Tunnel hole slag performance prediction method based on mother rock mineral composition

Abstract

The invention relates to the technical field of rock and soil analysis, in particular to a tunnel hole slag performance prediction method based on mother rock mineral composition. The method comprises the following steps: obtaining an initial type and a classification boundary corresponding to each performance data; establishing an initial prediction model corresponding to each type of performance data; obtaining a training set, a verification set and a test set of the initial prediction model; training the initial prediction model by using the training set and the verification set to obtain a first prediction model; testing the prediction model by using the test set to obtain prediction accuracy; updating the classification boundary or the initial prediction model according to the prediction accuracy, and training again to obtain a second prediction model; and if the prediction accuracy corresponding to the second prediction model reaches a preset standard, stopping updating the classification boundary or the initial prediction model, and if the prediction accuracy does not reach the preset standard, continuing updating until the preset standard is reached, thereby obtaining a final prediction model. According to the method, the hole slag performance can be accurately predicted.

Description

Technical Field

[0001] The invention relates to the technical field of rock and soil analysis, and in particular to a method for predicting tunnel slag properties based on parent rock mineral composition. Background Art

[0002] In recent years, my country's railway construction has developed rapidly. During the tunnel excavation process, a large amount of tunnel slag was generated. The storage and treatment of tunnel slag caused a large amount of land resources to be wasted and damaged the ecological environment along the railway. At the same time, due to the shortage of sand and gravel resources in my country and the difficulty in transporting raw materials, railway construction faces a severe shortage of concrete raw materials. If the tunnel slag is prepared into sand and gravel aggregates and admixtures and used locally for national railway construction, the problem of raw material shortage can be effectively alleviated. However, my country has a vast territory, different geological conditions in different regions, and the overall geology is relatively complex. The properties of the tunnel slag parent rock are varied, including not only granite, gneiss, and diorite layers, but also slate, mudstone and other weak strata and a large amount of fault crushed rock and sand.

[0003] Due to the wide variety of tunnel slag parent rocks, complex mineral composition, and large performance fluctuations, it is difficult to explore all the properties of railway tunnel slag parent rocks through experimental means. In contrast, the parent rock performance prediction model is of great significance in clarifying the performance characteristics of railway tunnel parent rocks. It can accurately predict the physical, chemical, and mechanical characteristics of parent rocks based on information such as rock type and elemental composition. However, traditional parent rock performance prediction methods usually have low accuracy and a small scope of application. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a method for predicting tunnel slag properties based on the mineral composition of the parent rock. The technical solution adopted is as follows:

[0005] One embodiment of the present invention provides a method for predicting tunnel slag properties based on parent rock mineral composition, the method comprising:

[0006] Acquire a parent rock information data set, wherein the parent rock information data set includes a parent rock number, mineral composition, physical and chemical properties, and mechanical properties; and pre-process the parent rock information data set;

[0007] Classify each performance data of physical and chemical properties and mechanical properties, obtain the initial type and classification boundary corresponding to each performance data; establish the initial prediction model corresponding to each performance data;

[0008] Based on the parent rock information data set, a training set, a validation set and a test set of the initial prediction model corresponding to each performance data are obtained; the initial prediction model is trained using the training set and the validation set to obtain a first prediction model; the prediction model is tested using the test set to obtain a prediction accuracy;

[0009] The classification boundary or the initial prediction model is updated according to the prediction accuracy, and the second prediction model is trained again; if the prediction accuracy corresponding to the second prediction model reaches the preset standard, the updating of the classification boundary or the initial prediction model is stopped; if it does not reach the preset standard, the updating continues until it reaches the preset standard to obtain the final prediction model.

[0010] Preferably, the physical and chemical properties and mechanical properties include:

[0011] The physical and chemical properties and mechanical properties are obtained by experiments; the physical and chemical properties include porosity, average pore size, density, harmful element content, pH, water absorption, thermal expansion coefficient and thermal conductivity; the mechanical properties include hardness, compressive strength, shear strength, tensile strength, elastic modulus, Poisson's ratio, crushing value of gravel and crushing value of machine-made sand.

[0012] Preferably, the preprocessing of the parent rock information data set includes:

[0013] Preprocessing the parent rock information data set includes data cleaning, data screening and data conversion of the data in the parent rock information data set.

[0014] Preferably, the mineral composition comprises:

[0015] The mineral composition includes the lithology of the parent rock and its three major rock types, and the element content of the parent rock; the parent rock is classified according to the lithology of the parent rock to obtain the three major rock types to which the lithology of the parent rock belongs.

[0016] Preferably, the initial type corresponding to each performance data includes:

[0017] For an initial type of performance data, the number of performance data in each initial type is the same as the number of performance data in other initial types.

[0018] Preferably, the initial prediction model includes:

[0019] The initial prediction model includes an input layer, a hidden layer and an output layer; the number of hidden layers is a first preset number, and the number of neurons in the hidden layers is a second preset number.

[0020] Preferably, the prediction accuracy is:

[0021]

[0022] Among them, δ represents the prediction accuracy; TP represents true positive examples; FP represents false positive examples; TN represents true negative examples; and FN represents false negative examples.

[0023] Preferably, updating the classification boundary or the initial prediction model according to the prediction accuracy includes:

[0024] The initial prediction model includes updating the activation function, hidden layer and neurons; when updating the classification boundary, activation function, hidden layer and neurons, only one item is updated at a time;

[0025] Updating the classification boundary includes: setting a first update strategy, wherein the first update strategy is specifically to increase or decrease the classification boundary under the condition that the amount of data of each category is within a preset range;

[0026] The activation function update includes: each time the update selects an activation function from the preset activation functions as the activation function of the initial prediction model;

[0027] The update of the hidden layer includes: adding one hidden layer each time the update is performed, provided that the number of hidden layers does not exceed the set number of layers;

[0028] The neuron update includes: under the condition that the number of neurons in each hidden layer does not exceed the set number, each update increases or decreases the set number of neurons in each hidden layer.

[0029] The embodiments of the present invention have at least the following beneficial effects: the present invention obtains a parent rock information data set, and pre-processes the parent rock information data set to improve the quality of the parent rock information data set, and then the training quality of the subsequent prediction model and the prediction accuracy; further, each performance data in the physical and chemical properties and mechanical properties is classified to obtain an initial type and a classification boundary, which is more in line with the application scenario and improves the training quality of the prediction model; further, a training set, a validation set and a test set are divided, and the initial training model is trained using the training set and the validation set to obtain a first prediction model, and the prediction accuracy of the first prediction model is obtained to guide the subsequent update of the classification boundary and the initial prediction model to obtain the optimal model; the classification boundary or the initial prediction model is updated according to the prediction accuracy, and the second prediction model is trained again; if the prediction accuracy corresponding to the second prediction model reaches a preset standard, the updating of the classification boundary or the initial prediction model is stopped; if the preset standard is not reached, the updating is continued until the preset standard is reached to obtain a final prediction model, and the robustness and accuracy of the prediction model can be improved by updating the classification boundary or the initial prediction model and training again. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0031] Figure 1 A method flow chart of a method for predicting tunnel slag properties based on parent rock mineral composition provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the tunnel slag performance prediction method based on the parent rock mineral composition proposed by the present invention, its specific implementation method, structure, characteristics and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0033] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0034] The specific scheme of the method for predicting the performance of tunnel slag based on the mineral composition of the parent rock provided by the present invention is described in detail below with reference to the accompanying drawings.

[0035] Example:

[0036] The main application scenarios of the present invention are: machine learning technology has shown significant advantages in predicting rock properties, especially in processing large amounts of complex data and mining the inherent laws between multiple factors. The accuracy and stability of machine learning models are better than traditional prediction methods, and they have broad application prospects in predicting parent rock properties. Therefore, it is urgent to develop a parent rock property prediction model based on machine learning methods.

[0037] See also Figure 1 , which shows a flow chart of a method for predicting tunnel slag properties based on parent rock mineral composition provided by an embodiment of the present invention, the method comprising the following steps:

[0038] Step S1, obtaining a parent rock information data set, wherein the parent rock information data set includes a parent rock number, mineral composition, physical and chemical properties, and mechanical properties; and preprocessing the parent rock information data set.

[0039] When constructing a prediction model for the performance of parent rock, it is necessary to obtain parent rock data. Specifically, it is necessary to investigate and test the parent rock, and obtain the parent rock mineral composition, physical and chemical performance data, and mechanical performance data based on the investigation and testing. The parent rock mineral composition, physical and chemical performance data, and mechanical performance data are obtained based on experiments. At the same time, the parent rock needs to be numbered to obtain the parent rock number; further, the parent rock needs to be classified according to the parent rock lithology, and the three main rock types to which the parent rock lithology belongs are obtained, namely igneous rock, sedimentary rock, and metamorphic rock. The parent rock type is represented by coding, where 1 represents igneous rock, 2 represents sedimentary rock, and 3 represents metamorphic rock. These numbers belong to categorical variables and only represent categories without size distinction. Mineral composition includes parent rock lithology and its three major rock types, and parent rock element content.

[0040] The parent rock number, mineral composition, physical and chemical properties and mechanical properties constitute the parent rock information data set, where the physical and chemical properties include porosity, average pore size, density, harmful element content, pH, water absorption, thermal expansion coefficient and thermal conductivity; mechanical properties include hardness, compressive strength, shear strength, tensile strength, elastic modulus, Poisson's ratio, crushing value of gravel and crushing value of machine-made sand. The performance data directly adopts the test result data, which are all numerical variables, that is, there are infinite possible values ​​in the definition domain, such as the value of porosity (unit %) can be any number in its definition domain (0, 100).

[0041] Furthermore, the parent rock information data set is preprocessed, specifically, the data in the parent rock information data set is cleaned, screened and transformed, and the blank data samples are removed through data cleaning to improve the quality of the data set. Thus, the preprocessed parent rock information data set can be obtained.

[0042] Step S2, classify each performance data of the physical and chemical properties and the mechanical properties, obtain the initial type and classification boundary corresponding to each performance data; and establish an initial prediction model corresponding to each performance data.

[0043] Furthermore, since the prediction of parent rock properties is mainly used in engineering scenarios, and in order to prevent overfitting during the training process, it is necessary to classify each performance data of the physical and chemical properties and mechanical properties. That is, the final predicted performance data is not the specific performance data value, but the interval or category to which it belongs.

[0044] When classifying each performance data, it is necessary to determine the classification boundary. If the adopted classification boundary causes the amount of data in each category after classification to differ too much, the knowledge learned by the model subsequently trained with these data is likely to be biased towards the category with larger data volume. The result is that the prediction of the model used for prediction is biased, making the model meaningless. Therefore, when classifying each performance data, adjustments should be made based on the statistical distribution of each performance data to ensure that the amount of data in each category is close. Taking density as an example, the data of this performance data is divided into 2.49-2.99, with a total of 123 performance data. This performance data is divided into three equal parts, that is, divided into three categories. There are 41 density data in each category. At this time, the classification boundaries are set to 2.63 and 2.71, that is, the data is divided into three intervals of 2.49-2.63, 2.63-2.71 and 2.71-2.99. When classifying each performance data, N+1 interval categories can be obtained through N classification boundaries, that is, N+1 initial types, and the proportion of each data is the proportion of the number of samples in each interval to the effective sample size of this performance.

[0045] Therefore, each performance data needs to be classified, and the amount of data in each type is the same. The classification boundary needs to be determined by the implementer based on the actual data distribution of each performance data. Therefore, each performance data in the physical and chemical properties and mechanical properties is initially classified to obtain the initial type and classification boundary corresponding to each performance data.

[0046] Furthermore, for each type of performance data, it is necessary to establish a corresponding prediction model, which is recorded as an initial prediction model. The initial prediction model includes an input layer, a hidden layer and an output layer; the number of hidden layers is a first preset number, and the number of neurons in the hidden layer is a second preset number; preferably, in the embodiment of the present invention, the first preset number is 1 and the second preset number is 20. It should be noted that the implementer can determine the values ​​of the first preset number and the second preset number according to actual conditions.

[0047] At this point, the initial category and classification boundary corresponding to each performance data can be obtained, and the initial prediction model corresponding to each performance data can also be obtained.

[0048] Step S3, based on the parent rock information data set, obtain the training set, verification set and test set of the initial prediction model corresponding to each performance data; use the training set and verification set to train the initial prediction model to obtain the first prediction model; use the test set to test the prediction model to obtain the prediction accuracy.

[0049] Furthermore, based on the parent rock information data set, the training set, validation set and test set of the initial prediction model corresponding to each performance data are obtained; among different types of performance data, the input of the initial prediction model corresponding to the performance data belonging to physical and chemical properties is the three major rock categories of parent rocks and one or more parent rock element contents, and the output is the initial type of each performance data in the physical and chemical properties; the input of the initial prediction model corresponding to the performance data belonging to mechanical properties is the three major rock categories of parent rocks, one or more parent rock element contents and porosity, and the output is the initial type of each performance data in the mechanical properties.

[0050] The parent rock information data set is divided into a training set, a validation set and a test set of the initial prediction model corresponding to each performance data according to a certain ratio. The initial prediction model corresponding to each performance data has a corresponding training set, a validation set and a test set.

[0051] The training set is the data source for model training. The validation set is used to evaluate the model performance during the training process and serve as a basis for adjustment to form a prediction model with good training performance. The test set is independent of the model training data and is used to test the accuracy and robustness of the trained model.

[0052] In the present invention, the amount of test set data is 20% of the total amount of data, and the remaining 80% of the data is used as training data, wherein the training data includes a training set and a validation set; at the same time, due to the problem of the amount of data, the number of training sets and validation sets is small, so k-fold cross validation is used in the training process to deal with the situation of insufficient data. Specifically, the data set of the remaining 80% of the data is divided into k equal parts of sub-samples. In each iteration, one of the sub-samples is used as a validation set, and the remaining k-1 sub-samples are used as training sets for model training. Subsequently, the prediction performance of the trained model on the validation set is evaluated. This process is repeated k times to ensure that each sub-sample has the opportunity to be used as a validation set. Finally, the results of these k iterations are averaged. The advantage of fold cross validation is that, in the case of a small amount of parent rock data in the data set, limited data can be more fully utilized for training and validation, while reducing the possibility that the evaluation results are affected by a specific data division method. In the present invention, the value of k is 10, and the implementer can adjust the fold number parameter according to actual conditions.

[0053] Take an initial prediction model corresponding to a performance data as an example, use its corresponding training set and validation set to train the initial prediction model, select the mean absolute error or mean square error as the loss function, and the implementer needs to determine which loss function to use according to the actual situation. At the same time, there will be a loss value during the training process, and the loss value decreases rapidly with the increase of the number of iterations, indicating that the initial prediction model has good learnability. Preferably, the number of iterations in the embodiment of the present invention is 400 times, and the implementer can adjust the number of iterations according to the actual situation. In the present invention, after the number of iterations reaches 400 times, the loss value tends to be flat, and the smaller the value of the loss function, the better the prediction performance of the model. After the training is completed, the first prediction model is obtained.

[0054] After obtaining the first prediction model, the data in the test set is input into the first prediction model to obtain the prediction result, and the confusion matrix is ​​used for accuracy evaluation. The relationship between the prediction result of the first prediction model on the test set and the actual category is displayed in the form of a matrix; the rows of the confusion matrix represent the actual category corresponding to the sample, the columns represent the predicted category of the sample predicted by the first prediction model, and each cell contains the number of samples matching the corresponding category; the typical composition includes true positive examples (TP), false positive examples (FP), true negative examples (TN) and false negative examples (FN), and then the prediction accuracy is calculated. The prediction accuracy is:

[0055]

[0056] Among them, δ represents the prediction accuracy; TP represents true positive examples; FP represents false positive examples; TN represents true negative examples; and FN represents false negative examples.

[0057] In this way, the initial prediction model for each performance data is trained, the first prediction model corresponding to each performance data is obtained, and then the prediction accuracy of the first prediction model is obtained.

[0058] Step S4, update the classification boundary or the initial prediction model according to the prediction accuracy, and train again to obtain the second prediction model; if the prediction accuracy corresponding to the second prediction model reaches the preset standard, stop updating the classification boundary or the initial prediction model; if it does not reach the preset standard, continue updating until it reaches the preset standard to obtain the final prediction model.

[0059] Due to the difference in performance data, if the same prediction model is used for prediction, the prediction accuracy will not meet the requirements, and the classification of performance data will also affect the final prediction results. Therefore, it is necessary to update the classification boundaries and the initial prediction model according to the prediction accuracy corresponding to the prediction model after each training. The update of the initial prediction model includes updating the activation function, hidden layer and neurons.

[0060] If there are too many classification intervals for each performance data, the model complexity will increase, which is not conducive to model training. Therefore, when updating the classification boundaries, it is necessary to adjust them according to the statistical distribution of each output variable to ensure that the amount of data in each category is close to

[0061] The activation function plays a very important role in the model training process. Its main function is to add a nonlinear operation to all hidden layers and output layers, making the output of the neural network more complex and more expressive. Commonly used activation functions are Sigmoid, Tanh and ReLU. Different activation functions have different sensitivities to different data types.

[0062] The hidden layer structure includes the number of hidden layers and the number of neurons in each layer. Adding hidden layers can help the model capture more complex features, but it may also lead to overfitting. Reducing hidden layers can reduce model complexity and prevent overfitting, but it may reduce the model's learning ability. Increasing the number of neurons can improve the model's learning ability, but too many neurons may lead to overfitting and increased computational costs.

[0063] Specifically, when updating the classification boundary, activation function, hidden layer and neuron, only one item is updated at a time, and the others are not updated. For example, if multiple classification boundaries are updated, the initial prediction model remains unchanged, that is, the activation function, hidden layer and neuron remain unchanged;

[0064] For the update of classification boundaries, it includes: setting a first update strategy, which is specifically: under the condition that the amount of data of each category is within a preset interval, increasing or decreasing the classification boundary. It should be noted that the preset interval is 15% to 50% of the amount of data of each performance, that is, the amount of data in each classification interval is not less than 15% and not more than 50%. In addition, the step size of increasing or decreasing the classification boundary needs to be determined by the implementer according to the actual situation;

[0065] For the activation function update, it includes: in the embodiment of the present invention, the preset activation function is that there are three activation functions: Sigmoid, Tanh and ReLU. Each time the update is performed, one activation function is selected from the three activation functions as the activation function of the initial prediction model, and then training is performed.

[0066] For the hidden layer update, it includes: when the hidden layer is updated, the number of hidden layers cannot exceed the set number of layers, and one hidden layer is added each time it is updated, where the set number of layers is 3;

[0067] Neuron update, that is, updating the number of neurons in each hidden layer, includes: under the condition that the number of neurons in each hidden layer does not exceed the set number, each update increases or decreases the set number of neurons in each hidden layer, where the set number is 10. For example, if there are two layers of neurons in the model, the number of neurons in both layers is increased by 10.

[0068] After each update of the first prediction model, it is necessary to use the training data for training to obtain the second prediction model, and then calculate the prediction accuracy of the updated second prediction model, and so on, after each update, it is necessary to train, and then calculate the prediction accuracy of the trained prediction model, until the prediction accuracy reaches the preset standard, and the final prediction model is obtained. Preferably, the preset standard in the embodiment of the present invention is 80%. Finally, the final prediction model corresponding to each performance data of the physical and chemical properties and mechanical properties can be obtained, wherein the optimal parameters and accuracy of the final prediction model of some types of performance data are shown in Table 1.

[0069] Table 1

[0070]

[0071] In this way, the final prediction model corresponding to each performance data can be obtained, and then the performance of the parent rock can be predicted.

[0072] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0073] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting tunnel slag properties based on parent rock mineral composition, characterized in that: The method includes: Acquire a parent rock information data set, wherein the parent rock information data set includes a parent rock number, mineral composition, physical and chemical properties, and mechanical properties; and pre-process the parent rock information data set; Classify each performance data of physical and chemical properties and mechanical properties, obtain the initial type and classification boundary corresponding to each performance data; establish the initial prediction model corresponding to each performance data; Based on the parent rock information data set, a training set, a validation set and a test set of the initial prediction model corresponding to each performance data are obtained; the initial prediction model is trained using the training set and the validation set to obtain a first prediction model; the prediction model is tested using the test set to obtain a prediction accuracy; The classification boundary or the initial prediction model is updated according to the prediction accuracy, and the second prediction model is trained again; if the prediction accuracy corresponding to the second prediction model reaches the preset standard, the updating of the classification boundary or the initial prediction model is stopped; if it does not reach the preset standard, the updating continues until it reaches the preset standard to obtain the final prediction model.

2. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The physical and chemical properties and mechanical properties include: The physical and chemical properties and mechanical properties are obtained by experiments; the physical and chemical properties include porosity, average pore size, density, harmful element content, pH, water absorption, thermal expansion coefficient and thermal conductivity; the mechanical properties include hardness, compressive strength, shear strength, tensile strength, elastic modulus, Poisson's ratio, crushing value of gravel and crushing value of machine-made sand.

3. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The preprocessing of the parent rock information data set includes: Preprocessing the parent rock information data set includes data cleaning, data screening and data conversion of the data in the parent rock information data set.

4. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The mineral composition comprises: The mineral composition includes the lithology of the parent rock and its three major rock types, and the element content of the parent rock; the parent rock is classified according to the lithology of the parent rock to obtain the three major rock types to which the lithology of the parent rock belongs.

5. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The initial type corresponding to each performance data includes: For an initial type of performance data, the number of performance data in each initial type is the same as the number of performance data in other initial types.

6. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The initial prediction model comprises: The initial prediction model includes an input layer, a hidden layer and an output layer; the number of hidden layers is a first preset number, and the number of neurons in the hidden layers is a second preset number.

7. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The prediction accuracy is: Among them, δ represents the prediction accuracy; TP represents true positive examples; FP represents false positive examples; TN represents true negative examples; and FN represents false negative examples.

8. The method for predicting tunnel slag properties based on parent rock mineral composition according to claim 1, characterized in that: The updating of the classification boundary or the initial prediction model according to the prediction accuracy includes: The initial prediction model includes updating the activation function, hidden layer and neurons; when updating the classification boundary, activation function, hidden layer and neurons, only one item is updated at a time; Updating the classification boundary includes: setting a first update strategy, wherein the first update strategy is specifically to increase or decrease the classification boundary under the condition that the amount of data of each category is within a preset range; The activation function update includes: each time the update selects an activation function from the preset activation functions as the activation function of the initial prediction model; The update of the hidden layer includes: adding one hidden layer each time the update is performed, provided that the number of hidden layers does not exceed the set number of layers; The neuron update includes: under the condition that the number of neurons in each hidden layer does not exceed the set number, each update increases or decreases the set number of neurons in each hidden layer.

Images