Foundation pit prediction model training method, foundation pit monitoring method and system

Abstract

The present application belongs to the technical field of federated learning, and discloses a foundation pit prediction model training method, a foundation pit monitoring method and a system. The training method comprises: local training of the edge device, so that the model predicts the foundation pit state of the future period based on the foundation pit information of the current period; the local training loss is a comprehensive loss considering the data fitting loss and the physical constraint loss; the edge device also maps the foundation pit geological information into a geological embedding vector, and sends the local model update parameter and the geological embedding vector to the cloud device; the cloud server clusters all the edge devices based on the similarity of the geological embedding vectors, aggregates the same local model update parameters to obtain the corresponding global model update parameters; during the aggregation calculation, the data quality evaluation score of the edge device is used as the weight coefficient of the corresponding local model update parameter. Based on the above method, the accuracy of the model prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field related to federated learning, and more particularly, relates to a foundation pit prediction model training method, a foundation pit monitoring method and system. BACKGROUND

[0002] Foundation pit monitoring is a key link to ensure the safety of engineering construction. By continuously collecting and analyzing multi-source data such as foundation pit displacement, settlement and stress, the future deformation trend and potential risks can be predicted, so that preventive measures can be taken in advance to avoid engineering accidents. At present, the existing technology attempts to introduce large models or deep learning methods to process the foundation pit monitoring task, which has improved the automation level of data processing to a certain extent and reduced the cost and time of manual analysis. However, in practical application, this kind of method still has several significant defects, which leads to limited prediction accuracy and makes it difficult to fully meet the reliability and accuracy requirements in complex engineering environments.

[0003] Therefore, how to effectively improve the accuracy of foundation pit prediction while maintaining the advantages of automation and efficient processing is still a key problem to be solved in the current technical field. SUMMARY

[0004] In view of the above defects or improvement needs of the prior art, the present application provides a foundation pit prediction model training method, a foundation pit monitoring method and system, which aims to improve the accuracy of foundation pit prediction.

[0005] To achieve the above-mentioned purpose, the following technical solutions are proposed.

[0006] According to a first aspect of the present application, a foundation pit prediction model training method based on federated learning is provided, and each round of federated training includes the following steps:

[0007] The cloud server issues global model update parameters to the edge device;

[0008] The edge device determines a local model based on the global model update parameters it receives, and performs local training on it using local data, so that the model can predict the future foundation pit state based on the current period of foundation pit information, and obtain local model update parameters; the foundation pit state includes foundation pit horizontal displacement and foundation pit lateral earth pressure; the local training loss is a comprehensive loss considering data fitting loss and physical constraint loss, wherein the data fitting loss represents the similarity deviation between the prediction result and the true label, and the physical constraint loss represents the degree of violation of mechanical equilibrium of the foundation pit support structure based on the prediction result; the edge device also maps the foundation pit geological information extracted during the collection of local data into a geological embedding vector, and sends the local model update parameters and the geological embedding vector to the cloud device;

[0009] The cloud server clusters all edge devices based on the similarity of the geological embedding vectors of the edge devices, aggregates the local model update parameters of the edge devices belonging to the same cluster set to obtain global model update parameters corresponding to the cluster set, and uses the data quality evaluation score of the edge device as the weight coefficient of the corresponding local model update parameter when performing the aggregation calculation, wherein the data quality evaluation score of the edge device is the cumulative of the local data integrity score when the edge device performs the current round of local training, the geological matching degree score between the edge device and the cluster center of the cluster set to which the edge device belongs, and the physical constraint consistency score obtained by the edge device verifying the local model trained by the edge device using the local verification set thereof, the smaller the physical constraint loss, the higher the physical constraint consistency score, the obtained global model update parameters are used as the global model update parameters issued to the edge devices in the next round of federated training, and the global model update parameters corresponding to different cluster sets are distributed to the edge devices belonging to the corresponding cluster set.

[0010] According to a second aspect of the present application, a cloud-edge collaborative foundation pit monitoring method is provided, comprising:

[0011] The edge device obtains the foundation pit information of the current period from the collection terminal and inputs it into the local lightweight model to obtain the prediction result of the foundation pit state of the future period, wherein the foundation pit state includes the horizontal displacement of the foundation pit and the lateral earth pressure of the foundation pit, and the local lightweight model is the local foundation pit monitoring model obtained by using the training method described above; the local foundation pit monitoring model outputs the probability of different prediction values of the foundation pit state, and the prediction value corresponding to the maximum probability is taken as the prediction result of the foundation pit state; the edge device also calculates the information entropy H(x) of the foundation pit state x based on the probability of different prediction values of the foundation pit state x output by the local foundation pit monitoring model, and if any of the information entropy H(x) or the prediction result exceeds the corresponding threshold value, it is determined that there is a risk, and a risk emergency operation is triggered; the risk emergency operation includes informing the collection terminal to increase the collection frequency and sending the high-frequency foundation pit monitoring data collected by the collection terminal to the cloud server, wherein the high-frequency foundation pit monitoring data are the data collected after increasing the collection frequency.

[0012] The cloud server runs the full-amount physical large language model deployed by it to perform in-depth analysis on the received high-frequency foundation pit monitoring data, and generates a risk report based on the analysis result.

[0013] According to a third aspect of the present application, a cloud-edge collaborative foundation pit monitoring system is provided, comprising edge devices and a cloud server, and the edge devices and the cloud server are used to cooperatively implement the method described above.

[0014] Overall, compared with the prior art, the above technical solutions conceived by the present application mainly have the following beneficial effects:

[0015] 1. In the local model training of the present application, the introduced loss considers the physical constraint loss, which will produce a high penalty gradient, forcing the model parameters to converge to the physical manifold that meets the "coordination of force and deformation" in the back propagation, so that the model learns the essential laws of geotechnical mechanics, not just statistical fitting at the digital level, that is, it can make the predicted displacement cumulative value and soil pressure change as much as possible to meet the equilibrium equation of the elastic foundation beam, reduce the prediction results that violate physical common sense, and thus improve the prediction accuracy. Moreover, in the cloud aggregation of the present application, the edge devices are first aggregated based on geological characteristics, and the foundation pits with similar geological characteristics are classified into the same cluster set, and then the local model update parameters in each cluster set are aggregated to obtain the global model update parameters corresponding to the cluster set, and then the global model update parameters aggregated by the cluster are correspondingly issued to each edge device in the cluster aggregation, so that the interference of heterogeneous geological data can be avoided, which is conducive to model learning and improves the prediction accuracy of the model. In addition, in the cloud aggregation of the present application, the cumulative score of the local data integrity score, the geological matching degree score and the physical constraint consistency score is used as the weight coefficient of the corresponding local model update parameter. The above dynamic adaptive weight fusion strategy can give higher voice to high-quality and high-matching construction sites, so as to accelerate the model convergence speed and improve the model prediction accuracy. Overall, the foundation pit prediction model training method based on federated learning proposed in the present application can effectively improve the prediction accuracy of the model.

[0016] 2. Further, in specific embodiments, a "work condition text-collected data" multi-modal encoding vector is constructed as an input sample, and by inputting the construction work condition text, the model will consider the influence of the construction work condition when making a prediction, and identify that the abnormality may come from external factors such as construction disturbance rather than a dangerous situation, thereby reducing the false positive rate and improving the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a step flow chart of the foundation pit prediction model training method based on federated learning in an embodiment of the present application.

[0018] Figure 2 is a step flow chart of the foundation pit monitoring method based on cloud-edge collaboration in an embodiment of the present application.

[0019] Figure 3 is a system architecture diagram of cloud-edge collaboration in an embodiment of the present application.

[0020] Figure 4 is an information interaction diagram of cloud-edge collaboration in combination with model training and application stages in an embodiment of the present application.

[0021] Figure 5 is a flow chart of data processing by the edge device in an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application. In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0023] In a first aspect, the present application claims a pit prediction model training method based on federated learning.

[0024] As Figure 1 The figure in the drawing is a step flow chart of the pit prediction model training method based on federated learning in an embodiment of the present application. The step sequence in the drawing is only one possible way provided, but is not limited thereto. The method will be described in detail below in combination with Figure 1

[0025] S11, the cloud server issues global model update parameters to the edge device.

[0026] S12, the edge device determines a local model based on the global model update parameters received by it, and uses local data to perform local training on it, so that the model predicts the pit state in the future period based on the pit information in the current period, and obtains local model update parameters; the pit state includes pit horizontal displacement and pit lateral earth pressure; the local training loss is a comprehensive loss considering data fitting loss and physical constraint loss, wherein the data fitting loss represents the similarity deviation between the prediction result and the true label, and the physical constraint loss represents the degree of violation of mechanical balance of the pit support structure based on the prediction result; the edge device also maps the pit geological information extracted during the collection of local data into a geological embedding vector, and sends the local model update parameters and the geological embedding vector to the cloud device.

[0027] At the edge device, two types of operations need to be performed, the first type of operation is to obtain local data to train a local model and obtain local model update parameters; the second type of operation is to extract contemporaneous pit geological information and map it into a geological embedding vector. After obtaining the local model update parameters and the geological embedding vector, they are packaged and sent to the remote server.

[0028] The process of obtaining the geological embedding vector is as follows.

[0029] ​Specifically, the edge device can extract the foundation pit geological information from the geological survey report of the foundation pit in the current period, including cohesion, internal friction angle, compression modulus, water content and lithology code, and then map it to a geological embedding vector through a pre-trained geological encoder. The pre-trained geological encoder can use a multi-layer perceptron (MLP) or a variational autoencoder (VAE) to be pre-trained using public geological survey datasets, so that it can map discrete geological parameters to a high-dimensional continuous vector space in which samples with similar geological properties are closer. For example, a foundation pit located in a coastal soft soil area, the edge device extracts geological feature information by analyzing the geological survey report:

[0030] ;

[0031] In the formula, g is the geological feature information.

[0032] Then, the pre-trained geological encoder is used to map the above geological feature information G to a standardized embedding vector, which is the geological embedding vector.

[0033] The process of obtaining local data and training a local model is as follows.

[0034] Specifically, the collection terminal is the sensing arm of the Internet of Things, and the edge device can obtain the foundation pit monitoring data collected by each collection terminal to obtain local data. The foundation pit monitoring data includes, for example, foundation pit displacement, soil pressure and soil inclination.

[0035] The edge device obtains the collected foundation pit monitoring data by communicating with the collection terminal, thereby constructing local data for local training. The goal of training the local model is to enable the local model to predict future period foundation pit information based on current period foundation pit information, so as to perform risk control in advance and avoid causing major problems.

[0036] In an embodiment, the current period foundation pit information input into the model is a multi-modal encoding vector, and the construction process of the multi-modal encoding vector includes:

[0037] Obtain the foundation pit monitoring data collected by the collection terminal in the current period, and encode the foundation pit monitoring data at each collection time according to the three elements of value, rate of change and time interval into a sensor time sequence Token sequence;

[0038] Obtain the construction condition text information in the current period and convert it to a context vector through a pre-trained text encoder;

[0039] The context vector is concatenated as a prefix prompt before the sensor time sequence Token sequence to obtain a multi-modal encoding vector.

[0040] In a specific operation, the system first encodes the parsed sensor time series data into a time series Token sequence according to the three elements of value, rate of change and time interval, and converts the obtained construction condition text including excavation depth, support erection state and surrounding load condition into a context vector (Context Embedding) through a pre-trained text encoder; then, the context vector is spliced in front of the sensor time series Token sequence as a prefix prompt to obtain a multi-modal encoding vector and input the local model for local training. Understandably, during training, the multi-modal encoding vector of the current period is input as an input sample, and each input sample corresponds to a sample label. In the present application, the sample label is the future period of the foundation pit state, including the future period of the foundation pit horizontal displacement and the foundation pit lateral earth pressure, etc. During training, the future period is relative to the current period, and the current period input sample and the sample label of the future period are both from the historical data.

[0041] In the above embodiment, the "working condition text-acquisition data" multi-modal encoding vector is constructed as an input sample, and by inputting the construction condition text, the model can consider the influence of the construction condition when making a prediction, identify that the abnormality may come from external factors such as construction disturbance rather than a dangerous situation, thereby reducing the false positive rate and improving the prediction accuracy.

[0042] In the present application, the local training loss is a comprehensive loss considering the data fitting loss and the physical constraint loss. Compared with the traditional model training loss, the present application considers the physical constraint loss for the special scene of foundation pit monitoring. The physical constraint loss represents the degree of violation of the mechanical balance of the foundation pit support structure calculated based on the prediction result. The present application considers the physical constraint loss when training the local model. The physical constraint loss will produce a high penalty gradient, forcing the model parameters to converge to the physical manifold that meets the "force and deformation coordination" in the back propagation, so that the model learns the essential laws of geotechnical mechanics, not just statistical fitting at the digital level, i.e. the predicted displacement cumulative value and the soil pressure change can as much as possible meet the balance equation of the elastic foundation beam, reducing the prediction results that violate the physical common sense, thereby improving the prediction accuracy. In other words, by explicitly embedding the elastic foundation beam balance differential equation as a strong constraint in the model training, and constructing a composite loss function to force the AI model to output in accordance with the basic laws of geotechnical mechanics, this mechanism can effectively prevent the traditional "black box" model from outputting absurd predictions that violate physical common sense (such as "severe deformation under no external force") in extreme working conditions (such as support failure, piping), greatly reducing the risk of false positives and false negatives, and ensuring the safety boundary of the project.

[0043] In an embodiment, the calculation formula of the comprehensive loss is:

[0044] ;

[0045] wherein, is the comprehensive loss, is the data fitting loss, is the physical constraint loss, and λ is a weight coefficient of the physical constraint loss set. When the model prediction result violates the mechanical equilibrium condition, is increased, forcing the model to converge in the latent space that meets the physical law.

[0046] wherein, the physical constraint loss is calculated by the following formula:

[0047] ;

[0048] wherein, T is the number of time points in the future period predicted by the model, t is a time point, , and are the foundation pit horizontal displacement and the foundation pit lateral earth pressure at time point t predicted by the model, respectively, is the foundation pit support structure balance differential operator constructed based on the elastic foundation beam theory;

[0049] For any foundation pit horizontal displacement u and foundation pit lateral earth pressure p, the calculation formula of the foundation pit support structure balance differential operator is:

[0050]

[0051] wherein, EI is the bending stiffness of the support pile, k is the horizontal reaction force coefficient of the soil, b is the pile spacing in the foundation pit, represents the derivative of the foundation pit horizontal displacement u along the depth direction z of the foundation pit.

[0052] Based on the above local data and local training loss, the local model can be fine-tuned, and the parameter update increment Δθ is calculated as the local model update parameter to be uploaded to the cloud server.

[0053] After obtaining the geological embedding vector and the local model update parameter, the local model update parameter and the encrypted geological embedding vector are packaged and uploaded to the cloud end through the MQTT protocol. The original sensor values and engineering drawings are not involved in the upload, effectively protecting the project privacy.

[0054] S13, the cloud server clusters all edge-end devices based on the similarity of the geological embedding vectors of the edge-end devices, aggregates the local model update parameters of the edge-end devices belonging to the same cluster set to obtain global model update parameters corresponding to the cluster set; when performing the aggregation calculation, the data quality evaluation score of the edge-end device is used as the weight coefficient of the corresponding local model update parameter, and the data quality evaluation score of the edge-end device is the cumulative of the local data integrity score when the edge-end device performs the current round of local training, the geological matching degree score between the edge-end device and the cluster center of the cluster set to which the edge-end device belongs, and the physical constraint consistency score obtained by the edge-end device using its local verification set to verify the local model trained by the edge-end device. The smaller the physical constraint loss is, the higher the physical constraint consistency score is. The obtained global model update parameters are used as the global model update parameters distributed to the edge-end devices in the next round of federated training, and the global model update parameters corresponding to different cluster sets are distributed to the edge-end devices belonging to the corresponding cluster sets.

[0055] In the cloud server, two types of operations are mainly performed, first, the edge-end devices are clustered, and then the edge-end devices belonging to the same class are aggregated according to the clustering result.

[0056] Specifically, after the cloud server receives the geological embedding vectors uploaded by each edge-end device, the cosine similarity between the geological embedding vectors is calculated, and the clustering of the edge-end devices is realized. In an embodiment, a clustering threshold can be set , and the edge-end devices with a cosine similarity higher than the clustering threshold are clustered into the same cluster. In a specific operation, the clustering threshold can be set to = 0.85.

[0057] Wherein, the calculation formula of the cosine similarity between any two geological embedding vectors is:

[0058] .

[0059] Through the clustering operation, the foundation pits with similar geological features can be clustered into the same class, for example, all foundation pits can be divided into "high water content soft soil group", "dense sand and pebble group", etc. In this way, the interference of heterogeneous geological data can be avoided.

[0060] After clustering, the local model update parameters of the edge-end devices belonging to the same cluster set are aggregated to obtain global model update parameters corresponding to the cluster set.

[0061] ​In the prior art, the cloud server directly aggregates the local model update parameters uploaded by each edge device to obtain global model update parameters, and then distributes the global model update parameters to each edge device. The disadvantage of this approach is that heterogeneous geological data may interfere with each other, which is not conducive to model learning.

[0062] In the present application, the edge devices are first aggregated based on geological characteristics, and the foundation pits with similar geological characteristics are classified into the same cluster set, and then the local model update parameters in each cluster set are aggregated to obtain global model update parameters corresponding to the cluster set, and then the global model update parameters aggregated by the cluster are correspondingly distributed to each edge device in the cluster aggregation. In this way, the interference of heterogeneous geological data is avoided, which is conducive to model learning.

[0063] In the prior art, when performing aggregation calculation, the local model update parameters subjected to aggregation are directly weighted and summed, and the weight coefficients of each local model update parameter are the same. The disadvantage of this approach is that high-quality model parameters cannot be fully utilized.

[0064] In the present application, the data quality evaluation score of the edge device is used as the weight coefficient of the corresponding local model update parameter, and the data quality evaluation score of the edge device is the cumulative score of the local data integrity score when the edge device performs local training in the current round, the geological matching degree score between the edge device and the cluster center of the cluster set to which the edge device belongs, and the physical constraint consistency score obtained by the edge device when verifying the local model trained by the edge device using its local verification set. The higher the local data integrity score, the more complete the data during local training, the higher the physical constraint consistency score, the more the local model conforms to the mechanical equilibrium constraint, and the higher the reliability of the local model update parameter. The higher the geological matching degree score, the more the corresponding local model update parameter conforms to the prediction of the geological characteristics of the cluster, and therefore, the cumulative score of the local data integrity score, the geological matching degree score and the physical constraint consistency score is used as the weight coefficient of the corresponding local model update parameter. The above dynamic adaptive weight fusion strategy can give higher voice to high-quality and high-matching construction sites, thereby accelerating the model convergence speed and improving the model prediction accuracy.

[0065] In specific operation, the data quality evaluation score of the local model update parameter of the i-th edge device The expression can be represented as:

[0066] ;

[0067] In the formula, , , respectively represent the local data integrity score, the geology matching degree score, the physical constraint consistency score of the i-th edge device, 、 、 respectively represent the local data integrity score, the geology matching degree score, the physical constraint consistency score of the i-th edge device, 、 、 are the weighted summation coefficients of

[0068] For example, the coefficients are α = 0.4, β = 0.4, γ = 0.2, and α + β + γ = 1.

[0069] wherein the local data integrity score characterizes the integrity of the local data, and the calculation formula can be represented as:

[0070] ;

[0071] In the formula, MR represents the data missing rate of the i-th edge device in the current training period.

[0072] wherein the geology matching degree score characterizes the similarity between the i-th edge device and its aggregation cluster center, and the calculation formula can be represented as:

[0073] ;

[0074] In the formula, is the geology embedding vector of the i-th edge device, is the cluster center of the cluster to which the i-th edge device belongs.

[0075] wherein the physical constraint consistency score is related to the physical constraint loss, the smaller the physical constraint loss, the higher the physical constraint consistency score, and the calculation formula can be represented as:

[0076] ;

[0077] In the formula, is the physical constraint loss of the local model of the i-th edge device on the local verification set.

[0078] After obtaining the above data quality evaluation scores , the local model update parameters in each cluster set can be aggregated, and the aggregation formula is:

[0079] ;

[0080] In the formula, wherein is the global model update parameter, represents the data quality evaluation score of the i-th edge device, denotes the local model update parameter of the i-th edge device, that is, the parameter increment of the local model of the i-th edge device, and C denotes the cluster set to which the i-th edge device belongs.

[0081] After obtaining the global model update parameter, it is distributed to each edge device in the corresponding cluster set, and the edge device can update the local model based on the received global model update parameter, and the update formula is:

[0082] ;

[0083] In the formula, η is a learning rate, for example, it can be 0.1, is the local model parameter before updating based on the global model update parameter, is the local model parameter after updating based on the global model update parameter.

[0084] In an embodiment, the cloud server has a full-amount physical large language model, and in the initial stage of federated learning, the initial local model of the edge device is obtained by knowledge distillation based on the full-amount physical large language model as a teacher model.

[0085] Based on the above steps, one federated training is completed, multiple federated trainings are performed, and federated learning is completed, so that each edge device can obtain a finally trained local model, that is, a trained local foundation pit prediction model.

[0086] Subsequently, the obtained local foundation pit prediction model can be applied to actual inference.

[0087] In a second aspect, the present application claims to protect a foundation pit monitoring method based on cloud edge collaboration.

[0088] As Figure 2 shown is a step flow chart of the foundation pit monitoring method based on cloud edge collaboration in an embodiment of the present application, as Figure 3 shown is a system architecture diagram of cloud edge collaboration in an embodiment of the present application, as Figure 4 shown is an information interaction diagram of the combination model training and application stage of cloud edge collaboration in an embodiment of the present application. The following Figures 2-4 introduces the monitoring method.

[0089] S21, the edge device obtains the foundation pit information of the current period from the collection terminal and inputs the local lightweight model thereof to obtain the prediction result of the foundation pit state of the future period. The foundation pit state includes the horizontal displacement of the foundation pit and the lateral earth pressure of the foundation pit. The local lightweight model is a local foundation pit monitoring model. The local foundation pit monitoring model outputs the probability of different prediction values of the foundation pit state. The prediction value corresponding to the maximum probability is taken as the prediction result of the foundation pit state. The edge device also calculates the information entropy H(x) of the foundation pit state x based on the probability of different prediction values of the foundation pit state x output by the local foundation pit monitoring model. If any of the information entropy H(x) or the prediction result exceeds the corresponding threshold value, it is judged that there is a risk, and a risk emergency operation is triggered. The risk emergency operation includes informing the collection terminal to increase the collection frequency and sending the high-frequency foundation pit monitoring data collected by the collection terminal to the cloud server. The high-frequency foundation pit monitoring data are the data collected after the collection frequency is increased.

[0090] Specifically, the edge device is in communication connection with the collection terminal. The collection terminal collects the foundation pit monitoring data of the corresponding foundation pit in real time and transmits the data to the edge device.

[0091] In an embodiment, the collection terminal can include a multi-source heterogeneous sensor group, a collection control module and a signal transmission module.

[0092] The multi-source heterogeneous sensor group can include a displacement meter, a soil pressure cell and an inclinometer. The displacement meter can be a magnetostrictive displacement sensor. The communication adopts the RT1010 protocol. The range is 0-50mm. The accuracy reaches ±0.1mm. The displacement meter is specifically arranged on the crown beam and the top of the supporting pile around the foundation pit, and is used to capture millimeter-level horizontal and vertical displacement micro-variation. The soil pressure cell can be a vibrating wire soil pressure gauge. The communication adopts the Modbus-RTU protocol. The range is 0-1000kPa. The accuracy is ±1% F.S. The soil pressure cell is buried in the contact surface of the supporting pile and the soil (at depths of-5m, -10m and -15m), and is used to monitor the lateral earth pressure in real time. The inclinometer can be a MEMS dual-axis inclinometer. The communication adopts the RS485 protocol. The range is ±5°. The accuracy is ±0.01°. The inclinometer is installed at the waist beam of the side wall of the foundation pit, and is used to monitor the inclination angle of the deep soil.

[0093] The collection control module (MCU) can be built-in with a low-power ARM Cortex-M4 processor, and supports a "dual-mode" collection logic. In the default mode, the collection is performed at a regular frequency (1 time / 1 hour) of a primary foundation pit. Meanwhile, the listening state is maintained, and the switching instruction from the edge is received at any time. Once the instruction is received, the high-frequency collection mode (up to 1 time / 1 minute) is switched to.

[0094] The signal transmission module can adopt a dual-mode signal transmission module, integrating an SX1278 (LoRa) and an ESP32 (WiFi) dual-mode communication chip, and having a "service adaptive" protocol switching function. When the monitoring data is a statistical value (such as a single displacement floating point number), the number of bytes is less than 1024 bytes (about 64 bytes), and the LoRa mode is automatically locked. The working frequency band is 470-510MHz, and the transmission interval is 100ms. The strong penetration is used to solve the problem of signal blind area of the underground-2 layer of the foundation pit. When the edge end triggers the waveform return instruction due to "high information entropy", and needs to upload the original vibration waveform data (the number of bytes is greater than 1024 bytes, about 2048 bytes / second) without compression, the WiFi mode is automatically awakened. Using the high transmission rate of 150Mbps, the millisecond-level uploading of large-capacity data is realized, and the analysis materials for the cloud full-amount physical model are provided.

[0095] Specifically, the edge device has a multi-modal interface, which can not only obtain the foundation pit monitoring data of the collection terminal, but also can pull the electronic construction log text (such as "excavator operation", "support erection") in real time, provide the context for the model, and constitute the construction condition text information of the current period.

[0096] Specifically, the edge device has an analysis module for processing the obtained foundation pit monitoring data. The analysis module can be developed based on an NVIDIA Jetson or a high-performance NPU embedded platform. The analysis module has a protocol analysis function and a data standardization function. The built-in protocol recognition algorithm based on feature matching can automatically identify the sensor protocol type by scanning the frame header structure (such as the address code of Modbus, the start symbol of RT1010), the baud rate feature and the CRC / LRC check mode of the input signal, and then encapsulate the parsed heterogeneous data into a standard JSON format containing device_id (device ID), timestamp (time stamp accurate to milliseconds), type (physical quantity type), value (numerical value) and unit (unit) for subsequent model calling.

[0097] For example, the edge device parses the Modbus message [02 04 00 00 00 01 31 CA] uploaded by the earth pressure cell into a stress value of 120kPa, and converts the JSON data returned by the inclinometer into a standard format. At the same time, the electronic construction log text information of the current foundation pit is obtained synchronously, for example, the working condition text at the current time is "the excavator in the south area is excavating, the depth is-5m", so as to prepare the semantic context data for subsequent multi-modal analysis.

[0098] Specifically, the edge device has a data preprocessing module for filling in missing values and correcting abnormal values of the standard format data output by the parsing module, and converting the construction condition text information and the foundation pit monitoring data of the current period into multi-modal encoding vectors as the foundation pit information of the current period. The specific conversion process can be referred to the foregoing description.

[0099] As Figure 5A flow chart of data processing by the edge device in an embodiment of the application is shown. Taking a municipal foundation pit monitoring project in a certain city as an example, the soil layer of the foundation pit is silty clay. At time t, the analysis module receives a set of original monitoring data containing different protocols. Specifically, the displacement meter uploads the sequence [10.5, 11.0, -, -, 12.0, 13.5] (unit: mm) using the RT1010 protocol, and there is continuous data loss at the 3rd and 4th time points; at the same time, the soil pressure cell uploads the sequence [105, 108, 190, 112, 115] (unit: kPa) using the Modbus protocol, and the value at the 3rd time point deviates significantly from the normal state. The data preprocessing module first performs missing value filling. For the continuous two missing points at the 3rd and 4th time points in the displacement meter data sequence, since the number of missing points is small (3 points), the system uses linear interpolation algorithm for repair: according to the valid value 11.0 mm at the 2nd time point and the valid value 12.0 mm at the 5th time point, the time step increment is calculated, and then the value of the 3rd missing point is calculated as 11.0+(12.0-11.0) / 3=11.33 mm, and the value of the 4th missing point is calculated as 11.0+2×(12.0-11.0) / 3=11.67 mm, thereby completing the continuity of the time series data. Subsequently, the system performs abnormal value correction based on physical perception. For the 190 kPa value collected by the soil pressure cell at the 3rd time point, the system no longer relies solely on simple threshold judgment, but calls the built-in physical constraint module for verification. The system combines the physical characteristic parameters of silty clay, analyzes according to the soil body ultimate stress constraint equation, finds that the stress sudden increase (from 108 kPa to 190 kPa) at this time point does not accompany the corresponding displacement mutation, which seriously violates the physical constitutive relationship of stress-strain coordination, and therefore determines that the value is an abnormal value of the sensor caused by electromagnetic interference rather than a real danger. The preprocessing module selects the mean of the valid values (108 kPa and 112 kPa) before and after the abnormal point, and corrects the value of the point to 110 kPa. After completing the data cleaning, the system standardizes the data to unify the distribution. For the displacement meter sequence [10.5, 11.0, 11.33, 11.67, 13.5], the system converts it to a standard score (Z-Score) sequence according to the historical statistical distribution, and the result is [-1.0, -0.5, -0.17, 0.17, 2.0]; for the soil pressure cell sequence [105, 108, 110, 112, 115], the system calculates the median as 110 and the interquartile range (IQR) as 7, and after processing using the robust standardization formula, the sequence [-0.71, -0.29, 0, 0.29, 0.71] is obtained. Finally, the system performs "working condition-data" dual-mode prompt fine tuning coding.The preprocessing module does not simply encode the time series data into pseudo text, but obtains the current time construction log text "excavator working near the monitoring point" through the interface, and converts the text into a context vector (Context Embedding) by using a pre-trained text encoder; at the same time, the above-mentioned standardized sensor numerical sequence is encoded into a time series Token sequence. The system splices and fuses the context vector as a prefix prompt with the time series Token sequence, for example, the displacement mutation value 13.5mm (standardized value 2.0) at t_5 time, because the prefix prompt contains the semantic features of "excavator operation", the model's attention mechanism automatically identifies the high correlation between the mutation and construction vibration, so as to correctly attribute it to construction disturbance in subsequent analysis, effectively eliminate abnormal interference, and ensure the quality and physical interpretability of the model input.

[0100] Specifically, the edge device has a model running module, which deploys a local foundation pit monitoring model, which is used to input the construction condition text information of the current period into the local foundation pit monitoring model to obtain the prediction result of the foundation pit state in the future period. Specifically, the predicted foundation pit state includes the horizontal displacement and lateral earth pressure of the foundation pit in the future period, and can also include the predicted inclination of the foundation pit.

[0101] Specifically, the edge device has a risk judgment module. The local foundation pit monitoring model outputs the probability of different predicted values of the foundation pit state, and the predicted value corresponding to the maximum probability is taken as the prediction result of the foundation pit state. The risk judgment module is used to calculate the information entropy H(x) of the foundation pit state x based on the probability of different predicted values of the foundation pit state x output by the local foundation pit monitoring model. If any of the information entropy H(x) or the prediction result exceeds the corresponding threshold value, it is judged that there is a risk, and a risk emergency operation is triggered. Wherein, if the information entropy H(x) is too large, it means that the prediction result has strong uncertainty, and the model has insufficient cognition of the current working condition, and the risk emergency operation needs to be started; when the predicted value of the foundation pit state exceeds the threshold value specified by the industry standard, it means that there is a risk, and the risk emergency operation also needs to be started. The risk emergency operation includes informing the collection terminal to increase the collection frequency and sending the foundation pit monitoring data collected after increasing the collection frequency to the cloud server.

[0102] For example, the risk judgment module calculates the information entropy of the model prediction result in real time. If the prediction information entropy output by the model exceeds the preset threshold (for example, H(x) > 1.0), or the displacement rate breaks through the safety threshold (for example, 0.5 mm / h) according to the specification, the system will immediately trigger the risk emergency operation, send an instruction to the on-site terminal, and increase the sampling frequency of the risk point and the sensors near the risk point from 1 time / hour to 1 time / minute, and continue to collect high-frequency data until the information entropy falls back or the risk is removed, so as to capture the dynamic evolution trend of the micro risk. The collected high-frequency data is uploaded to the cloud server based on the MQTT protocol by the communication module of the edge device.

[0103] S22, the cloud server runs the full-amount physical large language model deployed by it to perform deep analysis on the received high-frequency foundation pit monitoring data, and generates a risk report based on the analysis result.

[0104] Specifically, after the acquisition terminal increases the collection frequency, the high-frequency data collected by the acquisition terminal is transmitted to the edge device, which does not process the data and directly forwards it to the cloud server for in-depth analysis. Since the cloud server has more storage resources and more powerful computing capability, it can undertake heavy computing tasks, and therefore can deploy a full-amount physical large language model with a huge number of parameters to perform comprehensive analysis and reasoning. In an embodiment, the loss function used to train the full-amount physical large language model is also a comprehensive loss considering data fitting loss and physical constraint loss. The specific form of the loss can be referred to the above description.

[0105] After receiving the high-frequency data uploaded by the edge device, the full-amount physical large language model is run to perform deep attribution analysis and physically consistent review on the preliminary judgment result of the edge. For example, the edge device analyzes that the displacement rate at a certain point reaches 75% of the warning value and is accompanied by abnormal distribution of soil pressure, triggering the risk emergency operation. Subsequently, the cloud server calls the full-amount physical large language model to perform more accurate and detailed prediction, and can generate a risk report based on the prediction result, for example, a structured risk report containing the risk location (such as the middle of the south side of the foundation pit), the risk level, the abnormal index (excessive displacement), and the physical attribution analysis. The structured risk report further includes a natural language explanation text generated by the cloud full-amount model based on the geotechnical mechanics attribution chain, for example: “the displacement of point CX-05 on the south side of the foundation pit accelerates, mainly due to the 30% attenuation of the support axial force at -10 m, resulting in excessive elastic deformation of the supporting pile. It is recommended to immediately tighten and strengthen monitoring”, and the report is pushed to the project manager's mobile APP within 1 minute through the API interface, realizing the rapid transmission of risk information.

[0106] In an embodiment, the edge device can periodically train the local model using newly collected data, and the training method is the federated learning method as described above, which will not be described here.

[0107] Thirdly, the present invention also relates to a cloud-edge collaborative foundation pit monitoring system, which includes an edge device and a cloud server, the edge device and the cloud server being used to collaboratively implement the methods mentioned in the first and second aspects above.

[0108] In summary, this invention considers two main aspects. First, physical constraint loss generates a high penalty gradient, forcing model parameters to converge towards a physical manifold that conforms to the "force-deformation coordination" principle during backpropagation. This allows the model to learn the essential laws of geotechnical mechanics, rather than merely performing statistical fitting at the numerical level. Specifically, it ensures that the predicted cumulative displacement and earth pressure changes conform as closely as possible to the equilibrium equations of an elastic foundation beam, reducing predictions that violate physical principles and thus improving prediction accuracy. Second, during cloud aggregation, this invention first aggregates edge devices based on geological features, grouping foundation pits with similar geological characteristics into the same cluster. Then, it aggregates the local model update parameters within each cluster to obtain global model update parameters corresponding to that cluster. Finally, it distributes these global model update parameters to each edge device within that cluster. This avoids interference from heterogeneous geological data, facilitates model learning, and improves prediction accuracy. Furthermore, when performing cloud aggregation, this invention uses the cumulative score of local data integrity, geological matching, and physical constraint consistency as the weight coefficients for the corresponding local model update parameters. This dynamic adaptive weight fusion strategy allows high-quality, highly matched construction sites to gain greater influence, thereby accelerating model convergence and improving prediction accuracy. Therefore, the federated learning-based foundation pit prediction model training method proposed in this invention can effectively improve model prediction accuracy. Using the foundation pit prediction model obtained through this training method as a lightweight large model for each edge device can improve the accuracy of local predictions, thus enabling more precise identification of foundation pit risks.

[0109] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. It should be noted that the terms "in one embodiment," "for example," and "again" are intended to illustrate the present invention and are not intended to limit the present invention.

[0110] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for training a foundation pit prediction model based on federated learning, characterized in that, Each round of federated training thereof comprises the following steps: The cloud server sends global model update parameters to the edge device; The edge device determines a local model based on the global model update parameters received by it, and performs local training on the local model using local data, so that the model predicts the foundation pit state in the future period based on the foundation pit information in the current period, and obtains local model update parameters; the foundation pit state includes foundation pit horizontal displacement and foundation pit lateral earth pressure; the local training loss is a comprehensive loss considering data fitting loss and physical constraint loss, wherein the data fitting loss represents the similarity deviation between the prediction result and the true label, and the physical constraint loss represents the degree of violation of mechanical equilibrium of the foundation pit support structure based on the prediction result; The edge device also maps the foundation pit geological information extracted during the collection of local data into a geological embedding vector, and sends the local model update parameters and the geological embedding vector to the cloud device; The cloud server clusters all edge devices based on the similarity of the geological embedding vectors of the edge devices, and aggregates the local model update parameters of the edge devices belonging to the same cluster set to obtain the global model update parameters corresponding to the cluster set; when performing aggregation calculation, the data quality evaluation score of the edge device is used as the weight coefficient of the corresponding local model update parameter, and the data quality evaluation score of the edge device is the cumulative of the local data integrity score when the edge device performs local training in the current round, the geological matching degree score between the edge device and the cluster center of the cluster set to which the edge device belongs, and the physical constraint consistency score obtained by the edge device when verifying the local model trained by it using its local verification set; the smaller the physical constraint loss, the higher the physical constraint consistency score, and the obtained global model update parameters are used as the global model update parameters sent to the edge device in the next round of federated training, and the global model update parameters corresponding to different cluster sets are distributed to the edge devices belonging to the corresponding cluster set; Wherein, at the edge device, the calculation formula of the comprehensive loss is: ； In the formula, is a comprehensive loss, is a data fitting loss, is a physical constraint loss, and λ is a weight coefficient of the set physical constraint loss; wherein the physical constraint loss is calculated as: ； In the formula, T is the number of time periods in the future period predicted by the model, t is the time, , are the horizontal displacement of the foundation pit and the lateral earth pressure of the foundation pit at time t predicted by the model, respectively, is the balance differential operator of the foundation pit support structure based on the elastic foundation beam theory; For any foundation pit horizontal displacement u and lateral earth pressure p, the balance differential operator of the foundation pit supporting structure is The calculation formula is: In the formula, EI is the bending stiffness of the support pile, k is the horizontal reaction force coefficient of the soil, b is the distance between the piles in the foundation pit, denotes the derivative of the horizontal displacement u of the foundation pit along the depth direction z of the foundation pit. 2.The foundation pit prediction model training method of claim 1, wherein, At the edge device, the current period foundation pit information input into the model is a multi-modal encoding vector, and the construction process of the multi-modal encoding vector includes: Obtain the foundation pit monitoring data collected by the collection terminal in the current period, and encode the foundation pit monitoring data at each collection time into a sensor time Token sequence according to three elements of value, rate of change and time interval; Obtain the construction working condition text information in the current period and convert it into a context vector through a pre-trained text encoder; The context vector is concatenated as a prefix prompt before the sensor time Token sequence to obtain a multi-modal encoding vector. 3.The foundation pit prediction model training method of claim 2, wherein, The foundation pit monitoring data collected by the collection terminal includes foundation pit displacement, earth pressure and soil inclination. 4.The foundation pit prediction model training method of claim 1, wherein, The foundation pit geological information includes cohesion, internal friction angle, compression modulus, water content and lithology code. 5.The foundation pit prediction model training method of claim 1, wherein, At the cloud server, all edge devices are clustered based on the similarity of the geological embedding vectors of the edge devices, including: The cosine similarity between the geological embedding vectors is calculated two by two, and the edge devices whose cosine similarity is higher than a set clustering threshold are clustered into the same class. The cosine similarity between the geological embedding vectors is calculated two by two, and the edge devices whose cosine similarity is higher than a set clustering threshold are clustered into the same class. 6.The foundation pit prediction model training method of claim 1, wherein, At the cloud server, the local model update parameters of the edge devices belonging to the same cluster set are aggregated to obtain the global model update parameters corresponding to the cluster set, and the aggregation formula is: ； In the formula, is a global model update parameter, denotes the data quality evaluation score of the i-th edge device, denotes the local model update parameter of the i-th edge device, and C denotes the cluster set to which the i-th edge device belongs. wherein the score is calculated by the formula: ； , , are the local data integrity score, the geology matching score, the physical constraint consistency score of the i-th edge device, respectively, , , denote the weighted sum coefficients of , , , respectively. Score The formula for the calculation is: ； In the formula, MR represents the data missing rate of the i-th edge device in the current training period; Score The formula for the calculation is: ； In the formula, is a geological embedding vector of the ith edge device, is a cluster center of a cluster to which the ith edge device belongs. Score The formula for the calculation is: ； In the formula, is the physical constraint loss of the local model of the ith edge device on the local validation set. 7.The foundation pit prediction model training method of claim 1, wherein, The cloud server has a full-amount physical large language model, and in the initial stage of federated learning, the initial local model of the edge device is obtained by knowledge distillation from the full-amount physical large language model as a teacher model.

8. A cloud-edge collaboration based foundation pit monitoring method, characterized in that, The method comprises the following steps: The edge device obtains the foundation pit information of the current period from the collection terminal and inputs it into the local lightweight model to obtain the prediction result of the foundation pit state in the future period. The foundation pit state includes the horizontal displacement of the foundation pit and the lateral earth pressure of the foundation pit. The local lightweight model is a local foundation pit monitoring model obtained by using the training method according to any one of claims 1 to 6. The local foundation pit monitoring model outputs the probability of different prediction values of the foundation pit state. The prediction value corresponding to the maximum probability is taken as the prediction result of the foundation pit state. The edge device also calculates the information entropy H(x) of the foundation pit state x based on the probability of different prediction values of the foundation pit state x output by the local foundation pit monitoring model. If any of the information entropy H(x) or the prediction result exceeds the corresponding threshold value, it is judged that there is a risk, and a risk emergency operation is triggered. The risk emergency operation includes informing the collection terminal to increase the collection frequency and sending the high-frequency foundation pit monitoring data collected by the collection terminal to the cloud server. The high-frequency foundation pit monitoring data are the data collected after increasing the collection frequency. The cloud server runs the full-amount physical large language model deployed by it to perform deep analysis on the received high-frequency foundation pit monitoring data, and generates a risk report based on the analysis result.

9. A cloud-edge collaborative foundation pit monitoring system, characterized in that, The method comprises the following steps: The edge device comprises a first processor, a first communication module and a first memory. The first memory stores a computer program. The first processor executes the program to realize the edge-side steps of any one of claims 1-7. The cloud server comprises a second processor, a second communication module and a second memory. The second processor executes the program to realize the cloud-side steps of any one of claims 1-7. The edge device and the cloud server perform data interaction through a network.

Images