Gravel pile residual vibration data monitoring method for sea reclamation operation
By establishing a data acquisition and control system and introducing time series prediction methods, the accuracy problem of traditional crushed stone pile vibration data monitoring was solved, realizing automatic evaluation and real-time optimization of crushed stone pile construction quality, and improving construction safety and efficiency.
Patent Information
- Application Number
- CN202511110871.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-21
AI Technical Summary
Traditional monitoring of vibration data from crushed stone piles relies on manual methods, resulting in poor data accuracy and reliability. There is an urgent need to introduce intelligent algorithms to achieve automatic monitoring and prediction, thereby ensuring the safety and construction quality of land reclamation operations.
A data acquisition and control system was established to record vibration data in real time and introduce time series prediction methods. Quality assessment was performed through radial basis function neural networks. Data augmentation and kernel synthesis methods were combined to expand the data volume, train the time series prediction model, and generate construction quality assessment results.
It improves the accuracy and efficiency of vibration data monitoring, ensures construction quality, enables real-time optimization of the construction process and fault early warning, and guarantees the safety and efficiency of construction.
Smart Images
Figure CN120995209A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of engineering construction data monitoring technology, and in particular to a method for monitoring the vibration data of crushed stone piles used in land reclamation operations. Background Technology
[0002] During the construction of land reclamation projects, the resulting foundations are mostly deep, soft soil layers containing a large amount of silt, which have poor bearing capacity and stability. Therefore, vibro-compaction stone pile technology is often used to replace part of the soil in the soft soil layer with dense stone piles, forming a composite foundation with the soil between the piles. This significantly improves the bearing capacity of the foundation, enabling it to withstand the loads of subsequent buildings and facilities.
[0003] Vibro-compaction stone pile technology utilizes the vibration of a vibro-compactor and the action of high-pressure water flow to form holes in soft soil foundations. Crushed stone is then filled into the holes, and the vibratory compaction action of the vibro-compactor densifies the stone, forming a stone pile. Monitoring the vibration data of the stone piles during the pile formation process is a crucial aspect of construction quality control. By monitoring this data, it is ensured that the vibration data of the vibro-compactor at each construction depth meets design requirements, thereby ensuring sufficient compaction of the crushed stone and forming a high-quality pile. Based on the monitored vibration data, combined with other construction parameters, problems encountered during construction can be analyzed, allowing for timely adjustments to construction techniques and parameters, optimizing the construction process, and improving construction efficiency.
[0004] Traditional vibration monitoring of crushed stone piles primarily relies on manual methods. Construction workers record vibration data at the land reclamation site using data acquisition equipment and judge whether the vibration is normal based on experience. This method depends heavily on the experience and skill level of the construction workers, and the accuracy and reliability of the data are relatively poor. There is an urgent need to improve the detection methods by introducing intelligent algorithms into the data acquisition and control system to achieve automatic monitoring of vibration data, predict future trends, and promptly send fault warning information to ensure the safety of land reclamation operations. Summary of the Invention
[0005] To address the technical problems of the prior art, this application provides a method for monitoring vibration data of vibratory crushed stone piles used in land reclamation operations. This method collects vibration data of vibratory crushed stone piles in real time during land reclamation operations, introduces a time series prediction method to predict the future trend of the vibration data, evaluates the construction quality of the crushed stone piles based on the prediction and the real-time collected vibration data, and optimizes construction parameters based on the evaluation results, thereby improving the accuracy and efficiency of vibration data monitoring for crushed stone piles.
[0006] This application provides a method for monitoring vibration data of crushed stone piles used in land reclamation operations, including:
[0007] (1) Establish a data acquisition and control system and use data acquisition equipment to record the vibration data of vibratory crushed stone piles in real time during the land reclamation operation;
[0008] (2) Based on the historical vibration data of crushed stone piles stored in the data acquisition and control system, the data volume is expanded by data augmentation method, and the time series prediction model is trained using the enhanced vibration data of crushed stone piles.
[0009] (3) The real-time collected vibration data of crushed stone piles is preprocessed, input into the pre-trained time series prediction model, and the prediction results of the vibration data of crushed stone piles are output.
[0010] (4) A radial basis function neural network is introduced to process the predicted and real-time collected vibration data of the crushed stone piles, automatically generate the quality assessment results of the crushed stone piles, and transmit them to the construction personnel through the data acquisition and control system.
[0011] Furthermore, the data acquisition and control system includes a sensor module, a data transmission module, a data storage module, a data analysis and processing module, and a user interface.
[0012] The sensor module is equipped with data acquisition devices including current sensors, frequency sensors, depth sensors, and vibration time sensors to collect vibration data of the vibratory crushed stone pile in real time, including compaction current, vibration frequency, vibration depth, and vibration time.
[0013] The data transmission module supports wireless communication, transmitting the collected vibration data of the crushed stone piles to other modules in the data acquisition and control system via a communication protocol.
[0014] A database is established in the data storage module to store the vibration data of the crushed stone piles, and a timestamp is assigned to each type of data;
[0015] The data processing and analysis module embeds an intelligent algorithm model to predict the vibration data of crushed stone piles and evaluate the quality of crushed stone piles;
[0016] The user interface is used to display the quality assessment results of the data processing and analysis module and send relevant information to construction workers involved in land reclamation operations.
[0017] Furthermore, methods for data enhancement of historical crushed stone pile vibration data stored in the data storage module of the data acquisition and control system include time series mixing and kernel synthesis.
[0018] Before data augmentation, historical crushed stone pile vibration data is processed into time series data, with each type of crushed stone pile vibration data containing continuous time information and aligned in time.
[0019] The method for data augmentation based on time series mixing is as follows: Extract data records x of the same sequence length from the time series data of each type of crushed stone pile vibration data. i and x j The starting point for data extraction is randomly generated. By weighting and merging two data records, a new time series data is calculated.
[0020]
[0021] Where λ∈[0,1] and conforms to the Beta distribution; for each type of crushed stone pile vibration data, a certain number of new time series data are synthesized by time series mixing.
[0022] Kernel synthesis uses Gaussian processes to synthesize time series. A Gaussian process is a probabilistic model that captures patterns and structures in time series data through kernel functions. By randomly combining different kernel functions, kernel synthesis generates time series with different patterns, enriching data complexity and further improving the model's generalization ability.
[0023] The process of generating time series data of crushed stone pile vibration data by kernel synthesis includes: randomly selecting several kernel functions from the kernel function library, the kernel function types being linear kernels and periodic kernels; combining these kernel functions using binary operators, and using the combined kernel function to define the prior distribution of the Gaussian process; and sampling from the prior distribution of the Gaussian process to generate the synthesized time series.
[0024] Furthermore, the time-series data of the original and enhanced crushed stone pile vibration data are used as the dataset for the time-series prediction model, and the samples in the dataset are quantized to achieve the classification-based time-series prediction task:
[0025] Quantization is achieved through a discretization mapping function, which processes the continuous vibration data of crushed stone piles into categorical variables. Based on the specific value of the vibration data at each time point, the interval is used as its category label. The difference between the vibration data value at each time point and the minimum value of the interval is calculated as an additional label. The label value of each time series sample of crushed stone pile vibration data includes the specific value, the interval difference, and the category label.
[0026] The time series prediction model employs a hierarchical classification auxiliary network. Based on a hierarchical structure, the classifier is trained at two levels: fine-grained and coarse-grained. Each level divides the range of values in the oscillation data into a different number of categories, thereby obtaining high-entropy features with multi-granular representation. Specifically, fine-grained features contain a hierarchical structure with more categories, providing the model with accurate quantitative information; coarse-grained features correspond to a hierarchical structure with fewer categories, used to improve classification accuracy.
[0027] The model employs an Informer-based backbone network to extract features, which are then mapped to multi-granularity and temporal features via linear layers. Uncertainty-aware classifiers are trained at both coarse-grained and fine-grained levels using class labels. Backpropagation optimizes the multi-granularity features to improve classification performance. Temporal features are used to capture temporal information for time series prediction. The model maintains consistency between coarse-grained and fine-grained levels through hierarchical consistency loss, mitigating boundary effects.
[0028] Furthermore, by collecting and analyzing historical crushed stone pile construction records, a mapping relationship between crushed stone pile vibration data and quality status is established, classifying quality status into "qualified" and "unqualified" categories. A pre-trained model is obtained by training a radial basis function neural network using historical crushed stone pile vibration data and quality category labels. This model is then used to process real-time collected and predicted crushed stone pile vibration data, outputting quality assessment results.
[0029] Radial basis function neural networks (RBNs) are a type of feedforward neural network that consists of an input layer, hidden layers, and an output layer. They generate features by calculating the distance between the input data and the center point using the radial basis functions of the hidden layer neurons. A Gaussian function is used as the radial basis function K(x), expressed by the formula:
[0030]
[0031] Where x is the input data of the hidden layer, c is the center point of the radial basis function, and σ is the variance; the center point is calculated from the training data using a clustering algorithm, and the variance is calculated by the following formula:
[0032]
[0033] Where h is the number of training data, c max This represents the maximum distance between the center points. The weights from the hidden layer to the output layer of the radial basis function neural network are calculated using the least squares method. Combined with the kernel function from the input layer to the hidden layer and the weights from the hidden layer to the output layer, the entire network is trained. The training process of radial basis function neural networks is simpler and more computationally efficient than that of deep neural networks, enabling rapid processing of vibration data from crushed stone piles and real-time monitoring of land reclamation operations.
[0034] The present invention discloses the following technical effects:
[0035] This invention provides a method for monitoring vibration data of crushed stone piles used in land reclamation operations. Based on a data acquisition and control system, it introduces intelligent algorithms to predict the vibration data of crushed stone piles and automatically evaluate the construction quality of the piles. The invention establishes a data acquisition and control system that collects vibration data of crushed stone piles in real time through data acquisition equipment and constructs it into time series data, recording multiple vibration data points for each time stamp. Considering the limited data volume, data augmentation methods, time series mixing, and kernel synthesis are applied to synthesize a time series to expand the data volume of crushed stone pile vibration data. This data is used to train a time series prediction model, enabling the model to generalize to predict different time series patterns, thereby improving the model's robustness and prediction accuracy. This invention applies a time series prediction model combined with classification and regression tasks to predict the time series of vibration data from crushed stone piles. Based on the range of vibration data values, the data is quantified into different categories. An uncertainty-aware classifier is trained at both fine-grained and coarse-grained levels, and the hierarchical consistency loss between the two levels of features is calculated. The time series prediction model is then trained using the mean squared error loss commonly used in time series prediction. The model employs a multi-level classification structure to extract multi-granular features from the crushed stone pile vibration time series data, obtaining more accurate time-series information and improving prediction accuracy. Furthermore, based on the predicted and real-time collected crushed stone pile vibration data, a radial basis function neural network is introduced to automatically evaluate the quality of the vibration data, and the quality evaluation report is promptly delivered to construction personnel. This neural network is fast, simple in structure, and can quickly process data, ensuring the smooth progress of land reclamation operations. Attached Figure Description
[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.
[0037] Figure 1 This is a flowchart illustrating a method for monitoring vibration data of crushed stone piles used in land reclamation, as provided in an embodiment of this application.
[0038] Figure 2 This is a schematic diagram of the structure of the time series prediction model provided in the embodiments of this application.
[0039] Figure 3 This is a schematic diagram of the backbone network structure in the time series prediction model provided in the embodiments of this application. Detailed Implementation
[0040] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.
[0041] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0042] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.
[0043] Example 1: This application provides a method for monitoring vibration data of crushed stone piles used in land reclamation operations, such as... Figure 1 As shown, the method includes:
[0044] Step S10: Establish a data acquisition and control system and use data acquisition equipment to record the vibration data of vibratory crushed stone piles during the land reclamation operation in real time.
[0045] In this embodiment, the data acquisition and control system includes a sensor module, a data transmission module, a data storage module, a data analysis and processing module, and a user interface.
[0046] The sensor module is equipped with data acquisition devices including current sensors, frequency sensors, depth sensors, and vibration time sensors to collect vibration data of the vibratory crushed stone pile in real time, including compaction current, vibration frequency, vibration depth, and vibration time.
[0047] The data transmission module supports wireless communication, transmitting the collected vibration data of the crushed stone piles to other modules in the data acquisition and control system via a communication protocol.
[0048] A database is established in the data storage module to store the vibration data of the crushed stone piles, and a timestamp is assigned to each type of data;
[0049] The data processing and analysis module embeds an intelligent algorithm model to predict the vibration data of crushed stone piles and evaluate the quality of crushed stone piles;
[0050] The user interface is used to display the quality assessment results of the data processing and analysis module and send relevant information to construction workers involved in land reclamation operations.
[0051] Step S20: Based on the historical vibration data of crushed stone piles stored in the data acquisition and control system, the data volume is expanded using data augmentation methods, and the time series prediction model is trained using the augmented vibration data of crushed stone piles.
[0052] In this embodiment, the method for data enhancement of historical crushed stone pile vibration data stored in the data storage module of the data acquisition and control system includes time series mixing and kernel synthesis.
[0053] Before data augmentation, historical crushed stone pile vibration data is processed into time series data, with each type of crushed stone pile vibration data containing continuous time information and aligned in time.
[0054] The method for data augmentation based on time series mixing is as follows: Extract data records x of the same sequence length from the time series data of each type of crushed stone pile vibration data. i and x j The starting point for data extraction is randomly generated. By weighting and merging two data records, a new time series data is calculated.
[0055]
[0056] Where λ∈[0,1] and conforms to the Beta distribution; for each type of crushed stone pile vibration data, 200 new time series data are synthesized by time series mixing.
[0057] The process of generating time series data of crushed stone pile vibration data using kernel synthesis includes: randomly selecting several kernel functions from a kernel function library, with the kernel function types being linear kernels and periodic kernels.
[0058] The linear kernel function is expressed by the formula: k l (t,t′)=α(t-∈)(t′-∈), where α is the amplitude parameter of the linear kernel function, t and t′ represent two different timestamps, and ∈ is the offset;
[0059] The periodic kernel function can be expressed by the following formula:
[0060]
[0061] Where β is the amplitude parameter of the periodic kernel function, γ is the length scale parameter, and p is the period.
[0062] These kernel functions are combined using binary operators, and the combined kernel function is used to define the prior distribution of the Gaussian process. The combined composite kernel function is expressed by the formula: k c (t,t′)=k l (t,t′)+μk p (t,t′), where μ represents the weighting coefficient, and the composite kernel function is defined as the variance of the Gaussian process, with the mean of the Gaussian process being 0.
[0063] A synthetic time series of length l is generated by sampling from the prior distribution of a Gaussian process. It is a vector representing the synthetic data at a given timestamp.
[0064] The time series data of the original and enhanced crushed stone pile vibration data are used as the dataset for the time series prediction model. The samples in the dataset are quantized to achieve the classification-based time series prediction task.
[0065] Quantization is achieved through a discretization mapping function, which processes the continuous vibration data of crushed stone piles into categorical variables. Based on the specific value of the vibration data at each time point, the interval is used as its category label. The difference between the vibration data value at each time point and the minimum value of the interval is calculated as an additional label. The label value of each time series sample of crushed stone pile vibration data includes the specific value, the interval difference, and the category label.
[0066] Step S30: Preprocess the real-time collected vibration data of the crushed stone piles, input it into the pre-trained time series prediction model, and output the prediction results of the vibration data of the crushed stone piles.
[0067] In this embodiment, the detailed steps for obtaining the pre-trained time series prediction model include:
[0068] The time series data of the enhanced and quantized crushed stone pile vibration data were used as the dataset, and 80% of it was selected as the training set and 20% as the test set.
[0069] Build a time series forecasting model and set the hyperparameters in the model:
[0070] The model uses a hierarchical classification auxiliary network, which trains classifiers at two levels, fine-grained and coarse-grained, based on a hierarchical structure. Each level divides the range of values of the vibration data into different numbers of categories, thereby obtaining high-entropy features with multi-granularity representation: In this embodiment, the number of categories in the fine-grained level is 4, and the number of categories in the coarse-grained level is 2.
[0071] The model uses an Informer-based backbone network to extract features and maps them to multi-granularity features and temporal features through linear layers. In this embodiment, the encoder and decoder of the backbone network are stacked with two layers of Informer structure, and the normalization method in the structure is layer normalization.
[0072] An uncertainty-aware classifier is trained at both coarse-grained and fine-grained levels using category labels, and multi-granular features are optimized through backpropagation. Temporal features are used to capture temporal information for time series prediction. The model maintains consistency between the coarse-grained and fine-grained levels through hierarchical consistency loss.
[0073] ADAM is used as the default optimizer. An initial learning rate and a dynamic adjustment strategy are set. The model is trained based on stochastic gradient descent.
[0074] The overall code of the time series prediction model is implemented based on PyTorch. The model is trained using an NVIDIA RTX 3090. When the loss function converges, the model parameters at this point are retained as the pre-trained time series prediction model.
[0075] Step S40: A radial basis function neural network is introduced to process the predicted and real-time collected vibration data of the crushed stone piles, automatically generate the quality assessment results of the crushed stone piles, and transmit them to the construction personnel through the data acquisition and control system.
[0076] In this embodiment, by collecting and analyzing historical crushed stone pile construction records, a mapping relationship between crushed stone pile vibration data and quality status is established, classifying quality status into "qualified" and "unqualified" categories. A pre-trained model is obtained by training a radial basis function neural network using historical crushed stone pile vibration data and quality category labels. This model is then used to process the real-time collected and predicted crushed stone pile vibration data, outputting quality assessment results.
[0077] Example 2: This application provides a method for monitoring vibration data of crushed stone piles used in land reclamation operations. The detailed structure of the applied time series prediction model is as follows: Figure 2 As shown:
[0078] Using the time series data of historical crushed stone pile vibration as the input features of the model, the input features are first fed into the backbone network for feature extraction. The backbone network adopts an Informer structure, such as... Figure 3 As shown:
[0079] Informer introduces the ProbSparse self-attention mechanism, self-attention distillation technology, and generative decoder to solve the problems of high computational complexity, large memory consumption, and slow inference speed of traditional Transformer when processing long sequences.
[0080] The ProbSparse self-attention mechanism reduces unnecessary computation through probabilistic sparsity. It focuses only on the most important time points in the sequence, calculates attention weights by randomly sampling a portion of the sequence, and then supplements them with statistical methods to obtain the complete attention weights.
[0081] Self-attention distillation is used to solve the problem of excessive model parameters caused by long input sequences, reduce the input of cascaded layers, and extract the dominant attention. After ProbSparse self-attention, convolution and pooling operations are applied for dimensionality reduction to compress the attention, reducing the sequence dimension of the attention output of each layer to half of the original.
[0082] Generative decoders can predict the entire long-time sequence at once, without having to generate predictions for each time step as in traditional Transformer structures, which greatly improves inference speed and reduces error accumulation.
[0083] The backbone network employs three linear layers at its back end to map the output features of the backbone model into fine-grained features φ, coarse-grained features θ, and temporal features η. The fine-grained and coarse-grained features are then fed into uncertain perceptual classifiers at different levels to generate high-entropy features.
[0084] Traditional softmax-based classifiers often incorrectly assign high confidence scores to inaccurate predictions. This problem becomes even more pronounced when the goal of time series prediction models is to classify time-step values into different categories. To address this issue and improve robustness across classification levels, an evidence-based uncertainty estimation method is introduced to enhance the accuracy of uncertainty assessment.
[0085] This invention utilizes the Dirichlet distribution parameters, i.e., the conjugate prior of the class distribution, to calculate the confidence quality b and the overall uncertainty quality u for different classes. The uncertainty-aware classifier employs a softplus activation function to ensure that the output values are non-negative, and these output values are used as evidence vectors. Let K be a matrix, and K be the number of classes; use these evidence vectors to construct the parameters of the Dirichlet distribution:
[0086]
[0087] Where k is the category index. E represents the Dirichlet intensity. j This represents evidence of the j-th category.
[0088] Ultimately, the probability distribution p k It can be calculated as follows: The uncertainty for each training sample is calculated using the confidence quality. For the i-th training sample, 1-b i This represents category uncertainty; the uncertainty perception coefficient is defined as ω. i =(1-b i )⊙o i o i Let be the one-hot encoding of the true class label of the i-th sample, ⊙ be the Hadamard product, and let ⊙ be the uncertainty-aware loss function. Defined as:
[0089]
[0090] Where ψ(·) is the gamma function, λ UA and λ KL Represents the balance coefficient. S is the uncertainty perception coefficient indicating whether the i-th sample belongs to the k-th class. i It is the Dirichlet intensity of the i-th sample. Let p be the Dirichlet parameter of the i-th sample belonging to the k-th class, KL[·] denotes the Kourbach-Leibler (KL) divergence, Dir denotes the Dirichlet distribution, and p i Let be the probability distribution of the i-th sample. Dir(p i |1) Approximately uniform distribution.
[0091] To ensure the continuity of extracted features, predictions are made within each classification interval, and the mean squared error between the predicted difference and the difference label is calculated. The difference represents the difference between the value before quantization and the minimum value of the interval:
[0092]
[0093] Where I(·) represents a conditional indicator function, and c k and Δy k These are the category label and difference label for the k-th class, respectively. The hierarchical loss is the difference predicted by the model for the k-th class; the hierarchical loss is ultimately expressed as the sum of two loss functions with different granularities:
[0094]
[0095] Where f and c represent fine-grained level and coarse-grained level, respectively.
[0096] Because time series data are continuous, directly classifying time step values can lead to boundary effects—misclassification near class boundaries. To address this, a hierarchical consistency loss is introduced. The goal is to accurately assign values near the fine-grained category boundary to their corresponding coarse-grained category. The loss function is represented by the KL divergence between the output evidence vectors of the two granularity level classifiers.
[0097]
[0098] Among them, e f and e cThese represent the evidence vectors output by the fine-grained and coarse-grained uncertainty-aware classifiers, respectively. This approach ensures that model predictions remain consistent across different levels, effectively mitigating boundary effects.
[0099] Furthermore, the fine-grained feature φ and coarse-grained feature value θ output from the linear layer are multiplied to generate a hierarchical aware attention matrix A. Then, the attention matrix A is multiplied by the temporal feature η and added to the output feature of the backbone network before being input to the predictor through the linear layer. This process can be expressed by the following formula:
[0100]
[0101] Attention(θ,φ,η)=η·Softmax(θ·φ)
[0102] in, W is the predicted value output by the predictor. f W represents the linear layer weights, Attention represents the attention calculation process, F represents the output features of the backbone network, b is the bias, and Softmax represents the softmax activation function; the mean squared error between the predicted value and the true label is calculated.
[0103]
[0104] Where N represents the total number of samples, Y i This represents the actual numerical value of the i-th sample. The overall loss function of the final time series prediction model represents the predicted value of the i-th sample. Represented as:
[0105]
[0106] In this embodiment, ρ1 and ρ2 are loss weight hyperparameters determined through grid search. The values of ρ1 and ρ2 are selected from {1, 0.1, 0.01}. The parameters of the time series prediction model are updated by backpropagation based on the overall loss function until the loss function converges. The model parameters at this point are then retained as the pre-trained time series prediction model.
[0107] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
Claims
1. A method for monitoring vibration data of crushed stone piles used in land reclamation operations, characterized in that, The method includes: (1) Establish a data acquisition and control system and use data acquisition equipment to record the vibration data of vibratory crushed stone piles in real time during the land reclamation operation; (2) Based on the historical vibration data of crushed stone piles stored in the data acquisition and control system, the data volume is expanded by data augmentation method, and the time series prediction model is trained using the enhanced vibration data of crushed stone piles. (3) The real-time collected vibration data of crushed stone piles is preprocessed, input into the pre-trained time series prediction model, and the prediction results of the vibration data of crushed stone piles are output. (4) A radial basis function neural network is introduced to process the predicted and real-time collected vibration data of the crushed stone piles, automatically generate the quality assessment results of the crushed stone piles, and transmit them to the construction personnel through the data acquisition and control system.
2. The method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, characterized in that, In step (1), the data acquisition and control system includes a sensor module, a data transmission module, a data storage module, a data analysis and processing module, and a user interface: The sensor module is equipped with data acquisition devices including current sensors, frequency sensors, depth sensors, and vibration time sensors to collect vibration data of vibratory crushed stone piles in real time, including compaction current, vibration frequency, vibration depth, and vibration time. The data transmission module supports wireless communication, transmitting the collected vibration data of the crushed stone piles to other modules in the data acquisition and control system via a communication protocol. A database is established in the data storage module to store the vibration data of the crushed stone piles, and a timestamp is assigned to each type of data; The data processing and analysis module embeds an intelligent algorithm model to predict the vibration data of crushed stone piles and evaluate the quality of crushed stone piles; The user interface is used to display the quality assessment results of the data processing and analysis module and send relevant information to construction workers involved in land reclamation operations.
3. The method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, characterized in that, In step (2), before data augmentation, the historical crushed stone pile vibration data is processed into time series data so that each type of crushed stone pile vibration data contains continuous time information and is aligned in time. The method for data augmentation based on time series mixing is as follows: Extract data records x of the same sequence length from the time series data of each type of crushed stone pile vibration data. i and x j The starting point for data extraction is randomly generated. By weighting and merging two data records, a new time series data is calculated. Where λ∈[0,1] and conforms to the Beta distribution; for each type of crushed stone pile vibration data, a certain number of new time series data are synthesized by time series mixing.
4. The method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, characterized in that, Step (2), the process of generating the time series of vibration data of crushed stone piles using a kernel synthesis-based data augmentation method, includes: Several kernel functions are randomly selected from the kernel function library. The kernel function types are linear kernels and periodic kernels: The linear kernel function is expressed by the formula: k l (t,t′)=α(t-∈)(t′-∈), where α is the amplitude parameter of the linear kernel function, t and t′ represent two different timestamps, and ∈ is the offset; The periodic kernel function can be expressed by the following formula: Where β is the amplitude parameter of the periodic kernel function, γ is the length scale parameter, and p is the period; These kernel functions are combined using binary operators, and the combined kernel function is used to define the prior distribution of the Gaussian process. The combined composite kernel function is expressed by the formula: k c (t,t′)=k l (t,t′)+μk p (t,t′), μ represents the weighting coefficient, and the composite kernel function is defined as the variance of the Gaussian process, with the mean of the Gaussian process being 0; A synthetic time series of length l is generated by sampling from the prior distribution of a Gaussian process. It is a vector representing the synthetic data at a given timestamp.
5. The method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, wherein step (2) is characterized in that, The time series data of the original and enhanced crushed stone pile vibration data are used as the dataset for the time series prediction model. The samples in the dataset are quantized to achieve the classification-based time series prediction task. Quantization is achieved through a discretization mapping function, which processes the continuous vibration data of crushed stone piles into categorical variables. Based on the specific value of the vibration data at each time point, the interval is used as its category label. The difference between the vibration data value at each time point and the minimum value of the interval is calculated as an additional label. The label value of each time series sample of crushed stone pile vibration data includes the specific value, the interval difference, and the category label.
6. The method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, characterized in that, In step (3), the detailed steps for obtaining the pre-trained time series prediction model include: The time series data of the enhanced and quantized crushed stone pile vibration data were used as the dataset, and 80% of it was selected as the training set and 20% as the test set. A time series prediction model was built, the hyperparameters in the model were set, ADAM was used as the default optimizer, the initial learning rate was set and the dynamic adjustment strategy was set, and the model was trained based on stochastic gradient descent. When the loss function converged, the model parameters at this time were retained as the pre-trained time series prediction model.
7. A method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 6, characterized in that, The time series prediction model uses a hierarchical classification auxiliary network. Based on the hierarchical structure, the classifier is trained at two levels: fine-grained and coarse-grained. Each level divides the range of values of the residual data into different numbers of categories, thereby obtaining high-entropy features with multi-granularity representation. The model uses an Informer-based backbone network to extract features and maps them to multi-granularity features and temporal features through linear layers. Uncertainty-aware classifiers are trained at both coarse-grained and fine-grained levels using class labels. Multi-granularity features are optimized through backpropagation to obtain accurate quantitative information. Temporal features are used to capture temporal information for time series prediction; the model maintains consistency between coarse-grained and fine-grained levels through hierarchical consistency loss.
8. A method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 1, characterized in that, In step (4), by collecting and analyzing historical construction records of crushed stone piles, a mapping relationship between crushed stone pile vibration data and quality status is established, and the quality status is divided into two categories: "qualified" and "unqualified". The radial basis function neural network is trained using historical crushed stone pile vibration data and quality category labels to obtain a pre-trained model, which is used to process the real-time collected and predicted crushed stone pile vibration data and output the quality assessment results.
9. A method for monitoring vibration data of crushed stone piles used in land reclamation operations as described in claim 8, characterized in that, The radial basis function neural network is a feedforward neural network containing an input layer, hidden layers, and an output layer. It generates features by calculating the distance between the input data and the center point using the radial basis functions of the hidden layer neurons. A Gaussian function is used as the radial basis function K(x), expressed by the formula: Where x is the input data of the hidden layer, c is the center point of the radial basis function, and σ is the variance; the center point is calculated from the training data using a clustering algorithm, and the variance is calculated by the following formula: Where h is the number of training data, c max It is the maximum distance between the center points; the weights from the hidden layer to the output layer of the radial basis function neural network are calculated by the least squares method, and the training of the entire network is completed by combining the kernel function from the input layer to the hidden layer and the weights from the hidden layer to the output layer.