Real-time monitoring and early warning system for geotechnical engineering data based on deep learning

Abstract

The present application relates to the technical field of geotechnical engineering safety monitoring and deep learning, and specifically discloses a real-time monitoring and early warning system for geotechnical engineering data based on deep learning. The present application solves the problem that the traditional method is difficult to analyze and identify abnormalities in real time, causing risk information lag, and the fixed threshold model is not adaptive to the evolution law of rock mass, resulting in early warning misjudgment or omission in the case of sudden environment, and improves the real-time and accuracy of the conversion of monitoring data into safety decision information.

Description

Technical Field

[0001] This invention relates to the fields of geotechnical engineering safety monitoring and deep learning technology, and in particular to a real-time monitoring and early warning system for geotechnical engineering data based on deep learning. Background Technology

[0002] Geotechnical engineering safety monitoring is a crucial technical means to ensure the safe construction and operation of tunnels, slopes, and underground engineering projects. Existing monitoring systems primarily rely on distributed sensor networks for multi-source data acquisition. However, in practical engineering applications, monitoring data exhibits massive volume, high dimensionality, and strong nonlinearity. Traditional data processing methods, limited by computational architecture and algorithm capabilities, struggle to complete real-time analysis and anomaly identification within a limited timeframe, resulting in a significant lag in the extraction of potential risk information. Simultaneously, existing early warning models are mostly constructed based on fixed thresholds or empirical statistical methods, lacking sufficient dynamic adaptability to the evolution of soil and rock masses under complex geological conditions. When faced with abrupt changes in stress or seepage fields, false alarms or missed warnings frequently occur. These problems hinder the timely transformation of monitoring data into safety decision-making information, failing to meet the comprehensive requirements of modern geotechnical engineering for real-time performance and accuracy.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a real-time monitoring and early warning system for geotechnical engineering data based on deep learning. The technical solution of this system is as follows: The multi-source data preprocessing module is used to acquire multi-source monitoring data of geotechnical engineering collected by a distributed sensor network, perform time alignment on the multi-source monitoring data and channelization processing according to physical quantity type to generate multi-channel synchronous time series data. The spatiotemporal feature encoding module is used to process the multi-channel synchronous time series data using a spatiotemporal attention encoding network. It captures the global dependencies between the channels of each physical quantity through a self-attention mechanism and extracts the local temporal evolution patterns of each channel through temporal convolution to generate a spatiotemporal fusion feature tensor. An anomaly detection module is used to process the spatiotemporal fusion feature tensor using a variational autoencoder, calculate the reconstruction probability as an anomaly deviation, and determine the anomaly state at the current monitoring time based on the comparison result of the anomaly deviation and the dynamic anomaly threshold. The threshold adaptive adjustment module is used to generate the adjustment amount of the dynamic anomaly threshold by using a deep reinforcement learning agent based on the spatiotemporal fusion feature tensor, the anomaly deviation amount and historical early warning feedback, and to adaptively correct the dynamic anomaly threshold when the stress field or seepage field changes abruptly. The graded early warning output module is used to generate graded early warning information based on the abnormal state determination result of the abnormality identification module.

[0005] The technical solution of this invention performs time alignment and channelization processing on geotechnical engineering monitoring data, uses a spatiotemporal attention coding network to capture the global dependencies and temporal patterns of each channel to generate fusion features, uses a variational autoencoder to reconstruct probabilities to identify anomalies, and uses a deep reinforcement learning agent to adjust the threshold when the stress field or seepage field changes abruptly based on historical early warning feedback. This solves the problems of traditional methods, which are difficult to analyze and identify anomalies in real time, resulting in delayed risk information. Furthermore, fixed threshold models are not adaptable to the evolution of geotechnical bodies and may lead to misjudgment or omission of early warnings in abrupt environments. This improves the real-time performance and accuracy of the transformation of monitoring data into safety decision-making information.

[0006] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0008] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of a deep learning-based real-time monitoring and early warning system for geotechnical engineering data according to the present invention. Detailed Implementation

[0009] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0010] Figure 1 This diagram illustrates the structure of an embodiment of a real-time monitoring and early warning system for geotechnical engineering data based on deep learning, provided by the present invention. Figure 1 As shown, this deep learning-based real-time monitoring and early warning system for geotechnical engineering data includes: The multi-source data preprocessing module 110 is used to acquire multi-source monitoring data of geotechnical engineering collected by a distributed sensor network, perform time alignment on the multi-source monitoring data and channelization processing according to physical quantity type to generate multi-channel synchronous time series data.

[0011] Distributed sensor networks refer to collaborative data acquisition networks formed by deploying multiple sensor nodes at spatial locations on a geotechnical engineering site and connecting them through communication links. For example, in a tunnel project, a total of 86 sensor nodes were deployed along the tunnel arch, sidewalls, and face to collect stress, displacement, seepage pressure, and temperature data. Each node transmitted the data to the field acquisition station via an RS485 bus, thus forming the distributed sensor network for this tunnel.

[0012] Geotechnical engineering refers to engineering activities that use rock and soil as the main research objects or working media, encompassing tunnel excavation, slope stabilization, and underground space construction. For example, a section of a highway traversing a mountainous area has an artificial slope as high as 62m. This slope requires stability monitoring during both the construction and operation phases, which is a typical example of geotechnical engineering. Multi-source monitoring data refers to the raw data set reflecting the state of multiple physical quantities of rock and soil, collected by various types of sensors in a distributed sensor network. For example, in the aforementioned tunnel project, the data collected by 86 sensor nodes within the same hour includes stress data arrays from vibrating wire stress gauges, displacement data arrays from laser displacement gauges, seepage pressure data arrays from piezometers, and temperature data arrays from thermistor thermometers. These four types of data together constitute the multi-source monitoring data for this tunnel project.

[0013] Among them, multi-channel synchronous time series data refers to multi-dimensional time series data formed by dividing multi-source monitoring data into multiple channels according to the type of physical quantity and aligning the data of each channel in time. For example, stress data of the aforementioned tunnel project is used as the first channel, displacement data as the second channel, seepage pressure data as the third channel, and temperature data as the fourth channel. The data sequences of the four channels are all resampled to a unified timestamp with an interval of 0.5s through cubic spline interpolation, resulting in multi-channel synchronous time series data with a set of corresponding values ​​for each of the four channels at each time point.

[0014] The spatiotemporal feature encoding module 120 is used to process the multi-channel synchronous time series data using a spatiotemporal attention encoding network. It captures the global dependencies between the channels of each physical quantity through a self-attention mechanism and extracts the local temporal evolution patterns of each channel through temporal convolution to generate a spatiotemporal fusion feature tensor.

[0015] Among them, the spatiotemporal attention coding network refers to a deep learning network structure that simultaneously extracts the global spatial dimension dependency and the local temporal dimension evolution pattern in the input sequence by stacking multiple spatiotemporal attention layers. For example, for the four-channel synchronous time series data of the aforementioned tunnel project, a coding network containing two spatiotemporal attention layers is constructed. The expansion factor of the first layer is 2, and the expansion factor of the second layer is 4. The self-attention sublayer in the network learns the cross-channel correlation between the arch stress change and the sidewall displacement, and the temporal convolution sublayer learns the local trend of the seepage pressure slowly increasing over time. This network is the spatiotemporal attention coding network.

[0016] The self-attention mechanism refers to an attention calculation method that models global dependencies by calculating the correlation weights between any two positions in the input sequence. For example, in the aforementioned spatiotemporal attention encoding network, the self-attention mechanism calculates the correlation between the 100th time step of the first channel stress sequence and the 100th time step of the second channel displacement sequence, the 100th time step of the third channel seepage pressure sequence, and the 100th time step of the fourth channel temperature sequence, and outputs a weighted representation, capturing the cross-physical quantity coupling relationship of displacement response caused by stress abrupt changes. Global dependencies refer to the mutual influence and constraint relationships between different physical quantity channels in multi-channel time series data at any time step. For example, in the aforementioned tunnel engineering, the excavation of the tunnel face leads to the release of stress at the crown. The stress data channel experiences a sudden drop in stress at the moment of excavation, the displacement data channel experiences an increase in settlement at the same moment, and the seepage pressure data channel experiences a change in seepage pressure several minutes later. The self-attention mechanism captures the interdependent relationships between stress, displacement, and seepage pressure across different time steps, forming a global dependency.

[0017] Temporal convolution refers to a one-dimensional convolution operation that extracts local temporal features from a time series by sliding the convolution kernel along the time dimension. For example, in the aforementioned spatiotemporal attention coding network, a temporal convolution with a kernel size of 5 is used to slide the calculation on the displacement data channel sequence, extracting the local trend of displacement change within each time window. After temporal convolution of displacement data from ten consecutive time steps, a local feature reflecting the rate of displacement change during that period is obtained. Local temporal evolution pattern refers to the local temporal representation of the changing trend and fluctuation characteristics of data within a short time window, extracted by temporal convolution operations in the time series of a single physical quantity channel. For example, in the aforementioned tunnel engineering, the seepage pressure data channel shows a monotonically increasing trend in seepage pressure values ​​within 50 seconds after tunnel excavation. The temporal convolution operation extracts the change pattern of "slow and continuous increase," which serves as the local temporal evolution pattern for this channel.

[0018] Among them, the spatiotemporal fusion feature tensor refers to the unified multidimensional feature representation obtained by weighting the feature tensor corresponding to the global dependency relationship and the feature tensor corresponding to the local temporal evolution mode through a gated fusion mechanism. For example, in the aforementioned tunnel project, the global dependency feature tensor A output by the self-attention sublayer and the local temporal evolution mode feature tensor C output by the temporal convolution sublayer are calculated through a gated fusion mechanism to obtain a spatiotemporal fusion feature tensor F containing 512 feature dimensions. Tensor F simultaneously encodes the cross-channel coupling information of stress and displacement and the local upward trend information of seepage pressure.

[0019] The anomaly identification module 130 is used to process the spatiotemporal fusion feature tensor using a variational autoencoder, calculate the reconstruction probability as an anomaly deviation, and determine the anomaly state at the current monitoring time based on the comparison result of the anomaly deviation and the dynamic anomaly threshold.

[0020] Among them, variational autoencoder refers to a generative deep learning model in which the encoder network maps the input data to the latent variable distribution, and the decoder network reconstructs the input data from the latent variables. For example, the spatiotemporal fusion feature tensor F of the aforementioned tunnel project is input into the encoder network, which outputs the mean vector and variance vector of the latent variable distribution. The latent variables are then sampled from the latent variable distribution and reconstructed data with the same structure as the original multi-channel synchronous time series data is generated by the decoder network. This model is variational autoencoder.

[0021] The reconstruction probability refers to the average log-likelihood of the original input data under the reconstructed data after multiple samplings of latent variables from the latent variable distribution and generation of reconstructed data through the decoder network in a variational autoencoder. For example, in the aforementioned tunnel project, 50 latent variables were sampled from the latent variable distribution to generate 50 sets of reconstructed data. The log-likelihood of the four-channel synchronous time series data under each set of reconstructed data was calculated, and the average of the 50 log-likelihood values ​​was taken as the reconstruction probability. The anomaly deviation refers to the quantitative indicator used to characterize the degree to which the current monitoring data deviates from the normal state after numerical transformation of the reconstruction probability calculated by the variational autoencoder. For example, in the aforementioned tunnel project, the average reconstruction probability under normal operating conditions was -2.3. When slight deformation occurred in the arch, the reconstruction probability dropped to -5.8, at which point the anomaly deviation was -5.8.

[0022] The dynamic anomaly threshold refers to a threshold parameter used to identify abnormal states, which is adjusted in real time by the deep reinforcement learning agent based on the state of the soil and rock mass and early warning feedback information. For example, in the aforementioned tunnel project, the dynamic anomaly threshold was maintained at -4.0 during the stable period. When the rainy season arrived and seepage pressure fluctuated, the agent adjusted the dynamic anomaly threshold from -4.0 to -3.2. The abnormal state at the current monitoring time refers to the binary judgment result of normal or abnormal obtained by comparing the abnormal deviation with the dynamic anomaly threshold at a certain sampling time point. For example, at the 13820th sampling time of the aforementioned tunnel project, the abnormal deviation was -5.8, and the current dynamic anomaly threshold was -4.0. Since -5.8 is less than -4.0, the current monitoring time is determined to be an abnormal state.

[0023] The threshold adaptive adjustment module 140 is used to generate an adjustment amount for the dynamic anomaly threshold using a deep reinforcement learning agent, based on the spatiotemporal fusion feature tensor, the anomaly deviation, and historical early warning feedback. This adjustment adaptively corrects the dynamic anomaly threshold when there are sudden changes in the stress field or seepage field. The reward function of the deep reinforcement learning agent is negatively weighted and positively correlated with the number of false alarms and missed alarms in the historical early warning feedback, ensuring that the adjustment of the dynamic anomaly threshold is driven by both the statistical distribution characteristics of the anomaly deviation and the historical costs of the early warning service.

[0024] Here, a deep reinforcement learning agent refers to a decision-making model that employs an actor-critic architecture, interacts with the soil and rock monitoring environment, and continuously optimizes its strategy based on reward signals. For example, the deep reinforcement learning agent deployed in the aforementioned tunnel project includes an actor network for generating adjustments to dynamic anomaly thresholds, a critic network for evaluating the value of the adjustments, and the agent updates its network parameters based on false alarms or missed alarms received after each warning. Historical warning feedback refers to records of confirmation results returned by the system after manual review or on-site verification of warning information issued during past monitoring periods. For example, the aforementioned tunnel project issued 15 warnings in the past 30 days, of which 12 were confirmed by on-site verification as local deformation of the surrounding rock, and 3 were false alarms caused by sensor interference from construction vibrations. These 15 records constitute historical warning feedback.

[0025] The adjustment amount refers to the continuous value output by the deep reinforcement learning agent at each moment to correct the dynamic anomaly threshold. For example, at the 13820th sampling moment of the aforementioned tunnel project, the deep reinforcement learning agent calculates and outputs an adjustment amount of +0.5 based on the current spatiotemporal fusion features and the anomaly deviation amount at the previous moment, adjusting the dynamic anomaly threshold from -4.0 to -3.5.

[0026] Among them, the stress field refers to the spatial distribution of stress state at various points within the rock and soil mass; for example, in the aforementioned tunnel project, after the excavation of the tunnel face, the stress field in the crown area redistributed, and the stress value in the crown compressive stress concentration zone increased from 1.2 MPa before excavation to 2.8 MPa. The seepage field refers to the spatial distribution of groundwater seepage state within the rock and soil mass; for example, in the aforementioned tunnel project, after the groundwater level rose by 3m during the rainy season, the seepage field changed accordingly, and the seepage velocity in the soil in front of the tunnel face increased from 0.05m per day to 0.3m per day.

[0027] Adaptive correction refers to the process by which a deep reinforcement learning agent automatically adjusts the dynamic anomaly threshold based on real-time state input to adapt to changes in the stress field or seepage field. For example, in the aforementioned tunnel project, a sudden change in the seepage field caused an increase in the fluctuation amplitude of the seepage pressure data channel. The deep reinforcement learning agent outputs adjustment values ​​of +0.2, +0.3, and +0.4 at three consecutive sampling times, gradually correcting the dynamic anomaly threshold from -4.5 to -3.6.

[0028] The graded early warning output module 150 is used to generate graded early warning information based on the abnormal state determination result of the abnormality identification module 130.

[0029] The abnormal state determination result refers to the identification information output by the abnormality identification module 130, indicating whether the current monitoring time is normal or abnormal. For example, in the aforementioned tunnel project, at the 13820th sampling time, the abnormality identification module 130 outputs an "abnormal" identifier as the abnormal state determination result after comparison. The graded early warning information refers to the early warning output generated after classifying the abnormal state determination result and the duration of the abnormality according to preset levels. For example, in the aforementioned tunnel project, after the abnormal state lasted for 15 minutes, the graded early warning output module 150 generated an orange early warning information, notifying the tunnel construction manager via SMS and platform pop-up window.

[0030] The technical solution of this embodiment performs time alignment and channelization processing on geotechnical engineering monitoring data, uses a spatiotemporal attention coding network to capture the global dependencies and temporal patterns of each channel to generate fusion features, uses a variational autoencoder to reconstruct probabilities to identify anomalies, and uses a deep reinforcement learning agent to adjust the threshold when the stress field or seepage field changes abruptly based on historical early warning feedback. This solves the problems of traditional methods, which are difficult to analyze and identify anomalies in real time, resulting in delayed risk information. Furthermore, the fixed threshold model is not adaptable to the evolution law of geotechnical bodies and may lead to false or missed early warnings in abrupt environments. This improves the real-time performance and accuracy of the transformation of monitoring data into safety decision-making information.

[0031] In one optional approach, the multi-source monitoring data includes stress monitoring data, displacement monitoring data, seepage pressure monitoring data, and temperature monitoring data. The physical quantity types correspond to stress type, displacement type, seepage pressure type, and temperature type. The channelization process involves constructing an independent data channel from the multi-source monitoring data of each physical quantity type. The time alignment process involves resampling the data sequence of each independent data channel using cubic spline interpolation, so that the data sequences of all independent data channels are unified to the same timestamp sequence and the sampling frequency is consistent.

[0032] Among them, stress monitoring data refers to measurement data collected by stress sensors that reflect the magnitude of stress within the soil or rock mass or on the support structure; for example, the stress value sequence measured by the vibrating wire stress gauge installed on the arch in the aforementioned tunnel project, recording 7200 stress values ​​per hour, constitutes the stress monitoring data for this tunnel. Displacement monitoring data refers to measurement data collected by displacement sensors that reflect the change in position of points on or inside the soil or rock mass; for example, the displacement value sequence measured by laser displacement gauges installed on the arch and sidewalls in the aforementioned tunnel project, recording the arch settlement and sidewall convergence displacement daily, constitutes the displacement monitoring data for this tunnel. Seepage pressure monitoring data refers to measurement data collected by piezometers that reflect the magnitude of pore water pressure within the soil or rock mass; for example, the pore water pressure value sequence measured by the piezometer installed in the borehole ahead of the tunnel face in the aforementioned tunnel project constitutes the seepage pressure monitoring data for this tunnel. Temperature monitoring data refers to measurement data collected by temperature sensors that reflect changes in the temperature of soil or the environment; for example, the temperature value sequence measured by the thermistor thermometer installed inside the lining in the aforementioned tunnel project is used to monitor the heat of hydration of concrete and seasonal temperature changes, constituting the temperature monitoring data of this tunnel.

[0033] Among them, stress type refers to the stress category identifier in the multi-source monitoring data, classified according to physical quantity attributes; for example, in the aforementioned tunnel project's multi-source monitoring data, there are four channels, and the physical quantity type of the first channel is marked as stress type, corresponding to all stress data collected by the vibrating wire stress gauges of the arch and sidewalls. Displacement type refers to the displacement category identifier in the multi-source monitoring data, classified according to physical quantity attributes; for example, in the aforementioned tunnel project's multi-source monitoring data, there are four channels, and the physical quantity type of the second channel is marked as displacement type, corresponding to all displacement data collected by the laser displacement gauge. Seepage pressure type refers to the seepage pressure category identifier in the multi-source monitoring data, classified according to physical quantity attributes; for example, in the aforementioned tunnel project's multi-source monitoring data, there are four channels, and the physical quantity type of the third channel is marked as seepage pressure type, corresponding to all seepage pressure data collected by the piezometer. Temperature type refers to the temperature category identifier in the multi-source monitoring data, classified according to physical quantity attributes; for example, in the aforementioned tunnel project's multi-source monitoring data, there are four channels, and the physical quantity type of the fourth channel is marked as temperature type, corresponding to all temperature data collected by thermistor thermometers.

[0034] Independent data channels refer to logical storage units where monitoring data of the same physical quantity type are stored separately as continuous data sequences in multi-channel data organization. For example, in the aforementioned tunnel project, all stress monitoring data were placed in independent data channel one in chronological order, and all displacement monitoring data were placed in independent data channel two in chronological order. The data structures of the two channels are independent of each other and have the same length. Cubic spline interpolation refers to a numerical method that uses piecewise cubic polynomial functions to perform curve fitting and interpolation calculations on discrete data points to obtain a continuous and smooth interpolation curve. For example, in the aforementioned tunnel project, the original sampling interval of the stress monitoring data was uneven. With a target interval of 0.5s, cubic spline interpolation function was used to interpolate the original stress data to obtain stress values ​​at times such as 0.0s, 0.5s, 1.0s, and 1.5s, aligning the stress data sequence with the displacement data sequence, seepage pressure data sequence, and temperature data sequence on the timestamp.

[0035] In the above-mentioned optional methods, the monitoring data of stress, displacement, seepage pressure and temperature are further divided into independent data channels according to the type of physical quantity. Cubic spline interpolation is used to resample the data sequence of each channel, so that the monitoring data from different sources are synchronously aligned under a unified timestamp and consistent sampling frequency, providing a data basis for spatiotemporal feature extraction.

[0036] In one alternative embodiment, the multi-source data preprocessing module 110 is specifically used for: Before generating the multi-channel synchronous time series data, outlier removal and wavelet thresholding are performed sequentially on the data sequence of each independent data channel. The outlier removal adopts the Laida criterion.

[0037] The Raida criterion refers to a statistical discrimination method that uses the mean and standard deviation of a data sequence as a benchmark to identify and remove values ​​that deviate from the mean by more than three times the standard deviation. For example, in the aforementioned tunnel project, the stress value of a sampling point in the stress monitoring data channel is eight times the normal stress value range. The deviation of this data point exceeds three times the standard deviation. The multi-source data preprocessing module 110 identifies it as an outlier according to the Raida criterion and replaces it with the interpolation results of the two points before and after it.

[0038] In the above-mentioned optional approach, before channelization and time alignment, outlier removal and wavelet thresholding are performed sequentially on the data sequence of each independent data channel to reduce the interference of sensor noise and random errors on feature encoding, thereby improving the quality of input data and the reliability of anomaly identification.

[0039] In one alternative embodiment, the spatiotemporal attention coding network includes a positional coding layer and multiple stacked spatiotemporal attention layers. Each spatiotemporal attention layer contains a multi-head self-attention sublayer and a temporal convolution sublayer. The temporal convolution sublayer employs dilated causal convolution, and the dilation factor increases exponentially with the number of stacked spatiotemporal attention layers.

[0040] The location encoding layer refers to a network layer at the input of the spatiotemporal attention encoding network used to add temporal and location information to multi-channel synchronous time series data. For example, in the aforementioned tunnel project, after the four-channel synchronous time series data enters the spatiotemporal attention encoding network, the location encoding layer uses sine and cosine functions to generate encoding vectors corresponding to time steps, encoding the 0th time step as a 512-dimensional location vector, which is then superimposed on the four-channel data of the 0th time step, enabling the network to perceive the chronological order of the time series. Multiple stacked spatiotemporal attention layers refer to connecting two or more spatiotemporal attention layers sequentially, with the output of the previous layer serving as the input of the next layer, forming a deep feature extraction network. For example, in the aforementioned tunnel project, three spatiotemporal attention layers are stacked: the first layer outputs primary spatiotemporal features, the second layer extracts intermediate spatiotemporal features based on the primary features, and the third layer further extracts advanced spatiotemporal features. These three stacked layers constitute the core of the deep spatiotemporal attention encoding network.

[0041] The multi-head self-attention sublayer refers to a network sublayer in the spatiotemporal attention layer where multiple self-attention heads compute self-attention representations in different subspaces in parallel, and then the outputs of all attention heads are concatenated. For example, the aforementioned tunnel project used eight attention heads, each independently computed self-attention in a 64-dimensional subspace of four-channel data, and the outputs of the eight attention heads were concatenated to form a 512-dimensional feature representation, which served as the output of the multi-head self-attention sublayer. The temporal convolutional sublayer refers to a network sublayer in the spatiotemporal attention layer where the output of the multi-head self-attention sublayer is convolved along the temporal dimension to extract local temporal information. For example, in the aforementioned tunnel project, the temporal convolutional sublayer used a one-dimensional dilated causal convolution with a kernel size of 5, convolving each dimension of the 512-dimensional feature sequence along the temporal direction. The dilation factors increased with the number of layers to 2, 4, and 8, respectively, extracting local temporal evolution patterns at different time scales.

[0042] Dilated causal convolution refers to a one-dimensional convolution method that introduces a dilation factor on the basis of causal convolution, causing the receptive field of the convolution kernel to increase exponentially, and only uses information from the current time and previous time steps. For example, in the aforementioned tunnel project, the dilation factor of the first time convolution sublayer is 2, the convolution kernel size is 5, the receptive field covers 9 time steps, and only uses multi-channel data from the nth sampling time and previous time steps, without introducing information from future time steps.

[0043] In the above optional approach, an encoding network is further constructed by a positional encoding layer and multiple stacked spatiotemporal attention layers. Each layer contains a multi-head self-attention sub-layer and a dilated causal convolution sub-layer. The dilation factor increases exponentially with the number of layers. This captures the global dependencies between channels while extracting local temporal evolution patterns, thereby enhancing the feature representation capability.

[0044] In one alternative approach, the kernel size of the temporal convolutional sub-layer is larger than the number of attention heads in the multi-head self-attention sub-layer, and the spatiotemporal feature encoding module 120 is specifically used for: The local temporal evolution pattern output by the temporal convolutional sublayer and the global dependency relationship output by the multi-head self-attention sublayer are summed through residual connections, and then input into the next stacked spatiotemporal attention layer after layer normalization.

[0045] Here, kernel size refers to the number of data points covered by the kernel along the time dimension in a one-dimensional convolution operation; for example, in the aforementioned tunnel project, the kernel size of the temporal convolution sub-layer is set to 5, and each convolution operation covers the multi-channel feature data of 5 adjacent sampling times. The number of attention heads refers to the number of attention heads that perform self-attention computation in parallel in the multi-head self-attention sub-layer; for example, in the aforementioned tunnel project, the number of attention heads in the multi-head self-attention sub-layer is set to 8, and the 512 dimensions of the input features are divided into 8 64-dimensional subspaces, with each attention head independently performing attention computation in its corresponding subspace.

[0046] In the above-mentioned optional approach, the local temporal evolution pattern of the output of the temporal convolutional sub-layer and the global dependency of the output of the multi-head self-attention sub-layer are further summed through residual connections, and then input into the next layer after layer normalization. Under the configuration that the kernel size of the temporal convolutional sub-layer is larger than the number of attention heads, multi-scale feature stable transfer and layer-by-layer fusion are achieved.

[0047] In one alternative embodiment, the variational autoencoder includes an encoder network and a decoder network, and the anomaly detection module 130 is specifically used for: Multiple latent variables are sampled from the latent variable distribution of the variational autoencoder, and reconstructed data is generated through the decoder network. The average log-likelihood of the multi-channel synchronous time series data under the reconstructed data is calculated as the reconstruction probability. When the reconstruction probability is less than the dynamic anomaly threshold, the current monitoring time is determined to be an abnormal state.

[0048] In this context, the latent variable distribution refers to the parameterized probability distribution output by the encoder network in a variational autoencoder, used to approximate the posterior distribution of the input data. For example, in the aforementioned tunnel project, the encoder network maps the spatiotemporal fusion feature tensor into a mean vector and a variance vector, which together parameterize a 512-dimensional multivariate Gaussian distribution; this distribution is the latent variable distribution. A latent variable is a low-dimensional latent representation vector randomly sampled from the latent variable distribution. For example, in the aforementioned tunnel project, a 512-dimensional random vector is sampled from the 512-dimensional multivariate Gaussian distribution parameterized by the mean and variance vectors. This random vector, after passing through the decoder network, can reconstruct an output with the same structure as the original input data; this random vector is the latent variable.

[0049] Reconstructed data refers to the output data with the same dimension as the original input data generated after the latent variables are input into the decoder network in the variational autoencoder. For example, in the aforementioned tunnel project, starting from 512-dimensional latent variables, the decoder network outputs reconstructed data with the same dimension as the original four-channel synchronous time series data through transpose convolution and upsampling operations. This includes reconstructed stress data sequence, reconstructed displacement data sequence, reconstructed seepage pressure data sequence, and reconstructed temperature data sequence.

[0050] In the above-mentioned optional approach, based on the encoder and decoder of the variational autoencoder, multiple latent variables are sampled from the latent variable distribution to generate reconstructed data. The average log-likelihood of the multi-channel synchronous time series data under the reconstructed data is used as the reconstruction probability. The abnormal state at the current monitoring time is quantitatively determined by comparing it with the dynamic anomaly threshold.

[0051] In one alternative approach, the deep reinforcement learning agent adopts an actor-critic architecture. The state space of the deep reinforcement learning agent includes a statistical feature vector obtained by global average pooling of the spatiotemporal fusion feature tensor, the abnormal deviation at the current monitoring time, and the number of false alarms and missed alarms accumulated in historical early warning feedback. The action space is the continuous adjustment amount of the dynamic anomaly threshold. The reward function is negatively weighted and positively correlated with the number of false alarms and the number of missed alarms.

[0052] The actor-critic architecture refers to a dual-network learning architecture in deep reinforcement learning, consisting of a policy function network that outputs actions and a value function network that evaluates the value of actions. For example, in the aforementioned tunnel project, the actor network receives the statistical feature vector and abnormal deviation amount at the current monitoring time as state input, and outputs the mean and variance of the dynamic abnormality threshold adjustment amount. The critic network receives the same state and the adjustment amount output by the actor network, and outputs the value estimate corresponding to this adjustment amount. The two networks alternately update to form the actor-critic architecture.

[0053] The number of false alarms refers to the cumulative number of times, in the historical early warning feedback records, the early warning information was confirmed as normal after on-site verification; for example, the aforementioned tunnel project issued 15 early warnings in the past 30 days, of which 3 were confirmed by on-site personnel to have stable and undeformed surrounding rock, resulting in 3 false alarms. The number of missed alarms refers to the cumulative number of times, in the historical early warning feedback records, abnormal events were found during on-site inspections but did not trigger an early warning; for example, the aforementioned tunnel project also had 2 instances in the past 30 days where on-site inspections found significant settlement in the tunnel arch but did not trigger an early warning, resulting in 2 missed alarms.

[0054] The continuous adjustment amount refers to the continuous real number output by the deep reinforcement learning agent used to correct the dynamic anomaly threshold. For example, in the aforementioned tunnel project, the deep reinforcement learning agent outputs an adjustment amount of +0.5, which adjusts the dynamic anomaly threshold from -4.0 to -3.5, with an adjustment range of 0.5.

[0055] Among the above-mentioned optional approaches, an actor-critic architecture is further adopted to construct a deep reinforcement learning agent. The global average pooling statistical feature vector, the current abnormal deviation, and the number of historical false positives and false negatives are incorporated into the state space. The threshold adjustment is continuously adjusted as the action space, and the negative weighted sum of the number of false positives and false negatives is used as the reward function to optimize the threshold adjustment strategy.

[0056] In one alternative approach, when the deep reinforcement learning agent generates the adjustment amount for the dynamic anomaly threshold, the adjustment amount is determined by the following formula: ; in, For a moment The adjustment amount of the dynamic abnormal threshold. To adjust the step size factor, For a moment Abnormal deviation, This is the cumulative error term for historical early warning feedback. This is the error penalty coefficient. This is the error attenuation factor. The mutation response coefficient is... For a moment The statistical feature vector obtained by global average pooling from the spatiotemporal fusion feature tensor. For a moment Statistical eigenvectors, The L2 norm is represented by the cumulative error term of the historical early warning feedback, which is calculated based on the weighted sum of the number of false alarms and the number of missed alarms within a preset time window in the past.

[0057] It should be noted that the adjustment formula for the dynamic anomaly threshold utilizes the saturation characteristics of the hyperbolic tangent function to nonlinearly compress the difference between the anomaly deviation and the historical early warning feedback cumulative error term, generating a bounded base value adjustment component. Simultaneously, an exponential decay term is introduced to rapidly decrease the amplitude of the base value adjustment component as the absolute value of the historical early warning feedback cumulative error term increases. The L2 norm is then used to quantify the change amplitude between statistical feature vectors at adjacent monitoring times. After compression by the hyperbolic tangent function, this becomes the abrupt response compensation component. The base value adjustment component and the abrupt response compensation component are added to obtain the adjustment amount. The function of the above formula is as follows: when abrupt changes occur in the stress field or seepage field, the L2 norm of the statistical feature vectors at adjacent times increases, and the abrupt response compensation component provides an additional threshold adjustment amount, enabling the dynamic anomaly threshold to quickly respond to rapid changes in the soil and rock mass state. When the historical early warning feedback cumulative error term is large, the exponential decay term suppresses the amplitude of the base value adjustment component, avoiding drastic threshold fluctuations when the system's early warning accuracy is low, thereby improving the adaptive adjustment stability of the dynamic anomaly threshold in complex soil and rock environments.

[0058] In the above-mentioned optional methods, based on the abnormal deviation, the cumulative error term of historical early warning feedback, and the L2 norm change of the statistical feature vector at adjacent times, the threshold adjustment amount is generated by adjusting the step size factor, error penalty coefficient, error decay factor, and sudden change response coefficient, so that the threshold correction takes into account both the historical error accumulation and the sudden change response of the stress field or seepage field.

[0059] In one alternative embodiment, the tiered early warning output module 150 is specifically used for: The warning levels are divided into yellow warning, orange warning and red warning. A yellow warning is generated when the reconstruction probability is less than the dynamic anomaly threshold and the duration is less than a first preset duration. An orange warning is generated when the reconstruction probability is less than the dynamic anomaly threshold and the duration is greater than or equal to the first preset duration. A red warning is generated when the reconstruction probability is less than the product of the dynamic anomaly threshold and a preset scaling factor.

[0060] Among them, a yellow warning refers to the lowest level warning in the graded warning information, indicating a minor or initial abnormality; for example, in the aforementioned tunnel project, the abnormal deviation first fell below the dynamic abnormality threshold at the 13820th sampling time and lasted for only 3 minutes, the graded warning output module 150 generated a yellow warning, prompting monitoring personnel to pay attention to the trend of tunnel arch displacement. An orange warning refers to a medium level warning in the graded warning information, indicating a persistent abnormality requiring countermeasures; for example, in the aforementioned tunnel project, if the abnormal state persisted for 15 minutes without returning to normal, the graded warning output module 150 upgraded the warning level from yellow to orange, notifying the construction team to prepare to evacuate the workers at the tunnel face. A red warning refers to the highest level warning in the graded warning information, indicating a severe abnormality requiring immediate emergency measures; for example, in the aforementioned tunnel project, if the abnormal deviation suddenly dropped to -12.7, reaching below the product of the dynamic abnormality threshold -4.0 and the preset scaling factor 0.5 (-2.0), the graded warning output module 150 generated a red warning, triggering an audible and visual alarm and notifying all personnel in the tunnel to evacuate immediately.

[0061] The first preset duration refers to the duration threshold used to distinguish between yellow and orange warnings. For example, in the aforementioned tunnel project, the first preset duration is set to 10 minutes. When the abnormal deviation amount remains below the dynamic abnormal threshold for 10 minutes, the warning level is upgraded from yellow to orange. The preset scaling factor refers to the abnormal deviation amount multiplier used to determine the triggering condition for a red warning. For example, in the aforementioned tunnel project, the preset scaling factor is set to 0.5. When the abnormal deviation amount is less than the product of the dynamic abnormal threshold and 0.5, a red warning is directly triggered.

[0062] In the above-mentioned optional methods, the warning level is further divided into three levels: yellow, orange and red. Based on the duration of the reconstruction probability being lower than the dynamic anomaly threshold and the comparison result of the product of the reconstruction probability and the preset scaling factor, different levels of warning information output are triggered respectively, so as to realize the fine distinction and graded response of the severity of the anomaly.

[0063] In an alternative approach, the spatiotemporal feature encoding module 120 is specifically used to fuse the global dependency and the local temporal evolution pattern through a gating fusion mechanism, wherein the spatiotemporal fusion feature tensor is determined by the following formula: ; in, For spatiotemporal fusion feature tensor, The global dependency feature tensor output by the multi-head self-attention sublayer. The local temporal evolution pattern feature tensor output by the temporal convolutional sublayer. , , , The weight matrix is ​​a learnable matrix. and For bias vectors, It is the sigmoid activation function. The hyperbolic tangent activation function is used. This represents element-wise product.

[0064] It should be noted that the spatiotemporal fusion feature tensor formula performs a nonlinear transformation on the linear combination of the global dependency feature tensor and the local temporal evolution mode feature tensor using the sigmoid activation function to generate a gated weight matrix with values ​​between 0 and 1. At the same time, it performs a nonlinear transformation on another linear combination of the global dependency feature tensor and the local temporal evolution mode feature tensor using the hyperbolic tangent activation function to generate a candidate fusion feature tensor with values ​​between -1 and 1. The gated weight matrix and the candidate fusion feature tensor are then multiplied element-wise to achieve selective filtering and fusion of candidate fusion features by the gated signal. The above formula can adaptively allocate fusion weights between global dependencies and local temporal evolution patterns based on the input characteristics of different physical quantity channels and different time steps. When the soil and rock mass is in a stable period, the gating weights tend to enhance the contribution of the local temporal evolution pattern to capture the slow creep trend. When the stress field or seepage field undergoes abrupt changes, the gating weights tend to enhance the contribution of the global dependencies to capture cross-channel coupling anomalies, thereby generating a spatiotemporal fusion feature tensor that accurately reflects the spatiotemporal correlation state of multiple physical quantities in the soil and rock mass.

[0065] In the above-mentioned optional approach, the global dependency feature tensor output by the multi-head self-attention sub-layer and the local temporal evolution pattern feature tensor output by the temporal convolution sub-layer are further fused by gating fusion. The spatiotemporal fusion feature tensor is generated by using sigmoid and hyperbolic tangent activation functions and learnable weight matrices to enhance the fusion expression of global and local features.

[0066] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A real-time monitoring and early warning system for geotechnical engineering data based on deep learning, characterized in that, include: The multi-source data preprocessing module is used to acquire multi-source monitoring data of geotechnical engineering collected by a distributed sensor network, perform time alignment on the multi-source monitoring data and channelization processing according to physical quantity type to generate multi-channel synchronous time series data. The spatiotemporal feature encoding module is used to process the multi-channel synchronous time series data using a spatiotemporal attention encoding network. It captures the global dependencies between the channels of each physical quantity through a self-attention mechanism and extracts the local temporal evolution patterns of each channel through temporal convolution to generate a spatiotemporal fusion feature tensor. An anomaly detection module is used to process the spatiotemporal fusion feature tensor using a variational autoencoder, calculate the reconstruction probability as an anomaly deviation, and determine the anomaly state at the current monitoring time based on the comparison result of the anomaly deviation and the dynamic anomaly threshold. The threshold adaptive adjustment module is used to generate the adjustment amount of the dynamic anomaly threshold by using a deep reinforcement learning agent based on the spatiotemporal fusion feature tensor, the anomaly deviation amount and historical early warning feedback, and to adaptively correct the dynamic anomaly threshold when the stress field or seepage field changes abruptly. The graded early warning output module is used to generate graded early warning information based on the abnormal state determination result of the abnormality identification module.

2. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 1, characterized in that, The multi-source monitoring data includes stress monitoring data, displacement monitoring data, seepage pressure monitoring data, and temperature monitoring data. The physical quantity types correspond to stress type, displacement type, seepage pressure type, and temperature type. The channelization process involves constructing an independent data channel for each type of multi-source monitoring data. The time alignment process involves resampling the data sequence of each independent data channel using cubic spline interpolation, so that the data sequences of all independent data channels are unified to the same timestamp sequence and the sampling frequency is consistent.

3. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 2, characterized in that, The multi-source data preprocessing module is specifically used for: Before generating the multi-channel synchronous time series data, outlier removal and wavelet thresholding are performed sequentially on the data sequence of each independent data channel. The outlier removal adopts the Laida criterion.

4. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 1, characterized in that, The spatiotemporal attention coding network includes a positional coding layer and multiple stacked spatiotemporal attention layers. Each spatiotemporal attention layer contains a multi-head self-attention sub-layer and a temporal convolution sub-layer. The temporal convolution sub-layer employs dilated causal convolution, and the dilation factor increases exponentially with the number of stacked spatiotemporal attention layers.

5. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 4, characterized in that, The kernel size of the temporal convolutional sub-layer is larger than the number of attention heads in the multi-head self-attention sub-layer, and the spatiotemporal feature encoding module is specifically used for: The local temporal evolution pattern output by the temporal convolutional sublayer and the global dependency relationship output by the multi-head self-attention sublayer are summed through residual connections, and then input into the next stacked spatiotemporal attention layer after layer normalization.

6. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 1, characterized in that, The variational autoencoder includes an encoder network and a decoder network, and the anomaly detection module is specifically used for: Multiple latent variables are sampled from the latent variable distribution of the variational autoencoder, and reconstructed data is generated through the decoder network. The average log-likelihood of the multi-channel synchronous time series data under the reconstructed data is calculated as the reconstruction probability. When the reconstruction probability is less than the dynamic anomaly threshold, the current monitoring time is determined to be an abnormal state.

7. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 1, characterized in that, The deep reinforcement learning agent adopts an actor-critic architecture. The state space of the deep reinforcement learning agent includes a statistical feature vector obtained by global average pooling of the spatiotemporal fusion feature tensor, the abnormal deviation at the current monitoring time, and the number of false alarms and missed alarms accumulated in the historical early warning feedback. The action space is the continuous adjustment of the dynamic anomaly threshold. The reward function is negatively weighted and positively correlated with the number of false alarms and the number of missed alarms.

8. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 7, characterized in that, When the deep reinforcement learning agent generates the adjustment amount for the dynamic anomaly threshold, the adjustment amount is determined by the following formula: ; in, For a moment The adjustment amount of the dynamic abnormal threshold. To adjust the step size factor, For a moment Abnormal deviation, This is the cumulative error term for historical early warning feedback. This is the error penalty coefficient. This is the error attenuation factor. The mutation response coefficient is... For a moment The statistical feature vector obtained by global average pooling from the spatiotemporal fusion feature tensor. For a moment Statistical eigenvectors, The L2 norm is represented by the cumulative error term of the historical early warning feedback, which is calculated based on the weighted sum of the number of false alarms and the number of missed alarms within a preset time window in the past.

9. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 6, characterized in that, The hierarchical early warning output module is specifically used for: The warning levels are divided into yellow warning, orange warning and red warning. A yellow warning is generated when the reconstruction probability is less than the dynamic anomaly threshold and the duration is less than a first preset duration. An orange warning is generated when the reconstruction probability is less than the dynamic anomaly threshold and the duration is greater than or equal to the first preset duration. A red warning is generated when the reconstruction probability is less than the product of the dynamic anomaly threshold and a preset scaling factor.

10. The real-time monitoring and early warning system for geotechnical engineering data based on deep learning according to claim 5, characterized in that, The spatiotemporal feature encoding module is specifically used to fuse the global dependency and the local temporal evolution pattern through a gating fusion mechanism. The spatiotemporal fusion feature tensor is determined by the following formula: ; in, For spatiotemporal fusion feature tensor, The global dependency feature tensor output by the multi-head self-attention sublayer. The local temporal evolution pattern feature tensor output by the temporal convolutional sublayer. , , , The weight matrix is ​​a learnable matrix. and For bias vectors, It is the sigmoid activation function. The hyperbolic tangent activation function is used. This represents element-wise product.

Images