Underground engineering acceleration data information clustering and hazard source identification method based on FCNN-RNN

Abstract

This invention discloses a method for clustering and identifying hazard sources in underground engineering acceleration data based on FCNN-RNN. The method proposes a neural network model specifically for data information identification and extraction in underground engineering monitoring and early warning information. The first part of the model is mainly based on convolutional neural networks, constructing three sizes of dilated convolutional kernels to maximize the extraction of multi-scale information in the spatial domain from underground engineering monitoring data. This solves the problem of the model's insensitivity to subtle defect information and reduces the computational cost. The back-end of the model is mainly based on RNN networks. Since RNN networks can extract hidden time-series features between data, and monitoring information at different time periods reflects the development patterns of different types of damage and defects in underground engineering, the addition of RNN networks significantly improves accuracy.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology and relates to a multi-scale data information extraction and recognition system and method, specifically to a method for clustering underground engineering acceleration data information and identifying hazards based on FCNN-RNN. Background Technology

[0002] With the rise of deep learning methods, especially the outstanding performance of convolutional neural networks in the field of image processing, image information extraction and recognition based on deep learning has become a research hotspot. However, in the field of underground engineering, the application of such methods is scarce.

[0003] Meanwhile, in underground engineering, the monitoring information from sensors at different time periods reflects the development patterns of different types of damage and defects. Taking structural deformation as an example, the deformation amount in the future is related to the deformation amount in the previous period. Furthermore, similar projects or similar types of damage also exhibit certain regularities. In the field of image processing, existing technologies often only consider the spatial characteristics of data at a specific point in time, ignoring the unique temporal variations of the monitoring data. Therefore, while existing technologies can identify the location of dangerous deformation or damage, they cannot judge the development trend of the damage or deformation, nor can they accurately warn of potential dangers.

[0004] The mechanisms of damage and disasters in underground engineering projects are complex and influenced by numerous factors, requiring the deployment of a large number of sensors for monitoring. However, the diverse types and massive amounts of monitoring data significantly increase the difficulty of processing them. To address these challenges, a novel data analysis and processing method is urgently needed. This method should be able to handle multi-scale spatial data, including minute data points, while also considering the temporal correlation of monitoring information. It should enable temporal monitoring and early warning of dangerous deformations and damage, maximizing the value of the data and truly activating its potential. Summary of the Invention

[0005] In view of the technical problems existing in the background technology, the present invention provides a method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN, which can continuously and quickly identify defects and damages in underground engineering, greatly improve efficiency, and enhance the reliability and stability of underground engineering.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A system for clustering and identifying hazard sources from underground engineering acceleration data based on FCNN-RNN includes an information acquisition module, a convolutional module, a recurrent neural network module, and a visual output module, wherein:

[0008] The information acquisition module is located at relevant positions in the underground project and is used to acquire monitoring information;

[0009] The convolutional module uses VGG-16 as the main framework to construct a multi-scale global FCN network. The pooling layer Pool5 is removed, and a multi-scale convolutional module is constructed using dilated convolution. Three dilated convolutional channels with dilation rates of 1, 2, and 5 are connected in parallel to perceive information from receptive fields at different scales and extract multi-scale features.

[0010] The recurrent neural network module is an RNN neural network model. The result of feature fusion from the convolution module is input into the RNN neural network model to extract time features.

[0011] The visual output module performs deconvolution on the output of the RNN neural network model, and performs skip connections and feature fusion with the pool4 node of the convolution module. The fused result is then deconvolved twice through a deconvolution layer and skipped with the pool3 node of the convolution module to correct the feature vanishing phenomenon caused by layer iteration. The loss function is calculated through the SoftMax layer for normalization, thereby realizing global feature perception and prediction.

[0012] A method for multi-scale data information extraction and recognition using the above system includes the following steps:

[0013] Step 1: Set up multiple information collection points for collecting monitoring information of underground engineering, and use the data of underground structural defects as a dataset;

[0014] Step 2: Input the data into a multi-scale convolutional module constructed from dilated convolutions to extract multi-scale features;

[0015] Step 3: Use the add algorithm to fuse the features of the pooled results from the three channels;

[0016] Step 4: Input the result of feature fusion into the RNN neural network model to extract time features;

[0017] Step 5: Input the temporal and spatial features extracted by the multi-scale convolutional module and the RNN neural network model into the deconvolutional layer for aggregation.

[0018] Step 6: Calculate the loss value between the output value and the input true value of each pixel by normalizing the output result of the SoftMax layer, and backpropagate it to the previous layer. Iterate continuously with the FCNN-RNN mixture to realize the feature perception of the FCNN-RNN mixture.

[0019] Step 7: Visualize the feature perception and prediction results.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] 1. This invention proposes a neural network model specifically for data information identification and extraction in underground engineering monitoring and early warning systems. The first part of the model is based on a convolutional neural network (CNN), constructing three sizes of dilated convolutional kernels instead of the traditional CNN using only one size kernel per convolutional layer. This maximizes the extraction of multi-scale spatial information from underground engineering monitoring data, solving the problem of the model's insensitivity to subtle defect information and reducing parameter computation. The back-end of the model is based on an reactive neural network (RNN). Since RNNs can extract hidden temporal series features between data points, and monitoring information from different time periods reflects the development patterns of different types of damage and defects in underground engineering, adding an RNN significantly improves accuracy. Therefore, after extracting spatial features through multi-scale convolution in the first part, the data is input into the RNN for temporal feature extraction, enabling the prediction of the development of underground engineering defects and damage in the future.

[0022] 2. Based on the characteristics of underground engineering, this invention employs a multi-scale analysis and identification strategy. Scale 1 provides damage warnings to indicate whether structural damage has occurred; Scale 2 locates the damage; and Scale 3 determines the degree and cause of the damage. This three-scale identification strategy is visualized and presented to staff and researchers, allowing users to intuitively understand the defects and damage in underground engineering projects. Attached Figure Description

[0023] Figure 1 The architecture of a method for clustering and identifying hazard sources from acceleration data in underground engineering based on FCNN-RNN;

[0024] Figure 2 This is a recurrent neural network structure. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0026] This invention provides a clustering and hazard identification system for underground engineering acceleration data based on FCNN-RNN, such as... Figure 1 As shown, the system includes an information acquisition module, a convolution module, a recurrent neural network module, and a visual output module, wherein:

[0027] 1. Information Acquisition Module: This module includes various sensors. Monitoring personnel will place the sensors at relevant locations in the underground project to acquire monitoring information, which is mainly acceleration data.

[0028] 2. Convolutional Module: Using VGG-16 as the main framework, a multi-scale global FCN network is constructed. The pooling layer Pool5 is removed, and a multi-scale module is constructed using dilated convolutions. Three dilated convolutional channels with dilation rates of 1, 2, and 5 are connected in parallel, and the feature information obtained from the three channels is combined.

[0029] The dilated convolution process is defined as follows:

[0030]

[0031] In the formula, x[m,n] is the input feature map, y[m,n] is the output feature map, w[i,j] is the dilated filter, r is the dilation rate, m and n are the length and width of the feature map pixels, and i and j are the length and width of the convolution kernel, respectively. The dilation rate determines the sampling interval of the convolution kernel. When r>1, it means that r-1 holes are introduced between the convolution kernels. When r=1, the dilated convolution is a normal convolution.

[0032] When the dilation rate is r and the kernel size is k×k, the receptive field size of the dilated convolution is:

[0033] k+(k+1)(r-1)

[0034] Using a regular convolution with a stride of 1, when the number of convolutional layers is M layer The size of the sensing field is:

[0035] (k-1)×M labyer +1

[0036] In contrast, dilated convolution can effectively increase the receptive field. When the number of convolutional layers is 1, dilated convolution expands the receptive field of ordinary convolution from 3 to 7 and 15, thereby effectively reducing the amount of parameter computation.

[0037] 3. Recurrent Neural Network Module: RNNs have certain advantages, as they can consider the temporal information existing between data, breaking the limitations of fixed-size input and output of ordinary neural networks. Recurrent neural networks typically consist of three parts: an input layer, an output layer, and hidden layers. This invention inputs the result of feature fusion into the RNN network to extract temporal features.

[0038] Considering the unique temporal variation characteristics of monitoring data, a targeted modeling scheme is needed. In underground engineering, the monitoring information from sensors at different time periods reflects the development patterns of different types of damage and defects. Taking structural deformation as an example, the deformation amount in the future period is related to the deformation amount in the previous period. Furthermore, similar projects or similar types of damage exhibit certain regularities. This invention uses an RNN prediction model to perform a more comprehensive predictive analysis of structural damage information in the future period based on the results of feature fusion.

[0039] The method for training an RNN neural network model is as follows:

[0040] The RNN neural network model consists of an input layer x, a hidden layer s, and an output layer o, with a loop operation on the hidden layer s.

[0041] Figure 2 This is the unfolded diagram of the hidden layer before and after time t, where t-1 and t+1 represent the time states before and after time t, x and o represent the input and output respectively, and S t The data and information stored at time t are calculated using the following formula:

[0042] S t =f(UX) t +WS t-1 )

[0043] In the formula, U represents the weight of the monitoring data input at this moment, W represents the weight of the monitoring data input at the previous moment, and V represents the weight of the output data. U, V, and W are randomly initialized, and S0 = 0 is set. Then, at time t = 1, the calculation steps of the RNN neural network model are as follows:

[0044] h1=Ux1+Ws0

[0045] s1 = f(h1)

[0046] o1 = g(Vs1)

[0047] In the formula, f represents the activation function and g represents the classification function.

[0048] Following the chronological order, the calculation steps for the next time step can be obtained. Therefore, the formulas for calculating the input and output in the RNN neural network model are as follows:

[0049] h t =Ux t +Ws t-1

[0050] s t =f(h) t )

[0051] o t =g(Vs) t )

[0052] In the formula, x t The input representing the characteristics of the monitored data at time t, and correspondingly, x t-1 and x t+1 h represents the input for monitoring data characteristics at times t-1 and t+1. t O represents the hidden state of the model at time t. t This represents the output of the model at time t;

[0053] The monitoring information features are used for training, and the output value O is obtained after each calculation. t Each will produce an error value e compared to the expected value. t Therefore, the total error can be expressed as:

[0054]

[0055] To find the optimal parameters U, V, and W for the model, the error needs to be minimized. The function that measures the total error E is called the loss function. The network model is continuously tuned to optimize the parameters and reduce the loss. Then, gradient descent is used to update the U, V, and W parameters through backpropagation, thereby reducing the error value at the output end and obtaining the optimal weight parameters. The obtained weight parameter table is then loaded into the RNN neural network.

[0056] 4. Visual Output Module: This module performs deconvolution on the output of the RNN network module and establishes skip connections and feature fusion with the pool4 node. By selecting deconvolution layers of different lengths to perform secondary deconvolution on the fused result and establishing skip connections with the pool3 node, it corrects the feature vanishing phenomenon caused by layer iteration. Normalization is performed using a SoftMax layer, and the loss function is calculated and iterated with the FCNN-RNN hybrid network to achieve global feature perception and prediction.

[0057] This invention divides the identification scale of damage risk according to the characteristics of underground engineering and adopts a three-scale step-by-step identification strategy. Scale 1 provides a damage warning to see if the structure has been damaged; Scale 2 locates the damage; and Scale 3 gives a judgment on the degree and cause of the damage.

[0058] Finally, the feature perception and prediction results are loaded into the visual output module, which can directly identify and locate existing damaged areas and predict new types and locations of damage in underground engineering projects.

[0059] The present invention also provides a method for multi-scale data information extraction and recognition using the above-described system, the method comprising the following steps:

[0060] Step 1: Set up multiple data collection points to collect acceleration data monitoring information of underground engineering projects. Use 2000 collected and labeled images of underground structural defects as a dataset. This dataset contains deformation and damage data of underground engineering projects under different geological conditions, structural forms, climatic factors, engineering disturbances, and other important influencing factors. Based on this, to obtain more data, data augmentation is performed through rotation, mirroring, random cropping, and adding salt-and-pepper noise, expanding the dataset to 16000 images. Divide the dataset into training and test sets. Use the labeled training set to train and fine-tune the FCNN-RNN hybrid network, and use the test set to validate the network.

[0061] Step 2: Input the data into a multi-scale convolutional module constructed from dilated convolutions. Divide the data into three channels and select dilated convolutional layers with dilation rates r of 1, 2, and 5, respectively, with a kernel size of 3×3 for each. Each scale is an independent branch, sensing information from the receptive field at different scales and extracting multi-scale features. Different dilation rates control the multi-scale module to learn features at different scales, preserving the original feature information while fusing feature information from multiple scales, thus solving the problem of the model's insensitivity to small defect information. The parallel multi-scale module avoids structural redundancy caused by too many convolutional layers and reduces the amount of parameter computation.

[0062] Step 3: Use the add algorithm to fuse the features of the pooled results from the three channels.

[0063] Step 4: Establish a suitable RNN neural network model based on the characteristics of the monitoring data, including the number of neurons in its input layer, output layer, and hidden layer, the number of hidden layers, the learning rate, activation function, etc. Input the result after feature fusion into the RNN neural network model to extract time features.

[0064] Step 5: Input the temporal and spatial features extracted by the multi-scale convolutional module and the RNN neural network model into the deconvolutional layer for aggregation.

[0065] Step 6: Calculate the loss value between the output value and the input true value of each pixel by normalizing the output result of the SoftMax layer, and backpropagate it to the previous layer. As the FCNN-RNN hybrid network iterates continuously, the network realizes feature perception.

[0066] The formula for calculating the SoftMax layer is:

[0067]

[0068] In the formula, Z i This is the normalization factor. The formula calculates the probability of each label category individually and outputs the predicted value of the feature map.

[0069] Step 7: Visualize the feature perception and prediction results. This invention divides the model's recognition scale according to the characteristics of underground engineering, adopting a three-scale step-by-step recognition strategy. Scale 1 provides a damage warning, checking if the structure has been damaged; Scale 2 locates the damage; and Scale 3 determines the degree and cause of the damage. Based on the existing image information, the damaged areas are identified and located, and new types and locations of damage in underground engineering are predicted.

[0070] In this invention, the first half of the model is mainly based on a convolutional neural network, with certain modifications, to extract multi-scale information of underground engineering monitoring data in space, thus solving the problem of the model's insensitivity to small defect information; the back end of the model is mainly based on an RNN network to extract temporal features, enabling the prediction of the development of underground engineering defects and damage in the future.

Claims

1. A method for clustering and identifying hazard sources from underground engineering acceleration data based on FCNN-RNN, characterized in that... The method utilizes a system consisting of an information acquisition module, a convolution module, a recurrent neural network module, and a visual output module to achieve clustering of acceleration data information and hazard source identification in underground engineering projects, wherein: The information acquisition module is located at relevant positions in the underground project and is used to acquire monitoring information; The convolutional module uses VGG-16 as the main framework to construct a multi-scale global FCN network. The pooling layer Pool5 is removed, and a multi-scale convolutional module is constructed using dilated convolution. Three dilated convolutional channels with dilation rates of 1, 2, and 5 are connected in parallel to perceive information from receptive fields at different scales and extract multi-scale features. The recurrent neural network module is an RNN neural network model. The result of feature fusion from the convolution module is input into the RNN neural network model to extract time features. The visual output module performs deconvolution on the output of the RNN neural network model, and performs skip connections and feature fusion with the pool4 node of the convolution module. The fusion result is deconvolved twice through the deconvolution layer and skipped with the pool3 node of the convolution module to correct the feature disappearance phenomenon caused by layer iteration. The loss function is calculated by normalization through the SoftMax layer to achieve global feature perception and prediction. The method includes the following steps: Step 1: Set up multiple information collection points for collecting monitoring information of underground engineering, and use the data of underground structural defects as a dataset; Step 2: Input the data into a multi-scale convolutional module constructed from dilated convolutions to extract multi-scale features; Step 3: Use the add algorithm to fuse the features of the pooled results from the three channels; Step 4: Input the result of feature fusion into the RNN neural network model to extract time features; Step 5: Input the temporal and spatial features extracted by the multi-scale convolutional module and the RNN neural network model into the deconvolutional layer for aggregation. Step 6: Calculate the loss value between the output value and the input true value of each pixel by normalizing the output result of the SoftMax layer, and backpropagate it to the previous layer. Iterate continuously with the FCNN-RNN mixture to realize the feature perception of the FCNN-RNN mixture. Step 7: Visualize the feature perception and prediction results.

2. The method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN according to claim 1, characterized in that... The monitoring information is acceleration data.

3. The method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN according to claim 1, characterized in that... The dilated convolution process is defined as follows: In the formula, For the input feature map, To output the feature map, For hole filters, The void ratio, and These are the length and width of the feature map pixels, respectively. and These represent the length and width of the convolution kernel, respectively.

4. The method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN according to claim 3, characterized in that... The void ratio determines the sampling interval of the convolution kernel. , indicating the introduction of between convolution kernels A void, when Dilated convolution is a normal convolution, with a kernel size of [size missing]. At that time, the receptive field size of the dilated convolution is: 。 5. The method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN according to claim 1, characterized in that... The method for training the RNN neural network model is as follows: RNN neural network models include an input layer. Hidden layer Output layer In the hidden layer There is a loop operation above; and express The state of time before and after a given moment. and They represent input and output respectively. Indicates time The data and information stored at a given time are calculated using the following formula: In the formula, This indicates the weight of the monitoring data input at this moment. This indicates the weight of the monitoring data input at the previous moment. The output data weights are randomly initialized. , , and set Then in At time 1, the computation steps of the RNN neural network model are as follows: In the formula, This represents the activation function. Represents the classification function; Following the chronological order, the calculation steps for the next time step can be obtained. Therefore, the formulas for calculating the input and output in the RNN neural network model are as follows: In the formula, represent The input of real-time monitoring data characteristics, accordingly, and Representative at and Input of real-time monitoring data characteristics Representative at Hidden states of the time-series model Representative at Output of the time-mapping model; The monitoring information features are used for training, and the output value is obtained after each calculation. Each will produce an error value compared to the expected value. Therefore, the total error is expressed as: The network model is continuously optimized by tuning parameters to reduce the loss, and then the gradient descent method is used to update it through backpropagation. , , The parameters are then used to reduce the error value at the output end, resulting in the optimal weight parameters. The obtained weight parameter table is then loaded into the RNN neural network.

6. The method for clustering and identifying hazard sources of underground engineering acceleration data based on FCNN-RNN according to claim 1, characterized in that... The specific steps of step 7 are as follows: the identification scale of damage risk is divided according to the characteristics of underground engineering, and a three-scale step-by-step identification strategy is adopted. Scale 1 provides a damage warning to see if the structure has been damaged. Scale 2 is used to locate the damage, and Scale 3 is used to determine the degree and cause of the damage. Finally, the feature perception and prediction results are loaded into the visual output module to identify and locate existing damaged parts and predict new types of damage and location information of underground engineering.

Images