Method for monitoring foundation pit through Internet of Things

Abstract

The invention relates to the technical field of civil construction, in particular to a method for monitoring a foundation pit through the internet of things, which comprises the following steps: collecting earth surface data of a construction area over the years; constructing a data driving model according to the earth surface data over the years, and obtaining a foundation pit deformation trend through the data driving model; predicting through a preset deep learning model according to the proportion of the weight of the foundation pit pre-filler to the weight of the finished building to obtain the settlement trend of the foundation pit; after construction of the foundation pit pre-filler is completed, a monitoring sensor is installed; data detected by the monitoring sensor is transmitted to the preset neural network for prediction, and the foundation pit deformation trend in the construction state and the foundation pit settlement trend in the construction state are obtained; comparing the deformation trend of the foundation pit with the deformation trend of the foundation pit in the construction state, and comparing the settlement trend of the foundation pit with the settlement trend of the foundation pit in the construction state; if a threshold is exceeded, an alarm is given. According to the invention, accurate monitoring of the foundation pit can be realized.

Description

Technical Field

[0001] This invention relates to the field of civil engineering construction technology, and more specifically to a method for monitoring foundation pits using the Internet of Things (IoT). Background Technology

[0002] With urban redevelopment and development, deep foundation pit projects are becoming increasingly common, with larger scales and dimensions. Examples include underground plazas, rail transit facilities, and tunnels. Deep foundation pit projects differ from general foundation pit projects, exhibiting strong individual characteristics. They are not only related to local engineering and hydrogeological conditions but also to the location, deformation resistance, and importance of adjacent buildings, structures, and municipal underground pipelines, as well as the surrounding site conditions. Therefore, deep foundation pit engineering is technically complex, involves a wide range of aspects, and is prone to accidents; thus, monitoring is essential during construction.

[0003] Currently, most deep foundation pit projects rely on manual monitoring, which suffers from drawbacks such as long data acquisition time, high manpower consumption, slow information feedback, and the inability to continuously monitor the deformation of the foundation pit support system. Conventional automated monitoring methods for deep foundation pits are constrained by factors such as visibility limitations and the influence of meteorological conditions on measurement accuracy, often resulting in incomplete monitoring during the process. Summary of the Invention

[0004] The purpose of this invention is to provide a method for monitoring foundation pits using the Internet of Things (IoT) to solve the above-mentioned problems and achieve accurate monitoring of foundation pits.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for monitoring foundation pits using the Internet of Things (IoT) includes the following steps: Collect historical surface data of the construction area; A data-driven model is constructed based on the historical surface data, and the deformation trend of the foundation pit is obtained through the data-driven model. The settlement trend of the foundation pit is predicted by using a pre-set deep learning model based on the ratio of the weight of the pre-filled material to the weight of the finished building. After the foundation pit prefilling is completed, monitoring sensors are installed; The data detected by the monitoring sensor is transmitted to the preset neural network for prediction, thereby obtaining the deformation trend and settlement trend of the foundation pit under construction conditions. Compare the deformation trend of the foundation pit with the deformation trend of the foundation pit under construction conditions, and compare the settlement trend of the foundation pit with the settlement trend of the foundation pit under construction conditions; if the threshold is exceeded, an alarm will be triggered.

[0006] Preferably, the step of predicting the historical surface data using a data-driven model to obtain the deformation trend of the foundation pit includes: A deep learning model is constructed and trained using historical land surface data to obtain a data-driven model. The deformation trend of the foundation pit is obtained by predicting the deformation data of the foundation pit under a previous fixed time through a data-driven model. The deformation trend of the foundation pit is used as the maximum deformation, and a deformation threshold is set based on the deformation trend of the foundation pit.

[0007] Preferably, the step of predicting the settlement trend of the foundation pit based on the ratio of the weight of the prefill material to the weight of the finished building using the preset deep learning model includes: Based on the expected state after the prefill material is filled into the foundation pit according to its weight, the pre-set deep learning model is used to predict the settlement trend of the filled foundation pit and obtain the predicted settlement trend. Based on the expected state of the finished building weight, the pre-set deep learning model is used to predict the settlement trend of the foundation pit after completion, and the settlement trend after completion is obtained. The settlement trend of the foundation pit is obtained based on the settlement trend of the filled foundation pit, the settlement trend of the foundation pit after completion, and the ratio of the weight of the prefill material to the weight of the finished building. Compare the stated settlement trend, predicted settlement trend, and completed settlement trend of the foundation pit; ensure that the stated settlement trend of the foundation pit < the predicted settlement trend < the completed settlement trend.

[0008] Preferably, the monitoring sensors include a settlement meter, a soil pressure sensor, a water level sensor, an inclinometer, an axial force gauge, and a pore water pressure gauge.

[0009] Preferably, the step of transmitting the data detected by the monitoring sensor to the preset neural network for prediction to obtain the deformation trend and settlement trend of the foundation pit under construction conditions includes: The data from the monitoring sensors, including settlement gauges, soil pressure sensors, water level sensors, inclinometers, axial force gauges, and pore water pressure gauges, are input into the preset neural network for prediction, thereby obtaining the deformation trend and settlement trend of the foundation pit under construction conditions.

[0010] Preferably, the step of comparing the deformation trend of the foundation pit with the deformation trend of the foundation pit under construction conditions, and comparing the settlement trend of the foundation pit with the settlement trend of the foundation pit under construction conditions; and triggering an alarm when the threshold is exceeded, includes: If the deformation trend of the foundation pit under the construction state is less than the deformation threshold, no alarm will be triggered; otherwise, an alarm will be triggered. If the settlement trend of the foundation pit under construction conditions is less than the settlement trend of the foundation pit, no alarm will be triggered; otherwise, an alarm will be triggered.

[0011] Preferably, the data-driven model includes a plurality of spatiotemporal blocks and an output layer connected in sequence, wherein the spatiotemporal blocks are provided with a spatial diffusion layer and a state transition layer connected in sequence, wherein the spatial diffusion layer adopts a graph convolutional network layer and the state transition layer adopts a graph convolutional gated recurrent unit.

[0012] Preferably, the preset neural network model adopts an encoder-decoder network structure, with different expected states as condition vectors input into the decoder structure, and the encoder's input data being historical subsidence trend data.

[0013] The present invention has the following technical effects: This invention provides an IoT-based method for monitoring foundation pits. Through multi-source data fusion and intelligent predictive analysis, it achieves accurate monitoring and early warning of foundation pit deformation and settlement trends. The method first collects historical surface data of the construction area to construct a data-driven model to predict foundation pit deformation trends. Then, combining the weight ratio of pre-filled material to finished building material, a pre-set deep learning model predicts foundation pit settlement trends, forming a baseline reference before construction. During construction, the system deploys various monitoring devices, including settlement meters, earth pressure sensors, water level sensors, and tiltmeters, to collect real-time foundation pit status data. This data is then input into a pre-set neural network for dynamic prediction, obtaining deformation and settlement trends under construction conditions.

[0014] By comparing real-time prediction results with preset thresholds, the system can issue timely alarms when deformation or settlement exceeds safe limits, effectively preventing risks in foundation pit engineering. This method innovatively combines historical data-driven modeling, multi-condition settlement prediction, and real-time IoT monitoring, overcoming the shortcomings of traditional manual monitoring such as low efficiency and slow feedback. It significantly improves the automation and accuracy of foundation pit monitoring, providing full-process, intelligent safety assurance for deep foundation pit construction, and has good engineering applicability and promotional value. Detailed Implementation

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments.

[0017] This embodiment provides a method for monitoring a foundation pit using the Internet of Things (IoT), including the following steps: Collect historical surface data of the construction area; A data-driven model is constructed based on the historical surface data, and the deformation trend of the foundation pit is obtained through the data-driven model. The settlement trend of the foundation pit is predicted by using a pre-set deep learning model based on the ratio of the weight of the pre-filled material to the weight of the finished building. After the foundation pit prefilling is completed, monitoring sensors are installed; The data detected by the monitoring sensor is transmitted to the preset neural network for prediction, thereby obtaining the deformation trend and settlement trend of the foundation pit under construction conditions. Compare the deformation trend of the foundation pit with the deformation trend of the foundation pit under construction conditions, and compare the settlement trend of the foundation pit with the settlement trend of the foundation pit under construction conditions; if the threshold is exceeded, an alarm will be triggered.

[0018] Further optimization of the plan involves using data-driven models to predict historical surface data and obtain the deformation trend of the foundation pit, including the following steps: After processing historical surface data through a data-driven model, the deformation trend of the foundation pit is predicted. The deformation trend of the foundation pit is used as the maximum deformation variable, and a deformation variable threshold is set based on the deformation trend of the foundation pit.

[0019] The prediction methods of data-driven models mainly include the following: Collect historical surface data, including historical surface deformation monitoring data (settlement, horizontal displacement), foundation pit design drawings, and geological survey reports.

[0020] Missing and obvious outliers in historical surface deformation monitoring data were processed. Missing values ​​were filled using interpolation, while obvious outliers were identified and removed based on statistical methods or engineering experience.

[0021] The historical surface deformation monitoring data of all the above monitoring points are aligned with a unified timestamp to form a regular spatiotemporal data. Under this spatiotemporal data, the constructed deep learning model is trained and optimized to generate a usable data-driven model. The training process is optimized in a data-driven manner, that is, the deep learning model is trained and its parameters are optimized using historical data to achieve data-driven operation. This allows the deep learning model to learn the data trends and characteristics in historical data. The training process can be implemented conventionally and will not be elaborated here.

[0022] The aforementioned deep learning model primarily uses surface deformation monitoring data from a previous fixed time period to predict data for subsequent fixed time periods. This involves using the previous fixed-time surface deformation monitoring data as the basis for constructing a graph structure. Based on the spatiotemporal data and corresponding spatial relationships, a graph structure is transformed into a corresponding graph structure. This graph structure is then used to comprehensively process the monitoring data across time and space, and within this graph structure representing spatial structural relationships, predictions are made for future deformation monitoring data. The surface deformation monitoring data from a previous fixed time period (e.g., 15 days) is acquired through data collection at different locations using relevant deformation sensor equipment.

[0023] For the graph structure described above, each monitoring point is set as a node in the graph structure, and a corresponding feature vector is assigned to each node. The feature vector includes the cumulative settlement at time t, the horizontal displacement at time t, the soil type at that point (e.g., clay, sand, using unique thermal encoding), and the shortest distance to the center point of the excavation face. For the connections between each node, corresponding edges in the graph structure are set, and an adjacency matrix is ​​constructed. Define the edges of the graph (construct the adjacency matrix A): Based on spatial proximity, emphasize spatial proximity: ; Based on node similarity, the similarity of deformation behavior is emphasized: ; in, This represents the elements of the adjacency matrix, corresponding to the relationships between nodes. Indicates the corresponding spatial distance. This indicates the Gaussian kernel bandwidth, which is preset. This indicates the similarity between nodes.

[0024] You can choose any one or two of the above methods to calculate the elements of the adjacency matrix using a weighted sum, with the first method being preferred.

[0025] The graph structure described above includes a set of nodes, the feature vectors of each node in the set at different times, and a set of edges (adjacency matrix).

[0026] To address the aforementioned graph structure processing, a deep learning network capable of simultaneously capturing spatial diffusion and temporal evolution is constructed. This deep learning network approach captures a model structure that considers the relationship between the starting point and the evolution point in foundation pit deformation.

[0027] The deep learning model is formed by stacking several spatiotemporal blocks. Each spatiotemporal block contains a spatial diffusion layer and a state transition layer. The spatial diffusion layer uses a graph convolutional network layer. Its function is to allow each node to aggregate information from its neighbors, simulating the propagation process of deformation in space. The state transition layer uses a graph convolutional gated recurrent unit. Its function is to simulate the evolution of each node's state (such as deformation development) after receiving information from its neighbors. The "gating" mechanism (update gate, reset gate) determines the degree of forgetting of historical states and the acceptance of new information.

[0028] The relevant data processing methods in the spatial diffusion layer are as follows: The input data for the spatiotemporal block is the node features output from layer (l-1). Where N represents the number of monitoring points (nodes) and Fin represents the dimension of the input features.

[0029] First, the adjacency matrix is ​​preprocessed to construct a normalized adjacency matrix with self-loops, which is used to control the strength of information propagation. The self-loops are added first to ensure that each node retains its own characteristics when aggregating neighbor information. ; Where A represents the adjacency matrix, and its elements are... Represents the identity matrix. This represents a self-loop adjacency matrix.

[0030] Calculate the degree (number of connections) of each node: ; in, This represents the total connection strength of node i.

[0031] Symmetric normalization is applied to the adjacency matrix to prevent changes in feature scale during propagation. ; in, This is the normalized adjacency matrix.

[0032] Then, relevant graph convolution operations are performed on the input node features to perform actual neighbor information aggregation and feature transformation.

[0033] During the neighbor information aggregation process, the normalized adjacency matrix is ​​multiplied by the input features: ; in, This represents the output characteristics of the spatiotemporal block of the previous layer. This represents the aggregated node feature matrix. Each node's new feature is a weighted average of its own features and those of all its neighboring nodes. The weights are determined by the adjacency matrix; nodes that are closer or more correlated have larger weights. This characterizes the potential diffusion of deformed states in other neighboring nodes caused by different nodes.

[0034] Feature dimension transformation is performed using a trainable weight matrix, and a bias term is added to each output feature: ; in, This represents the node features after graph convolution processing. This represents the graph convolution weight parameters. This represents the bias term, which is trainable and learns how to combine different input feature channels through weight parameters.

[0035] After graph convolution, nonlinear activation is performed using an activation function to increase nonlinearity and alleviate the gradient vanishing problem. ; This indicates the output data of the current spatial diffusion layer: The state transition layer employs a graph convolutional gated recurrent unit (GC-GRU), whose input is the output data of the spatial diffusion layer of the simultaneous empty block. The state transition layer is based on a graph convolutional gated recurrent unit (GC-GRU) and includes four core computational steps: First, perform a door reset calculation: ; Current spatial characteristics Perform graph convolution transformation on historical states Perform graph convolution transformation, add the two results, activate them using the Sigmoid activation function, and output a reset gating signal. This operation determines how many historical states to forget, which is used to calculate candidate states. Among the above... This represents the weights of the input data to the gate structure, the weights of the historical data to the gate structure, and the bias term. The subscripts r, z, and h indicate that the above parameters correspond to the reset gate, update gate, and candidate state, respectively. t represents the corresponding time step number, and l represents the corresponding layer number.

[0036] Next, perform the update gate calculation: ; Similar to the reset gate structure, but using independent weight parameters. This operation determines how much historical state is retained to control the state update ratio.

[0037] Then, candidate state calculation is performed: ; Perform graph convolution transformation on the current spatial features, and then process the reset historical state. Perform graph convolution transformation, add the two results together, activate them with tanh, and output candidate states. This operation generates a new state proposal based on the current input and filtered historical information.

[0038] Finally, update the status: ; Use updated gating signal The final new state is output by taking a weighted average of the historical states and candidate states. This operation integrates historical memories with new changes to generate the complete state at the current moment.

[0039] The above sequence is predicted using several spatiotemporal blocks, and the final predicted structure is output through an output layer such as a fully connected layer. This yields the foundation pit deformation trend, which is a normal deformation trend conforming to historical patterns. Under normal construction conditions, the deformation trend meets the requirements of historical patterns. The values ​​at the corresponding time points in this data are used as relevant thresholds for subsequent detection of abnormal foundation pit deformation when construction conditions do not meet requirements.

[0040] Further optimization of the plan involves using a pre-set deep learning model to predict the settlement trend of the foundation pit based on the ratio of the weight of the pre-filled material to the weight of the finished building. This includes the following steps: Based on the expected state after the pre-filled material is filled into the foundation pit according to its weight, a preset deep learning model is used to predict the settlement trend of the filled foundation pit, thereby obtaining the predicted settlement trend; this step is used to predict the expected settlement trend of the filling.

[0041] Based on the expected weight of the finished building, a pre-set deep learning model is used to predict the settlement trend of the foundation pit after completion, thus obtaining the settlement trend after completion; this step is used to predict the settlement trend after the completion of construction.

[0042] The settlement trend of the foundation pit is obtained based on the settlement trend of the filled foundation pit, the settlement trend of the foundation pit after completion, and the ratio of the weight of the pre-filled foundation pit to the weight of the finished building. Compare the settlement trend of the foundation pit, the predicted settlement trend, and the settlement trend upon completion; ensure that the settlement trend of the foundation pit is less than the predicted settlement trend, which is less than the settlement trend upon completion.

[0043] The preset deep learning model adopts an encoding-decoding structure. In the encoding process, the historical settlement trend of the monitoring point is first encoded based on a certain period of time (set according to the time corresponding to the expected state, with a fixed time). After encoding, the expected state under different loads is injected into the encoded content to perform relevant state evolution to generate the corresponding state evolution results. The state evolution results are used as the input of the decoder to generate the final settlement trend data. The settlement trend data will have different settlement trend prediction results due to different expected load states, in order to adapt to the different task requirements.

[0044] In this process, the settlement trend data sequence of historical foundation pits (data from a fixed preceding time period) is mapped by an encoder to represent latent variables that represent their local geomechanical state. For settlement monitoring point i in the foundation pit, the input is a sequence of settlement deformation data over its past k time steps: , where F represents the feature dimension (settlement value, displacement deformation, etc.); In the encoder, local short-term patterns (such as abrupt changes in deformation rate) in the sequence are captured through one-dimensional causal convolutional layers. This ensures that the length of the output sequence remains constant, preventing future information leakage.

[0045] The one-dimensional causal convolutional layer is followed by a gated recurrent unit (GRU). This GRU captures long-term dependencies in the sequence and generates the final sequence summary. The GRU outputs the final hidden state. This serves as the encoding for the sequence. The data processing method for the model structure of one-dimensional causal convolutional layers and gated recurrent units (GRUs) is implemented conventionally and will not be elaborated upon.

[0046] Then the output of GRU Mapping to a low-dimensional latent space using a linear transformation layer and the tanh activation function; ; in, The latent variable representing the local state of point i. This represents the weights and biases in the linear transformation layer.

[0047] Obtain the set of latent variables for all monitoring points .

[0048] The expected state of the load (the expected state after the prefill material is filled into the foundation pit, and the expected state of the finished building weight) is used as a global condition, affecting the local state of all monitoring points. Conditional affine transformation is used for fusion. First, the expected state (load weight) is encoded into a condition vector c using an MLP. Then, the condition vector c is used to generate a unique pair of transformation parameters for each latent variable. : First, perform a linear mapping on the condition vector c to obtain the mapping result. : ; After mapping, reshaping is performed using the reshaping function. Implementation: ; in, This indicates reshaping the tensor dimension. This indicates the batch dimension size, and N represents the total number of nodes. Indicates double the feature dimension; Segmentation after reshaping Along the last dimension of the tensor, the tensor is evenly divided into two smaller tensors; ; in, It is a scaling parameter matrix, containing the scaling parameters of the dz-dimensional hidden states of all N nodes, for each node i. B is the offset parameter matrix, containing the offset adjustment parameters for the dz-dimensional hidden states of all N nodes, for each node i. Scale parameters Depending on the expected state, the importance of certain state components can be increased or decreased. For example, under high load expectations, the state dimension related to "compressibility" can be increased. Offset parameter Based on the expected state, directly change the baseline of the state. For example, directly increase the value of the state dimension related to "settlement potential".

[0049] Z is the original hidden state matrix, with the shape of... , and All dimensions are ,according to and The elements of the original hidden state matrix are adjusted to obtain the corrected hidden state. : ; For the corrected hidden state, state evolution is performed through a spacetime interactor: The spatiotemporal interactor consists of several sequentially connected interaction layers. Within a single interaction layer: The corrected hidden state is calculated using an attention weight calculation layer: First, an adjacency matrix is ​​constructed based on the distances between different nodes. Different elements in the adjacency matrix represent the distances between different nodes, and these distances are spatial distances. The node distances in the adjacency matrix are then mapped and encoded using a multilayer perceptron to obtain the encoded distances. The representation of the mapping encoding is implemented using a multilayer perceptron (MLP). In the attention weight calculation layer: ; in, This represents the hidden state of a node, with subscripts i and j indicating the corresponding nodes i and j. This represents a trainable weight matrix used to generate queries and keys. This represents a learnable attention vector. This represents a distance encoding function that maps spatial distance to a high-dimensional space. This represents a vector concatenation operation. Let i represent the set of neighbors of node i. This represents the attention score. This represents the attention weight.

[0050] Then, based on the attention weights, the corresponding node aggregation information is calculated: ; in, This represents the message transformation weight matrix. This represents an activation function, such as ReLU or Tanh. This represents the aggregated message received by node i.

[0051] The aggregated message and the hidden features of the node itself are processed by a gated recurrent unit (GRU) to perform state evolution and obtain the evolved hidden state. : ; The network structure and data processing flow of the gated recurrent unit (GRU) will not be elaborated here, as they can be implemented using conventional techniques.

[0052] After generating the evolved hidden state, the process enters the interaction layer for repeated aggregation and evolution. This evolution occurs once or several times to aggregate information from more neighboring nodes. Multiple rounds of interaction generate an accurate evolution result from the multiple evolutionary results, and the final output hidden state is... ; The final hidden state described above is input into the decoder network structure. The decoder network decodes the hidden state to generate the final predicted future subsidence trend. The final hidden state after the interaction evolution is mapped back to the physical world to obtain the subsidence prediction trend result. The decoder adopts a fully connected neural network structure, wherein: First, feature extraction is performed using a fully connected mapping layer: : Represents the weight matrix. Let H represent the bias vector, and H represent the enhanced features.

[0053] Then perform a linear mapping: ; Represents the weight matrix. Let T represent the bias vector and T represent the time-spread feature.

[0054] Then, dimensional reshaping is performed: ; Where B represents the number of foundation pits, N represents the number of detection points, and m represents the prediction time step. Indicates the internal feature dimension.

[0055] in, This represents the reshaping result. Then, a fully connected layer is used for the final multi-step output prediction: ; in, This represents the final output of the settlement prediction result. Represents the weight matrix. This represents the bias vector.

[0056] After obtaining the settlement trend of the filled foundation pit, the settlement trend of the foundation pit after completion, and the ratio of the weight of the pre-filled foundation pit to the weight of the finished building, the settlement trend of the filled foundation pit and the settlement trend of the foundation pit after completion are used as basic data. The ratio of the weight of the pre-filled foundation pit to the weight of the finished building is used as the first weight parameter. The basic data is weighted and calculated to generate the foundation pit settlement trend. The initial data is calculated by multiplying the first weight parameter by the settlement trend of the filled foundation pit and the settlement trend of the foundation pit after completion by (1-first weight parameter). The initial data is adjusted by multiplying the initial data by a percentage (e.g., 90%, which is further reduced if the constraint conditions are not met). That is, the initial data is scaled by an adjustable scaling factor (e.g., the initial value is 90%, which is gradually reduced if the prediction result does not meet the constraint conditions) to generate the final foundation pit settlement trend. The settlement content at the time corresponding to the foundation pit settlement trend is used as the final threshold content.

[0057] The scheme has been further optimized, and the monitoring sensors include settlement gauges, soil pressure sensors, water level sensors, inclinometers, axial force gauges, and pore water pressure gauges.

[0058] Further optimization of the scheme involves transmitting data detected by monitoring sensors to a pre-set neural network for prediction, obtaining the deformation trend and settlement trend of the foundation pit under construction conditions. This includes the following steps: Data from monitoring sensors, including settlement gauges, soil pressure sensors, water level sensors, inclinometers, axial force gauges, and pore water pressure gauges, are input into a preset neural network for prediction, thereby obtaining the deformation trend and settlement trend of the foundation pit under construction conditions.

[0059] This can be achieved through a pre-set neural network. The aforementioned monitoring sensors can collect data from various aspects. All of these data are strongly correlated with the deformation trend and settlement trend of the foundation pit under construction conditions. Based on the data from various aspects, the neural network model can make detailed predictions of the aforementioned trends.

[0060] Further optimize the plan by comparing the deformation trend of the foundation pit with that of the foundation pit under construction, and comparing the settlement trend of the foundation pit with that of the foundation pit under construction; the steps to trigger an alarm when the threshold is exceeded include: If the deformation trend of the foundation pit during construction is less than the deformation threshold, no alarm will be triggered; otherwise, an alarm will be triggered. If the settlement trend of the foundation pit during construction is less than the settlement trend of the foundation pit, no alarm will be triggered; otherwise, an alarm will be triggered.

[0061] Furthermore, historical surface data of the construction area are collected, and in this embodiment, daily surface settlement values ​​for past periods are collected. Read design parameters: excavation depth of foundation pit, stiffness of support structure, depth of groundwater level, weight of soil, weight of prefill material, and weight of finished building. Use settlement gauge, soil pressure sensor, water level sensor, inclinometer, axial force gauge, and pore water pressure gauge to monitor the real-time deformation of foundation pit. The collection cycle for historical surface data of the construction area is within the same cycle as the current construction cycle, environment, and season; the historical surface data is mainly used to train the aforementioned preset neural network. In the preset neural network: the input data consists of different monitoring data collected by sensor devices.

[0062] The data is input into a one-dimensional convolutional neural network and processed through two layers of one-dimensional convolution. First layer: 64 convolutional kernels, kernel size 3; The second layer: 128 convolutional kernels, with a kernel size of 3; By using adaptive max pooling and compressing the length to 30, local change patterns in the sedimentation time series are extracted, such as sudden drops, gradual changes, and fluctuations. The output of the one-dimensional convolutional neural network is then input into a bidirectional LSTM recurrent neural network, which is configured with forward and reverse directions to read the sequence simultaneously. Two layers are stacked vertically, resulting in a total of four independent LSTM networks (two forward layers and two backward layers). Hidden unit: 128 internal state dimensions per layer and per direction; Temporal characteristics: For each time point, the forward and reverse directions are directly concatenated to obtain 128+128=256-dimensional data; Parameter characteristics: The design parameters, such as the excavation depth of the foundation pit, the stiffness of the support structure, the depth of the groundwater level, the weight of the soil, the weight of the prefill material, and the weight of the finished building, are passed through a fully connected layer to output a 64-dimensional vector. The activation function is ReLU, and Dropout (0.2) is added to prevent overfitting.

[0063] The temporal and parametric features are concatenated into a 320-dimensional array, processed through two fully connected layers. The first layer outputs a 128-dimensional array with ReLU activation. The second layer outputs a 2-dimensional array with no activation (regression task), showing the maximum deformation and cumulative settlement over the next 7 days under construction conditions. These two predicted values ​​are then compared to a set threshold for evaluation.

[0064] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the indicated orientation or positional relationship, and are only for the convenience of describing this invention, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0065] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method of monitoring a foundation pit by an Internet of Things, characterized by, The method comprises the following steps: Collecting ground surface data of construction area in previous years; Building a data-driven model according to the ground surface data in previous years, and obtaining a foundation pit deformation trend through the data-driven model; According to the ratio of the weight of the foundation pit pre-filler to the weight of the finished building, the preset deep learning model is used for prediction to obtain a foundation pit settlement trend; After the construction of the foundation pit pre-filler is completed, a monitoring sensor is installed; The data detected by the monitoring sensor is transmitted to the preset neural network for prediction to obtain a construction state foundation pit deformation trend and a construction state foundation pit settlement trend; The foundation pit deformation trend and the construction state foundation pit deformation trend are compared, and the foundation pit settlement trend and the construction state foundation pit settlement trend are compared; If the threshold is exceeded, an alarm is given.

2. The method of claim 1, wherein, The step of predicting the ground surface data in previous years through the data-driven model to obtain the foundation pit deformation trend comprises: Building a deep learning model, training the deep learning model through the ground surface data in previous years to obtain the data-driven model; The foundation pit deformation data at a fixed time in the previous sequence is predicted through the data-driven model to obtain the foundation pit deformation trend, which is used as the maximum deformation value, and a deformation threshold is set based on the foundation pit deformation trend.

3. The method of claim 2, wherein, The step of predicting the foundation pit settlement trend through the preset deep learning model according to the ratio of the weight of the foundation pit pre-filler to the weight of the finished building comprises: According to the expected state of the foundation pit after the foundation pit pre-filler is filled, the preset deep learning model is used to predict the filling foundation pit settlement trend to obtain a predicted settlement trend; According to the expected state of the weight of the finished building, the preset deep learning model is used to predict the post-completion foundation pit settlement trend to obtain a post-completion settlement trend; According to the filling foundation pit settlement trend, the post-completion foundation pit settlement trend and the ratio of the weight of the foundation pit pre-filler to the weight of the finished building, the foundation pit settlement trend is obtained; The foundation pit settlement trend, the predicted settlement trend and the post-completion settlement trend are compared to ensure that the foundation pit settlement trend is less than the predicted settlement trend which is less than the post-completion settlement trend.

4. The method of claim 3, wherein, The monitoring sensor comprises a settlement meter, a soil pressure detection sensor, a water level monitoring sensor, an inclinometer, an axial force meter and a pore water pressure meter.

5. The method of claim 4, wherein, The step of transmitting the data detected by the monitoring sensor to the preset neural network for prediction to obtain the construction state foundation pit deformation trend and the construction state foundation pit settlement trend comprises: The data of the monitoring sensor including the settlement meter, the soil pressure detection sensor, the water level monitoring sensor, the inclinometer, the axial force meter and the pore water pressure meter is input into the preset neural network for prediction to obtain the construction state foundation pit deformation trend and the construction state foundation pit settlement trend.

6. The method of claim 4, wherein, The step of comparing the foundation pit deformation trend and the construction state foundation pit deformation trend, and comparing the foundation pit settlement trend and the construction state foundation pit settlement trend; The step of comparing the foundation pit deformation trend and the construction state foundation pit deformation trend, and comparing the foundation pit settlement trend and the construction state foundation pit settlement trend; The step of comparing the foundation pit deformation trend and the construction state foundation pit deformation trend, and comparing the foundation pit settlement trend and the construction state foundation pit settlement trend; If the threshold is exceeded, an alarm is given. If the construction state foundation pit deformation trend is less than the deformation threshold, no alarm is given, otherwise an alarm is given; If the construction state foundation pit settlement trend is less than the foundation pit settlement trend, no alarm is given, otherwise an alarm is given.

7. The method of claim 2, wherein, The data-driven model comprises a plurality of spatio-temporal blocks and an output layer connected in sequence, wherein the spatio-temporal blocks are provided with a spatial diffusion layer and a state transition layer connected in sequence, wherein the spatial diffusion layer adopts a graph convolution network layer, and the state transition layer adopts a graph convolution gated recurrent unit.

8. The method of claim 3, wherein, The preset neural network model adopts an encoder-decoder network structure, different expected states are input into the decoder structure as a condition vector, and the input data of the encoder is historical settlement trend data.