A method and system for predicting and warning displacement strain of a circular ground wall of a terminal dumper room

Abstract

This invention relates to the field of civil engineering safety monitoring technology, specifically to a method and system for predicting and warning the displacement and strain of a circular diaphragm wall in a wharf tipper room. This invention introduces a physical information loss function into a deep learning model, embedding the displacement-strain compatibility relationship under the small deformation assumption of elasticity as a soft constraint into the training process, thus constructing a dual-path mechanism of data-driven and physical constraint-based collaborative optimization. The model adopts a CNN-LSTM architecture, where the CNN module extracts local high-order spatiotemporal features from multi-sensor time-series data, the LSTM layer captures long-term dependencies across windows, and the static feature encoding of engineering parameters and a multi-step differential prediction output strategy are combined to achieve accurate prediction of future radial displacement, circumferential displacement, radial strain, and circumferential strain. This invention significantly improves prediction accuracy and maintains good stability and generalization ability even with noisy, missing, or extrapolated data, significantly enhancing the reliability and practicality of engineering predictions.

Description

Technical Field

[0001] This invention relates to the field of civil engineering safety monitoring technology, specifically to a method and system for predicting and warning of displacement and strain of the circular diaphragm wall in a wharf tippler room. Background Technology

[0002] With the large-scale development of deep foundation pit projects such as large port terminals and heavy-duty tippler rooms, circular diaphragm walls (hereinafter referred to as "diaphragm walls") are a key support structure, and their safety and stability during service are crucial to ensuring the overall project safety. Under the combined effects of complex soil and water pressures, construction dynamic loads, and environmental factors, diaphragm wall structures are prone to cumulative displacement and strain, directly affecting structural safety and service life. To ensure project safety, structural health monitoring and safety early warning technology uses a sensor network to monitor the displacement and strain response of the wall in real time and assess its safety status. The core of this technology lies in the accurate prediction of the diaphragm wall's future short-term mechanical response, achieving advanced prediction of millimeter-level displacement and micro-strain level strain, providing a decision-making buffer period for project early warning and intervention measures.

[0003] Currently, the prediction of the mechanical response of diaphragm wall support structures is mainly based on numerical simulation, data-driven methods, or a combination of these methods.

[0004] Numerical simulation methods rely on the principles of geotechnical mechanics and structural mechanics to construct finite element or finite difference models. By inputting material parameters such as elastic modulus, Poisson's ratio, cohesion, and internal friction angle, as well as boundary conditions and load information, they solve equilibrium equations to predict displacement-strain response. These methods have clear mechanisms and a certain interpretability for new working conditions. However, under complex engineering conditions, factors such as the spatiotemporal variability of soil parameters, the interaction between structure and soil, and dynamic construction loads make constitutive modeling and load quantification difficult, and model assumptions are prone to deviating from reality, leading to decreased prediction accuracy. Furthermore, three-dimensional nonlinear coupled solutions are computationally expensive and time-consuming, making it difficult to meet the requirements of extremely short-term real-time prediction and rapid early warning in circular diaphragm wall scenarios.

[0005] Data-driven approaches rely on large amounts of historical or real-time monitoring data, using machine learning algorithms (such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks) to uncover hidden temporal and spatial patterns within the data, constructing a mapping relationship between monitoring inputs and mechanical response outputs. These methods do not require explicit physical models and can adaptively learn nonlinear, strongly coupled response patterns. However, traditional deep learning models are highly dependent on data quality and quantity. When the data contains excessive noise, insufficient sample coverage, or missing sensor data, overfitting or poor generalization ability is likely. Furthermore, the "black box" nature of the model weakens the physical interpretability of the prediction results, leading to low trust among engineers.

[0006] Hybrid approaches are based on the physical laws of mechanical response and employ a collaborative mechanism of "physics-guided + data-driven" to retain the interpretability of physical methods while leveraging data-driven approaches to handle uncertainties and nonlinearities. However, these models are structurally complex, requiring simultaneous coordination between physical and data-driven models. Furthermore, the fusion strategy for physical information and data-driven approaches needs to be customized for specific engineering scenarios, resulting in limited versatility. Summary of the Invention

[0007] The purpose of this invention is to propose a method and system for predicting and warning the displacement and strain of the circular diaphragm wall in a wharf tipper room. By constructing a convolutional recurrent neural network that integrates physical information, it achieves simultaneous and accurate prediction of four components: radial displacement, circumferential displacement, radial strain, and circumferential strain. Combined with a multi-level early warning mechanism, it provides efficient and reliable intelligent decision support for the safety monitoring of deep foundation pit projects.

[0008] According to a first aspect of the present disclosure, a method for predicting and warning of displacement strain of a circular diaphragm wall in a wharf tippler room is provided, comprising the following steps: Multi-source monitoring data and engineering parameters of circular diaphragm walls are collected. The multi-source data are aligned and fused according to timestamps to form a multi-feature dimension time series dataset. The fused data is then normalized to construct supervised learning samples containing historical input and future output labels. A convolutional recurrent neural network (PI-CNN-LSTM) prediction model that integrates physical information is constructed. The supervised learning samples are input into the model for training. The trained model is used to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. The convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network, while the time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, before being encoded into dynamic context vectors. The static feature vectors and dynamic context vectors are concatenated and fused together before decoding to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. Based on the prediction results, multi-level safety warnings are issued by combining multi-level static warning thresholds, dynamic thresholds for rate of change, and remaining safety time.

[0009] In one embodiment, the convolutional neural network includes multiple sequentially connected convolutional feature extraction blocks, each including a convolutional layer, a pooling layer, and a Dropout layer; the convolutional layer uses a preset number of convolutional kernels to slide along the time dimension to extract spatial correlations and short-term fluctuation patterns between multiple sensors; the pooling layer progressively compresses the time window length; through the stacking of multiple convolutional feature extraction blocks, spatiotemporal features from low to high order are extracted on the progressively compressed feature map.

[0010] In one embodiment, the displacement-strain compatibility physical loss is constructed based on the small deformation assumption of elasticity. The circumferential normal strain of the circular diaphragm wall is jointly determined by the circumferential expansion and contraction caused by the radial displacement and the circumferential displacement gradient. The radial normal strain is the rate of change of the radial displacement along the radial direction, specifically expressed as: For fixed ,along Direction For fixed ,along Direction in This is a circumferential displacement. For radial displacement, For circumferential strain, For radial strain, The radius of the circular diaphragm wall is... The sum of all predicted points in a training batch. and These are the theoretically calculated values ​​for radial and circumferential strain. and These are the predicted values ​​for radial and circumferential strains. This represents the loss of displacement-strain compatibility.

[0011] In one embodiment, the specific method for decoding the predicted output is as follows: the concatenated fusion vector is nonlinearly transformed using a fully connected layer with a ReLU activation function, and then mapped to a vector space of the prediction step size dimension through a linear fully connected layer without an activation function. The output is reshaped into normalized difference prediction values ​​of the four target components within the future prediction step size time. After the difference prediction values ​​are denormalized, they are accumulated step by step from the absolute value of the last known historical moment to reconstruct the absolute value prediction of the future sequence.

[0012] In one embodiment, the multi-level static early warning value is determined by: setting a three-level early warning threshold, comparing the predicted value with the design allowable value according to a preset proportional coefficient, and determining its corresponding early warning level; wherein, the determination condition for a three-level early warning is: Level I, also known as the attention level: Level II, or early warning level: Level III, also known as alarm level: in, for time Predicted values ​​at monitoring points For design allowable values, , , This is the warning level coefficient, and .

[0013] In one embodiment, the dynamic threshold for the rate of change is determined by: calculating the rate of change of the predicted value at the current time relative to the previous time; when the rate of change exceeds the preset allowable value for the rate of change, a Level II warning is triggered even if the absolute value of the predicted value does not exceed the multi-level static warning threshold. The rate of change is: in, for time The rate of change of the monitoring points Design allowable values ​​for the rate of change. One monitoring cycle.

[0014] In one embodiment, the remaining safe time is obtained by: predicting the remaining time required for the response value of each monitoring point to reach the warning threshold at each level, based on the predicted value and its rate of change. in The time required to reach the warning thresholds at each level for or These correspond to Level II and Level III warnings, respectively.

[0015] According to a second aspect of the present disclosure, a system for predicting and warning the displacement and strain of a circular diaphragm wall in a wharf tipper room is provided, comprising: The data processing module collects multi-source monitoring data and engineering parameters of the circular diaphragm wall, aligns and fuses the multi-source data according to timestamps to form a multi-feature dimension time series dataset, and normalizes the fused data to construct supervised learning samples containing historical input and future output labels. The response prediction module constructs a convolutional recurrent neural network (PI-CNN-LSTM) prediction model that integrates physical information. The supervised learning samples are input into the model for training. The trained model is used to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. In the loss calculation module, the convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network, and the time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, and then encoded into dynamic context vectors. The static feature vectors and dynamic context vectors are concatenated and fused, and then decoded to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. The graded early warning module, based on the prediction results, combines multi-level static early warning thresholds, dynamic thresholds for rate of change, and remaining safe time to provide multi-level safety early warnings.

[0016] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory. When the processor executes the program, it implements the method for predicting and warning the displacement strain of a circular diaphragm wall in a dock tipper room.

[0017] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting and warning the displacement strain of a circular diaphragm wall in a dock tipper room.

[0018] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention addresses the challenge of predicting the mechanical response of circular diaphragm walls under complex construction and service conditions. It constructs and trains a PI-CNN-LSTM model and integrates a safety early warning method. By inputting real-time multi-source monitoring data into the trained model, it can simultaneously output prediction results for four components: radial displacement, circumferential displacement, radial strain, and circumferential strain. This overcomes the limitation of traditional methods that can only predict a single physical quantity, achieving real-time prediction and dynamic risk assessment of the mechanical response of the circular diaphragm wall support structure of a wharf tipper, providing comprehensive data support for engineering safety monitoring.

[0019] 2. This invention organically integrates physical information loss into the training process of the CNN-LSTM network architecture. Through a dual-objective optimization mechanism of data fitting and physical law constraints, it guides the model to not only learn the statistical regularities of historical data from complex and noisy monitoring data, but also to ensure that the prediction results follow basic mechanical principles. This mechanism effectively ensures the physical rationality of the prediction results, significantly enhances the model's generalization ability and engineering credibility when data is scarce or when extrapolating operating conditions, and provides high-quality and highly reliable predictive input for subsequent real-time early warning.

[0020] 3. This invention forms a complete end-to-end prediction framework through a collaborative link of spatial feature extraction, temporal feature mining, and physical loss joint optimization. CNN transforms multi-source spatiotemporal data into structured features, LSTM delves into its temporal evolution patterns, and physical loss serves as a regularization term to ensure the learning process does not deviate from physical rules. The three components and constraints effectively complement each other, significantly improving overall prediction performance and engineering practicality. Attached Figure Description

[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0022] Figure 1 This is a diagram illustrating the overall framework of the PI-CNN-LSTM model. Figure 2 This is a schematic diagram of data windowing in the embodiment; Figure 3 This is a schematic diagram of the PI-CNN-LSTM model structure; Figure 4 Block diagram of displacement and strain prediction and early warning system for circular diaphragm wall of dock tipper room. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0027] Example 1: like Figure 1 As shown in the figure, this embodiment provides a method for predicting and warning of displacement strain of the circular diaphragm wall in a dock tipper room, including the following steps: S1. Collect multi-source monitoring data and engineering parameters of the circular diaphragm wall, align and fuse the multi-source data according to the timestamp to form a multi-feature dimension time series dataset, and normalize the fused data to construct supervised learning samples containing historical input and future output labels; Specifically, the multi-source monitoring data includes: radial displacement, circumferential displacement, radial strain, circumferential strain, earth pressure and pore water pressure, and construction load; engineering parameters include excavation depth, diaphragm wall radius, and diaphragm wall thickness. The multi-source data are aligned and fused according to timestamps to form a unified multi-feature dimension time-series dataset.

[0028] To eliminate the adverse effects of differences in dimensions and orders of magnitude between different physical quantities on model training and prediction accuracy, the fused multi-feature time-series data is processed using the Min-Max normalization method. This normalization method linearly maps all sample data of each feature to the [0,1] interval, and its formula is as follows: In the formula, The data is after normalization. This is the original data. The minimum value of this feature across all samples. This represents the maximum value of all samples for this feature. After normalization, all physical quantities are on the same numerical scale, avoiding problems such as unbalanced gradient updates, slow convergence, or getting stuck in local optima during training due to different units. Simultaneously, this invention fits the minimum and maximum parameters of the normalizer (MinMaxScaler) on the training set and saves them; in the test set and subsequent actual engineering predictions, the parameters of the training set are used for the same normalization transformation. The predicted values ​​output by the model can be restored to the original engineering units after inverse normalization. The inverse normalization formula is: Since the monitoring data is a time series, a sliding window process is used to construct supervised learning samples. The model input samples contain data from continuous historical time points, such as... Figure 2 As shown in the figure Indicates the window length. Indicates the step size of the sliding window. To predict the step size, 1 t 2 t , ..., mt Data representing multiple features at a given time. t Indicates the time. Window size represents the amount of sample data, that is, the number of data samples within the window. For input, future Displacement-strain data of the window Y This is for "labels". Substitute the input data into the model to obtain predicted values ​​of the mechanical response for subsequent time periods. The window can be clicked... The step length slides backward, and so on.

[0029] This embodiment's experimental dataset is constructed based on historical and real-time multi-source monitoring time-series data of the circular diaphragm wall of a dock tipper room throughout the entire construction and operation phases, covering more than three independent datasets of different circular diaphragm wall projects. Each dataset consists of two parts: key project parameters and time-series data. The key project parameters are shown in Table 1; the time-series data are used to generate input-output mapping pairs using the sliding window method for subsequent model training and testing. Input Features Within the historical window Each time step Multidimensional fusion of feature data, outputting labels For the future The displacement-strain data for each time step are represented as follows: in for Moment 3D fusion feature vector, for Moment 3D displacement-strain vector This represents the number of sensors for the cross-section of a circular foundation pit. Samples are arranged by step size. Sliding generation preserves spatiotemporal coupling characteristics. Key characteristics of the samples are shown in Table 2. End-to-end spatiotemporal mapping is achieved: a long input window captures slowly changing trends and periodic patterns, while a short output window focuses on extremely short-term dynamic responses, supporting the real-time risk warning requirements of engineering safety early warning systems.

[0030] Table 1 Engineering Parameters Table 2. Structure of the learning sample dataset S2. Construct a convolutional recurrent neural network (PI-CNN-LSTM) prediction model that integrates physical information, input the supervised learning samples into the model for training, and use the trained model to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. The overall structure of the PI-CNN-LSTM model is as follows: Figure 3 As shown, it mainly includes three core modules: spatiotemporal feature extraction, multi-step difference prediction output, and the connection between the two.

[0031] S2.1, the convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network. The time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, and then encoded into dynamic context vectors. The static feature vectors and the dynamic context vectors are concatenated and fused, and then decoded to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. Specifically, the input data consists of preprocessed multi-feature time-series data. Engineering parameters are extracted separately and fed into independent lightweight fully connected networks (Dense Layers) for encoding transformation to learn their high-level feature representations. Specifically, the original 3D engineering parameters are mapped and expanded into more expressive 8D feature vectors, outputting static feature vectors of shape [batch, 8]. The remaining time-series data are then processed in subsequent sub-layers. The Convolutional Neural Network (CNN) module consists of two concatenated "convolution-pooling-dropout" blocks. The first convolutional block uses 16 kernels of size 3, sliding along the time dimension and simultaneously performing convolution operations on all feature channels. The parameters of each kernel are independently and randomly initialized, and spatial correlations between different sensors and physical quantities at the same time step, as well as short-term fluctuation patterns within adjacent time steps, are extracted through parallel computation. Subsequently, a max-pooling layer compresses the time dimension to half its original length, and a dropout layer is used to suppress overfitting. The second convolutional block increases the number of kernels to 32, further combining basic patterns on the compressed feature map to generate more abstract high-order spatiotemporal features (such as multi-sensor phase differences or waveform evolution trends). Max pooling then compresses the time dimension to one-quarter of its original length. The entire CNN module outputs a compressed and enhanced high-order spatiotemporal feature sequence, significantly reducing the computational burden on subsequent LSTM while completing multi-level automated extraction from raw multi-sensor data to local high-order, short-term macroscopic features, providing the LSTM module with an information-rich and denoised refined representation.

[0032] The Long Short-Term Memory (LSTM) layer receives the high-order spatiotemporal feature sequence output by the CNN module and performs iterative computation along time steps. Its core lies in three gating mechanisms: the forget gate determines how much information from the previous time step's memory cells is discarded, allowing the model to adaptively fade irrelevant early perturbations; the input gate, together with the candidate cell state, controls how much new information from the current input is written into the memory cells, thus selectively remembering key events such as sudden changes in soil pressure and displacement acceleration; and the output gate determines the output content of the hidden state based on the current cell state. Memory cells act as information carriers, transmitting long-term dependencies (such as the accumulation of plastic deformation in a circular diaphragm wall) with almost no attenuation across multiple time steps, effectively overcoming the gradient vanishing problem of ordinary recurrent neural networks. The LSTM layer ultimately outputs the hidden state of the last time step, a dynamic context encoding vector of shape [batch, 32]. This vector integrates the temporal dynamics and memory effects within the complete historical window, providing a global generalized representation for subsequent fully connected prediction layers.

[0033] Subsequently, the static feature vector ([batch, 8]) is concatenated with the dynamic context encoding vector ([batch, 32]) output by the LSTM to obtain the fused vector ([batch, 40]), which is then input into the fully connected decoding layer for prediction.

[0034] The model employs a strategy of "predicting temporal differences and then accumulating them to obtain absolute values": the fusion vector first undergoes a nonlinear transformation through a fully connected layer with a ReLU activation function to further fuse information; then it passes through a linear fully connected layer without an activation function, mapping it to a vector space with a dimension equal to the prediction step size. The output of this linear layer, after reshaping, physically represents the normalized difference values ​​of the four target components (radial displacement, circumferential displacement, radial strain, and circumferential strain) at each future prediction time. In the forward computation of training and inference, the model first denormalizes these difference prediction values, and then, starting from the absolute value of the last known historical time step, it accumulates them step by step to reconstruct the absolute value prediction results of the future sequence. This design allows the model to learn relatively gradual changes, which is beneficial for improving training stability and the ability to capture long-term trends.

[0035] Preferably, the Physics-Informed Loss function aims to embed the fundamental physical laws governing the mechanical response of diaphragm walls into the training process of the deep learning model as soft constraints. Its core idea is to introduce an additional physical loss term into the total loss function, besides the data fitting loss, to penalize the portion of the model's predictions that violates preset physical principles. This invention calculates the displacement-strain compatibility physical loss during training: based on the small deformation assumption of elasticity, the circumferential normal strain of a circular diaphragm wall is jointly determined by the circumferential expansion and contraction caused by radial displacement and the circumferential displacement gradient, while the radial normal strain is the rate of change of radial displacement along the radial direction. The physical loss penalizes predictions that deviate from the above theoretical relationships, specifically as follows: For fixed ,along Direction For fixed ,along Direction in This is a circumferential displacement. For radial displacement, For circumferential strain, For radial strain, The radius of the circular diaphragm wall is... The sum of all predicted points in a training batch. and These are the theoretically calculated values ​​for radial and circumferential strain. and These are the predicted values ​​for radial and circumferential strains. This represents the loss of displacement-strain compatibility.

[0036] Total loss during model training for: in, The mean square error (MSE) is the difference between the predicted value and the actual monitored value. These are the weighting coefficients. These are actual monitoring values. These are the four-component predicted values ​​for the model. The optimizer uses this total loss for backpropagation to update all weight parameters of the model.

[0037] This invention constructs a dual-path optimization mechanism combining data-driven and physical constraints, enabling the PI-CNN-LSTM model to both fit historical data and satisfy fundamental mechanical laws when predicting complex multi-component coupled responses. This maintains the rationality and stability of predictions even under noisy, missing, or extrapolated data conditions, effectively mitigating the overfitting problem of purely data-driven models. Simultaneously, the introduction of physical constraints reduces reliance on massive amounts of high-quality training data, enhancing the model's generalization ability and engineering reliability.

[0038] S3. Based on the prediction results, multi-level safety warnings are issued by combining multi-level static warning thresholds, dynamic thresholds for rate of change, and remaining safety time.

[0039] Specifically, based on the predicted value With design allowable values The ratio is used to set a three-level early warning system, as shown in Table 3.

[0040] Table 3 Multi-level early warning value settings in, for time Predicted values ​​at monitoring points For design allowable values, , , This is the warning level coefficient, and When the warning level is Level I, the response value enters the range of concern, requiring enhanced monitoring and data analysis; when the warning level is Level II, the response value is close to the critical state, posing a clear risk, and intervention measures need to be prepared; when the warning level is Level III, the response value reaches or exceeds the safety limit, posing an immediate danger, and engineering measures need to be taken immediately.

[0041] Based on a dynamic threshold for the rate of change, abnormal acceleration in the response is detected. If the rate of change exceeds the threshold, even if the absolute value of the response does not exceed the static threshold, a Level II warning will be triggered, indicating an abnormal development trend. The calculation formula is as follows: in, for time The rate of change of the monitoring points Design allowable values ​​for the rate of change. One monitoring cycle.

[0042] The remaining safe time, based on the current rate of change of response, predicts the time required to reach the warning threshold at each level, providing time-based warning judgments. in The time required to reach the warning thresholds at each level for or These correspond to Level II and Level III warnings, respectively.

[0043] This embodiment uses a circular diaphragm wall project for a dock tipper room as an example and designs a corresponding experimental scheme. A finite element simulation software is used to establish a construction process model of the circular diaphragm wall and generate a corresponding experimental dataset. Five different diaphragm wall specifications are designed. The first four specifications are used to construct the training sample set, and the last one is used as the test set to evaluate model performance. Simultaneously, a CNN-LSTM model is selected as the benchmark for comparison. Details of the diaphragm wall specifications for each location are shown in Table 4.

[0044] Table 4 Specifications of Diaphragm Wall The following two evaluation index systems are adopted: 1) Mean Absolute Error (MAE) and Mean Squared Error (MSE), which are used to intuitively reflect the degree of deviation between the model's predicted values ​​and the actual values. The smaller the value, the higher the model's prediction accuracy; 2) Coefficient of Determination (R²) 2 The value of α is used to characterize the degree to which the model fits the data. The closer the value is to 1, the better the fit and the more reliable the prediction results. The relevant calculation formula is shown below. in, , and They represent the first The actual value, the predicted value, and the average of the actual values ​​at each time point; n Indicates the length of the time series data set.

[0045] Experimental results show that, under a prediction duration of 12 hours, the convolutional recurrent neural network method integrating physical information proposed in this invention achieves varying degrees of accuracy improvement compared to traditional models in predicting radial displacement, circumferential displacement, radial strain, and circumferential strain of the circular diaphragm wall of the dock tipper room. The average improvement exceeds 40%, with the improvement in circumferential displacement prediction accuracy being particularly significant. Table 5 presents a performance comparison between the benchmark model and the model of this patent.

[0046] Table 5 Comparison of Model Prediction Performance Example 2: like Figure 4 As shown, this embodiment provides a system for predicting and warning the displacement and strain of the circular diaphragm wall in a dock tipper room, including: The data processing module collects multi-source monitoring data and engineering parameters of the circular diaphragm wall, aligns and fuses the multi-source data according to timestamps to form a multi-feature dimension time series dataset, and normalizes the fused data to construct supervised learning samples containing historical input and future output labels. The response prediction module constructs a convolutional recurrent neural network (PI-CNN-LSTM) prediction model that integrates physical information. The supervised learning samples are input into the model for training. The trained model is used to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. In the loss calculation module, the convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network, and the time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, and then encoded into dynamic context vectors. The static feature vectors and dynamic context vectors are concatenated and fused, and then decoded to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. The graded early warning module, based on the prediction results, combines multi-level static early warning thresholds, dynamic thresholds for rate of change, and remaining safe time to provide multi-level safety early warnings.

[0047] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0048] Example 3: An electronic device is provided for running the aforementioned "Method for Predicting and Early Warning of Displacement Strain of a Circular Diaphragm Wall in a Dock Tilting Machine Room". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S3 of the method described in Embodiment 1, specifically including but not limited to: S1. Collect multi-source monitoring data and engineering parameters of the circular diaphragm wall, align and fuse the multi-source data according to the timestamp to form a multi-feature dimension time series dataset, and normalize the fused data to construct supervised learning samples containing historical input and future output labels; S2. Construct a convolutional recurrent neural network (PI-CNN-LSTM) prediction model that integrates physical information, input the supervised learning samples into the model for training, and use the trained model to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. S2.1, the convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network. The time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, and then encoded into dynamic context vectors. The static feature vectors and the dynamic context vectors are concatenated and fused, and then decoded to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. S3. Based on the prediction results, multi-level safety warnings are issued by combining multi-level static warning thresholds, dynamic thresholds for rate of change, and remaining safety time.

[0049] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0050] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S3 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0051] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0053] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for predicting and warning of displacement strain of a circular diaphragm wall in a dock tipper room, characterized in that, Includes the following steps: Multi-source monitoring data and engineering parameters of circular diaphragm walls are collected. The multi-source data are aligned and fused according to timestamps to form a multi-feature dimension time series dataset. The fused data is then normalized to construct supervised learning samples containing historical input and future output labels. A convolutional recurrent neural network prediction model that integrates physical information is constructed. The supervised learning samples are input into the model for training. The trained model is used to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. The convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network, while the time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, before being encoded into dynamic context vectors. The static feature vectors and dynamic context vectors are concatenated and fused together before decoding to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. Based on the prediction results, multi-level safety warnings are issued by combining multi-level static warning thresholds, dynamic thresholds for rate of change, and remaining safety time.

2. The method for predicting and warning the displacement and strain of the circular diaphragm wall in a wharf tipper room according to claim 1, characterized in that, The convolutional neural network comprises multiple sequentially connected convolutional feature extraction blocks, each including a convolutional layer, a pooling layer, and a Dropout layer. The convolutional layer uses a preset number of convolutional kernels to slide along the time dimension to extract spatial correlations and short-term fluctuation patterns among multiple sensors. The pooling layer progressively compresses the time window length. Through the stacking of multiple convolutional feature extraction blocks, spatiotemporal features from low to high order are extracted on the progressively compressed feature map.

3. The method for predicting and early warning of displacement strain of the circular diaphragm wall in a wharf tipper room according to claim 1, characterized in that, The displacement-strain compatibility physical loss is constructed based on the small deformation assumption of elasticity. The circumferential normal strain of the circular diaphragm wall is jointly determined by the circumferential expansion and contraction caused by the radial displacement and the circumferential displacement gradient. The radial normal strain is the rate of change of the radial displacement along the radial direction, specifically expressed as: in This is a circumferential displacement. For radial displacement, For circumferential strain, For radial strain, The radius of the circular diaphragm wall is... The sum of all predicted points in a training batch. and These are the theoretically calculated values ​​for radial and circumferential strain. and These are the predicted values ​​for radial and circumferential strains. This represents the loss of displacement-strain compatibility.

4. The method for predicting and warning the displacement and strain of the circular diaphragm wall of a wharf tipper room according to claim 1, characterized in that, The specific method for decoding the output prediction value is as follows: the concatenated fusion vector is nonlinearly transformed by a fully connected layer with a ReLU activation function, and then mapped to the vector space of the prediction step size dimension by a linear fully connected layer without an activation function. The output is reshaped into the normalized difference prediction value of the four target components within the future prediction step size time. After denormalizing the differential prediction values, the absolute values ​​of the future sequence are reconstructed by accumulating them step by step, starting from the absolute value of the last known historical moment.

5. The method for predicting and early warning of displacement strain of the circular diaphragm wall in a wharf tipper room according to claim 1, characterized in that, The multi-level static early warning value determination is as follows: A three-level early warning threshold is set, and the predicted value is compared with the design allowable value according to a preset proportional coefficient to determine its corresponding early warning level; wherein, the determination condition for a three-level early warning is: Level I, also known as the attention level: Level II, or early warning level: Level III, also known as alarm level: in, for time Predicted values ​​at monitoring points For design allowable values, , , This is the warning level coefficient, and .

6. The method for predicting and early warning of displacement strain of the circular diaphragm wall in a wharf tipper room according to claim 5, characterized in that, The dynamic threshold for the rate of change is determined by: calculating the rate of change of the predicted value at the current time relative to the previous time; when the rate of change exceeds the preset allowable value for the rate of change, a Level II warning is triggered even if the absolute value of the predicted value does not exceed the multi-level static warning threshold. The rate of change is: in, for time The rate of change of the monitoring points Design allowable values ​​for the rate of change. One monitoring cycle.

7. The method for predicting and warning the displacement and strain of the circular diaphragm wall of a wharf tipper room according to claim 5, characterized in that, The remaining safe time is obtained by predicting the remaining time required for the response value of each monitoring point to reach the warning threshold at each level, based on the predicted value and its rate of change. in The time required to reach the warning thresholds at each level for or These correspond to Level II and Level III warnings, respectively.

8. A system for predicting and warning the displacement and strain of a circular diaphragm wall in a dock tipper room, characterized in that, include: The data processing module collects multi-source monitoring data and engineering parameters of the circular diaphragm wall, aligns and fuses the multi-source data according to timestamps to form a multi-feature dimension time series dataset, and normalizes the fused data to construct supervised learning samples containing historical input and future output labels. The response prediction module constructs a convolutional recurrent neural network prediction model that integrates physical information. The supervised learning samples are input into the model for training. The trained model is used to simultaneously predict the four-component mechanical response of the circular diaphragm wall: radial displacement, circumferential displacement, radial strain, and circumferential strain. In the loss calculation module, the convolutional recurrent neural network prediction model separates engineering parameters and time-series data into two independent feature extraction paths. The engineering parameters are encoded into static feature vectors through a lightweight fully connected network, and the time-series data are sequentially processed by a convolutional neural network to extract spatial features, and by a long short-term memory network to extract temporal dependencies, before being encoded into dynamic context vectors. The static feature vectors and dynamic context vectors are concatenated and fused, and then decoded to output the predicted value. The total loss of the model training is composed of a weighted sum of the data fitting loss and the displacement-strain compatibility physical loss constructed based on the small deformation assumption of elasticity. The graded early warning module, based on the prediction results, combines multi-level static early warning thresholds, dynamic thresholds for rate of change, and remaining safe time to provide multi-level safety early warnings.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the method for predicting and warning the displacement strain of the circular diaphragm wall of a wharf tippler room as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for predicting and warning of displacement strain of the circular diaphragm wall of a dock tipper room as described in any one of claims 1-7.

Images