the rule. By multiplying the membership degree corresponding to each fuzzy variable in the rule antecedent, the initial excitation intensity of the rule can be obtained. The initial excitation intensity represents the functional weight of the rule. No. The initial excitation intensity of a rule is calculated from the membership degrees of all fuzzy variables in the rule's antecedents, using the following formula: That is, the weight of the fuzzy rule is obtained by performing an "AND" fuzzy logic operation on each fuzzy variable.
3. The deep convolutional fuzzy neural network method for medical data according to claim 1, characterized in that: Fuzzy results represent partial output prediction results The mean squared error is used to calculate the loss value, and the parameters of IP-FDCNN are updated using the backpropagation algorithm. The loss function is... .
4. A training method for a deep convolutional fuzzy neural network for medical data, characterized in that: The training method for the deep convolutional fuzzy neural network for medical data includes two stages: initialization of model parameters and iterative update of parameters. The training method adopts a hybrid learning approach, in which specific parameters are learned and updated through different algorithms during feedforward and backpropagation. The antecedent parameter of the fuzzy logic antecedent part is represented as follows: In each membership function ,in Indicates a fuzzy center. Indicates the fuzzy width; The antecedent parameters are initialized using a mesh-based initialization algorithm; Grid partitioning is a method of dividing data space by dividing the input data space into grid subspaces parallel to the axes based on the membership function of each feature. The initialization algorithm based on grid partitioning initializes the parameters of each membership function according to the predefined partitions created by the grid. For data samples falling within the grid, the fuzzy center... Initialize to the median of the data in this grid, fuzz width It is initialized as a linear approximation of the data scale in the grid, and its coefficients are derived from linear fitting; The parameters for the depthwise convolution computation are the weights and biases of the neurons in the convolutional and fully connected layers. The He-uniform algorithm is used to initialize these parameters, and the specific steps are as follows: A. Sample Formation Uniform distribution; B exist In the calculation, Indicates the first The number of input neurons in the layer; The consequent parameters of the fuzzy result representation are the parameters in the linear expression of the output of each rule, specifically: ; The results section of the rules is shown below; Before training the model, apply simple zero initialization to all consequent parameters, that is, assign all consequent parameters in the model to 0 at the beginning of training; The IP-DCFNN model is trained using a hybrid learning approach; The model updates the consequent parameters using the least squares estimation algorithm during the feedforward process and updates the antecedent and hidden layer parameters using the gradient descent algorithm during the backpropagation process. During backpropagation, the gradient is as follows: The calculation is shown, where express The loss value calculated in the middle, Indicates the first The layer is affected by the first The number of nodes affected by a node in a layer; The weights of nodes in the network layer are then updated in reverse order using the calculated gradient values.
5. The deep convolution for medical data as described in claim 4 The training method for fuzzy neural networks is characterized by: The training method for deep convolutional fuzzy neural networks for medical data employs a hybrid learning approach as follows: The antecedent parameters are located in the antecedent part of the fuzzy logic. Their initialization algorithm is the initialization algorithm of grid partitioning, their update timing is backpropagation, and their update algorithm is the gradient descent algorithm. The consequent parameters are located in the fuzzy result representation part. Their initialization algorithm is the zero initialization algorithm, their update timing is feedforward propagation, and their update algorithm is the least squares estimation algorithm. The node weights are located in the depthwise convolution calculation part. Their initialization algorithm is the He-normalization algorithm, their update timing is backpropagation, and their update algorithm is gradient descent. The node bias is located in the depthwise convolution computation part. Its initialization algorithm is the He-normalization algorithm, its update timing is backpropagation, and its update algorithm is the gradient descent algorithm.