Monitoring data prediction method, device and system, and storage medium

Abstract

The invention discloses a monitoring data prediction method, device and system, and a storage medium. The method comprises the following steps: S1, acquiring rock core acceleration monitoring data; s2, obtaining a Bayesian convolutional neural network prediction model according to the rock core acceleration monitoring data; and S3, inputting the experimental monitoring data of the engineering component into the Bayesian convolutional neural network prediction model for monitoring data prediction. By adopting the technical scheme of the invention, the prediction capability is greatly improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a monitoring data prediction method, device, system, and storage medium. Background Technology

[0002] Convolutional Neural Networks (CNNs) are a special type of feedforward neural network inspired by biological visual cognition mechanisms. Their local sensitivity and directional selectivity effectively reduce the complexity of the neural network. CNNs are primarily used for data classification, object recognition, and object detection. The prototype of CNNs was the Neocognitron proposed by Kunihiko Fukushima in 1980, which employed a hierarchical structure and local connections to achieve handwritten digit recognition. The modern CNN model originated with the classic LeNet-5 model proposed by Yann LeCun et al. in 1998. A typical CNN consists of five parts: an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. As a special type of feedforward neural network, CNNs reduce the number of model parameters through local connections and shared weights, thereby improving the efficiency of the convolutional layers in extracting features from the input data and capturing the spatial or temporal structural features of the input data. Compared to traditional neural networks, CNNs can automatically extract and learn effective features from data without the need for manually designing complex feature extraction algorithms. At the same time, by sharing the weights of the convolutional kernels, CNNs can significantly reduce the number of parameters in the trained model, thereby reducing computational costs and improving computational efficiency. Furthermore, CNNs also have stronger generalization capabilities.

[0003] However, when convolutional neural network models are used as time series prediction models for monitoring data, operators often struggle to select appropriate hyperparameter configurations. Relying solely on personal experience to adjust model hyperparameters frequently fails to achieve optimal predictive performance. Common hyperparameter setting methods include grid search and random search. Grid search is essentially an exhaustive method, enumerating all possible hyperparameter combinations to form a grid and finding the optimal set. Random search, on the other hand, evaluates a certain number of samples randomly selected from the hyperparameter space to find the optimal combination. Both grid search and random search are computationally intensive, and the parameter evaluations are independent, making it difficult to efficiently find the best parameters. In some cases, the parameter selection may even be less effective than hyperparameter combinations determined by human experience. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a monitoring data prediction method, device, system, and storage medium.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting monitoring data, comprising: Step S1: Obtain core acceleration monitoring data; Step S2: Based on the core acceleration monitoring data, obtain the Bayesian convolutional neural network prediction model; Step S3: Input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

[0006] Preferably, in step S2, the Bayesian optimization algorithm is used to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient.

[0007] Preferably, in step S2, Bayesian optimization uses a cost function as the objective function. The cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network. The network is trained using the hyperparameter combination input through the Bayesian optimization algorithm. The mean squared error (MSE) is calculated using the validation set data as samples as a performance evaluation metric. The MSE calculated by the cost function is: in, For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscript i Indicates the first i One sample.

[0008] The present invention also provides a monitoring data prediction device, comprising: The first processing module is used to acquire core acceleration monitoring data; The second processing module is used to obtain a Bayesian convolutional neural network prediction model based on the core acceleration monitoring data. The third processing module is used to input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

[0009] As a preferred option, the second processing module uses the Bayesian optimization algorithm to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient.

[0010] As a preferred approach, Bayesian optimization uses a cost function as the objective function. The cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network. The network is trained using a combination of hyperparameters input through the Bayesian optimization algorithm. The mean squared error (MSE) is calculated using the validation set data as a performance evaluation metric. The MSE calculated by the cost function is as follows: in, For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscript i Indicates the first i One sample.

[0011] The present invention also provides a monitoring data prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program performs a monitoring data prediction method when executed by the processor.

[0012] The present invention also provides a storage medium storing a computer program, which executes a monitoring data prediction method when running.

[0013] This invention integrates Bayesian optimization algorithms into the theoretical framework of convolutional neural networks (CNNs). It uses Bayesian optimization to optimize three hyperparameters in the CNN model: batch size, initial learning rate, and L2 regularization coefficient. This results in a Bayesian CNN prediction model that combines the excellent temporal and spatial feature extraction capabilities of CNNs with the efficient parameter tuning capabilities of Bayesian optimization, significantly improving its predictive ability. This invention uses the Bayesian CNN prediction model to predict experimental monitoring data for engineering components, achieving a significantly higher accuracy than single prediction models. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 This is a diagram of the CNN network structure. Figure 2 The diagram shows the experimental equipment; (a) is a PCB accelerometer, and (b) is a dual-channel PicoScope oscilloscope. Figure 3 These are experimental field test diagrams; where (a) is the field test diagram of the white specimen and (b) is the field test diagram of the gray specimen. Figure 4 Iterating Bayesian CNN parameters for acceleration testing of white rock cores; Figure 5 Comparison of Bayesian CNN training set results for acceleration testing of white rock cores; Figure 6 Comparison of Bayesian CNN test set results for acceleration testing of white rock cores; Figure 7 Histogram of Bayesian CNN error for acceleration testing of white rock core Figure 8 Fitting a Bayesian CNN training set for acceleration testing of white rock cores Figure 9 Fitting a Bayesian CNN test set for acceleration testing of white rock cores Figure 10 Iteration of Bayesian CNN parameters for acceleration testing of gray rock core group 1; Figure 11 Comparison of Bayesian CNN training set results for acceleration testing of gray rock core group 1; Figure 12 Comparison of Bayesian CNN test set results for acceleration testing of gray rock core group 1; Figure 13 The Bayesian CNN error histogram for acceleration testing of a set of gray rock cores; Figure 14 Fitting a Bayesian CNN training set for acceleration testing of gray rock core group 1; Figure 15 Fitting a Bayesian CNN test set for acceleration testing of gray rock core group 1; Figure 16 Iteration of Bayesian CNN parameters for acceleration testing of two sets of gray rock cores; Figure 17 Comparison of Bayesian CNN training set results for two sets of acceleration tests on gray rock cores; Figure 18 Comparison of Bayesian CNN test set results for two sets of acceleration tests on gray rock cores; Figure 19 Histograms of Bayesian CNN errors from two sets of acceleration tests on gray rock cores; Figure 20 Fitting a Bayesian CNN training set to two sets of acceleration tests in gray rock cores; Figure 21 Fitting a Bayesian CNN test set for two sets of acceleration tests in gray rock cores; Figure 22 This is a flowchart of the monitoring data prediction method according to an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1: like Figure 22 As shown, an embodiment of the present invention provides a monitoring data prediction method, including: Step S1: Obtain core acceleration monitoring data; Step S2: Based on the core acceleration monitoring data, a Bayesian convolutional neural network prediction model is obtained; wherein, the Bayesian optimization algorithm is used to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient. Step S3: Input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

[0019] As one embodiment of the present invention, a convolutional neural network (CNN) consists of five parts: an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The convolutional layer is the most crucial component of a CNN. The convolution operation involves using convolutional kernels, also known as filters, to scan various local regions of the input data, thereby extracting features. Each convolutional kernel can recognize a certain local pattern (e.g., edges, textures) in the external input. The output of a convolutional layer is one or more feature maps. For two-dimensional convolution, given an input... I and convolution kernel K The formula for convolution operation is: In the formula, This refers to the horizontal and vertical axes of the time series monitoring data input being respectively x and y point, This refers to the coordinates in the feature map output by the convolutional layer. The value of , in time series forecasting, refers to the model's performance at time t.t The given predicted value. This refers to the monitoring data observed at the lag time (tk) in time series forecasting; It is the position in the convolution kernel. The value of represents the weight of the "historical information to future value" mapping learned by the model in time series prediction, corresponding to the filter parameters of a CNN.

[0020] Pooling layers are primarily used for downsampling, reducing the spatial dimensionality of data, decreasing computational complexity, and preventing overfitting. Common pooling methods are max pooling and average pooling. Pooling operations select an optimal value (maximum or average) from each region using a sliding window. Max pooling retains the maximum value within each local region, essentially preserving the best match for that region. Max pooling doesn't focus on which specific location within the window matches, but rather on whether a match has occurred. For a 2×2 pooling region, max pooling selects the maximum value: In the formula, MP This refers to the maximum pooling value; It refers to all pixels within the region; max This refers to taking the maximum value.

[0021] Activation layers are mostly connected after convolutional and pooling layers. Common activation functions include ReLU (Rectified Linear Unit), Sigmoid, and Tanh. ReLU is commonly used in convolutional layers, while Sigmoid and Tanh are more often used in fully connected layers. The purpose of activation functions is to perform non-linear transformations on the output of convolutional layers. The ReLU function, or Rectified Linear Unit, outputs 0 if the input is negative, and the output remains unchanged if the input is positive. During the training of convolutional neural networks, it is common to encounter a situation where the training set loss function is small and the prediction accuracy is high, while the test set loss function is large and the prediction accuracy is low, indicating that the neural network model has overfitted. The addition of a Dropout layer can alleviate this overfitting problem to some extent. Dropout temporarily removes some neural network units from hidden layer nodes by randomly and equally probably removing them. This reduces the complex interactions between hidden nodes and avoids the phenomenon that some features extracted by the neural network have overly stringent exposure conditions, thus improving the model's robustness in prediction.

[0022] After features are extracted from convolutional and pooling layers and transformed nonlinearly by activation functions, these features are usually mapped to the final output space through fully connected layers for recognition and classification.

[0023] CNN models typically use loss functions and backpropagation to evaluate and optimize model prediction performance. The loss function reflects the error between the neural network's predicted values ​​and the true values. A smaller loss function indicates more accurate predictions from the network model. Backpropagation, first proposed by Rumelhart et al. in 1986, provides an effective optimization method for training neural networks. Based on gradient descent, backpropagation calculates the gradient of the loss function, transmits this gradient to the optimization function, and then updates the network weights to minimize the loss function. Common optimization algorithms include SGD (Stochastic Gradient Descent) and Adam. This invention employs the Adam gradient descent algorithm.

[0024] The Bayesian optimization algorithm (BO) is a global optimization algorithm based on the probabilistic Bayesian principle. The Bayesian theorem formula is: In the formula: f It is a black-box function with hyperparameters; t The total number of observation sets; The set of observed points; Let it be the decision vector; These are the observed values; This is the observation error; for y The likelihood function distribution; For the objective function f The prior function distribution; for f The marginal likelihood function distribution is used to optimize variable parameters; for f The posterior function distribution represents the confidence level of the unknown objective function after correcting the prior function distribution.

[0025] Bayesian optimization algorithm uses Gaussian process regression to automatically select an initial combination of hyperparameters for objective function evaluation, constructs a spatial probability model of hyperparameters, and selects the next set of hyperparameters for performance evaluation based on this model. After each iterative evaluation, the model is updated, and this iterative process is repeated to obtain the optimal combination of hyperparameters. The optimal values ​​for Bayesian optimization hyperparameter tuning are... In the formula: The optimal set of hyperparameters; z For hyperparameter combinations; Z For the experimental sample set; The target model for optimization.

[0026] This Bayesian optimization uses a cost function as its objective function. In terms of neural network parameter tuning, the cost function objective function is more efficient than Bayesian optimization algorithms based on grid search and random search. The cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network, using the hyperparameter combination input through the Bayesian optimization algorithm for network training. The mean squared error (MSE) is calculated using the validation set data as samples as a performance evaluation metric. The specific form of the cost function's MSE calculation is as follows: In the formula: For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscript i Indicates the first i One sample.

[0027] Bayesian optimization algorithms evaluate the cost function through multiple iterations, and the combination of hyperparameters that minimizes the mean squared error (MSE) in the validation set is the optimal model hyperparameter configuration. The cost function is more efficient than Bayesian optimization algorithms based on grid search and random search.

[0028] This invention first divides the original monitoring dataset into training and testing sets, and then analyzes and preprocesses the data. Parameters for cross-data prediction, such as the delay step size, are set. After setting, the data is normalized and flattened into one-dimensional data. Note that flattening the data into one dimension is a simplified method; it can also be flattened into two or three dimensions.

[0029] The Bayesian parameter optimization of this model uses a cost function as the objective function. A validation set is formed by selecting 20% ​​of the dataset for Bayesian iterative evaluation of hyperparameter performance. This model uses Bayesian optimization to adjust three hyperparameters: batch size, initial learning rate, and L2 regularization coefficient. The range of Bayesian tuning values ​​for batch size, initial learning rate, and L2 regularization coefficient is set, and the maximum number of Bayesian optimization iterations is selected.

[0030] like Figure 1 As shown, the CNN network structure of this invention sets the input layer size to [25, 1, 1], with a total of three convolutional layers, each with a kernel size of 3×1, and a ReLU activation layer after each convolutional layer; two pooling layers are set, each with a pooling size of 2×1 and a stride of 2; a Dropout layer is set after each convolutional pooling, for a total of three layers, using a dropout rate of 0.2 to prevent overfitting; finally, a fully connected layer and a regression output layer are set to output the training results.

[0031] After feature extraction via convolutional operations and training, performance evaluation metrics are output. To accurately quantify the performance and accuracy of the Bayesian convolutional neural network model, this invention selects four values ​​as model performance evaluation metrics: coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error. The coefficient of determination, root mean square error, mean absolute error, and mean absolute percentage error are respectively shown in the following formulas. In the formula, The actual value; This is a predicted value; n Indicates the number of predicted data points; R 2 The coefficient of determination is the highest accuracy of the model in predicting monitoring data. The closer the coefficient of determination is to 1, the higher the accuracy of the model in predicting monitoring data. R MSE This represents the root mean square error. M AE Indicates the mean absolute error. M APE The mean absolute percentage error represents the average absolute error. The smaller the root mean square error, mean absolute error, and mean absolute percentage error, the higher the accuracy of the model's prediction of the monitoring data.

[0032] Example: This invention uses the raw acceleration signal obtained from a rock core acceleration signal test experiment as a monitoring data prediction example based on a Bayesian convolutional neural network algorithm. The test employs the traditional impact echo method, which uses a PCB352C65 accelerometer sensor, such as... Figure 2 As shown in (a), the signal was acquired and stored using PicoScope 7 data acquisition software developed by PicoTech. This software can also display the waveform and spectrum of the signal in real time and provides multiple trigger modes to meet different testing needs. In the experiment, a PicoScope 2142 dual-channel PC-type digital oscilloscope was used to transmit the acquired signal to the PC. A PC oscilloscope consists of oscilloscope hardware and an oscilloscope program running on the PC. Essentially, a PC oscilloscope retains the core acquisition and processing parts of the oscilloscope, while the display, settings, and archiving are handled by the PC. This not only allows the Pico oscilloscope to perform all the functions of a traditional oscilloscope, but also makes it smaller and more portable. This device was manufactured by PicoTech in the UK and used with PicoScope 7 data acquisition software, such as... Figure 2 As shown in -(b).

[0033] Collection site Figure 3 As shown, firstly, an accelerometer was attached to the top of the core specimen using a coupling agent to facilitate signal acquisition and transmission. Then, the side of the core specimen was struck with a hammer to generate an acceleration signal. After the test began, the PicoScope 7 acquisition program was used to acquire the signal simultaneously with the hammering.

[0034] The complete analysis process of using a Bayesian optimized convolutional neural network to predict time series data from core acceleration data can be divided into the following steps. The first step is data preparation, which involves constructing supervised samples using a sliding window function, with the window length set as follows. L =30, with a stride of 1, meaning the 31st point is predicted every 30 historical data points, forming an N×30×1 tensor input. Secondly, to focus only on hyperparameters rather than structure during Bayesian optimization, a cost function is used as the objective function to lock the one-dimensional CNN: three convolutional layers followed by ReLU activation layers, each convolutional layer followed by a pooling layer for downsampling, then Dropout to suppress overfitting, and finally a fully connected layer to output the single-step prediction value; the entire model uses Adam as the optimizer and MSE as the loss function. Next, Bayesian optimization sets the network's mini-batch size, initial learning rate, and L2 regularization coefficient as the parameters to be optimized; the objective function uses 20% of the training data as the validation set. Bayesian optimization automatically selects hyperparameter combinations for objective function evaluation and selects the next set of hyperparameters for performance evaluation based on the evaluation results of this set of objective functions. After each iteration, the model is updated, and this iterative process is repeated 20 times to obtain the optimal hyperparameter combination.

[0035] After Bayesian optimization converges iteratively, the globally optimal hyperparameter combination is extracted, and the CNN is retrained using the complete training set, with the training process visualized in real time. Convolutional layers scan various local regions of the input core data to extract its temporal-spatial features. The data is then batch-normalized using ReLU activation. Next, pooling is performed, selecting an optimal value from each region using a sliding window, reducing the spatial dimension and computational cost. This is treated as a unit of the CNN, repeated three times. Dropout layers are added between each pooling and convolutional layer to prevent overfitting during training. After CNN training, the saved mapping parameters are read in, and the network output is inversely transformed to obtain the predicted future time values ​​in real physical dimensions, and R is calculated. MSE M AE M APE Evaluation metrics such as R² are used. Simultaneously, after training and prediction are completed, the following are plotted: Bayesian parameter iteration graph, comparison graph of prediction results between the training and test sets, error histogram, fitting effect graph of the training and test sets, 95% confidence interval graph, and comparison graph of prediction effects of different models.

[0036] (1) Acceleration prediction results of white core group: The Bayesian convolutional neural network monitoring data prediction model used for white rock core acceleration data prediction divides the first 80% of the original data into a training set and the last 20% into a test set. A delay step of 30 is set, and predictions are performed across one time point. That is, a data point is predicted from the first 30 data points as the 31st data point, and then data from the next time point is used for the next round of training and prediction. The data predicted in the previous round is used as the last data point in the current round of training and prediction. The dropout ratio is 0.4 for the first layer, 0.3 for the second layer, and 0.2 for the third layer.

[0037] The batch size was adjusted to [20, 64], the initial learning rate to [0.001, 0.1], and the L2 regularization coefficient to [0.0001, 0.005]. The maximum number of iterations was set to 20 to select the optimal hyperparameter configuration. Other parameters were manually set as follows: using the Adam gradient descent algorithm, the maximum number of training iterations was set to 500, the learning rate descent period was set to 300, and the learning rate descent factor was set to 0.2.

[0038] After 20 iterations of Bayesian hyperparameter optimization, such as Figure 4As shown, the optimal batch size for acceleration prediction from the white core No. 2 was found to be 61, the optimal initial learning rate was 0.0049309, and the optimal L2 regularization coefficient was 0.0048384. This hyperparameter combination was then input into a CNN network model for use.

[0039] After network training is complete, the comparison results of the actual values ​​and predicted values ​​on the training and test sets are output separately. The training set determination coefficient R0 is also shown. 2 It is 0.94671, such as Figure 5 As shown; the coefficient of determination R for the test set 2 It is 0.87447, such as Figure 6 As shown, the mean squared error (MSE) of the training set is 3.3788. 10 -4 The root mean square error (RMSE) was 0.018381, and the residual prediction deviation (RPD) was 4.3343; the mean square error of the test set (MSE) was 3.9599. 10 -6 The root mean square error (RMSE) was 0.0019899, ​​and the residual prediction error (RPD) was 3.0443. The model's predicted data generally matched the actual values, with significant errors only in the prediction of peak data.

[0040] The model performed statistical analysis on the prediction error of the test set and presented the results in the form of an error histogram, such as... Figure 7 As shown in the figure, the prediction errors of the model on the test set are relatively dense and high in the middle area of ​​the histogram, which means that the errors of most prediction samples are concentrated in a relatively small range. The largest error is only 0.000744, which reflects that the model can predict the monitoring data relatively accurately in most cases and has good stability.

[0041] Visual analysis of the comparison and fitting results between the training and test sets is performed separately, such as... Figure 8 and Figure 9 As shown, the mean absolute error (MAE) of the training set is 0.01103, while the MAE of the test set is 0.0016, with an overall accuracy of 91.293%. From the fitting results, the data distribution exhibits a clear linear fit, indicating that the prediction results of this Bayesian convolutional neural network-based monitoring data modeling and prediction method are highly consistent with the true values, and its prediction performance has good stability.

[0042] (2) Acceleration prediction results of the first group of gray rock cores: The Bayesian convolutional neural network monitoring data prediction model used for the first set of gray core acceleration data divided the first 75% of the original data into a training set and the last 25% into a test set. The delay step size was set to 25, and predictions were performed across one time point. The first layer dropout ratio was set to 0.3, and the subsequent two layers' dropout ratios were both set to 0.2. The batch size adjustment range was set to [20, 64], the initial learning rate adjustment range to [0.001, 0.1], and the L2 regularization coefficient adjustment range to [0.00001, 0.001]. The maximum number of iterations was 20 to select the optimal hyperparameter configuration. Other parameters were manually set as follows: the Adam gradient descent algorithm was used, the maximum number of training iterations was set to 500, the learning rate descent period was set to 300, and the learning rate descent factor was set to 0.2.

[0043] After 20 iterations of Bayesian hyperparameter optimization, such as Figure 10 As shown, the optimal batch size for acceleration prediction from the gray core No. 2 was found to be 47, the optimal initial learning rate was 0.017445, and the optimal L2 regularization coefficient was 1.2039. 10 -5 This hyperparameter combination is then input into the CNN network model for use.

[0044] The first group of network training for the gray rock cores has been completed, and the comparison results of the actual values ​​and predicted values ​​on the training and test sets are output respectively. Among them, the coefficient of determination R of the training set is... 2 It is 0.95492, such as Figure 11 As shown; the coefficient of determination R for the test set 2 It is 0.84065, such as Figure 12 As shown, the mean squared error (MSE) of the training set is 8.3481. 10 -7 The root mean square error (RMSE) was 0.00091368, and the residual prediction deviation (RPD) was 4.7302; the mean square error of the test set (MSE) was 6.203. 10 -7 The root mean square error (RMSE) was 0.00078759, and the residual prediction deviation (RPD) was 2.5181. The overall trend of the predicted data is accurate and stable, with small deviations in peak data, indicating excellent prediction performance.

[0045] The model performed statistical analysis on the prediction error on the test set, and the results are presented in the form of an error histogram, such as... Figure 13As shown in the figure, the model's prediction errors on the test set roughly conform to a normal distribution. The absolute value of the largest error is only 0.00032, which is smaller than the first group, indicating a smaller overall prediction error. The error distribution is mostly concentrated around the zero-error baseline. As the absolute value of the error increases, the number of samples decreases rapidly, indicating that the model has relatively few prediction samples with large errors, thus controlling the generation of extreme errors to a certain extent and demonstrating good robustness.

[0046] Visual analysis of the comparison and fitting results between the training and test sets is performed separately, such as... Figure 14 and Figure 15 As shown, the mean absolute error (MAE) of the training set is 0.0005964, and the MAE of the test set is 0.0006486, with an overall accuracy of 89.78%. From the fitting results, the data distribution exhibits a clear linear fit, without any obvious prediction bias or distortion. This indicates that the Bayesian convolutional neural network's monitoring data modeling and prediction method learns the data features sufficiently, has high training accuracy, and possesses good universality and stability.

[0047] (3) Acceleration prediction results of the second group of gray rock cores: The Bayesian convolutional neural network monitoring data prediction model used for the second set of gray core acceleration data divided the first 75% of the original data into a training set and the last 25% into a test set. The delay step size was set to 25, and predictions were performed across one time point. The dropout ratio for each layer was set to 0.2. The batch size adjustment range was set to [16, 64], the initial learning rate adjustment range to [0.0001, 0.05], and the L2 regularization coefficient adjustment range to [0.001, 0.05]. The maximum number of iterations was 20 to select the optimal hyperparameter configuration. Other parameters were manually set as follows: the Adam gradient descent algorithm was used, the maximum number of training iterations was set to 500, the learning rate descent period was set to 300, and the learning rate descent factor was set to 0.2.

[0048] After 20 iterations of Bayesian hyperparameter optimization, such as Figure 16 As shown, the optimal batch size for acceleration prediction from the gray core No. 2 was found to be 56, the optimal initial learning rate was 0.020197, and the optimal L2 regularization coefficient was 0.0010803. This hyperparameter combination was then input into a CNN network model for use.

[0049] The third group of network training is complete, and the comparison results of the actual values ​​and predicted values ​​on the training set and test set are output respectively. Among them, the coefficient of determination R of the training set is... 2 It is 0.96455, such as Figure 17 As shown; the coefficient of determination R for the test set 2 It is 0.88613, such as Figure 18 As shown, the mean squared error (MSE) of the training set is 5.2774. 10 -4 The root mean square error (RMSE) was 0.022973, and the residual prediction deviation (RPD) was 5.8154; the mean square error of the test set (MSE) was 1.6873. 10 -6 The root mean square error (RMSE) was 0.001299, and the residual prediction deviation (RPD) was 2.9635. The overall trend of the predicted data is accurate and stable, with small deviations in peak data, indicating excellent prediction performance.

[0050] The model performed statistical analysis on the prediction error on the test set, and the results are presented in the form of an error histogram, such as... Figure 19 As shown in the figure, the prediction error of the model on the test set roughly conforms to the characteristics of a normal distribution, with the largest error being only 0.000313. Compared to the first group, the overall prediction error is smaller. The error distribution is mostly concentrated around the zero error standard line, and as the absolute value of the error increases, the number of samples decreases rapidly, indicating that the number of prediction samples with large errors is relatively small, which to some extent controls the generation of extreme errors and demonstrates good robustness.

[0051] Visual analysis of the comparison and fitting results between the training and test sets is performed separately, such as... Figure 20 and Figure 21 As shown, the mean absolute error (MAE) is 0.0010625, and the overall accuracy reaches 92.534%. From the fitting results, the data distribution exhibits a clear linear fit, without any obvious prediction shift or distortion. This indicates that the Bayesian convolutional neural network's monitoring data modeling and prediction method learns the data features sufficiently, has high training accuracy, and possesses good universality and stability.

[0052] Monitoring data is crucial for analyzing and predicting the future trends of parameters such as strain and acceleration of engineering structures over time. To address the issues of low accuracy and slow training speed in prediction models, this invention proposes a monitoring data modeling and prediction method based on Bayesian convolutional neural networks (CNNs) for time series prediction of monitoring data from engineering facilities. First, the hyperparameters (batch size, initial learning rate, and L2 regularization coefficient) in the CNN are iteratively adjusted and optimized using a Bayesian optimization algorithm. Then, the optimized hyperparameter combination is input into the CNN for time series feature extraction and prediction. This model can automatically adjust the optimal hyperparameter configuration within a selected range and possesses powerful data feature extraction capabilities. Using engineering strain monitoring data as an example, this invention compares the prediction performance of this hybrid model with other models. Experimental results show that the model achieves a prediction accuracy of 90.13%, demonstrating good prediction performance.

[0053] Example 2: This invention also provides a monitoring data prediction device, comprising: The first processing module is used to acquire core acceleration monitoring data; The second processing module is used to obtain a Bayesian convolutional neural network prediction model based on the core acceleration monitoring data. The third processing module is used to input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

[0054] As one embodiment of the present invention, the second processing module uses the Bayesian optimization algorithm to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient.

[0055] In one embodiment of the present invention, Bayesian optimization uses a cost function as the objective function. The cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network. The network is trained using a combination of hyperparameters input through the Bayesian optimization algorithm. The mean squared error (MSE) is calculated using the validation set data as samples as a performance evaluation metric. The MSE calculated by the cost function is as follows: in, For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscripti Indicates the first i One sample.

[0056] Example 3: This invention also provides a monitoring data prediction system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a monitoring data prediction method when run by the processor.

[0057] Example 4: This invention also provides a storage medium storing a computer program that executes a monitoring data prediction method during runtime.

[0058] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting monitoring data, characterized in that, include: Step S1: Obtain core acceleration monitoring data; Step S2: Based on the core acceleration monitoring data, obtain the Bayesian convolutional neural network prediction model; Step S3: Input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

2. The monitoring data prediction method as described in claim 1, characterized in that, In step S2, the Bayesian optimization algorithm is used to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient.

3. The monitoring data prediction method as described in claim 2, characterized in that, In step S2, Bayesian optimization uses a cost function as the objective function. The cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network. The network is trained using the hyperparameter combination input through the Bayesian optimization algorithm. The mean squared error (MSE) is calculated using the validation set data as samples as a performance evaluation metric. The MSE calculated by the cost function is: in, For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscript i Indicates the first i One sample.

4. A monitoring data prediction device, characterized in that, include: The first processing module is used to acquire core acceleration monitoring data; The second processing module is used to obtain a Bayesian convolutional neural network prediction model based on the core acceleration monitoring data. The third processing module is used to input the experimental monitoring data of the engineering components into the Bayesian convolutional neural network prediction model to predict the monitoring data.

5. The monitoring data prediction device as described in claim 4, characterized in that, The second processing module uses the Bayesian optimization algorithm to optimize the three hyperparameters in the convolutional neural network model: batch size, initial learning rate, and L2 regularization coefficient.

6. The monitoring data prediction device as described in claim 5, characterized in that, Bayesian optimization uses the Cost function as the objective function. The Cost function allocates 20% of the training data as a validation set and constructs a network structure model identical to the main convolutional neural network. The network is trained using a combination of hyperparameters input through the Bayesian optimization algorithm. The mean squared error (MSE) is calculated using the validation set data as samples as a performance evaluation metric. The MSE calculated by the Cost function is: in, For hyperparameters; m The number of samples used for training on the validation set; For using hyperparameters and x The predicted value obtained from the prediction; y The true value from the original training samples; superscript i Indicates the first i One sample.

7. A monitoring data prediction system, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the monitoring data prediction method as described in any one of claims 1-3 when executed by the processor.

8. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, performs the monitoring data prediction method as described in any one of claims 1-3.

Images