Neural network model training method and system based on improved random configuration algorithm

Abstract

The application belongs to the technical field of neural network model training, and discloses a neural network model training method and system based on an improved random configuration algorithm, which comprises the following steps: under the condition of reducing the root mean square error of the neural network model output, screening suitable neurons as candidate hidden layer nodes through the inequality constraint condition of the random configuration algorithm; screening K neurons with the fastest training error reduction from the candidate hidden layer nodes, and selecting the neuron least related to the previous L-1 hidden layer nodes from the selected K neurons as the optimal hidden layer node; calculating the output weight through the least square method, and updating the structure of the random configuration network; and judging whether the network structure is completed by using the maximum allowable number of hidden layer nodes and the maximum allowable output error. The application can reduce the overall calculation amount of the algorithm while ensuring the prediction accuracy, and is suitable for application scenarios with high real-time requirements.

Description

Technical Field

[0001] This invention belongs to the field of neural network model training technology, and particularly relates to a neural network model training method and system based on an improved random allocation algorithm. Background Technology

[0002] Randomized Assigned Networks (RANs) are an incremental neural network model. Unlike other randomized neural network methods, RANs introduce a supervision mechanism during construction to allocate hidden layer parameters, ensuring their universal approximation properties. They offer advantages such as ease of implementation, fast convergence, and good generalization performance. Currently, they are widely used in process parameter prediction, fault diagnosis, and adaptive control. However, these applications require online real-time measurements, thus placing high demands on the prediction accuracy and computational speed of the network model.

[0003] Shortcomings of existing technology:

[0004] (1) It is difficult to set a suitable network structure for neural network models based on backpropagation of error. The training process requires adjusting the network connection weights and biases according to the gradient descent method. The training process is computationally intensive, time-consuming, and prone to getting stuck in local minima. It is very sensitive to parameter selection and has low efficiency.

[0005] (2) Randomly configured networks use a supervised mechanism to configure the input weights and biases of hidden layer nodes and calculate the output weights using the least squares method. However, the configuration process may result in repeated parameter traversal, which increases the computational load of the model, reduces the error convergence speed, and may not obtain good hidden layer output features, thus affecting the generalization performance of the model.

[0006] Based on the above analysis, the problems and defects of the existing technology are as follows: the existing neural network training process is computationally intensive, time-consuming, and prone to getting stuck in local minima. The parameter selection is very sensitive, resulting in low efficiency. It is also unable to obtain good hidden layer output features, which affects the generalization performance of the model.

[0007] The main shortcomings of existing technologies in industrial applications include the following:

[0008] 1. Real-time and efficiency issues: In many industrial applications, such as fault diagnosis and adaptive control, there are high requirements for real-time performance and response speed. However, neural network models based on error backpropagation may not be able to meet these real-time requirements due to their large computational demands, long training process, and tendency to get trapped in local minima.

[0009] 2. Structure and Parameter Selection: Traditional neural network models, such as backpropagation networks, require manual or experimental determination of suitable network structures, such as the number of layers and the number of nodes per layer. Furthermore, the initialization and selection of parameters significantly impact the model's convergence speed and performance, making network deployment and maintenance difficult in industrial environments.

[0010] 3. Local Minima Problem: Error backpropagation networks are prone to getting stuck in local minima during the learning process. This means that even after long training periods, the optimal model performance may not be achieved. In industrial applications, this can lead to performance degradation or instability of the system.

[0011] 4. Generalization Performance: Due to the potential for repeated parameter traversal and the inability to obtain satisfactory hidden layer output features during the configuration process, randomized network configuration can negatively impact the model's generalization performance. In industrial applications, the model needs to demonstrate strong predictive and responsive capabilities for unseen data or situations; otherwise, system failure or inefficiency may result.

[0012] To address the aforementioned shortcomings, the existing technologies urgently need to solve the following technical problems:

[0013] 1. Optimize training algorithms: In order to improve the training speed of the network and avoid getting stuck in local minima, it is necessary to study more efficient training algorithms or combine them with other optimization techniques, such as momentum and regularization.

[0014] 2. Adaptive Network Structure: Researching how to automatically or semi-automatically determine network structure to adapt to different applications and data.

[0015] 3. Enhance generalization performance: Enhance the network's generalization ability by introducing new network structures, activation functions, or training strategies.

[0016] 4. Reduce computational load: In order to meet the real-time requirements of industrial applications, it is necessary to study how to reduce the computational load of the network, such as model pruning, quantization and other techniques.

[0017] In summary, existing neural network technologies still face many unresolved technical problems and challenges in response to the characteristics and needs of industrial applications. Summary of the Invention

[0018] To address the problems existing in the prior art, this invention provides a neural network model training method and system based on an improved random allocation algorithm.

[0019] This invention is implemented as follows: a neural network model training method based on an improved random allocation algorithm, wherein the neural network model training method based on the improved random allocation algorithm includes:

[0020] First, suitable neurons are selected as candidate hidden layer nodes by using the inequality constraints of the random configuration algorithm.

[0021] Secondly, select the K neurons from the candidate hidden layer nodes that reduce the training error the fastest, and select the neuron that is least related to the first L-1 hidden layer nodes from the selected K neurons as the optimal hidden layer node;

[0022] Finally, the output weights are calculated using the least squares method, and the structure of the randomly configured network is updated. The network structure is then determined to be complete by using the maximum allowable number of hidden nodes and the maximum allowable output error.

[0023] Furthermore, the neural network model training method based on the improved random allocation algorithm specifically includes the following steps:

[0024] Step 1: Preprocess the sample data, set the parameters of the randomized network, and initialize the randomized network.

[0025] Step 2: Construct candidate hidden layer nodes according to the improved weight and bias definition method; substitute the obtained candidate hidden layer nodes into the inequality constraints and filter to obtain candidate hidden layer nodes that satisfy the inequality constraints.

[0026] Step 3: Perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraints to obtain the optimal candidate hidden layer nodes;

[0027] Step 4: Add the weights and biases of the optimal candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model.

[0028] Step 5: Determine whether the root mean square error of the current network model output is greater than the maximum expected output error tolerance. If it is greater, proceed to step 6; if it is less, proceed to step 7.

[0029] Step 6: Determine if the number of hidden layer nodes is less than the maximum allowed number of hidden layer nodes. If the number of hidden layer nodes is less than the maximum allowed number of hidden layer nodes, return to Step 2; if the number of hidden layer nodes is equal to the maximum allowed number of hidden layer nodes, proceed to Step 7.

[0030] Step 7: Training ends, outputting a neural network model that satisfies the constraints of the stochastic configuration theory.

[0031] Furthermore, the step of setting the parameters of the random configuration network and initializing the random configuration network includes:

[0032] Set the maximum allowed number of hidden layer nodes L max Maximum expected output error tolerance ∈, maximum number of candidate hidden layer nodes Tmax , the distribution of input weights γ={γ1, γ2,...}={[-λ1, λ1], [-λ2,-λ1]U[λ1, λ2],...,[-λ i+1 , -λ i ]U[λ i , λ i+1 ]}, where λ i It is a positive number, and λ i+1 >λ i Initialize the output error vector e0:=[y1, y2, ..., y N ] T The model output error scaling factor is 0 < r < 1, where T represents the ideal output, T = {t1, t2, ..., t}. N}, t i ={t i1 , t i2 , ..., t im} T N represents the number of samples; m represents the dimension of the output samples; Y represents the output samples, Y = {y1, y2, ..., y} N}, y i ={y i1 y i2 , ..., y im} T .

[0033] Furthermore, the construction of candidate hidden layer nodes based on the improved weight and bias definition method includes:

[0034] Randomly select input weight w from the distribution γ = {γ1, γ2, ...} of the input weights. L ={w L1 w L2 , ..., w Ld} T Bias Substitute into the activation function Candidate hidden layer nodes are obtained;

[0035] Where X represents the input sample, X = {x1, x2, ..., x} N}, x i ={x i1 x i2 , ..., x id} T d represents the dimension of the input sample.

[0036] Furthermore, the inequality constraints are as follows:

[0037]

[0038] in, nonnegative real number sequence {μ L} satisfies lim L→∞ μ L =0, and μ L ≤(1-r).

[0039] Furthermore, the step of performing secondary optimization on candidate hidden layer nodes that satisfy the inequality constraints based on semi-orthogonalization constraints to obtain the optimal candidate hidden layer nodes includes:

[0040] Define a set of variables From the candidate hidden layer nodes that satisfy the inequality constraint, select those that satisfy ξ. L The top K candidate nodes with a value greater than 0

[0041] Calculate the unit cosine values ​​of the K candidate nodes and the first L-1 nodes in the hidden layer.

[0042] Select from K candidate nodes The smallest node is selected as the optimal candidate hidden layer node.

[0043] Furthermore, the step of calculating the output weights of the neural network model based on the ideal output using the least squares method and updating the neural network model includes:

[0044] By comparing the output values ​​of each hidden layer node with the ideal output value, the output weights of the neural network model are calculated using the least squares method. The root mean square error of the neural network model's output is then calculated, and the model's output error e0 = e L The number of hidden layer nodes is L = L + 1.

[0045] Another object of the present invention is to provide a neural network model training system based on the improved random allocation algorithm for implementing the neural network model training method based on the improved random allocation algorithm, wherein the neural network model training system based on the improved random allocation algorithm includes:

[0046] The initialization module is used to preprocess the sample set data, set the parameters of the random configuration network, and initialize the random configuration network.

[0047] The node filtering module is used to construct candidate hidden layer nodes based on the improved weight and bias definition method; the obtained candidate hidden layer nodes are substituted into the inequality constraints to filter out candidate hidden layer nodes that satisfy the inequality constraints.

[0048] The node secondary optimization module is used to perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraints, so as to obtain the optimal candidate hidden layer nodes.

[0049] The model update module is used to add the weights and biases of the best candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model.

[0050] The output module is used to output a neural network model that satisfies the constraints of the random configuration theory.

[0051] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the neural network model training method based on the improved random allocation algorithm.

[0052] Another objective of this invention is to provide an information data processing terminal for implementing the neural network model training system based on the improved random configuration algorithm.

[0053] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0054] First, this invention uses an improved random allocation algorithm to train a neural network model. The inequality constraints of the random allocation algorithm are used to find suitable neurons as candidate hidden layer nodes. Unlike the original single constraint, the improved random allocation network of this invention adds correlation constraints on the hidden layer output.

[0055] This invention employs an improved stochastically configured network for training, adjusting the distribution of weight parameters to prevent repeated traversals and effectively improving the training speed of the network model. This invention also presents a method for defining the bias term of a stochastically configured network, linking it to the input data to improve the model's prediction accuracy.

[0056] This invention adds semi-orthogonal constraints to the original random configuration network, making the outputs of hidden layer nodes as uncorrelated as possible, reducing redundant nodes, enhancing the expressive power of the model, and thus improving generalization performance. When applied to industrial big data modeling training or prediction, it can improve the prediction accuracy of the model and has certain industrial application value.

[0057] Second, the randomized network introduces a supervision mechanism during the incremental construction process to randomly configure the input weights and biases of the hidden layer nodes. This results in fast training speed and high accuracy, without the need for iterative updates through backpropagation. This avoids problems such as getting trapped in local optima, gradient vanishing, or gradient explosion caused by updating network parameters using gradient descent.

[0058] This invention proposes an improved randomized network learning algorithm to optimize the original network model. Combined with the model framework designed in this invention, it can reduce the overall computational load of the algorithm while ensuring prediction accuracy. It is suitable for application scenarios with high real-time requirements and has good application prospects in application environments with limited computing resources, such as industrial artificial intelligence, smart healthcare, smart transportation, and autonomous driving.

[0059] Third, the application of each claim in industry has led to significant technological advancements. The following is an analysis of each claim and the industrial progress it has brought:

[0060] 1. A neural network model training method based on an improved random allocation algorithm:

[0061] Technological advancements: This requirement clarifies the outline and core ideas of the entire training method. By selecting suitable neurons, the selection process of hidden layer nodes is optimized, thereby improving the training efficiency and prediction accuracy of the neural network model. This can reduce computational costs and increase production efficiency in industry.

[0062] 2. Specific steps included:

[0063] Technological advancements: A detailed, structured training process is provided, making the method's implementation clearer and reducing its difficulty. For industrial applications, clear operating procedures can significantly reduce errors and improve work efficiency.

[0064] 3. Set the parameters for the random configuration network and initialize the random configuration network:

[0065] Technological advancements have provided a concrete and efficient method for network initialization, which is crucial for the convergence speed and final predictive performance of neural networks. In industrial applications, this can shorten model training time, thereby accelerating product development and innovation.

[0066] 4. Construct candidate hidden layer nodes based on the improved weight and bias definition method:

[0067] Technological advancements: This ensures the diversity and high quality of hidden layer nodes, providing richer feature representations for neural networks. In industry, richer feature representations can enhance the predictive power and generalization performance of models.

[0068] 5. Inequality constraints:

[0069] Technological advancement: A new constraint method has been introduced, which can more effectively screen out high-quality hidden layer nodes. In practical industrial operations, this can reduce unnecessary computation and improve the efficiency and accuracy of the model.

[0070] 6. Perform secondary optimization based on semi-orthogonality constraints:

[0071] Technological advancements: Semi-orthogonalization constraints further optimize the selection of hidden layer nodes, ensuring independence and diversity among nodes. This optimization improves model stability and prediction accuracy, making it highly valuable for industrial applications.

[0072] 7. Calculate the output weights of the neural network model using the least squares method:

[0073] Technological advancement: This method provides a fast and accurate way to calculate output weights. For industrial applications requiring real-time response, such as real-time control and monitoring, this method can provide rapid model output.

[0074] In summary, these claims not only provide a new and efficient method for training neural network models, but also bring significant technological advancements to industrial applications, including improved prediction accuracy, reduced computational costs, and accelerated model training. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the neural network model training method based on the improved random allocation algorithm provided in the embodiments of the present invention;

[0076] Figure 2 This is a flowchart of a neural network model training method based on an improved random allocation algorithm provided in an embodiment of the present invention. Detailed Implementation

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] like Figures 1-2 As shown, the neural network model training method based on the improved random allocation algorithm provided in this embodiment of the invention specifically includes the following steps:

[0079] S101, perform preprocessing of the sample set data, set the parameters of the random configuration network and initialize the random configuration network;

[0080] S102, construct candidate hidden layer nodes according to the improved weight and bias definition method; substitute the obtained candidate hidden layer nodes into the inequality constraints, and select candidate hidden layer nodes that satisfy the inequality constraints.

[0081] S103, perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraint to obtain the optimal candidate hidden layer nodes;

[0082] S104: Add the weights and biases of the optimal candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model.

[0083] S105, Determine whether the root mean square error of the current network model output is greater than the maximum expected output error tolerance. If it is greater, proceed to step S106; if it is less, proceed to step S107.

[0084] S106, determine whether the number of hidden layer nodes is less than the maximum allowable number of hidden layer nodes. If the number of hidden layer nodes is less than the maximum allowable number of hidden layer nodes, return to step S102; if the number of hidden layer nodes is equal to the maximum allowable number of hidden layer nodes, proceed to step S107.

[0085] S107, Training complete, outputting a neural network model that satisfies the constraints of the stochastic configuration theory.

[0086] The neural network model training system based on an improved random allocation algorithm provided in this embodiment of the invention includes:

[0087] The initialization module is used to preprocess the sample set data, set the parameters of the random configuration network, and initialize the random configuration network.

[0088] The node filtering module is used to construct candidate hidden layer nodes based on the improved weight and bias definition method; the obtained candidate hidden layer nodes are substituted into the inequality constraints to filter out candidate hidden layer nodes that satisfy the inequality constraints.

[0089] The node secondary optimization module is used to perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraints, so as to obtain the optimal candidate hidden layer nodes.

[0090] The model update module is used to add the weights and biases of the best candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model.

[0091] The output module is used to output a neural network model that satisfies the constraints of the random configuration theory.

[0092] The neural network model training method based on the improved random allocation algorithm provided in this embodiment of the invention includes the following steps:

[0093] Step 1, Given a sample set, input sample: X = {x1, x2, ..., x...} N}, x i ={x i1 x i2 , ..., x id} T Output sample: Y = {y1, y2, ..., y} N}, y i ={y i1 y i2 , ..., y im} T Ideal output: T = {t1, t2, ..., t} N}, t i ={t i1 , t i2 , ..., t im} T Where N is the number of samples, d is the dimension of the input samples, and m is the dimension of the output samples. Set the maximum allowable number of hidden layer nodes L. max Maximum expected output error tolerance ∈, maximum number of candidate hidden layer nodes T max , the distribution of input weights γ={γ1, γ2,...}={[-λ1, λ1], [-λ2,-λ1]U[λ1, λ2],...,[-λ i+1 , -λ i ]U[λ i , λ i+1 ]},(λ i It is a positive number, and λ i+1 >λ i ).

[0094] Step 2, initialize the output error vector e0 := [y1, y2, ..., y N ] T The model output error scaling factor is 0 < r < 1.

[0095] Step 3: Randomly select input weight w from γ = {γ1, γ2, ...} L ={w L1 w L2 , ..., w Ld} T bias Substitute into the activation function Candidate hidden layer nodes are obtained.

[0096] Step 4, T max Substituting the candidate hidden layer nodes into the inequality constraints of the random allocation algorithm, (nonnegative real number sequence {μ L} satisfies lim L→∞ μ L =0, and μ L ≤(1-r)), from which candidate hidden layer nodes that satisfy the inequality constraints are selected.

[0097] Step 5, define a set of variables Select those that satisfy ξ L The top K candidate nodes with a value greater than 0

[0098] Step 6: Calculate the unit cosine values ​​of these K candidate nodes and the first L-1 nodes in the hidden layer. Choose from The smallest node is selected as the optimal candidate node.

[0099] Step 7: Add the optimal candidate hidden layer nodes to the neural network model, directly calculate the output weight β of the neural network model based on the ideal output, calculate the root mean square error of the neural network model output, and update the model's output error e0 = e L And the number of hidden layer nodes L = L + 1.

[0100] Step 8: If the root mean square error ||e0||2 of the current network model output is greater than the maximum expected output error tolerance value ∈, and the number of hidden layer nodes L is less than the maximum allowable number of hidden layer nodes L max Then repeat steps 2 to 7; if the root mean square error ||e0||2 of the current network model output is less than the maximum expected output error tolerance value ∈, or the number of hidden layer nodes L is equal to the maximum allowable number L. max Then, the training ends, and we obtain a network model that satisfies the constraints of the stochastic configuration theory.

[0101] Steps 3 to 6 are the key points of this invention. Step 3 prevents repeated traversal in the same interval by changing the distribution of the weight parameters, thus accelerating error convergence. At the same time, a novel bias definition method is introduced to link it with the input data, thereby improving the prediction accuracy of the model.

[0102] The principles underlying the inequality constraints of the randomized allocation algorithm mentioned in step 4 are as follows:

[0103] Suppose that the vector space Γ in L2 space is dense, and at the same time Make 0 < ||g|| < b g Given a sequence of non-negative real numbers {μ}, where 0 < r < 1. L}, where lim L→+∞ μ L =0, μ L ≤(1-r). For hidden layer nodes L=1,2,..., given the following formula:

[0104]

[0105] If the random basis function g L The constructed output weight β L，q satisfy

[0106]

[0107] And it satisfies the following inequality constraints:

[0108]

[0109] Then we have lim L→+∞ ||ff L || = 0, where f is the output value of the actual data, f L It is the predicted output value of the model with L hidden layer nodes. That is, the constructed neural network model still has the global approximation property of the original model.

[0110] Step 5 aims to find candidate nodes that cause the model training error to decrease the fastest, thereby accelerating the convergence speed.

[0111] The principle underlying the secondary optimization of the candidate nodes proposed in step 6 is as follows:

[0112] Under the premise of error convergence, the candidate nodes that are least related to the first L-1 nodes in the hidden layer are found from the candidate nodes. Based on the algebraic properties of the model, the rank of the hidden layer output matrix is ​​maximized, the nodes are as unrelated as possible, redundant nodes are reduced, and the expressive power of the model is enhanced.

[0113] In step 7, the least squares method is used to calculate the output weight β. L，q The specific principle is explained below:

[0114] Suppose that the vector space Γ in L2 space is dense, and at the same time Make 0 < ||g|| < b g Given a sequence of non-negative real numbers {μ}, where 0 < r < 1. L}, where lim L→+∞ μ L =0, μ L ≤(1-r). For hidden layer nodes L=1,2,..., given the following formula:

[0115]

[0116] If the random basis function g L Constructed output weights satisfy

[0117]

[0118] And it satisfies the following inequality constraints:

[0119]

[0120] So there are That is, the constructed neural network model has global approximation properties.

[0121] Therefore, the suboptimal solution It can be obtained using the least squares method:

[0122]

[0123] Among them G L Let g be the matrix of the random basis function g, and T be the actual data output. It is the Moore-Penrose generalized inverse.

[0124] The neural network model training method based on the improved random allocation algorithm provided in this embodiment of the invention specifically includes the following steps:

[0125] Step 1: Preprocess the data.

[0126] Step 2, Randomly configure network initialization: Set the maximum allowed number of hidden layer nodes L max The initial node L = 0, the maximum expected output error tolerance is ∈, and the distribution of input weights is γ = {γ1, γ2, ...} = {[-λ1, λ1], [-λ2, -λ1] U[λ1, λ2], ..., [-λ i+1 , -λ i ]U[λ i , λ i+1 The maximum number of candidate hidden layer nodes T max Model output error initialization: e0 = T.

[0127] Step 3: Construct candidate nodes based on the improved weight and bias definition method: Randomly select input weights w from γ = {γ1, γ2, ...} L ={w L1 w L2 , ..., W Ld} T bias Substitute into the activation function Candidate hidden layer nodes are obtained.

[0128] Step 4: Supervised selection of candidate nodes based on inequality constraints: Set Will Tma Substituting x groups of candidate hidden layer nodes into the inequality constraints of the random allocation algorithm. Candidate hidden layer nodes that satisfy the inequality constraints are selected from these. When min{ξ L，1 ξ L，2 ,...,ξ L，q If}≥0, keep the selected w. L b LProceed to the next step; otherwise, when λ i ≤λ max , λ∈γ i+1 Return to step 3 and continue execution. If neither of the above two conditions is satisfied, let λ∈γ1, r∈{r1, r2, ..., r...} max Then return to step 3 and continue execution.

[0129] Step 5: Perform secondary optimization on the candidate nodes from Step 4 based on the semi-orthogonality constraint: select ξ from the retained candidate nodes. L The largest top K candidate nodes Calculate the unit cosine values ​​of these K candidate nodes and the first L-1 nodes in the hidden layer. Choose from The smallest node is selected as the optimal candidate node.

[0130] Step 6: Determine the optimal candidate node and add it to the hidden layer: Set the weights and biases of the optimal candidate node... It is added to the network structure as a new hidden layer node.

[0131] Step 7: Calculate the network output weights by combining the target output and the outputs of the hidden layer nodes: Update the neural network model, and calculate the model's output weights β using the least squares method by comparing the output values ​​of each hidden layer node with the ideal output value. * Finally, update the model output error e0 = e L And the number of hidden layer nodes L = L + 1. When the root mean square error of the model output ||e0||2>∈, and L≤L max If the condition is met, return to step 3; otherwise, the neural network model training ends. Apply the completed neural network model to the test set, obtain the output values ​​of the test set, and analyze the model's generalization performance.

[0132] Based on the improved random allocation algorithm for neural network model training described above, two specific industrial implementations and their solutions are listed below:

[0133] Example 1: Prediction of Parameters in Iron and Steel Smelting Process

[0134] Background: In the steelmaking process, accurate prediction of key parameters such as temperature, chemical composition and smelting time is crucial for improving production efficiency and quality control.

[0135] 1. Data preprocessing: First, the large amount of sensor data collected during the smelting process is preprocessed, including noise reduction, standardization and normalization.

[0136] 2. Initialize the network: Set the parameters of the random configuration network and initialize it according to step one above.

[0137] 3. Construct candidate hidden layer nodes: Using sensor data from the smelting process as input, construct candidate hidden layer nodes according to step two.

[0138] 4. Select the optimal hidden layer node: Following the descriptions in steps three, four, five, and six, select the optimal hidden layer node and update the neural network structure.

[0139] 5. Training and Prediction: After completing the construction of the neural network structure, it is trained using historical data from the smelting process, and then the model is used to predict key smelting parameters.

[0140] Example 2: Early Fault Diagnosis on Machine Equipment

[0141] Background: In many manufacturing and heavy industries, early detection of abnormalities or malfunctions in machinery and equipment can save significant maintenance costs and prevent production interruptions.

[0142] 1. Data collection: Using a variety of sensors, such as vibration sensors, temperature sensors and sound sensors, to collect the working data of the machine equipment periodically or in real time.

[0143] 2. Data preprocessing: Preprocess the sensor data, including noise reduction, filtering, and feature extraction.

[0144] 3. Initialize the network: Set the parameters of the random configuration network and initialize it according to step one above.

[0145] 4. Construct candidate hidden layer nodes: Using sensor data from the machine equipment as input, construct candidate hidden layer nodes according to step two.

[0146] 5. Select the optimal hidden layer node: Following the descriptions in steps three, four, five, and six, select the optimal hidden layer node and update the neural network structure.

[0147] 6. Training and Prediction: The model is trained using historical data, including normal operation data and known fault data. This model is then used to monitor the machine's operating status in real time, predicting and proactively identifying potential faults.

[0148] In these two embodiments, by applying a neural network model based on an improved randomized allocation algorithm, more accurate and faster predictions and diagnoses can be achieved, bringing benefits to industrial production.

[0149] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0150] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for training a neural network model based on an improved random allocation algorithm, characterized in that, include: Under the condition that the root mean square error of the neural network model output is reduced, suitable neurons are selected as candidate hidden layer nodes by means of the inequality constraints of the random allocation algorithm. Select the candidate hidden layer nodes that will cause the fastest reduction in training error. 10 neurons, selected from 10 neurons Select from the neurons that are previously The least relevant neuron in the hidden layer is selected as the optimal hidden layer node. The output weights are calculated using the least squares method, and the structure of the randomly configured network is updated. The network structure is then determined to be complete by using the maximum allowable number of hidden layer nodes and the maximum allowable output error. The neural network model training method based on the improved random allocation algorithm specifically includes the following steps: Step 1: Preprocess the sample set data, set the parameters of the randomized network and initialize the randomized network; preprocess the large amount of sensor data collected during the smelting process, including denoising, standardization and normalization. Step 2: Construct candidate hidden layer nodes according to the improved weight and bias definition method; substitute the obtained candidate hidden layer nodes into the inequality constraints, and screen out the candidate hidden layer nodes that satisfy the inequality constraints, using sensor data from the smelting process as input. Step 3: Perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraints to obtain the optimal candidate hidden layer nodes; Step 4: Add the weights and biases of the optimal candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model. Step 5: Determine whether the root mean square error of the current network model output is greater than the maximum expected output error tolerance. If it is greater, proceed to step 6; if it is less, proceed to step 7. Step 6: Determine if the number of hidden layer nodes is less than the maximum allowed number of hidden layer nodes. If the number of hidden layer nodes is less than the maximum allowed number of hidden layer nodes, return to Step 2; if the number of hidden layer nodes is equal to the maximum allowed number of hidden layer nodes, proceed to Step 7. Step 7: Training ends, outputting a neural network model that satisfies the constraints of the stochastic configuration theory; The steps of setting the parameters of the random configuration network and initializing the random configuration network include: Set the maximum allowed number of hidden layer nodes. Maximum expected output error tolerance Maximum number of candidate hidden layer nodes Distribution of input weights ,in It is a positive number, and Initialize the output error vector Model output error scaling factor ,in, This represents the ideal output. ; Indicates the number of samples; Indicates the dimension of the output samples. Indicates the output sample. ; The construction of candidate hidden layer nodes based on the improved weight and bias definition method includes: From the distribution of input weights In the process, input weights are randomly selected. Bias Substitute into the activation function Candidate hidden layer nodes are obtained; in, Indicates the input sample. ; Indicates the dimension of the input sample; The inequality constraints are as follows: ； in, nonnegative real number sequence satisfy ,and .

2. The neural network model training method based on the improved random allocation algorithm as described in claim 1, characterized in that, The step of performing secondary optimization on candidate hidden layer nodes that satisfy inequality constraints based on semi-orthogonalization constraints to obtain optimal candidate hidden layer nodes includes: Define a set of variables From the candidate hidden layer nodes that satisfy the inequality constraints, select those that satisfy the inequality constraints. The largest top K candidate nodes ; Calculate the K candidate nodes and the first hidden layer The unit cosine value of each node ,from Select from candidate nodes The smallest node is selected as the optimal candidate hidden layer node.

3. The neural network model training method based on the improved random allocation algorithm as described in claim 1, characterized in that, The process of calculating the output weights of the neural network model based on the ideal output using the least squares method and updating the neural network model includes: By comparing the output values ​​of each hidden layer node with the ideal output value, the output weights of the neural network model are calculated using the least squares method. The root mean square error of the neural network model's output is then calculated, and the model's output error is updated. and the number of hidden layer nodes .

4. A neural network model training system based on an improved random allocation algorithm, implementing the neural network model training method based on any one of claims 1-3, characterized in that, The neural network model training system based on the improved random allocation algorithm includes: The initialization module is used to preprocess the sample set data, set the parameters of the random configuration network, and initialize the random configuration network. The node filtering module is used to construct candidate hidden layer nodes based on the improved weight and bias definition method; the obtained candidate hidden layer nodes are substituted into the inequality constraints to filter out candidate hidden layer nodes that satisfy the inequality constraints. The node secondary optimization module is used to perform secondary optimization on the candidate hidden layer nodes that satisfy the inequality constraints according to the semi-orthogonalization constraints, so as to obtain the optimal candidate hidden layer nodes. The model update module is used to add the weights and biases of the best candidate hidden layer nodes as new hidden layer nodes to the hidden layer of the neural network, calculate the output weights of the neural network model based on the ideal output using the least squares method, and update the neural network model. The output module is used to output a neural network model that satisfies the constraints of the random configuration theory.

5. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the neural network model training method based on the improved random allocation algorithm as described in any one of claims 1-3.

6. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the neural network model training system based on the improved random configuration algorithm as described in claim 4.

Images