Neural network intelligent optimization method and system based on data node analysis

Abstract

The invention discloses a neural network intelligent optimization method and system based on data node analysis, and relates to the technical field of deep learning, and the method comprises the steps: obtaining basic data information, building a demand based on a neural network, dividing basic data, obtaining a training data set and a test data set, and obtaining a test data set according to the training data set; and obtaining an input category number and an output category number. According to the method, data categories are subjected to clustering analysis, the stability and reliability of data in the neural network construction process are ensured, errors caused by manual label adding are avoided, the number of nodes in each layer in the neural network is accurately controlled through the hidden node coefficient, the accuracy of the neural network is improved, and the method is suitable for large-scale popularization and application. The neural network hidden node condition is accurately evaluated through weight pruning in the training process, the neural network optimization is realized through pruning the hidden nodes, the neural network optimization efficiency is improved through the mean value coefficient, and the stability and reliability of the neural network are ensured.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a neural network intelligent optimization method and system based on data node analysis. Background Technology

[0002] Neural networks are the fundamental building blocks of artificial intelligence. Their architecture has evolved from the earliest perceptrons to today's recurrent neural networks, convolutional neural networks, generative adversarial neural networks, and so on, with each new architecture playing a crucial role in various fields. With the rise of deep learning, the architecture of neural networks has become increasingly complex, and the number of parameters they contain has increased exponentially. For example, the well-known DeepSeek has reached 32 billion internal parameters. The complex structure and massive number of parameters of these networks also place increasingly higher demands on computing power. Running such a large-scale model on some small embedded devices is clearly impractical. Many experts and scholars are also researching methods to construct small neural networks that meet accuracy requirements with limited parameters.

[0003] Currently, there are still problems with optimizing neural networks, such as the inability to accurately evaluate them, the tendency to increase the size of hidden nodes to improve prediction accuracy, the increasing complexity of the network internally, which affects the convenience and reliability of the neural network, the inability to deploy large-scale neural networks in the field, and the problem that adopting an edge-cloud collaborative architecture will face the pressure of transmitting massive amounts of data, resulting in low efficiency of the neural network and the inability to accurately adjust the neural network. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method and system for intelligent optimization of neural networks based on data node analysis. This technical solution solves the problems mentioned in the background technology, such as the inability to accurately evaluate neural networks, the tendency to improve prediction accuracy by continuously expanding the scale of hidden nodes in the neural network, the increasing complexity of the network, the inability to deploy large neural networks on-site, the inability to accurately adjust the neural network by adopting an edge-cloud collaborative architecture, and the resulting low efficiency and inability to accurately adjust the neural network.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A neural network intelligent optimization method based on data node analysis includes:

[0007] Acquire basic data information, which represents the data used for neural network training;

[0008] Based on the requirements for building neural networks, the basic data is divided to obtain a training dataset and a test dataset. The training dataset includes input parameters and output parameters.

[0009] Based on the training dataset, the DBSCAN algorithm is used to cluster the input and output parameters to obtain the number of input and output categories.

[0010] The sum of the number of input categories and the number of output categories is used as the total number of hidden layers in the neural network;

[0011] Based on the training dataset, obtain the total number of corresponding data features, where the total number of data features represents the total number of data features corresponding to all data in the training dataset;

[0012] The neural network is obtained based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features.

[0013] Preferably, obtaining the neural network based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features specifically includes:

[0014] Based on the total number of data features and the total number of hidden layers in the neural network, the product of the total number of data features and the total number of hidden layers in the neural network is taken as the number of hidden nodes in the first hidden layer of the neural network.

[0015] The hidden node coefficients are obtained based on the number of input categories, the number of output categories, and the total number of data features.

[0016] Based on the number of hidden nodes in the first hidden layer of the neural network, the number of hidden nodes in each hidden layer of the neural network is obtained according to the hidden node coefficients.

[0017] The neural network is constructed based on the number of hidden nodes corresponding to each hidden layer to obtain the initial neural network;

[0018] Based on the training dataset, the input and output categories are filtered to obtain the first feature data category information;

[0019] Based on the Adam algorithm, the neural network is trained using the first feature data category information and the training dataset to obtain the neural network;

[0020] Specifically, the number of hidden nodes in the neural network is:

[0021]

[0022] In the formula, n i Let ni represent the number of hidden nodes in the i-th hidden layer of the neural network, and n1 represent the number of hidden nodes in the first hidden layer of the neural network. I k is the number of input categories. o λ represents the number of output categories, m represents the total number of data features, and λ represents the total number of categories. n Represents the hidden node coefficients, ni-1 The number of hidden nodes in the (i-1)th layer of the neural network, where N is the total number of hidden layers in the neural network, and if... Then let

[0023] Preferably, obtaining the hidden node coefficients based on the number of input categories, the number of output categories, and the total number of data features specifically includes:

[0024] The absolute value of the difference between the number of input categories and the number of output categories is used as the data feature coefficient;

[0025] The absolute value of the difference between the total number of data features and the total number of hidden layers in the neural network is used as the cluster difference feature number;

[0026] The ratio of the number of cluster difference features to the total number of data features is used as the cluster difference coefficient;

[0027] The product of the clustering difference coefficient and the data feature coefficient is used as the number of hidden layers in the neural network.

[0028] Based on the total number of data features, obtain the minimum number of nodes in the standard hidden layer, where the minimum number of nodes in the standard hidden layer is half of the total number of data features;

[0029] The ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer is used as the neural network node difference coefficient.

[0030] The difference between the total number of hidden layers in a neural network and the number of hidden layers in the feature layer of the neural network is used as the baseline number of hidden layers in the neural network.

[0031] The hidden node coefficients are obtained based on the neural network node difference coefficients and the baseline hidden layer number of the neural network.

[0032] Specifically, the hidden node coefficients are:

[0033]

[0034] In the formula, λ n n represents the hidden node coefficients. min This represents the minimum number of nodes in the standard hidden layer. denoted as the number of hidden layers in the neural network.

[0035] Preferably, the step of filtering the input and output categories based on the training dataset to obtain the first feature data category information specifically includes:

[0036] Obtain input category information and output category information;

[0037] Based on the data topology, obtain the feature input category information corresponding to each output category;

[0038] Based on the output category information and the feature input category information, the sum of the number of feature input categories corresponding to all output categories is taken as the number of associated input categories;

[0039] The ratio of the number of output categories to the number of associated input categories is used as the data mapping coefficient;

[0040] Based on the data mapping coefficients, the sum of the number of input categories and the number of output categories is used as the maximum number of categories to obtain the number of calibrated input categories;

[0041] Based on the number of calibration input categories, obtain the number of calibration output categories;

[0042] Based on the number of calibrated input categories and the number of calibrated output categories, arbitrarily select a category from the input category information and the output category information as the first feature data category information;

[0043] Let w be the number of calibrated input categories, then:

[0044] w(1+μ)≤k I +k o ;

[0045] In the formula, μ is the data mapping coefficient, w is a constant, and the maximum value of w is taken as the number of input categories for calibration. Then the number of output categories for calibration is (w·μ).

[0046] Preferably, the step of training the neural network based on the first feature data category information and the training dataset to obtain the neural network specifically includes:

[0047] S1: Based on the first feature data category information, using any first feature data category as a basis, train the neural network on the training dataset to obtain the benchmark neural network;

[0048] S2: Based on the benchmark neural network, obtain the hidden layer information of the neural network, which includes the hidden node information of the neural network corresponding to each hidden layer of the neural network;

[0049] S3: Based on the hidden layer of the neural network, classify the hidden nodes of the neural network and the weights corresponding to each hidden node to obtain the set information of the hidden node of the neural network. The set of hidden nodes of the neural network represents the hidden nodes of the neural network in the same hidden layer of the neural network and the weights corresponding to each hidden node.

[0050] S4: Based on the information of the neural network hidden node set, cluster the weights of the neural network hidden nodes in each neural network hidden node set using the DBSCAN algorithm to obtain the node weight category information;

[0051] S5: Adjust the weight of each hidden node in the neural network based on the information of the set of hidden nodes, the weights of the hidden nodes, and the category of the node weights;

[0052] S6: Based on the first feature data category information, repeat steps S1 to S5 until all first feature data categories have been trained to obtain the neural network.

[0053] Preferably, adjusting the weights of each hidden node in the neural network based on the hidden node set information, the hidden node weights, and the node weight category information specifically includes:

[0054] The sum of the data mapping coefficient and 1 is used as the data category compression coefficient;

[0055] Based on the information of the neural network hidden node set, obtain the minimum and maximum neural network hidden node weights in each neural network hidden node set;

[0056] The product of the minimum hidden node weight corresponding to each set of hidden nodes in the neural network and the data class compression coefficient is taken as the minimum weight of the set of hidden nodes in the neural network, and the ratio of the maximum hidden node weight to the data class compression coefficient is taken as the maximum weight of the set of hidden nodes in the neural network.

[0057] Based on the information of the hidden node set of the neural network, the average weight of the hidden nodes in each neural network hidden node set is used as the feature node weight of each neural network hidden node set.

[0058] The ratio of the feature node weights to the weights of the largest hidden nodes in the neural network is used as the weight offset coefficient.

[0059] The ratio of the weight offset coefficient to the data category compression coefficient is used as the mean coefficient;

[0060] The average of the feature node weights of all hidden nodes in the neural network set is used as the baseline node weight.

[0061] The weights of each hidden node in the neural network are adjusted based on the mean coefficient, the baseline node weight, and the calibrated minimum weight.

[0062] Specifically, adjusting the weights of hidden nodes in the neural network involves:

[0063]

[0064] In the formula, n ijk θ represents the weights of the j-th hidden node in the i-th hidden layer of the neural network and the k-th hidden node in the (i-1)-th hidden layer of the neural network. i The number of node weight categories, ∈imin The λ represents the minimum calibrated weight corresponding to the hidden layer of the i-th neural network, μ is the weight of the baseline node, and λ is the minimum calibrated weight. μ Let n be the mean coefficient. ijk >∈ imax Then the weights of the hidden nodes in the neural network are fixed, and the fixed weights will not be adjusted during subsequent training. imax This represents the maximum calibrated weight corresponding to the hidden layer of the i-th neural network.

[0065] Furthermore, a neural network intelligent optimization system based on data node analysis is proposed to implement the optimization method described above, including:

[0066] The main control module is used to obtain the calibrated input category number based on the data mapping coefficients, taking the sum of the input category number and the output category number as the maximum category number, and obtaining the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, it arbitrarily selects a category from the input category information and the output category information as the first feature data category information. Based on the neural network hidden node set information, the neural network hidden node weights and node weight category information, it adjusts the weights of each neural network hidden node. Based on the first feature data category information and the training dataset, it trains the neural network to obtain the neural network.

[0067] The information acquisition module is used to acquire basic data information. Based on the requirements of neural network construction, the basic data is divided to acquire training datasets and test datasets. Based on the training dataset, the input and output parameters are clustered using the DBSCAN algorithm to acquire the number of input categories and the number of output categories. Based on the training dataset, the total number of corresponding data features is acquired. Based on the data topology, the feature input category information corresponding to each output category is acquired. Using the hidden layers of the neural network as the classification basis, the hidden nodes of the neural network and the weights corresponding to each hidden node are classified to acquire the set information of the hidden nodes of the neural network.

[0068] The evaluation module is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number, obtain the hidden node coefficient based on the neural network node difference coefficient and the neural network baseline hidden layer number, and take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient.

[0069] The display module interacts with the main control module and is used to output and display the training dataset, test dataset, hidden node coefficients, first feature data category information, and neural network.

[0070] Optionally, the main control module specifically includes:

[0071] The control unit is used to adjust the weight of each hidden node in the neural network according to the hidden node set information, the hidden node weights, and the node weight category information, and to train the neural network according to the first feature data category information and the training dataset to obtain the neural network.

[0072] An information receiving unit, which interacts with the information acquisition module and the evaluation module, is used to receive data and transmit it to the data recognition unit;

[0073] The data recognition unit is used to obtain the calibrated input category number by taking the sum of the input category number and the output category number as the maximum category number based on the data mapping coefficient, and to obtain the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, the unit arbitrarily selects a category from the input category information and the output category information as the first feature data category information.

[0074] Optionally, the information acquisition module specifically includes:

[0075] The first acquisition unit is used to acquire basic data information, divide the basic data based on the requirements of neural network construction, acquire training dataset and test dataset, and perform data clustering on input parameters and output parameters based on the DBSCAN algorithm according to the training dataset to acquire the number of input categories and the number of output categories.

[0076] The second acquisition unit is used to acquire the total number of corresponding data features based on the training dataset, acquire the feature input category information corresponding to each output category based on the data topology relationship, classify the hidden nodes of the neural network and the weights corresponding to each hidden node of the neural network based on the hidden layer of the neural network, and acquire the set information of the hidden nodes of the neural network.

[0077] Optionally, the evaluation module specifically includes:

[0078] The first evaluation unit is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, and the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number. Based on the neural network node difference coefficient and the neural network baseline hidden layer number, the hidden node coefficient is obtained.

[0079] The second evaluation unit is used to take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and to take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient.

[0080] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0081] This invention proposes a neural network intelligent optimization method and system based on data node analysis. By performing cluster analysis on data categories, the stability and reliability of data during the neural network construction process are ensured, avoiding errors introduced by manually adding labels. The accuracy of the neural network is improved by accurately controlling the number of nodes in each layer through hidden node coefficients. The condition of hidden nodes in the neural network is accurately evaluated through weight pruning during the training process. Neural network optimization is achieved by pruning hidden nodes. The mean coefficient is used to improve the optimization efficiency of the neural network, ensuring the stability and reliability of the neural network. Attached Figure Description

[0082] Figure 1 This is a flowchart of a neural network intelligent optimization method based on data node analysis proposed in this invention;

[0083] Figure 2 This is a flowchart of the process for obtaining X in this invention;

[0084] Figure 3 This is a flowchart of the process for obtaining X in this invention;

[0085] Figure 4 This is a flowchart of the process for obtaining X in this invention;

[0086] Figure 5 This is a block diagram of a neural network intelligent optimization system based on data node analysis proposed in this invention. Detailed Implementation

[0087] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0088] Reference Figure 1 - Figure 4 As shown in the figure, an intelligent optimization method for neural networks based on data node analysis in an embodiment of the present invention includes:

[0089] Acquire basic data information, which represents the data used for neural network training;

[0090] Based on the requirements for building neural networks, the basic data is divided to obtain a training dataset and a test dataset. The training dataset includes input parameters and output parameters.

[0091] Based on the training dataset, the DBSCAN algorithm is used to cluster the input and output parameters to obtain the number of input and output categories.

[0092] The sum of the number of input categories and the number of output categories is used as the total number of hidden layers in the neural network;

[0093] Based on the training dataset, obtain the total number of corresponding data features, where the total number of data features represents the total number of data features corresponding to all data in the training dataset;

[0094] The neural network is obtained based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features.

[0095] Specifically, the neural network is obtained based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features. This includes:

[0096] Based on the total number of data features and the total number of hidden layers in the neural network, the product of the total number of data features and the total number of hidden layers in the neural network is taken as the number of hidden nodes in the first hidden layer of the neural network.

[0097] The hidden node coefficients are obtained based on the number of input categories, the number of output categories, and the total number of data features.

[0098] Based on the number of hidden nodes in the first hidden layer of the neural network, the number of hidden nodes in each hidden layer of the neural network is obtained according to the hidden node coefficients.

[0099] The neural network is constructed based on the number of hidden nodes corresponding to each hidden layer to obtain the initial neural network;

[0100] Based on the training dataset, the input and output categories are filtered to obtain the first feature data category information;

[0101] Based on the Adam algorithm, the neural network is trained using the first feature data category information and the training dataset to obtain the neural network;

[0102] Specifically, the number of hidden nodes in the neural network is:

[0103]

[0104] In the formula, n i Let ni represent the number of hidden nodes in the i-th hidden layer of the neural network, and n1 represent the number of hidden nodes in the first hidden layer of the neural network. I k is the number of input categories.o λ represents the number of output categories, m represents the total number of data features, and λ represents the total number of categories. n Represents the hidden node coefficients, n i-1 The number of hidden nodes in the (i-1)th layer of the neural network, where N is the total number of hidden layers in the neural network, and if... Then let

[0105] In this scheme, the sum of the number of input and output categories is used as the total number of hidden layers in the neural network, establishing a correspondence between data categories and the number of hidden layers. This provides a foundation for subsequent neural network training. The number of hidden nodes in the first layer is determined by the total number of data features and the total number of hidden layers. Then, the hidden node coefficients are obtained by combining the number of input and output categories, and the number of hidden nodes is derived layer by layer. The resulting initial neural network closely matches the data features and network architecture requirements, laying an adaptive foundation for subsequent training and avoiding network inefficiency or overfitting caused by arbitrarily setting the number of nodes. Based on the training dataset, the input and output categories are selected to obtain the first feature data category information, allowing training to focus on key features, reducing redundant data interference, improving training targeting and efficiency, and making neural network learning more accurate.

[0106] It is understandable that hidden node coefficients are used to balance the goodness of fit between data categories and data nodes. However, in existing technologies, the coefficients are often determined based on conventional experience or by directly traversing the entire dataset to determine the number of data nodes. The former often fails to adjust the neural network according to the data conditions and cannot respond to differences in data dimensions and task types, resulting in a mismatch between the network architecture and the data, and failing to ensure the accuracy of the neural network. The latter often involves a large amount of data processing and consumes a lot of resources. Therefore, in this solution, the hidden node coefficients are associated with the number of input categories, the number of output categories, and the total number of features, so that the coefficients are strongly bound to the data complexity and task objectives, thereby achieving a precise, interpretable, and self-adaptive design of the network architecture.

[0107] Specifically, obtaining the hidden node coefficients based on the number of input categories, the number of output categories, and the total number of data features includes:

[0108] The absolute value of the difference between the number of input categories and the number of output categories is used as the data feature coefficient;

[0109] The absolute value of the difference between the total number of data features and the total number of hidden layers in the neural network is used as the cluster difference feature number;

[0110] The ratio of the number of cluster difference features to the total number of data features is used as the cluster difference coefficient;

[0111] The product of the clustering difference coefficient and the data feature coefficient is used as the number of hidden layers in the neural network.

[0112] Based on the total number of data features, obtain the minimum number of nodes in the standard hidden layer, where the minimum number of nodes in the standard hidden layer is half of the total number of data features;

[0113] The ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer is used as the neural network node difference coefficient.

[0114] The difference between the total number of hidden layers in a neural network and the number of hidden layers in the feature layer of the neural network is used as the baseline number of hidden layers in the neural network.

[0115] The hidden node coefficients are obtained based on the neural network node difference coefficients and the baseline hidden layer number of the neural network.

[0116] Specifically, the hidden node coefficients are:

[0117]

[0118] In the formula, λ n n represents the hidden node coefficients. min This represents the minimum number of nodes in the standard hidden layer. denoted as the number of hidden layers in the neural network.

[0119] This scheme constructs the hidden node coefficient calculation logic through multi-dimensional parameters such as the difference between the number of input and output categories (data feature coefficients) and the difference between the total number of data features and the total number of hidden layers (clustering difference feature count). By integrating data category, feature scale, and network architecture information, the coefficients are adapted to the complex characteristics of the data, making the number of hidden layer nodes in the neural network more aligned with task requirements and improving the network's ability to express data. A standard minimum number of hidden layer nodes (half the total number of features) is introduced to constrain the lower limit of the number of nodes and avoid underfitting due to too few nodes. The node difference coefficient (the ratio of the number of nodes in the first layer to the standard minimum number of nodes) is calculated and dynamically derived in conjunction with the baseline number of hidden layers. This allows the number of hidden layer nodes to decrease / adjust reasonably with network depth and data features, optimizing the hierarchical synergy of the network architecture. By obtaining the hidden node coefficients through the neural network node difference coefficients and the baseline number of hidden layers, the oscillations and slow convergence problems caused by unreasonable node settings during training are reduced, facilitating more efficient neural network training and shortening the model optimization cycle.

[0120] Specifically, the step of filtering the input and output categories based on the training dataset to obtain the first feature data category information includes:

[0121] Obtain input category information and output category information;

[0122] Based on the data topology, obtain the feature input category information corresponding to each output category;

[0123] Based on the output category information and the feature input category information, the sum of the number of feature input categories corresponding to all output categories is taken as the number of associated input categories;

[0124] The ratio of the number of output categories to the number of associated input categories is used as the data mapping coefficient;

[0125] Based on the data mapping coefficients, the sum of the number of input categories and the number of output categories is used as the maximum number of categories to obtain the number of calibrated input categories;

[0126] Based on the number of calibration input categories, obtain the number of calibration output categories;

[0127] Based on the number of calibrated input categories and the number of calibrated output categories, arbitrarily select a category from the input category information and the output category information as the first feature data category information;

[0128] Let w be the number of calibrated input categories, then:

[0129] w(1+μ)≤k I +k o ;

[0130] In the formula, μ is the data mapping coefficient, w is a constant, and the maximum value of w is taken as the number of input categories for calibration. Then the number of output categories for calibration is (w·μ).

[0131] In this scheme, based on data topology, the feature input categories corresponding to the output categories are identified, and the number of associated input categories and data mapping coefficients are calculated to uncover potential relationships between data. The first feature data category information selected through this logic removes redundant and irrelevant categories, retaining core features strongly correlated with the output. This makes the neural network training data more focused on key information and reduces noise interference. Based on the data mapping coefficients and combined with the number of input and output categories, the number of input / output categories is derived and calibrated, ensuring the number of input and output categories matches the network architecture. This dynamic matching makes the collaboration between the neural network's input / output layers and hidden layers more reasonable, avoiding network training bias caused by category imbalance, improving the model's adaptability to the task, and reducing the training data size and learning burden on the neural network by selecting the first feature data category information. Simultaneously, focusing on closely related categories allows the model to capture data patterns more quickly, accelerating training convergence, shortening the training cycle and saving computational resources while ensuring model performance.

[0132] Specifically, the neural network is trained based on the first feature data category information and the training dataset to obtain the neural network, which includes:

[0133] S1: Based on the first feature data category information, using any first feature data category as a basis, train the neural network on the training dataset to obtain the benchmark neural network;

[0134] S2: Based on the benchmark neural network, obtain the hidden layer information of the neural network, which includes the hidden node information of the neural network corresponding to each hidden layer of the neural network;

[0135] S3: Based on the hidden layer of the neural network, classify the hidden nodes of the neural network and the weights corresponding to each hidden node to obtain the set information of the hidden node of the neural network. The set of hidden nodes of the neural network represents the hidden nodes of the neural network in the same hidden layer of the neural network and the weights corresponding to each hidden node.

[0136] S4: Based on the information of the neural network hidden node set, cluster the weights of the neural network hidden nodes in each neural network hidden node set using the DBSCAN algorithm to obtain the node weight category information;

[0137] S5: Adjust the weight of each hidden node in the neural network based on the information of the set of hidden nodes, the weights of the hidden nodes, and the category of the node weights;

[0138] S6: Based on the first feature data category information, repeat steps S1 to S5 until all first feature data categories have been trained to obtain the neural network.

[0139] In this approach, a baseline network is first trained using a single feature category. Then, the hidden layer nodes and weights are decomposed layer by layer, and the node weight distribution is analyzed using DBSCAN clustering. This hierarchical and clustering method accurately identifies the weight patterns of hidden layer nodes, providing a basis for targeted weight adjustments. This solves the problem of "coarse parameter tuning and difficulty in adapting to complex networks" in traditional methods, improving the accuracy of parameter optimization. Using the hidden layer as the classification foundation, node set information is constructed, focusing on the correlation between nodes and weights within the same layer. Adjusting weights after clustering strengthens effective nodes and suppresses redundant nodes, making the hidden layer structure more compatible with data features, uncovering the value of nodes in feature learning, optimizing the network's ability to express data, reducing computational waste, and iteratively training by feature category, covering all first-feature data categories. This approach allows the network to fully learn the characteristics of different categories of data, avoiding insufficient generalization caused by single-category training. After multiple rounds of optimization, the model's prediction stability and robustness in multi-sample scenarios are significantly improved, contributing to the construction of a better neural network.

[0140] It should be noted that in this scheme, for the trained neural network, if n ijk If the value is 0, then delete the hidden node.

[0141] For the predicted data X P Predicted output:

[0142]

[0143] Where λP Here are the weight coefficients; f(*) represents the prediction result of the trained neural network, L2(C g X P ) represents C g and X P The Euclidean distance, C g Let g be the center of the class of the g-th input parameter, where g∈[1, k]. I ];

[0144] In this scheme, the product of the clustering difference coefficient and the data mapping coefficient is used as the model prediction bias coefficient, and the absolute value of the difference between 1 and the model prediction bias coefficient is used as the weight coefficient λ of the neural network's prediction result. P .

[0145] Specifically, based on the information of the neural network hidden node set, the neural network hidden node weights, and the node weight category information, the weights of each neural network hidden node are adjusted, including:

[0146] The sum of the data mapping coefficient and 1 is used as the data category compression coefficient;

[0147] Based on the information of the neural network hidden node set, obtain the minimum and maximum neural network hidden node weights in each neural network hidden node set;

[0148] The product of the minimum hidden node weight corresponding to each set of hidden nodes in the neural network and the data class compression coefficient is taken as the minimum weight of the set of hidden nodes in the neural network, and the ratio of the maximum hidden node weight to the data class compression coefficient is taken as the maximum weight of the set of hidden nodes in the neural network.

[0149] Based on the information of the hidden node set of the neural network, the average weight of the hidden nodes in each neural network hidden node set is used as the feature node weight of each neural network hidden node set.

[0150] The ratio of the feature node weights to the weights of the largest hidden nodes in the neural network is used as the weight offset coefficient.

[0151] The ratio of the weight offset coefficient to the data category compression coefficient is used as the mean coefficient;

[0152] The average of the feature node weights of all hidden nodes in the neural network set is used as the baseline node weight.

[0153] The weights of each hidden node in the neural network are adjusted based on the mean coefficient, the baseline node weight, and the calibrated minimum weight.

[0154] Specifically, adjusting the weights of hidden nodes in the neural network involves:

[0155]

[0156] In the formula, n ijk θ represents the weights of the j-th hidden node in the i-th hidden layer of the neural network and the k-th hidden node in the (i-1)-th hidden layer of the neural network. i The number of node weight categories, ∈ imin The λ represents the minimum calibrated weight corresponding to the hidden layer of the i-th neural network, μ is the weight of the baseline node, and λ is the minimum calibrated weight. μ Let n be the mean coefficient. ijk >∈ imax Then the weights of the hidden nodes in the neural network are fixed, and the fixed weights will not be adjusted during subsequent training. imax This represents the maximum calibrated weight corresponding to the hidden layer of the i-th neural network.

[0157] This scheme constructs data category compression coefficients by introducing data mapping coefficients, and derives and determines the minimum / maximum weights, feature node weights, and weight offset coefficients by combining the minimum / maximum weights and mean of the hidden node set. The multi-coefficient linkage provides clear data support for weight adjustments, eliminating the blindness of experience-based parameter tuning and enabling precise and differentiated control of hidden node weights. By setting weight thresholds, low-contribution and redundant hidden nodes are suppressed, simplifying the network structure. Reducing the participation of invalid parameters in calculations lowers the risk of overfitting and improves network computational efficiency, allowing the model to focus on feature learning from high-value nodes. Nodes exceeding their calibrated maximum weights are weighted and fixed, preventing over-adjustment during training from damaging effective feature representation. Layered processing adapts to the characteristics of network layers, ensuring stable training, helping the model converge to a better solution, and improving the generalization and prediction performance of the neural network.

[0158] Reference Figure 5 As shown, further, combining the above-mentioned neural network intelligent optimization method based on data node analysis, a neural network intelligent optimization system based on data node analysis is proposed, including:

[0159] The main control module is used to obtain the calibrated input category number based on the data mapping coefficients, taking the sum of the input category number and the output category number as the maximum category number, and obtaining the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, it arbitrarily selects a category from the input category information and the output category information as the first feature data category information. Based on the neural network hidden node set information, the neural network hidden node weights and node weight category information, it adjusts the weights of each neural network hidden node. Based on the first feature data category information and the training dataset, it trains the neural network to obtain the neural network.

[0160] The information acquisition module is used to acquire basic data information. Based on the requirements of neural network construction, the basic data is divided to acquire training datasets and test datasets. Based on the training dataset, the input and output parameters are clustered using the DBSCAN algorithm to acquire the number of input categories and the number of output categories. Based on the training dataset, the total number of corresponding data features is acquired. Based on the data topology, the feature input category information corresponding to each output category is acquired. Using the hidden layers of the neural network as the classification basis, the hidden nodes of the neural network and the weights corresponding to each hidden node are classified to acquire the set information of the hidden nodes of the neural network.

[0161] The evaluation module is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number, obtain the hidden node coefficient based on the neural network node difference coefficient and the neural network baseline hidden layer number, and take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient.

[0162] The display module interacts with the main control module and is used to output and display the training dataset, test dataset, hidden node coefficients, first feature data category information, and neural network.

[0163] The main control module specifically includes:

[0164] The control unit is used to adjust the weight of each hidden node in the neural network according to the hidden node set information, the hidden node weights, and the node weight category information, and to train the neural network according to the first feature data category information and the training dataset to obtain the neural network.

[0165] An information receiving unit, which interacts with the information acquisition module and the evaluation module, is used to receive data and transmit it to the data recognition unit;

[0166] The data recognition unit is used to obtain the calibrated input category number by taking the sum of the input category number and the output category number as the maximum category number based on the data mapping coefficient, and to obtain the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, the unit arbitrarily selects a category from the input category information and the output category information as the first feature data category information.

[0167] The information acquisition module specifically includes:

[0168] The first acquisition unit is used to acquire basic data information, divide the basic data based on the requirements of neural network construction, acquire training dataset and test dataset, and perform data clustering on input parameters and output parameters based on the DBSCAN algorithm according to the training dataset to acquire the number of input categories and the number of output categories.

[0169] The second acquisition unit is used to acquire the total number of corresponding data features based on the training dataset, acquire the feature input category information corresponding to each output category based on the data topology relationship, classify the hidden nodes of the neural network and the weights corresponding to each hidden node of the neural network based on the hidden layer of the neural network, and acquire the set information of the hidden nodes of the neural network.

[0170] The evaluation module specifically includes:

[0171] The first evaluation unit is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, and the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number. Based on the neural network node difference coefficient and the neural network baseline hidden layer number, the hidden node coefficient is obtained.

[0172] The second evaluation unit is used to take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and to take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient.

[0173] In summary, the advantages of this invention are as follows: by performing cluster analysis on data categories, the stability and reliability of data during the construction of the neural network are ensured, avoiding errors introduced by manually adding labels; hidden node coefficients are obtained through the neural network node difference coefficient and the baseline number of hidden layers in the neural network; the number of nodes in each layer of the neural network is accurately controlled through the hidden node coefficients, thereby improving the accuracy of the neural network; the condition of the hidden nodes in the neural network is accurately evaluated through weight pruning during the training process; neural network optimization is achieved through pruning hidden nodes; and the mean coefficient is used to improve the optimization efficiency of the neural network, thus ensuring the stability and reliability of the neural network.

[0174] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A neural network intelligent optimization method based on data node analysis, characterized in that, include: Acquire basic data information, which represents the data used for neural network training; Based on the requirements for building neural networks, the basic data is divided to obtain a training dataset and a test dataset. The training dataset includes input parameters and output parameters. Based on the training dataset, the DBSCAN algorithm is used to cluster the input and output parameters to obtain the number of input and output categories. The sum of the number of input categories and the number of output categories is used as the total number of hidden layers in the neural network; Based on the training dataset, obtain the total number of corresponding data features, where the total number of data features represents the total number of data features corresponding to all data in the training dataset; The neural network is obtained based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features.

2. The neural network intelligent optimization method based on data node analysis according to claim 1, characterized in that, The process of obtaining the neural network based on the training dataset, the total number of hidden layers in the neural network, and the total number of data features specifically includes: Based on the total number of data features and the total number of hidden layers in the neural network, the product of the total number of data features and the total number of hidden layers in the neural network is taken as the number of hidden nodes in the first hidden layer of the neural network. The hidden node coefficients are obtained based on the number of input categories, the number of output categories, and the total number of data features. Based on the number of hidden nodes in the first hidden layer of the neural network, the number of hidden nodes in each hidden layer of the neural network is obtained according to the hidden node coefficients. The neural network is constructed based on the number of hidden nodes corresponding to each hidden layer to obtain the initial neural network; Based on the training dataset, the input and output categories are filtered to obtain the first feature data category information; Based on the Adam algorithm, the neural network is trained using the first feature data category information and the training dataset to obtain the neural network; Specifically, the number of hidden nodes in the neural network is: In the formula, n i Let ni represent the number of hidden nodes in the i-th hidden layer of the neural network, and n1 represent the number of hidden nodes in the first hidden layer of the neural network. I k is the number of input categories. o λ represents the number of output categories, m represents the total number of data features, and λ represents the total number of categories. n Represents the hidden node coefficients, n i-1 The number of hidden nodes in the (i-1)th layer of the neural network, where N is the total number of hidden layers in the neural network, and if... Then let 3. The neural network intelligent optimization method based on data node analysis according to claim 2, characterized in that, The process of obtaining hidden node coefficients based on the number of input categories, the number of output categories, and the total number of data features specifically includes: The absolute value of the difference between the number of input categories and the number of output categories is used as the data feature coefficient; The absolute value of the difference between the total number of data features and the total number of hidden layers in the neural network is used as the cluster difference feature number; The ratio of the number of cluster difference features to the total number of data features is used as the cluster difference coefficient; The product of the clustering difference coefficient and the data feature coefficient is used as the number of hidden layers in the neural network. Based on the total number of data features, obtain the minimum number of nodes in the standard hidden layer, where the minimum number of nodes in the standard hidden layer is half of the total number of data features; The ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer is used as the neural network node difference coefficient. The difference between the total number of hidden layers in a neural network and the number of hidden layers in the feature layer of the neural network is used as the baseline number of hidden layers in the neural network. The hidden node coefficients are obtained based on the neural network node difference coefficients and the baseline hidden layer number of the neural network. Specifically, the hidden node coefficients are: In the formula, λ n n represents the hidden node coefficients. min This represents the minimum number of nodes in the standard hidden layer. denoted as the number of hidden layers in the neural network.

4. The neural network intelligent optimization method based on data node analysis according to claim 3, characterized in that, The process of filtering input and output categories based on the training dataset to obtain first feature data category information specifically includes: Obtain input category information and output category information; Based on the data topology, obtain the feature input category information corresponding to each output category; Based on the output category information and the feature input category information, the sum of the number of feature input categories corresponding to all output categories is taken as the number of associated input categories; The ratio of the number of output categories to the number of associated input categories is used as the data mapping coefficient; Based on the data mapping coefficients, the sum of the number of input categories and the number of output categories is used as the maximum number of categories to obtain the number of calibrated input categories; Based on the number of calibration input categories, obtain the number of calibration output categories; Based on the number of calibrated input categories and the number of calibrated output categories, arbitrarily select a category from the input category information and the output category information as the first feature data category information; Let w be the number of calibrated input categories, then: w(1+μ)≤k I +k o ; In the formula, μ is the data mapping coefficient, w is a constant, and the maximum value of w is taken as the number of input categories for calibration. Then the number of output categories for calibration is (w·μ).

5. The neural network intelligent optimization method based on data node analysis according to claim 4, characterized in that, The step of training the neural network based on the first feature data category information and the training dataset to obtain the neural network specifically includes: S1: Based on the first feature data category information, using any first feature data category as a basis, train the neural network on the training dataset to obtain the benchmark neural network; S2: Based on the benchmark neural network, obtain the hidden layer information of the neural network, which includes the hidden node information of the neural network corresponding to each hidden layer of the neural network; S3: Based on the hidden layer of the neural network, classify the hidden nodes of the neural network and the weights corresponding to each hidden node to obtain the set information of the hidden node of the neural network. The set of hidden nodes of the neural network represents the hidden nodes of the neural network in the same hidden layer of the neural network and the weights corresponding to each hidden node. S4: Based on the information of the neural network hidden node set, cluster the weights of the neural network hidden nodes in each neural network hidden node set using the DBSCAN algorithm to obtain the node weight category information; S5: Adjust the weight of each hidden node in the neural network based on the information of the set of hidden nodes, the weights of the hidden nodes, and the category of the node weights; S6: Based on the first feature data category information, repeat steps S1 to S5 until all first feature data categories have been trained to obtain the neural network.

6. The neural network intelligent optimization method based on data node analysis according to claim 5, characterized in that, The step of adjusting the weights of each hidden node in the neural network based on the hidden node set information, the hidden node weights, and the node weight category information specifically includes: The sum of the data mapping coefficient and 1 is used as the data category compression coefficient; Based on the information of the neural network hidden node set, obtain the minimum and maximum neural network hidden node weights in each neural network hidden node set; The product of the minimum hidden node weight corresponding to each set of hidden nodes in the neural network and the data class compression coefficient is taken as the minimum weight of the set of hidden nodes in the neural network, and the ratio of the maximum hidden node weight to the data class compression coefficient is taken as the maximum weight of the set of hidden nodes in the neural network. Based on the information of the hidden node set of the neural network, the average weight of the hidden nodes in each neural network hidden node set is used as the feature node weight of each neural network hidden node set. The ratio of the feature node weights to the weights of the largest hidden nodes in the neural network is used as the weight offset coefficient. The ratio of the weight offset coefficient to the data category compression coefficient is used as the mean coefficient; The average of the feature node weights of all hidden nodes in the neural network set is used as the baseline node weight. The weights of each hidden node in the neural network are adjusted based on the mean coefficient, the baseline node weight, and the calibrated minimum weight. Specifically, adjusting the weights of hidden nodes in the neural network involves: In the formula, n ijk θ represents the weights of the j-th hidden node in the i-th hidden layer of the neural network and the k-th hidden node in the (i-1)-th hidden layer of the neural network. i The number of node weight categories, ∈ imin The λ represents the minimum calibrated weight corresponding to the hidden layer of the i-th neural network, μ is the weight of the baseline node, and λ is the minimum calibrated weight. μ Let n be the mean coefficient. ijk >∈ imax Then the weights of the hidden nodes in the neural network are fixed, and the fixed weights will not be adjusted during subsequent training. imax This represents the maximum calibrated weight corresponding to the hidden layer of the i-th neural network.

7. A neural network intelligent optimization system based on data node analysis, used to implement the optimization method as described in any one of claims 1-6, characterized in that, include: The main control module is used to obtain the calibrated input category number based on the data mapping coefficients, taking the sum of the input category number and the output category number as the maximum category number, and obtaining the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, it arbitrarily selects a category from the input category information and the output category information as the first feature data category information. Based on the neural network hidden node set information, the neural network hidden node weights and node weight category information, it adjusts the weights of each neural network hidden node. Based on the first feature data category information and the training dataset, it trains the neural network to obtain the neural network. The information acquisition module is used to acquire basic data information. Based on the requirements of neural network construction, the basic data is divided to acquire training datasets and test datasets. Based on the training dataset, the input and output parameters are clustered using the DBSCAN algorithm to acquire the number of input categories and the number of output categories. Based on the training dataset, the total number of corresponding data features is acquired. Based on the data topology, the feature input category information corresponding to each output category is acquired. Using the hidden layers of the neural network as the classification basis, the hidden nodes of the neural network and the weights corresponding to each hidden node are classified to acquire the set information of the hidden nodes of the neural network. The evaluation module is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number, obtain the hidden node coefficient based on the neural network node difference coefficient and the neural network baseline hidden layer number, and take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient. The display module interacts with the main control module and is used to output and display the training dataset, test dataset, hidden node coefficients, first feature data category information, and neural network.

8. The neural network intelligent optimization system based on data node analysis according to claim 7, characterized in that, The main control module specifically includes: The control unit is used to adjust the weight of each hidden node in the neural network according to the hidden node set information, the hidden node weights, and the node weight category information, and to train the neural network according to the first feature data category information and the training dataset to obtain the neural network. An information receiving unit, which interacts with the information acquisition module and the evaluation module, is used to receive data and transmit it to the data recognition unit; The data recognition unit is used to obtain the calibrated input category number by taking the sum of the input category number and the output category number as the maximum category number based on the data mapping coefficient, and to obtain the calibrated output category number based on the calibrated input category number and the calibrated output category number. Based on the calibrated input category number and the calibrated output category number, the unit arbitrarily selects a category from the input category information and the output category information as the first feature data category information.

9. The neural network intelligent optimization system based on data node analysis according to claim 7, characterized in that, The information acquisition module specifically includes: The first acquisition unit is used to acquire basic data information, divide the basic data based on the requirements of neural network construction, acquire training dataset and test dataset, and perform data clustering on input parameters and output parameters based on the DBSCAN algorithm according to the training dataset to acquire the number of input categories and the number of output categories. The second acquisition unit is used to acquire the total number of corresponding data features based on the training dataset, acquire the feature input category information corresponding to each output category based on the data topology relationship, classify the hidden nodes of the neural network and the weights corresponding to each hidden node of the neural network based on the hidden layer of the neural network, and acquire the set information of the hidden nodes of the neural network.

10. The neural network intelligent optimization system based on data node analysis according to claim 7, characterized in that, The evaluation module specifically includes: The first evaluation unit is used to take the ratio of the number of hidden nodes in the first hidden layer of the neural network to the minimum number of nodes in the standard hidden layer as the neural network node difference coefficient, and the difference between the total number of hidden layers in the neural network and the number of feature hidden layers in the neural network as the neural network baseline hidden layer number. Based on the neural network node difference coefficient and the neural network baseline hidden layer number, the hidden node coefficient is obtained. The second evaluation unit is used to take the sum of the number of feature input categories corresponding to all output categories as the number of associated input categories based on the output category information and feature input category information, and to take the ratio of the number of output categories to the number of associated input categories as the data mapping coefficient.

Images