Electric tower foundation instability early warning method based on space-time attention mechanism

Abstract

The invention provides an electric pole tower foundation instability early warning method based on a space-time attention mechanism, and relates to the technical field of geological disaster monitoring and early warning, and the method comprises the steps: collecting monitoring data of four corners of an electric pole tower base; preprocessing the obtained settlement monitoring data to obtain input characteristics; inputting the input features into a pre-trained settlement prediction model to output a settlement prediction result; performing early warning on the instability state of the electric tower foundation according to the settlement prediction result and an early warning mechanism; the settlement prediction model comprises a space part, a time part and a full connection layer; a convolutional neural network and a convolutional attention module of the space part extract space features and endow the space features with space attention weights; the LSTM of the time part and the time attention module are used for extracting time features and endowing time attention weights; and the full connection layer maps a final settlement prediction result according to a feature obtained by splicing the spatial feature and the time feature. According to the scheme, the accuracy of electric tower foundation instability prediction and early warning can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster monitoring and early warning, and in particular to an electric pole tower foundation instability early warning method based on a spatiotemporal attention mechanism. Background Art

[0002] In the process of mining, it is inevitable that the original geological environment will be destroyed. After the minerals rich in the ground are mined, the stress and strain of the surrounding rock of the ore body will change, breaking the original balance and inducing the problem of foundation settlement. If the foundation settlement is not monitored and predicted during the mining and construction process, the consequences may be the deterioration of land resources at the least, and the deformation and collapse of the mine at the worst, which will directly threaten the foundation settlement of the mining area and affect the stability and safety of the infrastructure of human life in the range. Especially for electric poles and towers, when the foundation of the mining area settles, the foundation of the electric poles and towers may become unstable, and the instability of the foundation of the electric poles and towers will not only affect the normal power supply in the range, but also may cause major safety accidents. Therefore, it is necessary to monitor and warn the instability of the foundation of the electric poles and towers.

[0003] Traditional prediction methods for foundation settlement include empirical formula method, numerical simulation method, theoretical analysis method, etc. However, traditional prediction methods have problems such as difficult to determine parameters, complex models, and weak nonlinearity. Without solving these problems, it was subsequently proposed to apply machine learning to the prediction of foundation settlement. Since the data of foundation settlement monitoring belongs to time series data, the current machine learning-based foundation settlement prediction method mainly uses LSTM network to process sequence data with long-term dependent characteristics. However, the settlement data of the monitoring point is not only related to the time dimension, but also to the spatial dimension. It is difficult for LSTM to learn the spatial characteristics between different data, which leads to the low accuracy of the existing foundation settlement instability prediction and early warning. Summary of the invention

[0004] In view of this, in order to address the above shortcomings, it is necessary to propose a tower foundation instability warning method based on spatiotemporal attention mechanism to improve the accuracy of tower foundation instability prediction and warning.

[0005] The present invention provides a method for early warning of foundation instability of electric towers based on a spatiotemporal attention mechanism, the method comprising:

[0006] Collect monitoring data from the four corners of the tower base to obtain settlement monitoring data;

[0007] Preprocessing the settlement monitoring data to obtain input features suitable for inputting into a settlement prediction model;

[0008] Inputting the input features into a pre-trained settlement prediction model, and outputting a settlement prediction result of the electric pole tower;

[0009] According to the settlement prediction results and early warning mechanism, early warning is given for the instability of the foundation of the electric tower;

[0010] Among them, the settlement prediction model includes a spatial part, a temporal part and a fully connected layer; the spatial part includes a convolutional neural network and a convolutional attention module, which are used to extract the spatial features of the data and dynamically assign spatial attention weights to the input data; the temporal part includes a long short-term memory network and a temporal attention module, which are used to extract the temporal features of the data and dynamically assign temporal attention weights to the input data; the fully connected layer is used to map the features obtained by splicing the spatial features and the temporal features to produce the final settlement prediction results.

[0011] Preferably, the monitoring data of the four corners of the collector tower base includes:

[0012] The displacement, velocity and acceleration data of the electric towers are collected through the Global Navigation Satellite System (GNSS) and the Fiber Distributed Radiation (OFDR) system respectively;

[0013] Among them, the GNSS sensors are installed at the four corners of the middle of the pole tower, and the height is set at 1 / 3 of the total height of the pole tower; the OFDR optical cable is installed on the base of the pole tower in a surrounding manner and surrounds a full circle.

[0014] Preferably, the preprocessing of the settlement monitoring data includes:

[0015] For the data collected by GNSS and OFDR, the sliding window method is used to detect abnormal data points that exceed ±D times the mean, and the cubic spline interpolation method is used to process the outliers to obtain outlier processed data;

[0016] For the four groups of GNSS outlier processing data and OFDR outlier processing data, the linear interpolation method is used to expand the data to the monitoring data of 9 points in a 3*3 grid to obtain the expanded monitoring data;

[0017] Performing standardization processing on the expanded monitoring data to obtain standardized data;

[0018] The standardized data is constructed as a three-dimensional tensor, and a time series is constructed in chronological order to obtain the input feature; wherein, the three channels of the three-dimensional tensor are displacement, velocity and acceleration respectively; the three-dimensional tensor is used to input the spatial part of the settlement prediction model, and the time series is used to input the time part of the settlement prediction model.

[0019] Preferably, the training process of the settlement prediction model includes:

[0020] Obtain historical settlement monitoring data of power tower bases;

[0021] Preprocessing the historical settlement monitoring data to obtain three-dimensional tensor sample data and time series sample data;

[0022] The three-dimensional tensor sample data is used as the input of the spatial part of the model, and the time series sample data is used as the input of the time part. Model training is performed based on the back propagation algorithm to obtain the settlement prediction model.

[0023] Preferably, the loss function during model training is:

[0024]

[0025] In the formula, is the model prediction value, y i is the true value, n is the number of samples, λ is the regularization coefficient, θ j is the model parameter of the jth layer;

[0026] Three models are obtained by training using three different sets of training conditions, and the performance of the three models is evaluated using root mean square error and mean absolute percentage error, and the model with the best evaluation result is determined as the settlement prediction model; wherein the initial learning rates and regularization coefficients during the training of the three models are different.

[0027] Preferably, when training the model, the process of the spatial part includes:

[0028] The convolutional neural network moves the filter on the input feature F to perform a convolution operation to extract the convolution feature matrix X;

[0029] The convolution feature matrix X obtained by the convolutional neural network is input into the convolutional attention module, which dynamically assigns weights to different channels and positions to obtain spatial feature output.

[0030] Preferably, the convolutional attention module dynamically assigns weights to different channels and positions, including:

[0031] The following calculation formula is used for dynamic weight allocation:

[0032]

[0033] In the formula, represents matrix multiplication, X c is the channel attention result, X′ is the spatial output feature of the final output of the spatial part, and A c is the channel attention weight, A s is the spatial attention weight;

[0034] Among them, the channel attention weight A c and the spatial attention weight As Obtained by:

[0035]

[0036] In the formula, σ is the sigmoid function, AvgPool is the average pooling, MaxPool is the maximum pooling, and f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

[0037] Preferably, when training the model, the process of the time portion includes:

[0038] The time series sample data is input into a long short-term memory network, and weights are dynamically assigned to different environmental factors in a single time step through a factor attention mechanism, and weights are assigned to hidden states of each time step through a time attention mechanism to obtain a time feature output.

[0039] Preferably, the dynamically allocating weights to different environmental factors in a single time step through a factor attention mechanism includes:

[0040] For time series data X=(X1,X2,…,X T )=(X 1 ,X 2 ,…,X M ) T ∈R T×M , T represents the number of time steps, M represents the number of features; in time step t, the kth feature matrix The attention weight of this factor at time t Calculated by the following formula:

[0041]

[0042] Where V f ∈R T , W f ∈R T×2p , U f ∈R T×T is the weight matrix of the multi-layer perceptron MLP, σ c is the activation function, b f ∈R T is the corresponding bias, h t-1 ∈R p and c t-1 ∈R p is the hidden state and unit state of the previous moment;

[0043] The weighted input matrix, hidden state, and cell state of the LSTM network at time t are updated as follows:

[0044]

[0045] In the formula, σ s represents the softmax function, and f is a long short-term memory network unit;

[0046] The weights are assigned to the hidden state of each time step through the temporal attention mechanism, including:

[0047] The weights are assigned to the hidden states at each time step by the following calculation:

[0048]

[0049] Where H is the hidden state matrix, h k is the hidden state of the LSTM network unit in the kth step, s is the size of each hidden state, β is the normalized attention weight matrix, β i is the attention weight corresponding to the hidden state in step i, A t is a fully connected neural network, h at It is the time feature output of the final output of the long short-term memory network.

[0050] Preferably, the step of giving an early warning of the instability of the electric pole tower foundation according to the settlement prediction result and the early warning mechanism includes:

[0051] Determine whether the settlement prediction result output by the settlement prediction model exceeds a preset alarm threshold; if so, issue an alarm signal and send the alarm information to the corresponding operation and maintenance engineer.

[0052] It can be seen from the above technical scheme that in the method for early warning of the foundation instability of the electric pole tower based on the spatiotemporal attention mechanism provided by the present invention, the monitoring data of the four corners of the electric pole tower base are first collected to obtain the settlement monitoring data, and then the settlement monitoring data is preprocessed to obtain the input features of the settlement prediction model; the input features are further input into the settlement prediction results of the transmission tower in the pre-trained settlement prediction model, so that the instability state of the electric pole tower foundation can be warned according to the settlement prediction results and the early warning mechanism. Among them, the settlement prediction model is composed of a spatial part, a temporal part and a fully connected layer, and the convolutional neural network and the convolutional attention module constituting the spatial part can extract the spatial features of the data and dynamically assign the spatial attention weight to the input data, while the long short-term memory network and the temporal attention module constituting the temporal part can extract the temporal features of the data and dynamically assign the temporal attention weight to the input data. In this way, the temporal characteristics and spatial characteristics of the foundation settlement monitoring data can be fully considered, so that the settlement deformation of the electric pole tower can be predicted more accurately, thereby improving the accuracy of the prediction and early warning of the foundation instability of the electric pole tower. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A flowchart of a method for early warning of power tower foundation instability based on a spatiotemporal attention mechanism is provided in an embodiment of the present invention.

[0054] Figure 2 A schematic diagram of collecting monitoring data for the four corners of a pole tower base provided in an embodiment of the present invention.

[0055] Figure 3 A schematic diagram of a three-dimensional tensor structure in an input convolutional neural network provided in an embodiment of the present invention.

[0056] Figure 4 A schematic diagram of the structure of a settlement prediction model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0058] like Figure 1 As shown, the present invention provides a method for early warning of power tower foundation instability based on a spatiotemporal attention mechanism, the method comprising the following steps:

[0059] Step 101: Collect monitoring data of the four corners of the base of the electric pole tower to obtain settlement monitoring data;

[0060] Step 102: preprocessing the settlement monitoring data to obtain input features suitable for inputting into a settlement prediction model;

[0061] Step 103: Input the input features into the pre-trained settlement prediction model, and output the settlement prediction result of the electric pole tower; wherein the settlement prediction model includes a spatial part, a temporal part, and a fully connected layer; the spatial part includes a convolutional neural network and a convolutional attention module, which are used to extract the spatial features of the data and dynamically assign spatial attention weights to the input data; the temporal part includes a long short-term memory network and a temporal attention module, which are used to extract the temporal features of the data and dynamically assign temporal attention weights to the input data; the fully connected layer is used to map the features obtained by splicing the spatial features and the temporal features to the final settlement prediction result;

[0062] Step 104: Based on the settlement prediction result and the early warning mechanism, an early warning is issued for the instability of the electric tower foundation.

[0063] For step 101, monitoring data of the four corners of the base of the electric pole tower is collected to obtain settlement monitoring data;

[0064] In this step, the global navigation satellite system GNSS and the fiber distributed scattering system OFDR are considered to collect the displacement, velocity and acceleration data of the power tower. The two complement each other to deal with data anomalies. Figure 2 As shown in the figure, GNSS sensors are installed at the four corners of the middle of the electric pole tower, with a total of 4 sensors. The height of the sensors is set at 1 / 3 of the total height of the electric pole tower, which is used to monitor the displacement changes of the electric pole tower under external forces such as wind load and ice load; the OFDR optical cable is installed on the base of the electric pole tower in a surrounding manner, and it is surrounded in a full circle, focusing on the data at the four corners of the base of the electric pole tower, which is used to monitor the cumulative displacement, velocity and acceleration data of the base. The GNSS sensor and OFDR optical cable upload the collected data to the cloud platform through the combination of 5G wireless communication module.

[0065] For step 102, the settlement monitoring data is preprocessed to obtain input features suitable for inputting into the settlement prediction model;

[0066] In this step, when preprocessing the settlement monitoring data, it can mainly include outlier detection and processing, data expansion and standardization. Specifically, it can be achieved through the following steps S21-S24:

[0067] S21: For the data collected by GNSS and OFDR respectively, the abnormal data points exceeding ±D times the mean are detected by the sliding window method, and the outliers are processed by the cubic spline interpolation method to obtain the outlier processed data;

[0068] In this step, consider using the sliding window method to detect outliers on the data. Set the window length to 10 data points, calculate the mean and standard deviation of the data within the sliding window, and consider data points that exceed ±3 times the mean as outliers. Use the cubic spline interpolation method to process. For n+1 data points (x0, y0), (x1, y1), …, (x n ,y n ), the cubic spline interpolation formula is as follows:

[0069] S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 , x i ≤x≤x i+1

[0070] Among them, the four coefficients in the interpolation formula are determined according to the following conditions:

[0071] Interpolation point conditions:

[0072]

[0073] First-order derivative continuity:

[0074] S i ′(x i )=S i ' +1 (x i+1 )

[0075] Second-order derivative continuity:

[0076] S i ″(x i )=S i ″ +1 (x i+1 )

[0077] Second-order derivative boundary conditions:

[0078]

[0079] S22: for the four groups of GNSS outlier processed data and OFDR outlier processed data, the data are expanded to monitoring data of 9 points in a 3*3 grid using a linear interpolation method to obtain expanded monitoring data;

[0080] S23: Standardize the expanded monitoring data to obtain standardized data;

[0081] In this step, consider standardizing the expanded monitoring data to adjust the data to a range of 0 mean and 1 standard deviation, thereby eliminating the influence of different dimensions. Specifically, the standardization can be performed using the following formula:

[0082]

[0084] Where x is the original value, x′ is the standardized value, μ and σ are the mean and variance of the selected data set, respectively.

[0085] S24: construct the standardized data into a three-dimensional tensor, and construct a time series in chronological order to obtain input features; wherein, the three channels of the three-dimensional tensor are displacement, velocity and acceleration; the three-dimensional tensor is used to input the spatial part of the settlement prediction model, and the time series is used to input the time part of the settlement prediction model.

[0086] In this step, the standardized data is constructed into a 3*3*3 three-dimensional tensor, such as Figure 3 The figure shows the three-dimensional tensor structure of the deformed data, which serves as the input of the spatial part of the model, that is, the input of the convolutional neural network and the convolutional attention module. At the same time, the time series sequence of the data is constructed in chronological order, and the time series sequence serves as the input of the time part of the model, that is, the input of the long short-term memory network and the time attention module.

[0087] For step 103, the input features are input into a pre-trained settlement prediction model, and the settlement prediction result of the electric tower is output;

[0088] In this step, if Figure 4 As shown in the figure, the settlement prediction model includes a spatial part, a temporal part and a fully connected layer; the spatial part includes a convolutional neural network and a convolutional attention module, which are used to extract the spatial features of the data and dynamically assign spatial attention weights to the input data; the temporal part includes a long short-term memory network and a temporal attention module, which are used to extract the temporal features of the data and dynamically assign temporal attention weights to the input data; the fully connected layer is used to map the features obtained by concatenating the spatial features and the temporal features to the final settlement prediction results.

[0089] Specifically, the training process of the settlement prediction model may include:

[0090] Obtain historical settlement monitoring data of power tower bases;

[0091] Preprocess the historical settlement monitoring data to obtain three-dimensional tensor sample data and time series sample data;

[0092] The three-dimensional tensor sample data is used as the input of the spatial part of the model, and the time series sample data is used as the input of the temporal part. The model is trained based on the back propagation algorithm to obtain the settlement prediction model.

[0093] When training the model, consider calculating the gradient through the normal propagation algorithm and updating the model parameters; the initial learning rate can be set to 0.018, and the learning rate decays by 0.8 every 30 iterations; use the following mean square error as the loss function, and prevent the model from overfitting through regularization:

[0094]

[0095] In the formula, is the model prediction value, y i is the true value, n is the number of samples, λ is the regularization coefficient, θ j is the model parameter of the jth layer;

[0096] Three models were obtained by training under three different training conditions, and the performance of the three models was evaluated by using the root mean square error and the mean absolute percentage error, and the model with the best evaluation result was determined as the settlement prediction model; among them, the initial learning rate and regularization coefficient of the three models were different during training. For example, during the training process, a drop rate of 0.3 was randomly set in some layers to further prevent overfitting; the number of iterations of model training can be set to 100 times, and the model parameters are updated each iteration. Three different training conditions were set according to the following parameters: the initial learning rates were 0.018, 0.015, and 0.02, respectively, and the learning rates were decayed every 30 iterations (the decay rates were 0.85, 0.7, and 0.85, respectively), and the regularization coefficients were set to 0.01, 0.02, and 0.005, respectively. Finally, the three trained models were tested on the test set, and the model performance was evaluated by root mean square error (RMSE) and mean absolute percentage error (MAPE). The model with the best performance in the two indicators was selected as the settlement prediction model by weighted average to predict the stability of the tower foundation.

[0097] Among them, the expressions of root mean square error RMSE, mean absolute percentage error MAPE and comprehensive evaluation index are as follows:

[0098]

[0099] In the formula, y i is the true value, is the model prediction value, n is the number of samples, α and β are weight coefficients, and α+β=1.

[0100] Furthermore, when training the model, the spatial part of the process includes:

[0101] The convolutional neural network moves the filter on the input feature F to perform a convolution operation to extract the convolution feature matrix X;

[0102] The convolution feature matrix X obtained by the convolutional neural network is input into the convolutional attention module, which dynamically assigns weights to different channels and positions to obtain spatial feature output.

[0103] In this embodiment, the convolutional neural network may include two convolutional layers, a pooling layer, and a fully connected layer. The number of filters in the first convolutional layer is 30, the kernel size is (1,2), the number of filters in the second convolutional layer is 64, the kernel size is (2,2), and the activation function is ReLU. On the convolutional layer, the output feature X is generated by performing a convolution operation by moving the filter on the input feature F. After the convolutional layer, adjacent elements of the feature matrix X are integrated by maximum pooling, average pooling, or other operations. The convolution calculation method can be as follows:

[0104]

[0105] In the formula, x i,j is an element of the matrix X, c is the number of channels of F, and f i,j is a submatrix of F, with the same size as the filter, i and j are the number of movement steps of the rows and columns of filter k respectively. When the filter moves on the input feature, the filter repeatedly acts on different submatrices f on the input feature map i,j , which is the parameter sharing mechanism of CNN, can reduce the number of parameters and the complexity of the model. The special mathematical operations of matrix data during convolution and pooling make CNN have advantages in local spatial feature extraction.

[0106] Furthermore, for the convolutional attention module CBAM, when the feature X output by CNN is input to CBAM, CBAM will dynamically assign weights to different channels and positions through learning, thereby inferring the channel feature map and spatial feature map. The entire processing process can be expressed as:

[0107]

[0108] In the formula, represents matrix multiplication, X c is the channel attention result, X′ is the spatial output feature of the final output of the spatial part, and A c is the channel attention weight, A s is the spatial attention weight;

[0109] Among them, the channel attention weight A c and the spatial attention weight A s Obtained by:

[0110]

[0111] In the formula, σ is the sigmoid function, AvgPool is the average pooling, MaxPool is the maximum pooling, and f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

[0112] It can be seen that in this embodiment, the channel attention weight pays differentiated attention to different channels of the feature map, performs average pooling and maximum pooling on the original feature map, thereby integrating the spatial information in each channel.

[0113] When training the model, the time part of the process includes:

[0114] The time series sample data is input into the long short-term memory network, and the weights are dynamically assigned to different environmental factors in a single time step through the factor attention mechanism, and the weights are assigned to the hidden state of each time step through the time attention mechanism to obtain the time feature output.

[0115] In this embodiment, the long short-term memory network LSTM can be 2 layers, and the number of hidden units in each layer is 64. In the structure of LSTM, the state body is replaced by a self-circulating memory block, which consists of an input gate, a forget gate, an output gate and a memory unit, providing long-term memory and forgetting functions for LSTM. The forget gate determines how much historical information can be stored in the memory unit, the input gate determines how much information can be stored in the memory unit at the current moment, and the output gate is used to convert the current unit state in LSTM to the next unit state. LSTM can be expressed as follows:

[0116]

[0117] In the formula, i t 、f t , o t 、c t They are the input gate, forget gate, output gate and four outputs of the storage unit, b i , b f , b o , b c are the corresponding bias vectors, w xi 、w xf 、w xo 、w xc and w xi 、w hf 、w ho 、w hc are the corresponding input matrix and hidden matrix respectively, σ is the sigmoid function, and tanh is the hyperbolic tangent function.

[0118] For the temporal attention module, it is specifically divided into two parts: factor attention mechanism and temporal attention mechanism. When dynamically allocating weights to different environmental factors in a single time step through the factor attention mechanism, it can include:

[0119] For time series data X=(X1,X2,…,X T )=(X 1 ,X 2 ,…,X M ) T ∈R T×M , T represents the number of time steps, M represents the number of features; the time series of the present invention may include cumulative displacement, velocity and acceleration. In time step t, the kth feature matrix The attention weight of this factor at time t Calculated by the following formula:

[0120]

[0121] Where V f ∈R T , W f ∈R T×2p , U f ∈R T×T is the weight matrix of the multi-layer perceptron MLP, σ c is the activation function, b f ∈R T is the corresponding bias, h t-1 ∈R p and c t-1 ∈R p is the hidden state and unit state of the previous moment;

[0122] The weighted input matrix, hidden state, and cell state of the LSTM network at time t are updated as follows:

[0123]

[0124] In the formula, σ s represents the softmax function, and f is a long short-term memory network unit;

[0125] The temporal attention mechanism selects hidden states in the entire memory cycle. First, the input data is the hidden state of the LSTM unit at different time steps, and the fully connected neural network processes and outputs the temporal attention weights. At each time step, the temporal attention weights are assigned to the hidden state. Specifically, the weight allocation method and final output of the temporal attention mechanism are as follows:

[0126] The weights are assigned to the hidden states at each time step by the following calculation:

[0127]

[0128] Where H is the hidden state matrix, h k is the hidden state of the LSTM network unit in the kth step, s is the size of each hidden state, β is the normalized attention weight matrix, β i is the attention weight corresponding to the hidden state in step i, A t is a fully connected neural network, h at It is the time feature output of the final output of the long short-term memory network.

[0129] Finally, the outputs of the spatial and temporal parts are integrated by splicing, and then passed through a fully connected layer to obtain the final output of the model, which is the settlement prediction data of the four corners of the electric tower.

[0130] After the model training is completed, the model is packaged using Docker container technology to ensure the consistency and portability of the model in different environments. Select appropriate computing resources, storage resources, and network resources on the cloud platform; install the dependencies required for the model to run, including the deep learning framework and Python environment. Upload the packaged model file to the cloud platform for storage, use the hosting service provided by the cloud platform to encapsulate the model so that it can receive input and return predicted output through HTTP requests, define API outages, and determine the format of input parameters and the structure of output responses.

[0131] Furthermore, continuous optimization of the model can also be considered to ensure the accuracy and robustness of the model prediction. Specifically, various indicators of ground subsidence are monitored regularly and wireless data transmission is adopted. The collected data is transmitted to the cloud platform through the 5G wireless communication module; the sliding window method is used to detect outliers in the data (the specific method has been introduced in the data collection section above); the incremental learning strategy is used to integrate new data into the model, and the new data is added to the training set on a weekly basis, and then the model is retrained. The trained model is also evaluated using the root mean square error (RMSE) and mean absolute percentage error (MAPE) to determine the reliability of the model, and the Python Scikit-learn library is used to design scripts to perform model evaluation.

[0132] The above checking of new data, processing of new data, model training, and model evaluation will be automated, and the above steps will be integrated into the continuous integration / continuous deployment (CI / CD) pipeline. The scheduled task scheduling Windows TaskScheduler is used to set the scheduled task, which is executed regularly every week, and then the CI / CD pipeline is triggered to process new data and train and evaluate the model. The version control system Git is used to track the change history of the model and data set to ensure that all the results of the automated operation, the changes made, and the decisions made are fully documented.

[0133] In step 104, an early warning is issued for the instability of the foundation of the electric tower according to the settlement prediction result and the early warning mechanism;

[0134] In this step, it can be determined whether the settlement prediction result output by the settlement prediction model exceeds the preset alarm threshold; if so, an alarm signal is issued and the alarm information is sent to the corresponding operation and maintenance engineer. After the model is deployed, the pre-processed measured data can be uploaded to the cloud platform and input into the model to obtain the output prediction result, that is, the displacement prediction data of the four corners of the pole tower base. In this way, by comparing the predicted data with the set threshold, if it exceeds the set threshold, an alarm signal is issued, and the cloud setting is used to notify the designated engineer by email or text message to take relevant emergency measures.

[0135] The present specification also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed in a computer, the computer is caused to execute a method in any one of the embodiments in the specification.

[0136] The present specification also provides a computing device, including a memory and a processor, wherein executable codes are stored in the memory, and when the processor executes the executable codes, a method in any embodiment of the present specification is implemented.

[0137] Since the device embodiments provided by the present invention are based on the same inventive concept as the method embodiments of this specification, the specific contents can be found in the description of the method embodiments of this specification and will not be repeated here.

[0138] The modules or units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. The above disclosure is only the preferred embodiment of the present invention, and of course it cannot be used to limit the scope of the rights of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiment are implemented, and the equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for early warning of power tower foundation instability based on spatiotemporal attention mechanism, characterized in that: The method includes: Collect monitoring data from the four corners of the tower base to obtain settlement monitoring data; Preprocessing the settlement monitoring data to obtain input features suitable for inputting into a settlement prediction model; Inputting the input features into a pre-trained settlement prediction model, and outputting a settlement prediction result of the electric pole tower; According to the settlement prediction results and early warning mechanism, early warning is given for the instability of the foundation of the electric tower; Among them, the settlement prediction model includes a spatial part, a temporal part and a fully connected layer; the spatial part includes a convolutional neural network and a convolutional attention module, which are used to extract the spatial features of the data and dynamically assign spatial attention weights to the input data; the temporal part includes a long short-term memory network and a temporal attention module, which are used to extract the temporal features of the data and dynamically assign temporal attention weights to the input data; the fully connected layer is used to map the features obtained by splicing the spatial features and the temporal features to produce the final settlement prediction results.

2. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 1 is characterized in that: The monitoring data of the four corners of the collector tower base include: The displacement, velocity and acceleration data of the electric towers are collected through the Global Navigation Satellite System (GNSS) and the Fiber Distributed Radiation (OFDR) system respectively; Among them, the GNSS sensors are installed at the four corners of the middle of the pole tower, and the height is set at 1 / 3 of the total height of the pole tower; the OFDR optical cable is installed on the base of the pole tower in a surrounding manner and surrounds a full circle.

3. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 2 is characterized in that: The preprocessing of the settlement monitoring data includes: For the data collected by GNSS and OFDR, the sliding window method is used to detect abnormal data points that exceed ±D times the mean, and the cubic spline interpolation method is used to process the outliers to obtain outlier processed data; For the four groups of GNSS outlier processing data and OFDR outlier processing data, the linear interpolation method is used to expand the data to the monitoring data of 9 points in a 3*3 grid to obtain the expanded monitoring data; Performing standardization processing on the expanded monitoring data to obtain standardized data; The standardized data is constructed as a three-dimensional tensor, and a time series is constructed in chronological order to obtain the input feature; wherein, the three channels of the three-dimensional tensor are displacement, velocity and acceleration respectively; the three-dimensional tensor is used to input the spatial part of the settlement prediction model, and the time series is used to input the time part of the settlement prediction model.

4. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 1 is characterized in that: The training process of the settlement prediction model includes: Obtain historical settlement monitoring data of power tower bases; Preprocessing the historical settlement monitoring data to obtain three-dimensional tensor sample data and time series sample data; The three-dimensional tensor sample data is used as the input of the spatial part of the model, and the time series sample data is used as the input of the time part. Model training is performed based on the back propagation algorithm to obtain the settlement prediction model.

5. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 4 is characterized in that: The loss function during model training is: In the formula, is the model prediction value, y i is the true value, n is the number of samples, λ is the regularization coefficient, θ j is the model parameter of the jth layer; Three models are obtained by training using three different sets of training conditions, and the performance of the three models is evaluated using root mean square error and mean absolute percentage error, and the model with the best evaluation result is determined as the settlement prediction model; wherein the initial learning rates and regularization coefficients during the training of the three models are different.

6. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 4 is characterized in that: When training the model, the spatial part of the process includes: The convolutional neural network moves the filter on the input feature F to perform a convolution operation to extract the convolution feature matrix X; The convolution feature matrix X obtained by the convolutional neural network is input into the convolutional attention module, which dynamically assigns weights to different channels and positions to obtain spatial feature output.

7. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 6 is characterized in that: The convolutional attention module dynamically assigns weights to different channels and positions, including: The following calculation formula is used for dynamic weight allocation: In the formula, represents matrix multiplication, X c is the channel attention result, X′ is the spatial output feature of the final output of the spatial part, and A c is the channel attention weight, A s is the spatial attention weight; Among them, the channel attention weight A c and the spatial attention weight A s Obtained by: In the formula, σ is the sigmoid function, AvgPool is the average pooling, MaxPool is the maximum pooling, and f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

8. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 4 is characterized in that: When training the model, the time portion of the process includes: The time series sample data is input into a long short-term memory network, and weights are dynamically assigned to different environmental factors in a single time step through a factor attention mechanism, and weights are assigned to hidden states of each time step through a time attention mechanism to obtain a time feature output.

9. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 8 is characterized in that: The factor attention mechanism dynamically assigns weights to different environmental factors in a single time step, including: For time series data X=(X1,X2,…,X T )=(X 1 ,X 2 ,…,X M ) T ∈R T×M , T represents the number of time steps, M represents the number of features; in time step t, the kth feature matrix The attention weight of this factor at time t Calculated by the following formula: Where V f ∈R T , W f ∈R T×2p , U f ∈R T×T is the weight matrix of the multi-layer perceptron MLP, σ c is the activation function, b f ∈R T is the corresponding bias, h t-1 ∈R p and c t-1 ∈R p is the hidden state and unit state of the previous moment; The weighted input matrix, hidden state, and cell state of the LSTM network at time t are updated as follows: In the formula, σ s represents the softmax function, and f is a long short-term memory network unit; The weights are assigned to the hidden state of each time step through the temporal attention mechanism, including: The weights are assigned to the hidden states at each time step by the following calculation: Where H is the hidden state matrix, h k is the hidden state of the LSTM network unit in the kth step, s is the size of each hidden state, β is the normalized attention weight matrix, β i is the attention weight corresponding to the hidden state in step i, A t is a fully connected neural network, h at It is the time feature output of the final output of the long short-term memory network.

10. The method for early warning of power tower foundation instability based on spatiotemporal attention mechanism according to claim 1, characterized in that: The step of giving an early warning of the instability of the electric pole tower foundation according to the settlement prediction result and the early warning mechanism includes: Determine whether the settlement prediction result output by the settlement prediction model exceeds a preset alarm threshold; if so, issue an alarm signal and send the alarm information to the corresponding operation and maintenance engineer.

Images