Tunnel construction early warning method, device and system based on multi-source feature fusion

Abstract

This invention provides a tunnel construction risk early warning method, device, and system based on multi-source feature fusion. By acquiring asynchronous monitoring data streams of surrounding rock deformation, support stress, and microseismic signals, a continuous-time hidden state trajectory generation network is constructed for each sequence to map discrete observations into continuous hidden state trajectories. A dynamic interaction coordination module is invoked to perform temporal coupling interaction processing on the three-source hidden state trajectories to capture the dynamic interaction response between heterogeneous trajectories. At the risk query time, the interaction-coordinated hidden state vector is synchronously extracted and a multi-head cross-attention fusion mechanism is used to generate a multi-source fused hidden state representation vector. Finally, a multi-task prediction network outputs the risk category, severity, and occurrence time window prediction results and generates an early warning command. This invention achieves deep interactive modeling and fusion early warning of multi-source asynchronous monitoring data within a continuous time frame, significantly improving the accuracy of tunnel construction risk early warning.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically, to a method, apparatus, and system for early warning of tunnel construction based on multi-source feature fusion. Background Technology

[0002] Tunnel construction early warning refers to the technology of using data such as surrounding rock deformation, support stress, and microseismic signals collected by various monitoring devices deployed in the tunnel construction area to identify and warn of construction safety risks in advance. Currently, common tunnel construction early warning methods typically align and resample data collected independently from various monitoring sources according to a unified time reference, extract time-domain statistical features, and then input them into an early warning model for risk identification. However, because the acquisition mechanisms and sampling frequencies of different monitoring sources are independent, the resulting monitoring sequences have asynchronous and non-uniform interval characteristics. Forced alignment and resampling can easily disrupt the temporal integrity of the original monitoring signals, leading to the loss of key dynamic features or misalignment of temporal correspondences during multi-source information fusion, making it difficult to reliably guarantee the accuracy and timeliness of the early warning results. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide at least one method, device and system for early warning of tunnel construction based on multi-source feature fusion.

[0004] According to one aspect of the present invention, a tunnel construction early warning method based on multi-source feature fusion is provided, comprising: acquiring asynchronous monitoring data streams collected by multi-source monitoring devices within a tunnel construction area, wherein the asynchronous monitoring data streams include a surrounding rock deformation monitoring sequence, a support stress monitoring sequence, and a microseismic signal monitoring sequence, each monitoring sequence having an independent and non-equally spaced timestamp; constructing continuous-time latent state trajectory generation networks for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence, and the microseismic signal monitoring sequence respectively, mapping the discrete observation values ​​of each monitoring sequence to the continuous posterior latent state trajectories of the corresponding monitoring sources, thereby obtaining a surrounding rock deformation latent state trajectory sequence, a support stress latent state trajectory sequence, and a microseismic signal latent state trajectory sequence; and calling a dynamic interactive coordination module to perform temporal coupling interactive processing on the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence, and the microseismic signal latent state trajectory sequence, capturing the dynamic interactive response between the different source latent state trajectories within a continuous time frame, and generating an interactively coordinated surrounding rock deformation latent state trajectory sequence. The system generates hidden state trajectory sequences for surrounding rock deformation, support stress, and microseismic signals after interactive coordination. At the risk query time, it simultaneously extracts the hidden state vectors of these sequences, resulting in hidden state vectors for surrounding rock deformation, support stress, and microseismic signals. These vectors are then fused asynchronously using a multi-head cross-attention fusion mechanism to generate a multi-source fused hidden state representation vector. Based on this multi-source fused hidden state representation vector, a multi-task prediction network simultaneously outputs the predicted distribution of risk categories, the predicted value of risk severity, and the predicted interval of risk occurrence time windows. Finally, it generates tunnel construction early warning instructions based on these predictions.

[0005] According to another aspect of the present invention, a tunnel construction early warning device is provided, comprising: a data acquisition module, used to acquire asynchronous monitoring data streams collected by multi-source monitoring equipment within the tunnel construction area, the asynchronous monitoring data streams including a surrounding rock deformation monitoring sequence, a support stress monitoring sequence, and a microseismic signal monitoring sequence, each monitoring sequence having an independent and non-equally spaced timestamp; a trajectory mapping module, used to construct continuous-time latent state trajectory generation networks for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence, and the microseismic signal monitoring sequence respectively, mapping the discrete observation values ​​of each monitoring sequence to the continuous posterior latent state trajectories of the corresponding monitoring sources, thereby obtaining a surrounding rock deformation latent state trajectory sequence, a support stress latent state trajectory sequence, and a microseismic signal latent state trajectory sequence; and a sequence generation module, used to call a dynamic interaction coordination module to perform temporal coupling interaction processing on the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence, and the microseismic signal latent state trajectory sequence, capturing the dynamic interaction response between the different source latent state trajectories within a continuous time frame, and generating an interactively coordinated surrounding rock deformation latent state trajectory sequence. The system comprises three components: a hidden state trajectory sequence, a coordinated support stress hidden state trajectory sequence, and a coordinated microseismic signal hidden state trajectory sequence; a cross-fusion module, used to obtain the risk query time and simultaneously extract the hidden state vectors of the coordinated surrounding rock deformation hidden state trajectory sequence, the coordinated support stress hidden state trajectory sequence, and the coordinated microseismic signal hidden state trajectory sequence at the risk query time, thus obtaining the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector; and an asynchronous feature synchronous fusion processing of the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector through a multi-head cross-attention fusion mechanism to generate a multi-source fused hidden state representation vector; and an instruction generation module, used to simultaneously output the risk category prediction distribution, the risk severity prediction value, and the risk occurrence time window prediction interval through a multi-task prediction network based on the multi-source fused hidden state representation vector, and generate tunnel construction early warning instructions based on the risk category prediction distribution, the risk severity prediction value, and the risk occurrence time window prediction interval.

[0006] According to another aspect of the present invention, a computer system is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code, which, when executed by the processor, causes the processor to perform the method described above.

[0007] This invention acquires monitoring sequences of surrounding rock deformation, support stress, and microseismic signals asynchronously collected by multi-source monitoring devices. It then constructs a continuous-time latent state trajectory generation network for each monitoring source, mapping discrete and non-uniformly spaced observations to continuous posterior latent state trajectories. This naturally resolves the challenge of asynchronous data time alignment within a continuous time frame, avoiding information distortion and error accumulation caused by resampling alignment operations. A dynamic interactive coordination module performs temporal coupling and interactive processing on the heterogeneous latent state trajectories, capturing the dynamic interactive responses between monitoring sources in real time along the continuous time dimension. This allows the trajectories to mutually correct each other and fully express the multi-source coupling correlation. At any risk query moment, the latent state vectors of each source are synchronously extracted. A multi-head cross-attention fusion mechanism achieves natural synchronization and deep fusion of asynchronous features, generating a fused latent state vector rich in multi-source collaborative representation capabilities, completely eliminating the dependence on fixed sampling rates and time alignment in traditional methods. Based on this fused vector, a multi-task prediction network simultaneously outputs risk category, severity, and occurrence time window predictions, enabling the early warning results to possess type discrimination, severity assessment, and time positioning capabilities, significantly improving the accuracy and timeliness of tunnel construction risk early warning. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of an application scenario provided by the present invention; Figure 2 This is a logical schematic diagram of a tunnel construction early warning method based on multi-source feature fusion provided by the present invention; Figure 3 This is a flowchart illustrating a tunnel construction early warning method based on multi-source feature fusion provided by the present invention. Figure 4 This is a schematic diagram of the structure of a tunnel construction early warning device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] To facilitate a clearer understanding of this invention, we will first introduce the application scenarios of the tunnel construction early warning method based on multi-source feature fusion, such as... Figure 1 As shown, the application scenario of this invention includes a computer system 10 and a sensor cluster. The sensor cluster may include one or more sensors; the number of sensors is not limited here. Figure 1 As shown, the sensor cluster may specifically include sensor 1, sensor 2, ..., sensor n; it can be understood that sensor 1, sensor 2, sensor 3, ..., sensor n can all be network connected to computer system 10 so that each sensor can interact with computer system 10 through network connection.

[0010] It is understood that computer system 10 can refer to a device that executes the tunnel construction early warning method based on multi-source feature fusion provided in the embodiments of the present invention. Computer system 10 can be, for example, a server, a single physical server, a server cluster consisting of at least two physical servers, or a distributed system. Sensors are selected according to actual conditions. For example, convergence meters can be selected for monitoring surrounding rock deformation, pressure cells can be selected for monitoring support stress, and detectors can be selected for monitoring microseismic signals. Each sensor and computer system 10 can be directly or indirectly connected via wired or wireless communication. Furthermore, the number of sensors and computer system 10 can be one or at least two; the present invention does not impose any limitations on this.

[0011] Further, please see Figure 2 and Figure 3 The tunnel construction early warning method based on multi-source feature fusion provided by this invention consists of... Figure 1 The method is executed by a computer system 10, wherein the tunnel construction early warning method based on multi-source feature fusion may include the following steps: Step S100: Obtain asynchronous monitoring data streams collected by multi-source monitoring equipment within the tunnel construction area. The asynchronous monitoring data streams include surrounding rock deformation monitoring sequences, support stress monitoring sequences, and microseismic signal monitoring sequences. Each monitoring sequence has an independent and non-equally spaced timestamp.

[0012] The tunnel construction area refers to the underground space where excavation and initial support work are underway, as defined by the tunnel engineering design drawings. Within this area, the original stress equilibrium of the rock mass is disrupted by the excavation and unloading process, leading to stress redistribution and deformation response in the surrounding rock. This is the core area of ​​focus for monitoring and early warning. Multi-source monitoring equipment is a collective term for various sensors and acquisition instruments systematically deployed within the tunnel construction area according to the monitoring design scheme. These instruments operate independently based on different physical measurement principles, encompassing three main categories: surrounding rock deformation monitoring equipment, support stress monitoring equipment, and microseismic signal monitoring equipment. Each of these devices possesses an independent internal clock system and data acquisition triggering logic, and there is no unified synchronous clock signal between them. Therefore, the resulting monitoring sequences naturally exhibit asynchronous characteristics in the time dimension.

[0013] Asynchronous monitoring data stream refers to a comprehensive data stream containing multiple monitoring sequences with inconsistent timestamp systems, continuously collected by the aforementioned multi-source monitoring devices according to their respective preset sampling strategies and aggregated in real time to the data server. This data stream does not require the sampling times of each sequence to be aligned, the sampling interval to be constant, or the data update frequency to be the same between different sequences, thus fully preserving the original asynchronous time characteristics of each monitoring source. The surrounding rock deformation monitoring sequence is a one-dimensional ordered observation dataset of the evolution of relative or absolute displacement within the surrounding rock over time, collected by deformation monitoring sensors deployed inside boreholes around the tunnel excavation profile according to their independently set sampling periods. Each observation value in the sequence is associated with a timestamp recorded by the internal clock of the sensor. The support stress monitoring sequence is a one-dimensional ordered observation dataset of the change in the stress state of the support structure over time, collected by stress-strain monitoring sensors installed on the flanges of the steel arch frame, inside the shotcrete layer, or at the ends of the anchor bolts, at their respective independent time intervals. The timestamp of each observation value in this sequence is independently generated by the internal clock of the corresponding sensor, and there is no synchronous correspondence between it and the timestamp of the surrounding rock deformation monitoring sequence. The microseismic signal monitoring sequence is a one-dimensional ordered observation dataset of the distribution of microseismic event characteristic parameters over time, obtained by drilling holes or deploying microseismic sensor arrays inside the tunnel surrounding rock or on the surface, through continuous waveform recording or microseismic event-triggered detection. The timestamps of this sequence depend on the actual occurrence time of the micro-fracture event and exhibit highly irregular interval distribution characteristics.

[0014] Independent and non-equal interval timestamp marks refer to the timestamps attached to each observation within each monitoring sequence being generated independently by the clock module of the monitoring device corresponding to the sequence, without referencing the clock reference of other monitoring devices. Furthermore, the difference between adjacent timestamps within the same monitoring sequence is not a preset fixed constant, but rather a time increment value that may be unequal and naturally formed based on the actual sampling time.

[0015] Step S200: Construct continuous-time latent state trajectory generation networks for the surrounding rock deformation monitoring sequence, support stress monitoring sequence, and microseismic signal monitoring sequence, respectively. Map the discrete observation values ​​of each monitoring sequence to the continuous posterior latent state trajectories of the corresponding monitoring sources to obtain the surrounding rock deformation latent state trajectory sequence, support stress latent state trajectory sequence, and microseismic signal latent state trajectory sequence.

[0016] In one embodiment, step S200 may specifically include the following steps S210 to S260: Step S210: The first variational inference encoder in the first continuous-time hidden state trajectory generation network receives the discrete observation values ​​of surrounding rock deformation at each time stamp in the surrounding rock deformation monitoring sequence and the corresponding time interval information, and outputs the sequence of hidden state trajectories of surrounding rock deformation represented by continuous-time hidden state trajectories.

[0017] The first continuous-time latent state trajectory generation network is a continuous-time latent state modeling network specifically designed for surrounding rock deformation monitoring sources. The core components of this network include two parts: the first variational inference encoder and the first neural differential equation decoder. The first variational inference encoder is responsible for inferring the latent state posterior distribution parameters from the discrete observation values ​​and time interval information of the surrounding rock deformation monitoring sequence. The first neural differential equation decoder is responsible for reconstructing the continuous-time latent state trajectory based on the posterior distribution parameters and generating the reconstructed observation sequence accordingly.

[0018] The first variational inference encoder adopts a time-aware sequence-to-hidden-state encoder architecture, specifically composed of two cascaded parts: a time-reverse sequence encoding layer based on gated cyclic units and a continuous-time hidden-state mapping layer. The discrete observations of surrounding rock deformation at each time stamp in the surrounding rock deformation monitoring sequence refer to all observation data samples collected and aggregated by the surrounding rock deformation monitoring equipment in step S100. Each sample contains a timestamp scalar and at least one surrounding rock deformation observation, which can be a single-point displacement value, a convergent displacement value, or a deep relative displacement value, etc. The time interval information refers to the time increment sequence formed by the difference between the timestamps of two adjacent discrete observations of surrounding rock deformation, as well as the absolute timestamp value corresponding to each observation. This time information characterizes the sparsity and distribution pattern of the observation data on the time axis and is an indispensable input element for continuous-time modeling.

[0019] The sequence of hidden state trajectories of surrounding rock deformation represented by continuous-time hidden state trajectories refers to the fact that the output of the first variational inference encoder is not a set of discrete hidden state vectors, but a set of parameters that can define a continuous-time trajectory. These parameters uniquely determine a smooth hidden state trajectory extending from the start time of the sequence to the end time of the sequence within the framework of neural differential equations. The tangent direction of this trajectory at any time is given by the right-hand side function of the neural differential equations.

[0020] In this embodiment, the specific processing flow of the first variational inference encoder is as follows. First, the gated recurrent unit time-reverse sequence encoding layer receives the observation sequence and time interval sequence of the surrounding rock deformation monitoring sequence. This gated recurrent unit layer consists of a 128-dimensional hidden state bidirectional gated recurrent unit. Its forward propagation direction processes the sequence along the time-increasing direction, and its backward propagation direction processes the sequence along the time-decreasing direction. After bidirectional processing, the forward hidden state and backward hidden state of each time step are concatenated along the feature dimension to obtain the bidirectional context encoding vector for each observation time. The update gate, reset gate, and candidate hidden state calculation of this gated recurrent unit all adopt the standard structure of fully connected linear transformation plus nonlinear activation, wherein the activation function adopts a combination of hyperbolic tangent function and S-shaped growth curve function. After the bidirectional context encoding vector is generated, the continuous-time hidden-state mapping layer takes the encoded vector sequence and the timestamp values ​​of each observation time as input. It employs a hidden-state trajectory parameter inference method based on a neural differential equation initial value problem solver. Specifically, this mapping layer includes a hidden-state mean mapping subnetwork and a hidden-state variance mapping subnetwork. Both subnetworks use a three-layer fully connected perceptron architecture, with each layer containing 128 hidden units. The activation function for the hidden layers is an exponential linear unit function, while the output layer has no activation function. The hidden-state mean mapping subnetwork receives the bidirectional context encoding vector of the last observation time and the timestamp value of that observation time, outputting a parameter representing the initial mean vector of the hidden-state trajectory. The hidden-state variance mapping subnetwork receives the same input and outputs a parameter representing the initial log-variance vector of the hidden-state trajectory. Together, they define the Gaussian posterior distribution of the initial time of the hidden-state trajectory.

[0021] Next, the dynamic evolution of the hidden state trajectory of surrounding rock deformation is defined by the right-hand side function of the neural differential equation. This right-hand side function is a four-layer fully connected neural network, with each layer containing 64-dimensional hidden units. The activation function is a parameterized modified linear unit function. The input is the hidden state vector and timestamp value at the current time, and the output is the time derivative of the hidden state vector. Using the initial hidden state vector sampled from the Gaussian posterior distribution at the initial time of the hidden state trajectory as the initial value, a numerical solver for the initial value problem of the ordinary differential equation, constructed using the Dormand-Prince method with adaptive step size, is called. Integration is performed over the time interval from the start to the end of the sequence to generate the hidden state vector value at any specified time. The first variational inference encoder packages the initial distribution parameters of the hidden state trajectory with the network parameters of the right-hand side function of the neural differential equation as a compact representation of the continuous-time hidden state trajectory, thus outputting a sequence of hidden state trajectories of surrounding rock deformation represented by continuous-time hidden state trajectories. During training, the network employs a reparameterization technique to sample the initial hidden state values ​​from the posterior distribution to ensure that the gradient can be backpropagated to the parameters of the variational inference encoder through the numerical solver.

[0022] Step S220: The second variational inference encoder in the second continuous-time hidden state trajectory generation network receives the discrete observation values ​​of support stress at each time stamp in the support stress monitoring sequence and the corresponding time interval information, and outputs the support stress hidden state trajectory sequence represented by the continuous-time hidden state trajectory.

[0023] The second continuous-time latent state trajectory generation network is a continuous-time latent state modeling network specifically designed for support stress monitoring sources. Its network topology is isomorphic to the first continuous-time latent state trajectory generation network, but its parameters are independent, ensuring that the latent state trajectories of each monitoring source learn their own unique dynamic modes in their respective independent parameter spaces. The second variational inference encoder also adopts a cascaded architecture of a time-aware gated cyclic unit time-reverse sequence encoding layer plus a continuous-time latent state mapping layer. The latent state dimension of its gated cyclic unit layer is still set to 128 dimensions, and the continuous-time latent state mapping layer contains independent latent state mean mapping subnetworks and latent state variance mapping subnetworks. The discrete observation values ​​of support stress at each time stamp in the support stress monitoring sequence refer to all observation data samples collected by the support stress monitoring equipment in step S100. Each sample contains a timestamp scalar and at least one support stress observation. The observation can be the stress value of the steel arch flange, the stress value inside the shotcrete layer, or the axial force value of the anchor bolt, etc. These observation values ​​reflect the degree of mechanical response of the support structure under the action of surrounding rock pressure. The generation process of the support stress hidden state trajectory sequence is completely parallel to the generation process of the surrounding rock deformation hidden state trajectory sequence in terms of computational logic: the gated cyclic unit time-reverse sequence encoding layer of the second variational inference encoder first performs bidirectional temporal encoding on the discrete observation value sequence and time interval sequence of support stress, extracts the bidirectional context encoding vector of each observation time, and then the continuous time hidden state mapping layer uses the encoding vector and timestamp value of the last observation time to deduce the Gaussian posterior distribution parameters of the initial time of the hidden state trajectory. The right-hand function of the second neural differential equation decoder adopts a four-layer fully connected network with the same architecture as the first continuous time hidden state trajectory generation network but with independent parameters. The initial hidden state vector obtained by sampling the distribution is used as the initial value. The sequence is integrated over the time interval by the same type of Dormand-Prince ordinary differential equation numerical solver, and the support stress hidden state trajectory sequence represented by the continuous time hidden state trajectory can be obtained.

[0024] Step S230: The third variational inference encoder in the third continuous-time hidden state trajectory generation network receives the discrete observation values ​​of microseismic signals at each time point in the microseismic signal monitoring sequence and the corresponding time interval information, and outputs the microseismic signal hidden state trajectory sequence represented by the continuous-time hidden state trajectory.

[0025] The third continuous-time hidden state trajectory generation network is a continuous-time hidden state modeling network specifically designed for microseismic signal monitoring sources. Its network topology is isomorphic to that of the first continuous-time hidden state trajectory generation network, but it has an independent set of parameters. The discrete observations of microseismic signals refer to the microseismic event characteristic parameter observation sequence obtained by the microseismic signal monitoring equipment in step S100 and processed by the source parameter extraction algorithm. Each observation contains a timestamp scalar and at least one microseismic event characteristic parameter. The characteristic parameter can be the magnitude value of the microseismic event, the three-dimensional coordinates of the source location, the apparent stress value, or the seismic moment value, etc. These parameters characterize the intensity and spatial distribution characteristics of micro-fracture activity inside the surrounding rock from different aspects. The generation process of the microseismic signal hidden state trajectory sequence adopts the same computational paradigm as the two monitoring sources mentioned above. That is, the gated cyclic unit time-reverse sequence encoding layer of the third variational inference encoder receives the discrete observation value sequence and time interval sequence of the microseismic signal, and outputs the context encoding vector of each observation time after bidirectional encoding. The continuous time hidden state mapping layer infers the initial posterior distribution parameters of the hidden state trajectory from the encoding vector and timestamp value of the last observation time. The right-hand function of the third neural differential equation decoder performs numerical integration based on the initial hidden state vector sampled by the distribution and the time interval to generate the microseismic signal hidden state trajectory sequence represented by the continuous time hidden state trajectory. It is worth noting that, since the sparsity and non-uniformity of microseismic signal monitoring sequences are usually much higher than those of the first two types of monitoring sequences, the third variational inference encoder adopts a time decay gating mechanism in the time interval input processing of the time-reverse sequence encoding layer of the gated cyclic unit. That is, when recursively encoding the observations, the proportion of the hidden state from the previous moment to the recursive input is exponentially decayed according to the time interval between adjacent observations. The larger the interval, the more significant the contribution weight of the hidden state from the previous moment decays, thus adapting to the highly sparse distribution characteristics of microseismic events on the time axis.

[0026] Step S240: Input the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal into the initial dynamic interactive coordination module, perform multi-round time-series interactive coding processing, update the hidden state of the third monitoring source based on the current hidden state of any two monitoring sources in each round of processing, and obtain the initial interactive coordination trajectory sequence set through cross-iteration update.

[0027] In one embodiment, step S240 may specifically include the following steps S241 to S246: Step S241: The first interactive coding layer of the initial dynamic interactive coordination module performs time alignment on the input hidden state trajectory sequence of surrounding rock deformation, hidden state trajectory sequence of support stress, and hidden state trajectory sequence of microseismic signal to generate time-synchronized hidden state trajectory sequences of surrounding rock deformation, support stress, and microseismic signal.

[0028] The first interactive coding layer is the functional layer responsible for time alignment and trajectory resampling in the initial dynamic interactive coordination module. Its core function is to unify the latent state trajectory sequences from three different time bases onto the same standard time grid, providing a three-source latent state vector sequence with the same time coordinates for subsequent pairwise cross-attention calculations. Time alignment processing involves determining a common standard time interval based on the time coverage of each of the three latent state trajectory sequences, generating a set of equally spaced standard time points within this interval using a uniform time step, and then utilizing the characteristic that each latent state trajectory sequence is a continuous time function that can be evaluated at any time. At each standard time point, the latent state vector values ​​corresponding to the three latent state trajectories are queried, resulting in three sets of latent state vector sequences that are strictly synchronized in time. The time-synchronized surrounding rock deformation latent state trajectory sequence refers to the latent state vector sequence obtained by querying the surrounding rock deformation latent state trajectory function on the standard time grid at each time step, with its timestamps fully aligned with the other two monitoring sources. The time-synchronized support stress latent state trajectory sequence refers to the latent state vector sequence obtained by querying the support stress latent state trajectory function on the same standard time grid. The time-synchronized microseismic signal latent state trajectory sequence refers to the sequence of latent state vectors obtained by querying the latent state trajectory function of the microseismic signal on the same standard time grid.

[0029] For example, the process of determining the standard time grid in the first interactive coding layer is as follows: The start and end times of the time coverage of each of the three hidden state trajectory sequences are obtained. The maximum value among the three start times is taken as the common start time, and the minimum value among the three end times is taken as the common end time. Within this common time interval, a sequence of equally spaced standard time points is generated using a preset standard sampling step size. This standard sampling step size is selected based on the time granularity characteristics of the construction monitoring data, and is a value of the least common divisor of the average sampling interval of each monitoring source. After the standard time point sequence is determined, for each standard time point, the hidden state trajectory function of the surrounding rock deformation, the hidden state trajectory function of the support stress, and the hidden state trajectory function of the microseismic signal are evaluated at that time point. The specific implementation of the evaluation operation is as follows: using the standard time point as the time input and the initial hidden state vector of each as the initial value, driven by the right-hand side function of each neural differential equation, the hidden state vector value at that time point is obtained by integrating from the initial time to the standard time point through the numerical solver of the ordinary differential equation. The three sets of time-synchronized hidden state trajectory sequences generated are completely consistent in sequence length and sampling time, providing a unified time coordinate system for attention calculation in subsequent interactive coding layers.

[0030] Step S242: Through the second interactive coding layer of the initial dynamic interactive coordination module, calculate the first unidirectional response intensity of the time-synchronized surrounding rock deformation latent state trajectory sequence to the time-synchronized microseismic signal latent state trajectory sequence, and update the time-synchronized microseismic signal latent state trajectory sequence accordingly.

[0031] The second interactive coding layer is the functional layer in the initial dynamic interactive coordination module that performs the first-direction unidirectional interactive coding. Its core operation is a cross-attention mechanism. This mechanism uses the time-synchronized hidden state trajectory sequence of surrounding rock deformation as the query source and the time-synchronized hidden state trajectory sequence of microseismic signals as the key and value source. By calculating the attention weight between the query vector and the key vector, it quantifies the response intensity of the hidden state of surrounding rock deformation to the hidden state of microseismic signals. Then, it uses this response intensity to perform weighted aggregation on the value vector of the hidden state of microseismic signals, thereby achieving the goal of updating the hidden state trajectory of microseismic signals with surrounding rock deformation information. The first unidirectional response intensity refers to the magnitude of the interactive influence of the hidden state of surrounding rock deformation on the hidden state of microseismic signals at each standard time point. This magnitude is reflected by the cross-attention weight matrix. The higher the value, the stronger the correlation between the hidden state pattern of surrounding rock deformation and the hidden state pattern of microseismic signals at that time.

[0032] Specifically, the internal structure of the second interactive coding layer includes a multi-head cross-attention sublayer and a residual connection and layer normalization sublayer. The multi-head cross-attention sublayer first maps the latent state vector at each moment in the time-synchronized rock deformation latent state trajectory sequence to a query vector using a query linear transformation matrix. It then maps the latent state vector at each moment in the time-synchronized microseismic signal latent state trajectory sequence to a key vector and a value vector, respectively, using a key linear transformation matrix and a value linear transformation matrix. All three transformation matrices are trainable fully connected linear transformation parameter matrices. In each attention head, the scaled dot product between the query vector and the key vector is calculated (i.e., the inner product of the query vector and the key vector divided by the square root of the key vector's dimension), yielding the original attention score matrix. A flexible maximum normalization operation is applied to each row of this score matrix to obtain the attention weight matrix. Each row of the attention weight matrix corresponds to the distribution of attention weights of the latent state of rock deformation at a given moment to the latent state of microseismic signals at various moments. The attention weight matrix is ​​used to perform a weighted summation of the value vector sequence to obtain the microseismic signal interaction response vector sequence modulated by the surrounding rock deformation information under this attention head. The interaction response vector sequences of all attention heads are concatenated along the feature dimension, and then mapped back to the original hidden state dimension through an output linear transformation matrix to obtain the output of the multi-head cross-attention sublayer. The residual connection and layer normalization sublayer adds the output to the time-synchronized microseismic signal hidden state trajectory sequence at each time point, and then performs layer normalization along the feature dimension to obtain the updated time-synchronized microseismic signal hidden state trajectory sequence. The layer normalization is calculated as follows: for the hidden state vector at each time point, first calculate its mean and variance along each feature dimension, then subtract the mean from the vector, divide by the sum of the variance and a small constant, multiply by a trainable scaling parameter vector, and add the offset parameter vector.

[0033] Step S243: Calculate the second unidirectional response intensity of the updated time-synchronized microseismic signal latent state trajectory sequence to the time-synchronized support stress latent state trajectory sequence through the third interactive coding layer of the initial dynamic interactive coordination module, and update the time-synchronized support stress latent state trajectory sequence accordingly.

[0034] The third interactive coding layer is structurally isomorphic to the second interactive coding layer but has an independent parameter matrix. This layer uses the updated, time-synchronized microseismic signal latent state trajectory sequence as the query source and the time-synchronized support stress latent state trajectory sequence as the key and value source. Through a multi-head cross-attention sublayer, residual connections, and layer normalization sublayer with the same architecture, it calculates the second unidirectional response intensity of the microseismic signal latent state to the support stress latent state, and uses this response intensity to perform a weighted update of the support stress latent state trajectory. The second unidirectional response intensity characterizes the degree of interactive influence of the microseismic signal latent state on the support stress latent state at each standard time point. The updated, time-synchronized support stress latent state trajectory sequence has simultaneously absorbed the indirect influence from the microseismic signal latent state and the deeper indirect influence of the surrounding rock deformation latent state transmitted through the microseismic signal, thus initially demonstrating the effect of cross-iteration.

[0035] Step S244: Calculate the third unidirectional response intensity of the updated time-synchronized support stress latent state trajectory sequence to the time-synchronized surrounding rock deformation latent state trajectory sequence through the fourth interactive coding layer of the initial dynamic interactive coordination module, and update the time-synchronized surrounding rock deformation latent state trajectory sequence accordingly.

[0036] The fourth interactive coding layer is structurally identical to the second interactive coding layer and has independent parameters. This layer uses the updated, time-synchronized support stress latent state trajectory sequence as the query source and the time-synchronized surrounding rock deformation latent state trajectory sequence as the key and value source. It calculates the third unidirectional response intensity of the support stress latent state to the surrounding rock deformation latent state through a multi-head cross-attention sublayer, and updates the time-synchronized surrounding rock deformation latent state trajectory sequence accordingly. Thus, the three monitoring sources complete a closed-loop cross-update in a single round of interactive coding: surrounding rock deformation affects microseismic signals, microseismic signals affect support stress, and support stress reacts to surrounding rock deformation, forming a complete three-source cyclical interactive pathway.

[0037] Step S245: The time-synchronized hidden state trajectory sequence of surrounding rock deformation, the time-synchronized hidden state trajectory sequence of support stress, and the time-synchronized hidden state trajectory sequence of microseismic signal obtained in this round of update are residually connected with the corresponding trajectory sequences in the previous round to generate the hidden state trajectory sequence of surrounding rock deformation with cumulative interaction effect, the hidden state trajectory sequence of support stress with cumulative interaction effect, and the hidden state trajectory sequence of microseismic signal with cumulative interaction effect.

[0038] In this context, residual connection refers to element-wise addition of the latent state trajectory sequences output after the current round of cross-iteration update with the latent state trajectory sequences of the corresponding monitoring sources from the previous round of interactive coding, along with the latent state vectors at each time point. The purpose of this design is to provide an identity mapping path for the cumulative propagation of interactive information during multi-round interactive coding, ensuring that the deep interactive coding layer does not lose the effective interactive features extracted by the shallow interactive coding layer while learning complex interactive patterns. The latent state trajectory sequence of surrounding rock deformation due to cumulative interactive effects refers to the sequence obtained after the current round of residual connection processing. This sequence not only includes the direct response information from the latent state of support stress introduced by the current round of cross-update but also retains the interactive response imprints accumulated on this trajectory sequence from all previous rounds of interactive coding through the residual path. The definitions of the latent state trajectory sequence of support stress due to cumulative interactive effects and the latent state trajectory sequence of microseismic signals due to cumulative interactive effects are similar.

[0039] Step S246: If the current interaction round does not reach the preset round limit, the trajectory sequences of the accumulated interaction effect are used as the input for the next round and the above interaction coding layer processing is repeated. If the limit is reached, the trajectory sequences of the final accumulated interaction effect are output as the initial interaction coordination trajectory sequence set.

[0040] The preset upper limit for the number of rounds is a hyperparameter controlling the iteration depth of the interactive coding, and its value is set based on the complexity of the physical coupling relationship between the three monitoring sources. When the trajectory sequences of the cumulative interaction effect are used as input for the next round, the first interactive coding layer re-executes the time alignment operation, and the second to fourth interactive coding layers re-execute the cross-attention update and residual connection in sequence. After each round, the degree of interaction coordination between the hidden state trajectories of the three sources deepens by one level. When the number of interaction rounds reaches the preset upper limit, the three hidden state trajectory sequences of the cumulative interaction effect output in the last round constitute the initial interactive coordination trajectory sequence set, which provides input for subsequent training loss metric calculation and network weight update.

[0041] Step S250: Construct a training loss metric based on the initial set of interactive coordination trajectory sequences. The training loss metric is a weighted combination of the reconstruction difference metric generated by each monitoring sequence after dimensional unification and the interaction consistency metric between the hidden state trajectory sequences of each monitoring source. The network weights of the first continuous time hidden state trajectory generation network, the second continuous time hidden state trajectory generation network, the third continuous time hidden state trajectory generation network and the initial dynamic interactive coordination module are jointly adjusted by minimizing the training loss metric.

[0042] The training loss metric is the objective function guiding the end-to-end joint optimization of the entire network system. This function is composed of a weighted combination of two components with different physical meanings. Minimizing this function value drives the network weights to converge in a direction that simultaneously considers reconstruction accuracy and interaction consistency. Dimensional unification refers to the fact that the physical quantities corresponding to the three monitoring sources—surrounding rock deformation, support stress, and microseismic signals—have drastically different dimensions and numerical magnitudes. For example, the dimension of surrounding rock deformation is length units, the dimension of support stress is force divided by area, and the microseismic magnitude is dimensionless logarithmic. Simply summing the reconstruction differences among the three will lead to the larger dimensional physical quantity dominating the gradient direction of the loss function. Therefore, each reconstruction difference metric must first be converted into a dimensionless relative index to eliminate the influence of dimensions before fair weighted merging can be performed. The reconstruction difference metric for each monitoring sequence is a comprehensive index measuring the degree of difference between the reconstructed monitoring sequences generated by each continuous-time hidden state trajectory generation network and its corresponding decoder and the original monitoring sequences. This index reflects the fidelity of the hidden state trajectory to the original observation data. The interaction consistency metric among the hidden state trajectory sequences of each monitoring source is an indicator that measures the degree of synchronization in the evolutionary trends of the three hidden state trajectory sequences after time synchronization. This indicator guides the network to learn a physically reasonable three-source coupling relationship during the interactive encoding process. That is, the surrounding rock deformation, support stress change, and microseismic activity should have a certain degree of mutual correspondence in their temporal evolution patterns, rather than drifting independently. Joint adjustment refers to calculating the partial derivative of the training loss metric with respect to all trainable parameters of all four network modules through the backpropagation algorithm, and updating all parameters simultaneously based on this partial derivative. This end-to-end joint optimization method enables the continuous-time hidden state trajectory generation network and the initial dynamic interaction coordination module to co-evolve during training. The generation process of the hidden state trajectory not only considers the reconstruction capability of its own monitoring source observation data, but also takes into account the interaction coordination needs with the hidden state trajectories of other monitoring sources.

[0043] In one implementation, step S250 may specifically include the following steps S251 to S256: Step S251: Input the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal from the initial interactive coordination trajectory sequence set into the corresponding generation and reconstruction decoder to generate the reconstructed surrounding rock deformation monitoring sequence, the reconstructed support stress monitoring sequence, and the reconstructed microseismic signal monitoring sequence.

[0044] The generative reconstruction decoder is an essential component of each continuous-time hidden-state trajectory generation network. It is responsible for mapping the hidden-state vectors on the continuous-time hidden-state trajectory back to the original observation space, generating a sequence of reconstructed observation values ​​that corresponds one-to-one with the original monitoring sequence at each time stamp. The generative reconstruction decoder for the surrounding rock deformation hidden-state trajectory sequence is a four-layer fully connected decoding network with 128 hidden units per layer. The activation function is an exponential linear unit function, and the number of units in the output layer is consistent with the observation dimension of the surrounding rock deformation monitoring sequence. The input is the hidden-state vector value of the surrounding rock deformation hidden-state trajectory at each time stamp of the original surrounding rock deformation monitoring sequence, and the output is the reconstructed discrete observation value of the surrounding rock deformation at that time stamp. The generative reconstruction decoder for the support stress hidden-state trajectory sequence also adopts a four-layer fully connected network architecture with independent parameters. The output layer dimension is consistent with the observation dimension of the support stress monitoring sequence. Its input is the hidden-state vector value of the support stress hidden-state trajectory at each time stamp of the original support stress monitoring sequence, and the output is the reconstructed discrete observation value of the support stress at that time stamp. The generator-reconstruction decoder corresponding to the microseismic signal latent state trajectory sequence adopts the same four-layer fully connected architecture with independent parameters. The output layer dimension is consistent with the observation dimension of the microseismic signal monitoring sequence. The input is the latent state vector value of the microseismic signal latent state trajectory at each time stamp of the original microseismic signal monitoring sequence, and the output is the reconstructed discrete observation value of the microseismic signal at that time stamp. The reconstructed surrounding rock deformation monitoring sequence is a one-dimensional sequence composed of the reconstructed discrete observation values ​​of surrounding rock deformation at all time stamps arranged in chronological order, and its sequence length is completely consistent with the original surrounding rock deformation monitoring sequence. The definition methods of the reconstructed support stress monitoring sequence and the reconstructed microseismic signal monitoring sequence are similar.

[0045] Step S252: Calculate the reconstruction difference of surrounding rock deformation based on the difference between the reconstructed surrounding rock deformation monitoring sequence and the original surrounding rock deformation monitoring sequence; calculate the reconstruction difference of support stress based on the difference between the reconstructed support stress monitoring sequence and the original support stress monitoring sequence; calculate the reconstruction difference of microseismic signal based on the difference between the reconstructed microseismic signal monitoring sequence and the original microseismic signal monitoring sequence; perform dimensionless processing on the reconstruction difference of surrounding rock deformation, the reconstruction difference of support stress, and the reconstruction difference of microseismic signal respectively; after eliminating the dimensional differences of each monitored physical quantity, the three reconstruction difference quantities after processing are weighted and summed to obtain the generated reconstruction difference metric.

[0046] The calculation of the reconstructed rock deformation difference uses the mean square error function. Specifically, for all timestamps in the original rock deformation monitoring sequence, the square of the difference between the reconstructed discrete observation value and the original discrete observation value of rock deformation at each timestamp is calculated. The sum of these squared differences across all timestamps is then divided by the total number of timestamps to obtain the original value of the reconstructed rock deformation difference. The reconstructed difference of support stress and the reconstructed difference of microseismic signals are calculated independently using the exact same mean square error function. The dimensionless processing is performed as follows: for the reconstructed rock deformation difference, it is divided by the variance of all observations in the original rock deformation monitoring sequence to obtain a dimensionless ratio of the reconstructed rock deformation difference. This ratio represents the proportion of the reconstruction error relative to the fluctuation of the original data. Similarly, for the reconstructed support stress difference, it is divided by the variance of all observations in the original support stress monitoring sequence; and for the reconstructed microseismic signal difference, it is divided by the variance of all observations in the original microseismic signal monitoring sequence. After variance normalization, the three dimensionless reconstruction differences are numerically freed from the influence of physical quantity dimensions and are all within comparable orders of magnitude. The generated reconstruction difference metric is obtained by weighting and summing the three processed dimensionless reconstruction differences according to a preset proportional weight, which reflects the relative emphasis on the reconstruction accuracy requirements of different monitoring sources.

[0047] Step S253: At the end of each round of interactive coding, extract the hidden state change trends of the time-synchronized surrounding rock deformation hidden state trajectory sequence, the time-synchronized support stress hidden state trajectory sequence, and the time-synchronized microseismic signal hidden state trajectory sequence at continuous time intervals, and construct an interactive consistency metric based on the degree of synchronization between the hidden state change trends.

[0048] The latent state change trend refers to the difference in latent state vectors between adjacent standard time points in each time-synchronized latent state trajectory sequence. This vector difference reflects the direction and rate of evolution of the latent state over continuous time. The method for extracting the latent state change trend is as follows: for a time-synchronized rock deformation latent state trajectory sequence, subtract the latent state vector of the previous standard time point from the latent state vector of the later standard time point to obtain a trend vector sequence representing the change trend. The length of this trend vector sequence is one less than the original sequence. The same processing is performed on the time-synchronized support stress latent state trajectory sequence and the time-synchronized microseismic signal latent state trajectory sequence. The interactive consistency measure is constructed based on the degree of synchronization between the latent state change trends. The degree of synchronization is quantified by calculating the cosine similarity between the trend vector sequences. Specifically, for each standard time interval, the cosine similarity between the rock deformation trend vector and the support stress trend vector, the cosine similarity between the rock deformation trend vector and the microseismic signal trend vector, and the cosine similarity between the support stress trend vector and the microseismic signal trend vector are calculated respectively. The average of these three cosine similarities over all time intervals is used as the original value of the interactive consistency measure. The closer the cosine similarity is to its maximum value, the more consistent the evolution direction of the hidden states of the two monitoring sources is within that time interval, and the higher the degree of interaction coupling. In order to make the interaction consistency measure and the generation-reconstruction difference measure both minimized in the loss function, the final value of the interaction consistency measure is the negative value of the original average cosine similarity plus a constant offset, so that the higher the interaction consistency, the smaller the value of this measure.

[0049] Step S254: Multiply the generated reconstruction difference metric and the interaction consistency metric by the preset interaction consistency constraint weights and then add them together to obtain the training loss metric.

[0050] The preset interaction consistency constraint weights are hyperparameters that control the relative contribution of the interaction consistency metric to the total training loss. These weights are determined through grid search based on the model performance on the validation set before training begins. The training loss metric is the direct sum of the generated reconstruction difference metric and the interaction consistency metric weighted by the interaction consistency constraint weights.

[0051] Step S255: Calculate the adjustment gradient of the training loss metric relative to the weights of each network in the first continuous-time hidden state trajectory generation network, the second continuous-time hidden state trajectory generation network, the third continuous-time hidden state trajectory generation network, and the initial dynamic interaction coordination module through backpropagation.

[0052] Backpropagation is the standard algorithm for calculating the partial derivatives of the loss function with respect to the parameters of each layer in neural network training. The calculation process starts from the training loss metric and propagates the gradient signal backward layer by layer along the forward computation path of the network. For the numerical solver of the ordinary differential equations involved in continuous-time hidden-state trajectory generation networks, backpropagation does not directly solve the computation graph within the solver for automatic differentiation. Instead, it uses the adjoint sensitivity method to calculate the gradient of the training loss metric with respect to the parameters of the right-hand side of the neural differential equation and the initial hidden-state distribution parameters. The adjoint sensitivity method efficiently calculates the required gradient by solving an augmented adjoint ordinary differential equation in reverse, avoiding the memory overhead of storing all intermediate states of the forward solver. The gradients of the fully connected matrix parameters, query key-value linear transformation parameters, and layer normalization scaling and offset parameters of each interactive coding layer within the initial dynamic interaction coordination module are automatically calculated using standard backpropagation.

[0053] Step S256: Update the weights of each network based on the adjusted gradient, and repeatedly perform generation reconstruction, difference metric calculation and weight update until the training loss metric stabilizes.

[0054] The adaptive moment estimation algorithm is used to update the network weights based on gradient adjustment. This algorithm maintains an independent adaptive learning rate for each trainable parameter and adjusts the update step size by comprehensively considering the first and second moment estimates of the gradient. It has the advantages of fast convergence speed and robust performance on non-stationary objective functions. The initial learning rate, first moment decay factor, and second moment decay factor of the adaptive moment estimation algorithm all adopt the default values ​​recommended in the original paper of the algorithm. Training adopts a mini-batch stochastic gradient descent organization method. Each mini-batch randomly samples several asynchronous monitoring data stream segments from different time intervals and different construction stages from the training dataset to ensure that the interaction patterns learned by the network have temporal and spatial generalization capabilities. The iterative execution of generation reconstruction, difference metric calculation, and weight update refers to the repeated execution of the complete process of steps S251 to S256. Each round of iteration is called a training iteration step. One training cycle is defined as the complete process of traversing all training data. After iterative optimization for multiple training cycles, when the change of the training loss metric is lower than the preset convergence threshold in multiple consecutive training cycles, the training loss metric is considered to have stabilized, and training stops. After training converges, all parameters of the first continuous-time hidden state trajectory generation network, the second continuous-time hidden state trajectory generation network, the third continuous-time hidden state trajectory generation network, and the initial dynamic interaction coordination module are fixed and saved as model weights for subsequent inference application stages.

[0055] Step S260: After the training loss metric converges, the initial dynamic interaction coordination module that has been trained is used as the dynamic interaction coordination module. The hidden state trajectory sequences of surrounding rock deformation, support stress, and microseismic signal output by the first continuous time hidden state trajectory generation network, the second continuous time hidden state trajectory generation network, and the third continuous time hidden state trajectory generation network are input into the dynamic interaction coordination module for temporal coupling interaction processing.

[0056] The initial dynamic interaction coordination module, after training, has fully grasped the reasonable interaction response mode between the hidden state trajectories of the three monitoring sources through the aforementioned end-to-end joint optimization. The parameters of each interaction coding layer within it have remained fixed after iterative adjustments using a large amount of construction monitoring data. At this point, this module is formally named the Dynamic Interaction Coordination Module to distinguish it from the initial module in the training phase. After directly inputting the hidden state trajectory sequences output by the three continuous-time hidden state trajectory generation networks into the Dynamic Interaction Coordination Module, the module will perform time alignment, cross-attention updates, and residual connection operations on the input trajectory sequences according to the same multi-round temporal interaction coding processing flow as in the training phase. It will output the hidden state trajectory sequence after deep interaction coupling coordination; this process is the core of the temporal coupling interaction processing.

[0057] Step S300: Call the dynamic interactive coordination module to perform time-series coupling interactive processing on the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal. Capture the dynamic interactive response between the hidden state trajectories of different sources in a continuous time frame, and generate the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signal.

[0058] Calling the dynamic interaction coordination module refers to inputting the three-source hidden state trajectory sequence obtained after processing the newly acquired asynchronous monitoring data stream in step S200 into the trained dynamic interaction coordination module during the inference application phase, triggering the module to execute forward inference calculations. Capturing the dynamic interaction response between heterogeneous hidden state trajectories within a continuous time frame is the core technical capability of the dynamic interaction coordination module. This means that the module's cross-attention mechanism does not calculate the interaction response at fixed discrete time steps, but rather on a continuous-time standard grid after time synchronization. Since the hidden state vector at each standard time point essentially comes from the evaluation of the continuous-time hidden state trajectory function, the intensity of the interaction response between different times can also be continuously captured through dense sampling of the standard grid, forming an interaction response function defined in continuous time. The rock deformation hidden state trajectory sequence after interaction coordination refers to the rock deformation hidden state trajectory sequence output after multiple rounds of cross-interactive encoding processing by the dynamic interaction coordination module. This sequence not only contains the inherent deformation dynamics mode of the surrounding rock itself, but also embeds the coupled information of the dynamic response of support stress changes and microseismic activity to rock deformation. The definitions of the support stress latent state trajectory sequence after interactive coordination and the microseismic signal latent state trajectory sequence after interactive coordination are similar. Together, the three constitute a complete latent state trajectory representation system after three-source interactive coordination.

[0059] In one embodiment, step S300 may specifically include the following steps S310~S360: Step S310: Input the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal into the trajectory time parameter unification layer of the dynamic interactive coordination module, extract the timestamps and time intervals of each hidden state trajectory sequence, and generate a three-source hidden state trajectory time alignment sequence with a unified time reference.

[0060] The trajectory time parameter unification layer is the input preprocessing layer of the dynamic interactive coordination module. This layer receives three consecutive time-series hidden state trajectory sequences, parses the timestamp information carried by each sequence, determines the overlapping interval of the time coverage of the three trajectories, generates a uniform standard time point sequence with equal spacing within the overlapping interval, and then uses a neural differential equation initial value problem solver to obtain the function values ​​of the three hidden state trajectory functions at each standard time point, resulting in three sets of time-strictly aligned hidden state vector sequences. Extracting the timestamps and time intervals of each hidden state trajectory sequence refers to reading the start timestamp, end timestamp, and adjacent time interval information of the initial observation sequence from the data structure of the three hidden state trajectory sequences. This information has already been retained as trajectory metadata when constructing the hidden state trajectory in step S200. The three-source hidden state trajectory time-aligned sequence with a unified time reference means that the three sets of hidden state vector sequences obtained after time alignment calculation have completely identical time coordinate axes. Each standard time point corresponds to three hidden state vectors, which come from the hidden state trajectories of surrounding rock deformation, support stress, and microseismic signals, respectively, providing input data with strict time-dimensional registration for the subsequent bidirectional cross-interactive coding layer.

[0061] Step S320: Input the time-aligned sequence of the three-source hidden state trajectories into the bidirectional cross-interactive coding layer of the dynamic interactive coordination module. Calculate the instantaneous response vector of the hidden state of surrounding rock deformation to the hidden state of support stress through the first directional coding branch, and calculate the instantaneous response vector of the hidden state of support stress to the hidden state of surrounding rock deformation through the second directional coding branch. After splicing, generate the bidirectional interactive response vector of surrounding rock deformation and support stress.

[0062] The bidirectional cross-interactive coding layer is the core computational layer of the dynamic interactive coordination module. This layer consists of multiple paired coding branches. Each paired branch performs bidirectional cross-attention calculations for two monitoring sources, generating bidirectional interactive response vectors. The first directional coding branch uses the hidden state of surrounding rock deformation as the active party and the hidden state of support stress as the passive party. It calculates the instantaneous response information received by the hidden state of surrounding rock deformation from the hidden state of support stress at each unified reference time through cross-attention calculations. The instantaneous response vector refers to the vector representing the instantaneous interactive influence of support stress on surrounding rock deformation, obtained by attention aggregation at a single unified reference time, using the hidden state vector of surrounding rock deformation at that time as the query and the hidden state vectors of support stress at that time and adjacent times as the key and value. This vector has the same feature dimension as the original hidden state vector of surrounding rock deformation. The second directional coding branch uses the hidden state of support stress as the active party and the hidden state of surrounding rock deformation as the passive party, calculating the instantaneous response vector received by the hidden state of support stress from the hidden state of surrounding rock deformation through cross-attention calculations. The bidirectional interactive response vector is formed by concatenating the first-direction instantaneous response vector and the second-direction instantaneous response vector along the feature dimension and then selectively filtering them through a gating fusion unit. This vector fully depicts the bidirectional interactive coupling state of the two monitoring sources at each unified reference time.

[0063] In one embodiment, step S320 may specifically include the following steps S321 to S325: Step S321: Obtain the hidden state vector of surrounding rock deformation, the hidden state vector of support stress, and the hidden state vector of microseismic signal for each time stamp in the three-source hidden state trajectory time alignment sequence. Generate a surrounding rock deformation query vector and a surrounding rock deformation key vector from the hidden state vector of surrounding rock deformation through the first linear projection layer. Generate a support stress query vector and a support stress key vector from the hidden state vector of support stress through the second linear projection layer.

[0064] The first linear projection layer is a trainable linear transformation module applied to the latent state vector of surrounding rock deformation. This module contains two independent fully connected linear transformation matrices: a query projection matrix, which maps the latent state vector of surrounding rock deformation to a query vector, and a key projection matrix, which maps the same latent state vector of surrounding rock deformation to a key vector. The output dimensions of the two projection matrices are equal, both being the preset internal dimension of the attention head multiplied by the number of attention heads. The second linear projection layer is a trainable linear transformation module applied to the latent state vector of support stress. Its structure is completely symmetrical with the first linear projection layer, containing a query projection matrix and a key projection matrix, which map the latent state vector of support stress to the support stress query vector and the support stress key vector, respectively. The output dimension is consistent with the first linear projection layer.

[0065] Step S322: Generate the interaction attention weight of the surrounding rock deformation on the support stress by the inner product of the surrounding rock deformation query vector and the support stress key vector. Obtain the support stress value vector by passing the support stress hidden state vector through the third linear projection layer. The support stress value vector is weighted and aggregated by the interaction attention weight to obtain the support stress interaction response vector received by the surrounding rock deformation as the output of the first direction encoding branch.

[0066] In the attention mechanism, the inner product of the surrounding rock deformation query vector and the support stress key vector represents the degree of matching between the two vectors. The larger the inner product value, the more correlated the latent state pattern corresponding to the surrounding rock deformation query vector is with the latent state pattern corresponding to the support stress key vector. The inner product is calculated for the surrounding rock deformation query vector at the same time and the support stress key vectors at all times. This is then scaled by dividing by the square root of the key vector dimension, and finally, a flexible maximum normalization operation is applied along the time dimension to obtain the interactive attention weight distribution of the surrounding rock deformation at that time with respect to the support stress at each time. The third linear projection layer is a value projection matrix applied to the support stress latent state vector, linearly mapping the support stress latent state vector to a value vector. The dimension of the value vector is the same as that of the query vector and the key vector. The weighted aggregation of the support stress value vector by interactive attention weights refers to, for the hidden state of surrounding rock deformation at each time step, using the corresponding interactive attention weight distribution as coefficients, weighting and summing the support stress value vectors at all time steps to obtain the comprehensive interactive response vector received by the surrounding rock deformation at that time step from the support stress information at all time steps. This response vector is the output of the first directional encoding branch. The above operation is performed independently on all attention heads. The outputs of each attention head are concatenated along the feature dimension and then mapped back to the original hidden state dimension through an output linear transformation matrix, completing the multi-head cross-attention aggregation.

[0067] Step S323: Generate the interaction attention weight of support stress on surrounding rock deformation by the inner product of the support stress query vector and the surrounding rock deformation key vector. Obtain the surrounding rock deformation value vector by passing the hidden state vector of surrounding rock deformation through the fourth linear projection layer. The surrounding rock deformation value vector is weighted and aggregated by the interaction attention weight to obtain the surrounding rock deformation interaction response vector received by the support stress as the output of the second direction encoding branch.

[0068] The computational logic of this step is completely symmetrical to step S322, only the roles of the active and passive parties are reversed. The inner product of the support stress query vector and the surrounding rock deformation key vector, after scaling and normalization to the maximum flexibility value, generates the interaction attention weight distribution of support stress on surrounding rock deformation. The fourth linear projection layer is a value projection matrix applied to the hidden state vector of surrounding rock deformation, mapping the hidden state vector of surrounding rock deformation to the surrounding rock deformation value vector. The surrounding rock deformation value vector is weighted and aggregated using the interaction attention weights to obtain the surrounding rock deformation interaction response vector received by the support stress, which serves as the output of the second directional encoding branch.

[0069] Step S324: The interaction response vector of support stress received by the surrounding rock deformation and the interaction response vector of surrounding rock deformation received by the support stress are spliced ​​along the channel dimension, and after being filtered by the first gated fusion unit, a bidirectional interaction response vector of surrounding rock deformation and support stress is generated.

[0070] Channel-dimensional concatenation refers to concatenating two identical interaction response vectors end-to-end along the feature dimension to form a concatenated vector with double the feature dimension. The first gated fusion unit is a small gated network whose function is to adaptively determine the information retention ratio of each feature channel in the concatenated vector based on its content, filtering out the feature components that are truly effective for subsequent interaction coordination. This gated fusion unit consists of a two-layer fully connected perceptron and an sigmoid growth curve activation function. The fully connected perceptron receives the concatenated vector and outputs a gate coefficient vector with the same dimension as the concatenated vector. Each element of the gate coefficient vector takes a value between zero and one. The gate coefficient vector is multiplied element-wise with the concatenated vector to obtain the filtered bidirectional interaction response vector. The gated fusion unit also includes a residual path directly connected to the input concatenated vector. The filtered vector is added to the identity mapping of the residual path and then subjected to layer normalization to output the final bidirectional interaction response vector of surrounding rock deformation and support stress.

[0071] Step S325: Based on the same processing structure as the first and second direction coding branches, generate a bidirectional interactive response vector of surrounding rock deformation and microseismic signal through the third and fourth direction coding branches, and generate a bidirectional interactive response vector of support stress and microseismic signal through the fifth and sixth direction coding branches.

[0072] The third directional coding branch uses the latent state of surrounding rock deformation as the active party and the latent state of microseismic signal as the passive party, calculating the interactive response vector of surrounding rock deformation received by the microseismic signal. The fourth directional coding branch uses the latent state of microseismic signal as the active party and the latent state of surrounding rock deformation as the passive party, calculating the interactive response vector of surrounding rock deformation received by the microseismic signal. The outputs of the two branches are then processed by channel splicing and filtered by the second gating fusion unit to generate a bidirectional interactive response vector of surrounding rock deformation and microseismic signal. The fifth and sixth directional coding branches use support stress and microseismic signal as the active and passive parties respectively, and generate a bidirectional interactive response vector of support stress and microseismic signal through the same processing structure and the third gating fusion unit. The linear projection layer parameters of all directional coding branches are independent of each other, and the parameters of the gating fusion units are also independent, ensuring that the modeling of each pair of interactive relationships has sufficient parameter capacity.

[0073] Step S330: Calculate the instantaneous response vector of the latent state of surrounding rock deformation to the latent state of microseismic signal through the third direction coding branch, calculate the instantaneous response vector of the latent state of microseismic signal to the latent state of surrounding rock deformation through the fourth direction coding branch, and generate a bidirectional interactive response vector of surrounding rock deformation and microseismic signal after splicing.

[0074] This step, within the overall framework of step S325, further clarifies the specific execution content of the third and fourth directional coding branches. Their calculation process completely corresponds to that of the first and second directional coding branches: the third directional coding branch uses the latent state vector of surrounding rock deformation as the query source and the latent state vector of microseismic signals as the key and value source, generating the interactive response vector of the microseismic signals received by the surrounding rock deformation through multi-head cross-attention; the fourth directional coding branch uses the latent state vector of microseismic signals as the query source and the latent state vector of surrounding rock deformation as the key and value, generating the interactive response vector of surrounding rock deformation received by the microseismic signals. The two response vectors are processed by channel splicing and the second gating fusion unit to output a bidirectional interactive response vector of surrounding rock deformation and microseismic signals.

[0075] Step S340: Calculate the instantaneous response vector of the hidden state of support stress to the hidden state of microseismic signal through the fifth direction coding branch, calculate the instantaneous response vector of the hidden state of microseismic signal to the hidden state of support stress through the sixth direction coding branch, and generate a bidirectional interactive response vector of support stress and microseismic signal after splicing.

[0076] The fifth directional encoding branch uses the latent state vector of the support stress as the query source and the latent state vector of the microseismic signal as the key and value source to generate the interactive response vector of the support stress received from the microseismic signal. The sixth directional encoding branch uses the latent state vector of the microseismic signal as the query source and the latent state vector of the support stress as the key and value source to generate the interactive response vector of the support stress received from the microseismic signal. The two response vectors are concatenated and filtered by the third gating fusion unit to output the bidirectional interactive response vector of the support stress and the microseismic signal.

[0077] Step S350: Input the bidirectional interactive response vectors of surrounding rock deformation and support stress, surrounding rock deformation and microseismic signals, and support stress and microseismic signals into the interactive response fusion layer, splice them in the channel dimension to generate a comprehensive vector of full-source interactive response, and generate interactive correction vectors that match the dimensions of each hidden state trajectory sequence through a fully connected mapping network.

[0078] The interactive response fusion layer is an aggregation layer responsible for integrating three pairs of bidirectional interactive response vectors. The comprehensive interactive response vector is formed by concatenating the bidirectional interactive response vectors of surrounding rock deformation and support stress, surrounding rock deformation and microseismic signals, and support stress and microseismic signals along the channel dimension. The feature dimension of this vector is three times the dimension of a single bidirectional interactive response vector. The fully connected mapping network is a four-layer fully connected perceptron. Its input is the comprehensive interactive response vector, and its output is an interactive correction vector with the same stacked dimension as the original three-source hidden state vectors. The first and second hidden layers of this fully connected mapping network both use exponential linear unit activation functions, each containing 128 hidden units. The third hidden layer is a linear layer without activation functions, mapping the feature dimension to the same length as the flattened stacked matrix of the three-source hidden state vectors. The output interactive correction vector is then split into three parts according to the original hidden state dimension, corresponding to the correction components of the surrounding rock deformation hidden state trajectory sequence, the support stress hidden state trajectory sequence, and the microseismic signal hidden state trajectory sequence, respectively.

[0079] Step S360: Add the correction component of the corresponding hidden state trajectory sequence of surrounding rock deformation in the interactive correction vector to the original hidden state trajectory sequence of surrounding rock deformation at each time step to obtain the interactively coordinated hidden state trajectory sequence of surrounding rock deformation; add the correction component of the corresponding hidden state trajectory sequence of support stress to the original hidden state trajectory sequence of support stress at each time step to obtain the interactively coordinated hidden state trajectory sequence of support stress; add the correction component of the corresponding hidden state trajectory sequence of microseismic signal to the original hidden state trajectory sequence of microseismic signal at each time step to obtain the interactively coordinated hidden state trajectory sequence of microseismic signal.

[0080] The time-by-time addition refers to adding the surrounding rock deformation correction component vector in the interactive correction vector to the hidden state vector of the original surrounding rock deformation hidden state trajectory sequence at each unified reference time point, element by element. This residual correction method ensures that the introduction of interactive coordination information does not completely cover the intrinsic dynamic information in the original hidden state trajectory, but rather superimposes interactive coupling correction while retaining the original information. The interactively coordinated surrounding rock deformation hidden state trajectory sequence is the new surrounding rock deformation hidden state trajectory sequence obtained after time-by-time addition. The hidden state vector at each time point of this sequence simultaneously carries the surrounding rock's own deformation dynamic information, the instantaneous interactive response information of the support stress, the instantaneous interactive response information of the microseismic signal, and the deep interactive information of high-order cross-coupling among the three.

[0081] In one implementation, after step S360, the following steps S370 to S3120 may also be included: Step S370: Input the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signal into the trajectory smoothing layer for temporal convolution smoothing to obtain the smoothed hidden state trajectory sequence of surrounding rock deformation, the smoothed hidden state trajectory sequence of support stress, and the smoothed hidden state trajectory sequence of microseismic signal.

[0082] The trajectory smoothing layer is a temporal convolutional filter layer. Its function is to locally smooth the interactively coordinated hidden state trajectory sequence generated by time-by-time addition, eliminating high-frequency jitter between hidden state vectors at adjacent time steps that may be caused by the independent generation of interactive correction vectors at each time step, thus improving the temporal smoothness of the trajectory. This trajectory smoothing layer employs a one-dimensional causal dilated convolution structure. The convolution kernel slides along the time axis, and the receptive field of each kernel covers the hidden state vectors at the current time step and several preceding time steps. The dilation rate of the dilated convolution increases progressively with the number of layers to expand the effective receptive field without increasing the number of parameters. Specifically, the trajectory smoothing layer consists of three stacked causal dilated convolution layers. The first dilated convolution layer has a dilation rate of one and a kernel width of three, with the number of output channels matching the dimension of the input hidden state. The second layer has a dilation rate of two and a kernel width of three. The third layer has a dilation rate of four and a kernel width of three. Each convolution layer is followed by an exponential linear unit activation function, and residual connections are used between layers. The three interconnected latent state trajectory sequences share the same convolution parameters of the same trajectory smoothing layer, but each undergoes independent temporal convolutional smoothing. The smoothed latent state trajectory sequence of surrounding rock deformation is the output sequence after three layers of causal void convolutional smoothing filtering, with a smoother transition between latent state vectors at adjacent time points. The smoothed latent state trajectory sequence of support stress and the smoothed latent state trajectory sequence of microseismic signals are processed similarly.

[0083] Step S380: Input the smoothed hidden state trajectory sequences into the trajectory consistency verification layer, calculate the similarity of the hidden state vectors of any two source hidden state trajectory sequences at the same time, and obtain the consistency scores of the surrounding rock deformation and support stress trajectories, the consistency scores of the surrounding rock deformation and microseismic signal trajectories, and the consistency scores of the support stress and microseismic signal trajectories.

[0084] The trajectory consistency verification layer is a functional layer used to quantitatively evaluate the rationality of coupling between the three hidden state trajectory sequences after interactive coordination. The hidden state vector similarity uses cosine similarity as a metric. Specifically, for any two hidden state vectors at the same unified reference time point, their inner product is calculated and divided by the product of their respective L2 norms. The resulting value ranges from negative one to positive one; the closer to positive one, the more consistent the directions of the two hidden state vectors at that time. The trajectory consistency score is a single-valued index obtained by averaging the cosine similarity values ​​over time at all unified reference time points. The consistency score for the surrounding rock deformation and support stress trajectories is obtained by averaging the cosine similarities of the hidden state trajectory sequences of the surrounding rock deformation and support stress at all time points. The consistency scores for the other two sets of trajectories are calculated similarly. This score reflects the degree of alignment of the hidden state trajectories of different monitoring sources in the overall evolution trend after interactive coordination.

[0085] Step S390: Determine whether the consistency scores of the three trajectories all fall within the preset reasonable consistency range. If any score exceeds the preset reasonable consistency range, generate a trajectory correction feedback signal based on the degree of exceedance.

[0086] The pre-defined reasonable consistency interval is a numerical range defined during the training phase by statistically analyzing the distribution range of consistency scores of the three-source hidden state trajectories under normal construction conditions from a large amount of historical monitoring data. The upper and lower bounds of this interval correspond to the statistical quantiles of the consistency scores under normal conditions. If the consistency score of a trajectory is lower than the lower bound of this interval, it indicates that the hidden state trajectories of the two corresponding monitoring sources have diverged in trend after interactive coordination, which may mean that unreasonable coupling information has been introduced during the interactive correction process. If the consistency score of a trajectory is higher than the upper bound of this interval, it indicates that the hidden state trajectories of the two monitoring sources are too similar, which may indicate over-coupling and mask the differentiated response characteristics that should exist between the monitoring sources. The specific method for generating trajectory correction feedback signals based on the degree of excess is as follows: the difference between the trajectory consistency score and the nearest boundary of the reasonable interval is calculated as the excess amount. This excess amount is mapped by a scaling factor to generate a correction intensity coefficient, which, along with the identifier of the excess monitoring source pair, is packaged into a trajectory correction feedback signal.

[0087] Step S3100: The trajectory correction feedback signal is sent back to the interactive response fusion layer, the mapping weights of the fully connected mapping network are adjusted to regenerate the interactive correction vector, and the time-by-time addition, trajectory smoothing and consistency verification are repeated until all trajectory consistency scores fall into the preset reasonable consistency range.

[0088] Feeding the trajectory correction feedback signal back to the interactive response fusion layer signifies the activation of a lightweight online adaptive adjustment mechanism during the inference phase. Adjusting the mapping weights of the fully connected mapping network does not involve retraining the entire network through backpropagation. Instead, it utilizes the correction strength coefficient and monitoring source pair identifier carried in the trajectory correction feedback signal to locate the output subspace of the fully connected mapping network associated with the anomaly-prone monitoring source pair. A small perturbation proportional to the correction strength coefficient is applied to the corresponding fully connected output weights of this subspace, with the perturbation direction ensuring that the consistency score of the monitoring source pair's trajectory moves towards a reasonable range in the next forward computation. Repeated execution of time-by-time addition, trajectory smoothing, and consistency verification means that after each adjustment of the fully connected mapping network's mapping weights, a complete interactive correction generation and verification loop is executed again starting from step S360. The adjusted consistency score is checked to see if it returns to a reasonable range. If it still exceeds the range, the mapping weights are fine-tuned. This loop is executed a limited number of times to ensure that the real-time performance of online inference is not significantly affected.

[0089] Step S3110: After the verification is passed, the smoothed surrounding rock deformation hidden state trajectory sequence is taken as the final interactively coordinated surrounding rock deformation hidden state trajectory sequence, the smoothed support stress hidden state trajectory sequence is taken as the final interactively coordinated support stress hidden state trajectory sequence, and the smoothed microseismic signal hidden state trajectory sequence is taken as the final interactively coordinated microseismic signal hidden state trajectory sequence.

[0090] A successful verification means that the consistency scores of the three trajectories have all fallen within the preset reasonable consistency range in the judgment of step S390, and no further feedback adjustment in step S3100 is required, or that the three scores have returned to the reasonable range after a limited number of feedback adjustments in step S3100. At this time, the three smoothed hidden state trajectory sequences output by the trajectory smoothing processing layer are considered to have reached the convergence state of interactive coordination and can be output as the final interactively coordinated hidden state trajectory sequences to the downstream processing steps.

[0091] Step S3120: Output the final interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the final interactively coordinated hidden state trajectory sequence of support stress, and the final interactively coordinated hidden state trajectory sequence of microseismic signal to the processing step of synchronously extracting hidden state vectors at the time of risk query.

[0092] This step submits the three-source hidden state trajectory sequence, which has undergone all interactive coordination, smoothing and verification processing, to the risk query and feature fusion stage in step S400.

[0093] Step S400: Obtain the risk query time. At the risk query time, simultaneously extract the hidden state vectors of the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signal to obtain the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal. Then, through a multi-head cross-attention fusion mechanism, asynchronous feature synchronous fusion processing is performed on the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal to generate a multi-source fused hidden state representation vector.

[0094] The risk query time refers to a specific point in time, designated by an external system or input by on-site safety management personnel, where the current tunnel construction safety risk status needs to be assessed. This time is a definite timestamp value, indicating the time segment of interest for the early warning judgment. Simultaneous extraction of latent state vectors at the risk query time is based on the characteristic that the latent state trajectory of continuous time can be calculated at any moment. For this specific time point, the latent state vector value corresponding to that moment is precisely calculated using an ordinary differential equation numerical solver on the final interactively coordinated sequence of surrounding rock deformation latent state trajectories, support stress latent state trajectories, and microseismic signal latent state trajectories. The three latent state vectors have natural synchronicity at this moment, even though their original monitoring sequences were acquired asynchronously. The surrounding rock deformation latent state vector is a fixed-dimensional latent state feature vector extracted from the interactively coordinated sequence of surrounding rock deformation latent state trajectories at the risk query time, integrating the surrounding rock's own deformation dynamics and multi-source interactive coupling information. The definitions of the support stress latent state vector and the microseismic signal latent state vector are similar. Asynchronous feature synchronous fusion processing refers to the process where, at the risk query time, although the three hidden state vectors are obtained synchronously, they originate from three monitoring sources with different physical mechanisms, and their semantic meanings in their respective feature spaces are not the same. A specialized fusion mechanism is needed to fuse these three heterogeneous feature vectors into a unified multi-source fusion hidden state representation vector. The multi-head cross-attention fusion mechanism is the core technical means to achieve this heterogeneous feature fusion. The multi-source fusion hidden state representation vector is a fusion representation that compresses all interactive and coordinated information from the three monitoring sources—surrounding rock deformation, support stress, and microseismic signals—at the risk query time into a single vector. This vector will serve as the unified input for the subsequent multi-task prediction network.

[0095] In one embodiment, step S400 may specifically include the following steps S410~S450: Step S410: Obtain the timestamp value of the risk query time specified externally or input by the user, and based on the timestamp value, query the corresponding hidden state vector in the interconnected and coordinated hidden state trajectory sequence of surrounding rock deformation to obtain the hidden state vector of surrounding rock deformation, query the corresponding hidden state vector in the interconnected and coordinated hidden state trajectory sequence of support stress to obtain the hidden state vector of support stress, and query the corresponding hidden state vector in the interconnected and coordinated hidden state trajectory sequence of microseismic signal to obtain the hidden state vector of microseismic signal.

[0096] The timestamp value for the risk query time, either externally specified or user-inputted, can originate from the timed risk inspection module of the tunnel construction safety monitoring system. This module automatically generates the timestamp value for the query time according to a preset inspection cycle. Alternatively, a focus time can be manually input by the on-site safety engineer through the human-computer interaction interface of the early warning system. The process of querying the corresponding hidden state vector based on this timestamp value leverages the ability of continuous-time hidden state trajectories to be evaluated at any time: using the queried timestamp value as a time parameter, driven by the right-hand side function of the neural differential equation corresponding to the interactively coordinated hidden state trajectory sequence of surrounding rock deformation and the initial hidden state vector, the result is integrated from the trajectory start time to the query time using an ordinary differential equation numerical solver. The final value obtained is the hidden state vector of surrounding rock deformation. The same integration evaluation operation is performed on the hidden state trajectory sequence of support stress and the hidden state trajectory sequence of microseismic signals, yielding the hidden state vectors of support stress and microseismic signals, respectively. The three vectors are strictly synchronized at the query time.

[0097] Step S420: Stack the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal into a three-source hidden state vector stacking matrix, input the first attention head of the multi-head cross-attention fusion mechanism, and use the hidden state vector of surrounding rock deformation as the query, the hidden state vector of support stress as the key, and the hidden state vector of microseismic signal as the value to perform attention aggregation to generate the output vector of the first attention head.

[0098] The three-source latent state vector stacking matrix is ​​a two-dimensional matrix consisting of three rows and multiple columns. It is formed by concatenating three one-dimensional latent state vectors as three row vectors, following a preset order of surrounding rock deformation, support stress, and microseismic signal along the row direction. The matrix has three rows and the number of columns equals the dimension of a single latent state vector. The multi-head cross-attention fusion mechanism is a three-head attention fusion architecture. Its overall design idea is to allow the latent state vector of each monitoring source to take turns as the query, and to perform cross-attention aggregation with the latent state vectors of the other two monitoring sources as the key and value, respectively. The three attention heads correspond to three different query, key, and value role allocation schemes. Finally, the outputs of the three heads are concatenated and fused to obtain the multi-source fused latent state representation vector. The first attention head uses the surrounding rock deformation latent state vector as the query, the support stress latent state vector as the key, and the microseismic signal latent state vector as the value. This role allocation allows the first attention head to focus on capturing the cross-source fusion features from the perspective of surrounding rock deformation, using support stress as the matching reference, and microseismic signal information as the actual aggregation content.

[0099] In one embodiment, step S420 may specifically include the following steps S421 to S425: Step S421: Detect the dimensions of the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal. If the dimensions of the three are inconsistent, map each hidden state vector to a unified target dimension through a dimension alignment mapping network to obtain the dimension-aligned hidden state vectors.

[0100] The dimension alignment mapping network is an alignment module consisting of a single-layer fully connected linear transformation. The three hidden state vectors are linearly projected through their respective independent dimension alignment mapping matrices. The unified target dimension of the projection output is set to the maximum value of the dimensions of the three original hidden state vectors. If the three original dimensions are already consistent, the dimension alignment mapping network in this step directly performs an identity mapping without changing the vector dimensions.

[0101] Step S422: If the dimensions are consistent, then each hidden state vector is used as a stacking row vector and stacked along the row direction in a preset order to generate a stacking matrix of three hidden state vectors.

[0102] Once the dimensions are consistent or aligned via step S421, the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal are converted into one-dimensional vectors of the same dimension and arranged in a fixed row order along the row direction to form a two-dimensional matrix with three rows of target dimension columns. This matrix is ​​the stacked matrix of the three hidden state vectors.

[0103] Step S423: Input the stacked matrix of the three hidden state vectors into the multi-head segmentation layer of the first attention head, and uniformly divide each hidden state vector into multiple sub-vector segments along the channel dimension to obtain the set of surrounding rock deformation sub-vector segments, the set of support stress sub-vector segments, and the set of microseismic signal sub-vector segments.

[0104] The multi-head segmentation layer uniformly divides each hidden state vector in the stacked matrix of three-source hidden state vectors into multiple sub-vector segments along its feature dimension, equal to the number of attention heads. The dimension of each sub-vector segment is the original feature dimension divided by the number of attention heads. If the first attention head contains four sub-attention heads, then each hidden state vector is uniformly divided into four sub-vector segments. The set of surrounding rock deformation sub-vector segments contains four sub-vector segments with the same dimension, and the set of support stress sub-vector segments and the set of microseismic signal sub-vector segments are similar.

[0105] Step S424: Map each sub-vector segment in the set of surrounding rock deformation sub-vector segments to generate a query sub-vector set through the query linear transformation of the first attention head; generate a bond sub-vector set through the bond linear transformation of each sub-vector segment in the set of support stress sub-vector segments; and generate a value sub-vector set through the value linear transformation of each sub-vector segment in the set of microseismic signal sub-vector segments.

[0106] The query linear transformation is a trainable linear projection matrix applied to the surrounding rock deformation sub-vector segment. Query linear transformation matrices from different sub-attention heads share parameters to reduce the number of parameters. The key linear transformation is a linear projection matrix applied to the support stress sub-vector segment, and the value linear transformation is a linear projection matrix applied to the microseismic signal sub-vector segment. Both also share parameters between sub-attention heads.

[0107] Step S425: Perform dot product attention operation on the query sub-vector, key sub-vector and value sub-vector at the corresponding positions to obtain multiple sub-attention output vectors, and concatenate the multiple sub-attention output vectors into the first attention head output vector.

[0108] The corresponding position refers to the Kth query sub-vector extracted from the set of surrounding rock deformation sub-vector segments, the Kth key sub-vector extracted from the set of support stress sub-vector segments, and the Kth value sub-vector extracted from the set of microseismic signal sub-vector segments under the same sub-attention head number. These three sub-vectors belong to the same sub-attention head. During the dot product attention operation, the inner product of the query sub-vector and the key sub-vector is calculated, and the result is scaled by dividing by the square root of the key sub-vector dimension. A flexible maximum normalization operation is applied to this scaled scalar to generate an attention weight. This attention weight is directly multiplied by the value sub-vector to obtain the sub-attention output vector of this sub-attention head. The sub-attention output vectors of all sub-attention heads are concatenated along the feature dimension, and then mapped through an output linear transformation matrix to obtain the output vector of the first attention head, whose dimension is consistent with the dimension of the original single hidden state vector.

[0109] Step S430: Input the stacked matrix of the three hidden state vectors into the second attention head. The second attention head uses the hidden state vector of the support stress as the query, the hidden state vector of the surrounding rock deformation as the key, and the hidden state vector of the microseismic signal as the value to perform attention aggregation to generate the output vector of the second attention head.

[0110] The internal structure of the second attention head is completely identical to that of the first attention head, including a multi-head segmentation layer, a query key-value linear transformation layer, and a dot-product attention aggregation layer. The only difference lies in the role allocation scheme: the query comes from the support stress, the key comes from the surrounding rock deformation, and the value still comes from the microseismic signal. This configuration allows the second attention head to start from the perspective of support stress, using the surrounding rock deformation as a matching bridge to aggregate information from the microseismic signal, producing fusion features different from those of the first attention head.

[0111] Step S440: Input the stacked matrix of the three hidden state vectors into the third attention head. The third attention head uses the hidden state vector of the microseismic signal as the query, the hidden state vector of the surrounding rock deformation as the key, and the hidden state vector of the support stress as the value to perform attention aggregation to generate the output vector of the third attention head.

[0112] The third attention head uses microseismic signals as the query source, surrounding rock deformation as the bond source, and support stress as the value source, aggregating support stress information from the perspective of microseismic activity. The role allocation of the three attention heads covers three different cross-source fusion paths, ensuring that the multi-source fusion latent state representation vector does not lose the interaction information between specific sources due to the preference of a single fusion path.

[0113] Step S450: Concatenate the output vectors of the first attention head, the second attention head, and the third attention head along the channel dimension to obtain a multi-head cross-attention concatenation vector. Then, perform dimensionality reduction and fusion on the multi-head cross-attention concatenation vector through a linear projection fusion layer to generate a multi-source fusion hidden state representation vector.

[0114] The multi-head cross-attention concatenation vector is a vector with a dimension three times that of the hidden state, obtained by concatenating the output vectors of the first, second, and third attention heads along the feature dimension. The linear projection fusion layer is a single-layer fully connected linear transformation matrix with an input dimension three times that of the hidden state and an output dimension consistent with the dimension of a single original hidden state vector. This linear transformation does not apply a nonlinear activation function; it directly compresses the dimension of the concatenated vector to the target dimension through matrix multiplication, generating a multi-source fused hidden state representation vector. This vector condenses all the interactive coordination information and cross-source fusion information of surrounding rock deformation, support stress, and microseismic signals at the time of risk query, and will serve as the sole input to the multi-task prediction network.

[0115] Step S500: Based on the multi-source fusion latent state representation vector, the risk category prediction distribution, risk severity prediction value, and risk occurrence time window prediction interval are simultaneously output through the multi-task prediction network. Tunnel construction early warning instructions are generated based on the risk category prediction distribution, risk severity prediction value, and risk occurrence time window prediction interval.

[0116] The multi-task prediction network is a hard-parameter shared multi-task learning architecture with a shared feature transformation layer and three independent prediction branches. The network takes a multi-source fused latent state representation vector as input, extracts high-level abstract features common to the three prediction tasks in the shared feature transformation layer, and then outputs three different prediction results in parallel through risk category prediction branches, risk severity prediction branches, and risk time window prediction branches. The risk category prediction distribution is a probability distribution vector output by the risk category prediction branch across all preset risk categories. This distribution indicates the probability that the current rock safety status belongs to each preset risk category, such as collapse, rockburst, large deformation, or water inrush. The risk severity prediction value is a continuous scalar output by the risk severity prediction branch, representing the potential severity of the harm if the risk occurs; a higher value indicates a more severe consequence. The risk occurrence time window prediction interval is a continuous time interval consisting of a start time and an end time, output by the risk time window prediction branch, predicting that the risk event has a high probability of occurrence within this interval. The combination of these three elements enables tunnel construction early warning instructions to simultaneously inform on-site personnel what types of risks may occur, how severe the risks are, and within what time window the risks may erupt, thereby supporting targeted, tiered response measures.

[0117] In one embodiment, step S500 may specifically include the following steps S510 to S560: Step S510: Input the multi-source fusion hidden state representation vector into the shared feature transformation layer of the multi-task prediction network for nonlinear feature transformation to generate a shared hidden layer representation vector. The shared hidden layer representation vector is used by the risk category prediction branch, the risk severity prediction branch, and the risk time window prediction branch.

[0118] The shared feature transformation layer is the backbone of the multi-task prediction network, consisting of three stacked fully connected perceptron layers. Each layer contains 192 hidden units, and the activation function is a parameterized linear unit function. Insertion-dropout regularization is used between layers to prevent overfitting. The shared feature transformation layer receives a multi-source fused hidden state representation vector as input, and after three layers of nonlinear feature transformation, outputs a shared hidden layer representation vector. This vector has the same dimension as the input vector, but has already completed general feature extraction for the three prediction tasks in a high-dimensional feature space. The risk category prediction branch, risk severity prediction branch, and risk time window prediction branch all build their own dedicated mapping layers based on the shared hidden layer representation vector, sharing the general feature parameters of the lower layers to reduce the risk of overfitting and improve the prediction generalization ability for each task.

[0119] Step S520: Input the shared hidden layer representation vector into the risk category prediction branch, and generate the risk category prediction distribution through a multi-layer fully connected mapping and normalized exponential output layer. The risk category prediction distribution contains the probability values ​​of each preset risk category.

[0120] In one implementation, step S520 may specifically include the following steps S521 to S526: Step S521: Input the shared hidden layer representation vector into the first fully connected mapping layer of the risk category prediction branch, and generate the first intermediate hidden layer representation vector through linear transformation and nonlinear activation. Its dimension is consistent with that of the shared hidden layer representation vector.

[0121] The first fully connected mapping layer consists of a fully connected linear transformation matrix and a parameterized modified linear unit activation function. The linear transformation matrix maps the shared hidden layer representation vector to an intermediate vector with unchanged dimensions. The non-linear activation function activates this intermediate vector element by element to generate the first intermediate hidden layer representation vector. The design of maintaining the unchanged dimensions aims to provide sufficient feature capacity for the risk category prediction branch without prematurely compressing information.

[0122] Step S522: Input the first intermediate hidden layer representation vector into the second fully connected mapping layer, and generate the second intermediate hidden layer representation vector through linear transformation and nonlinear activation. Its dimension is consistent with the total number of preset risk categories.

[0123] The second fully connected mapping layer maps the representation vector of the first intermediate hidden layer to a vector of the preset total number of risk categories through a fully connected linear transformation. It is also processed nonlinearly by a parameterized linear unit activation function to obtain the representation vector of the second intermediate hidden layer. Each component of this vector corresponds to the original score of a preset risk category.

[0124] Step S523: Input the second intermediate hidden layer representation vector into the normalized exponential output layer, perform exponential calculation and normalization on the values ​​of each dimension, generate the risk category prediction distribution, and the sum of the probability values ​​of all preset risk categories is the unit value.

[0125] The normalized exponential output layer calculates an exponential value with the natural constant as the base for each component of the representation vector of the second intermediate hidden layer. Then, it sums the exponential values ​​of all components to obtain the normalized denominator. The probability value corresponding to each component is obtained by dividing the exponential value of each component by this denominator. The unit value, or numerical value one, represents the normalization of the total probability.

[0126] Step S524: Match the risk category prediction distribution with the preset risk category label set to generate a risk category prediction distribution mapping table containing risk category names and corresponding probability values.

[0127] The preset risk category label set is a mapping table between preset risk category indices and risk category name texts. For example, index zero corresponds to rock instability and collapse, and index one corresponds to high-stress rockburst. The matching process involves associating the probability values ​​in the risk category prediction distribution, arranged in index order, with the risk category names with the same index in the label set, forming a mapping table consisting of risk category names and corresponding probability values.

[0128] Step S525: Extract the name of the risk category with the highest probability value from the risk category prediction distribution mapping table as the main risk factor, and compare the highest probability value with the preset risk category judgment threshold. If the threshold is exceeded, a risk category confirmation signal is triggered.

[0129] The primary risk factor is the name of the most likely risk type at the current risk query time. This name is obtained by retrieving the entry corresponding to the highest probability value from the mapping table. The preset risk category determination threshold is a probability threshold statistically calibrated based on historical early warning accuracy during the early warning system configuration phase. The risk category prediction is only confirmed as valid when the highest probability value is higher than this threshold, thus avoiding issuing unreliable early warnings when the model uncertainty is high.

[0130] Step S526: Based on the risk category confirmation signal, provide the main risk-causing factor and risk category prediction distribution mapping table to the early warning instruction assembly unit, so as to assemble the main risk-causing factor as the dominant risk category code into the tunnel construction early warning instruction.

[0131] The early warning instruction assembly unit is the post-processing unit in steps S550 and S560 responsible for combining various early warning information and encoding it into a communication instruction format. Encoding the primary risk-causing factor as the dominant risk category means mapping the text name of the primary risk-causing factor to its corresponding preset risk category index code, which is used as the fill value for the risk category field in the early warning instruction.

[0132] Step S530: Input the shared hidden layer representation vector into the risk severity prediction branch, and generate the risk severity prediction value through a multi-layer fully connected mapping and continuous value regression output layer.

[0133] The risk severity prediction branch is a continuous value regression task branch. Its multi-layer fully connected mapping consists of two fully connected layers. The first layer maps the shared hidden layer representation vector to a 64-dimensional hidden vector, with an exponential linear unit function as the activation function. The second layer maps the 64-dimensional hidden vector to a one-dimensional scalar output. The continuous value regression output layer does not have an activation function set and directly outputs the scalar value of the second layer as the raw value of the risk severity prediction. This raw value is transformed into a risk severity level with actual physical meaning through a preset linear mapping relationship, such as the probability of support structure failure or the equivalent radius of the surrounding rock instability range.

[0134] Step S540: Input the shared hidden layer representation vector into the risk time window prediction branch. The first fully connected mapping subnetwork generates the prediction value at the start time of the risk occurrence time window, and the second fully connected mapping subnetwork generates the prediction value at the end time of the risk occurrence time window. The two are combined to form the risk occurrence time window prediction interval.

[0135] The risk time window prediction branch comprises two structurally identical but parameter-independent fully connected mapping subnetworks. Both subnetworks take a shared hidden layer representation vector as input and each outputs a time prediction value. The first fully connected mapping subnetwork predicts the possible start time of the risk, and the second fully connected mapping subnetwork predicts the possible end time of the risk. The two prediction values ​​are expressed as time offsets relative to the risk query time, using the risk query time as a reference.

[0136] In one embodiment, step S540 may specifically include the following steps S541 to S546: Step S541: Input the shared hidden layer representation vector into the first fully connected mapping subnetwork and the second fully connected mapping subnetwork of the risk time window prediction branch, respectively. The network weights of the two networks are independent.

[0137] Both the first and second fully connected mapping subnetworks are three-layer fully connected perceptrons, each with thirty-two hidden units. The activation function is the hyperbolic tangent function, and the last layer is a linear layer without an activation function, outputting a one-dimensional scalar value. The weight matrices of the two subnetworks are independently and randomly initialized and trained independently, without using shared parameters.

[0138] Step S542: In the first fully connected mapping subnetwork, the shared hidden layer representation vector is linearly transformed and nonlinearly activated by the first starting time mapping layer to obtain the first starting time intermediate vector, and then linearly transformed by the second starting time mapping layer to obtain the predicted value of the starting time of the risk occurrence time window.

[0139] The first starting time mapping layer is a fully connected linear transformation layer with a hyperbolic tangent activation function, and the second starting time mapping layer is a purely linear fully connected layer. It outputs the predicted value of the starting time of the risk occurrence time window. This value is the time offset relative to the risk query time. A positive value indicates that it is after the query time.

[0140] Step S543: In the second fully connected mapping subnetwork, the shared hidden layer representation vector is linearly transformed and nonlinearly activated by the first end-time mapping layer to obtain the intermediate vector at the first end-time. Then, the predicted value at the end of the risk occurrence time window is obtained by linear transformation by the second end-time mapping layer.

[0141] The structure of the first end time mapping layer and the second end time mapping layer is the same as that of the first start time mapping layer and the second start time mapping layer, except that the parameters are independent, and the output is the predicted value at the end time of the risk occurrence time window.

[0142] Step S544: Obtain the timestamp units of the predicted values ​​at the start and end times. If the timestamp units are inconsistent with those at the time of risk query, call the timestamp unit conversion process to unify them to the same unit.

[0143] The timestamp unit can be seconds, minutes, or hours, depending on the time labeling method used during model training. Timestamp unit conversion is achieved through a simple unit conversion multiplication factor, transforming the start and end time predicted values ​​from the output unit to the unit consistent with the risk query time.

[0144] Step S545: Compare the converted start time prediction value with the end time prediction value. If the start time is later than the end time, swap the two values ​​to ensure that the start time is earlier than the end time.

[0145] This step is a post-processing verification to ensure the reasonableness of the time window. The swapping operation ensures that the start time of the output risk occurrence time window prediction interval is always before the end time.

[0146] Step S546: Combine the adjusted start time prediction value and end time prediction value into a risk occurrence time window prediction interval and provide it to the early warning instruction assembly unit.

[0147] The risk occurrence time window prediction interval consists of two timestamp values, the start time and the end time, representing the time range within which the predicted risk is concentrated. This interval is passed to the early warning instruction assembly unit for subsequent assembly.

[0148] Step S550: Extract the preset risk category with the highest probability value from the risk category prediction distribution as the dominant risk category. When the probability value of the dominant risk category exceeds the preset warning trigger threshold, generate a tunnel construction warning instruction based on the dominant risk category, the predicted value of risk severity, and the predicted interval of the risk occurrence time window. The tunnel construction warning instruction includes the dominant risk category code, the risk severity level, and the start and end times of the risk occurrence time window.

[0149] In one implementation, step S550 may specifically include the following steps S551 to S556: Step S551: Traverse the probability values ​​of each preset risk category in the risk category prediction distribution, and determine the maximum probability value and its corresponding preset risk category index through the maximum value index.

[0150] The traversal process compares each probability value in the order of the index, records the maximum probability value encountered and its index position, and obtains the global maximum probability value and the corresponding preset risk category index after the traversal is completed.

[0151] Step S552: Query the corresponding risk category name in the preset risk category label mapping table according to the preset risk category index as the dominant risk category name, and use the maximum probability value as the dominant risk category probability value.

[0152] This step is consistent with the matching logic in step S524, retrieving the text name corresponding to the index from the label mapping table.

[0153] Step S553: ​​Compare the probability value of the dominant risk category with the preset early warning trigger threshold, which is a probability benchmark calibrated based on the historical early warning records of tunnel construction.

[0154] The calibration process of the preset early warning trigger threshold is based on the tunnel construction historical early warning database. The output probability of the multi-task prediction network at several moments before the occurrence of real risk events in history is retrospectively statistically analyzed, and the probability value that can control the false alarm rate within an acceptable range while meeting the expected detection rate is selected as the threshold value.

[0155] Step S554: If the probability value of the dominant risk category exceeds the preset warning trigger threshold, a warning trigger signal is generated, and the warning trigger signal is bound to the name of the dominant risk category to form the dominant risk category code activation information.

[0156] The warning trigger signal is a Boolean internal signal indicating that the warning process has been activated. The dominant risk category code activation information includes the index code and name text of the dominant risk category, and will serve as the core field content of the warning instruction.

[0157] Step S555: If the time limit is not exceeded, generate a warning suppression signal and stop the process of generating tunnel construction warning instructions.

[0158] The warning suppression signal indicates that the risk prediction confidence level at the current query time is insufficient to trigger a substantive warning, and the generation process terminates here without consuming subsequent computing and communication resources.

[0159] Step S556: After generating the early warning trigger signal, the activation information of the dominant risk category code, the predicted value of the risk severity, and the predicted interval of the risk occurrence time window are packaged into an early warning information data packet, input into the instruction serialization module, and converted into a tunnel construction early warning instruction in byte stream format that conforms to the tunnel construction early warning communication protocol.

[0160] The instruction serialization module is an encoder that encodes structured early warning information data packets into byte stream communication instructions. Based on the data frame format definition of the tunnel construction early warning communication protocol, the encoder serializes and packages fields such as the dominant risk category code, risk severity level value, and the start and end times of the risk occurrence time window according to the byte order and data type specified in the protocol, generating tunnel construction early warning instructions that can be transmitted on industrial communication networks.

[0161] Step S560: Send the tunnel construction early warning instruction to the tunnel construction early warning terminal, and store the predicted distribution of risk categories, the predicted value of risk severity, and the predicted time window interval of risk occurrence into the early warning history database.

[0162] The tunnel construction early warning terminal consists of audible and visual alarms, displays, or handheld terminals installed near the tunnel construction work face, at the tunnel entrance monitoring center, and in the remote project management department. Upon receiving a tunnel construction early warning command, these terminals automatically trigger the corresponding alarm signal based on the dominant risk category code and risk severity level contained in the command. The early warning historical record database is a persistent storage system deployed on the monitoring center server. It uses a structured query language relational database or a time-series database. Each stored record contains fields such as the timestamp of the risk query, the probability values ​​of all categories in the risk category prediction distribution, the predicted risk severity value, the start and end times of the risk occurrence time window prediction interval, and whether an early warning signal was actually triggered. This provides historical data support for subsequent model performance backtesting analysis, early warning threshold recalibration, and construction safety situation assessment.

[0163] Please see Figure 4 This is a schematic diagram of a tunnel construction early warning device provided in an embodiment of the present invention. The aforementioned tunnel construction early warning device 300 can be a computer program (including program code) running on a network device; for example, the tunnel construction early warning device 300 is an application software. This device can be used to execute corresponding steps in the method provided in the embodiments of the present invention. Figure 3 As shown, the tunnel construction early warning device 300 may include: a data acquisition module 310, a trajectory mapping module 320, a sequence generation module 330, a cross-fusion module 340, and an instruction generation module 350.

[0164] Among them, the data acquisition module 310 is used to acquire asynchronous monitoring data streams collected by multi-source monitoring equipment in the tunnel construction area. The asynchronous monitoring data streams include surrounding rock deformation monitoring sequences, support stress monitoring sequences, and microseismic signal monitoring sequences. Each monitoring sequence has an independent and non-equal interval timestamp. The trajectory mapping module 320 is used to construct a continuous-time latent state trajectory generation network for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence and the microseismic signal monitoring sequence, respectively. It maps the discrete observation values ​​of each monitoring sequence to the continuous posterior latent state trajectory of the corresponding monitoring source, and obtains the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence and the microseismic signal latent state trajectory sequence. The sequence generation module 330 is used to call the dynamic interaction coordination module to perform time-series coupling interaction processing on the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress and the hidden state trajectory sequence of microseismic signal. It captures the dynamic interaction response between the hidden state trajectories of different sources in a continuous time frame and generates the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress and the interactively coordinated hidden state trajectory sequence of microseismic signal. The cross-fusion module 340 is used to obtain the risk query time. At the risk query time, it synchronously extracts the hidden state vectors of the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signal, so as to obtain the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal. Then, through the multi-head cross-attention fusion mechanism, the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal are asynchronously and synchronously fused to generate a multi-source fused hidden state representation vector. The instruction generation module 350 is used to generate tunnel construction early warning instructions based on the multi-source fusion hidden state representation vector, and simultaneously output the risk category prediction distribution, risk severity prediction value and risk occurrence time window prediction interval through a multi-task prediction network.

[0165] According to one embodiment of the present invention, Figure 3 The steps involved in the tunnel construction early warning method based on multi-source feature fusion shown can be derived from... Figure 4 The various modules in the tunnel construction early warning device shown are used to execute this.

[0166] According to one embodiment of the present invention, Figure 4 The various modules in the tunnel construction early warning device shown can be individually or entirely combined into one or more units, or one or more of these units can be further divided into at least two functionally smaller sub-units to achieve the same operation without affecting the technical effects of the embodiments of the present invention. The above modules are based on logical function division. In practical applications, the function of one module can also be implemented by at least two units, or the function of at least two modules can be implemented by one unit. In other embodiments of the present invention, the tunnel construction early warning device may also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented collaboratively by at least two units.

[0167] According to one embodiment of the present invention, the following can be executed by running on a general-purpose computer system, such as a computer, which includes processing components and storage components such as a central processing unit (CPU), random access memory (RAM), and read-only memory (ROM), a device capable of performing operations such as... Figure 2 The computer program (including program code) involved in each step of the corresponding method shown, to construct such... Figure 3 The tunnel construction early warning device shown herein, and the tunnel construction early warning method based on multi-source feature fusion for implementing embodiments of the present invention, are described. The aforementioned computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the same medium, and run therein.

[0168] This invention also provides a computer system, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the tunnel construction early warning method based on multi-source feature fusion provided in this invention.

[0169] Please see details. Figure 5 This is a schematic diagram of the structure of a computer system provided in an embodiment of the present invention. Figure 5 As shown, the computer system 1000 described above may include: a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer system 1000 may also include: a user interface 1003, and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the aforementioned processor 1001. Figure 5 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0170] exist Figure 5 In the computer system 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

[0171] It should be understood that the computer system 1000 described in the embodiments of the present invention can execute the foregoing text. Figure 3 The implementation principle and beneficial effects of the tunnel construction early warning method based on multi-source feature fusion described in the corresponding embodiments will not be repeated here.

Claims

1. A tunnel construction early warning method based on multi-source feature fusion, characterized in that, The method includes: Acquire asynchronous monitoring data streams collected by multi-source monitoring equipment in the tunnel construction area. The asynchronous monitoring data streams include surrounding rock deformation monitoring sequences, support stress monitoring sequences, and microseismic signal monitoring sequences. Each monitoring sequence has an independent and non-equally spaced timestamp. A continuous-time latent state trajectory generation network is constructed for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence, and the microseismic signal monitoring sequence, respectively. The discrete observation values ​​of each monitoring sequence are mapped to the continuous posterior latent state trajectories of the corresponding monitoring sources, thereby obtaining the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence, and the microseismic signal latent state trajectory sequence. The dynamic interaction coordination module is invoked to perform time-series coupling interaction processing on the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal. The dynamic interaction response between the hidden state trajectories of different sources is captured in a continuous time frame, and the interactive coordination sequence of surrounding rock deformation, the interactive coordination sequence of support stress, and the interactive coordination sequence of microseismic signal are generated. At the time of risk query, the hidden state vectors of the interconnected and coordinated hidden state trajectory sequence of surrounding rock deformation, the interconnected and coordinated hidden state trajectory sequence of support stress, and the interconnected and coordinated hidden state trajectory sequence of microseismic signal are extracted synchronously to obtain the hidden state vectors of surrounding rock deformation, support stress, and microseismic signal. The hidden state vectors of surrounding rock deformation, support stress, and microseismic signal are then subjected to asynchronous feature synchronous fusion processing through a multi-head cross-attention fusion mechanism to generate a multi-source fused hidden state representation vector. Based on the multi-source fusion hidden state representation vector, the multi-task prediction network simultaneously outputs the risk category prediction distribution, the risk severity prediction value, and the risk occurrence time window prediction interval, and generates a tunnel construction early warning instruction according to the risk category prediction distribution, the risk severity prediction value, and the risk occurrence time window prediction interval.

2. The method of claim 1, wherein, The method involves constructing continuous-time latent state trajectory generation networks for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence, and the microseismic signal monitoring sequence, respectively. This maps the discrete observations of each monitoring sequence to the continuous posterior latent state trajectories of the corresponding monitoring sources, resulting in the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence, and the microseismic signal latent state trajectory sequence, including: The first variational inference encoder in the first continuous-time hidden state trajectory generation network receives the discrete observation values ​​of surrounding rock deformation at each time point in the surrounding rock deformation monitoring sequence and the corresponding time interval information, and outputs the sequence of hidden state trajectories of surrounding rock deformation represented by continuous-time hidden state trajectories. The second variational inference encoder in the second continuous-time hidden state trajectory generation network receives the discrete observation values ​​of support stress at each time point in the support stress monitoring sequence and the corresponding time interval information, and outputs the support stress hidden state trajectory sequence represented by continuous-time hidden state trajectory. The third variational inference encoder in the third continuous-time hidden state trajectory generation network receives the discrete observation values ​​of microseismic signals at each time point in the microseismic signal monitoring sequence and the corresponding time interval information, and outputs a sequence of microseismic signal hidden state trajectories represented by continuous-time hidden state trajectories. The hidden state trajectory sequence of the surrounding rock deformation, the hidden state trajectory sequence of the support stress, and the hidden state trajectory sequence of the microseismic signal are input into the initial dynamic interactive coordination module to perform multi-round time-series interactive coding processing. In each round of processing, the hidden state of the third monitoring source is updated based on the current hidden state of any two monitoring sources, and the initial interactive coordination trajectory sequence set is obtained through cross-iteration update. A training loss metric is constructed based on the initial interactive coordination trajectory sequence set. The training loss metric is a weighted combination of the reconstruction difference metric generated by each monitoring sequence after dimensional unification and the interaction consistency metric between the hidden state trajectory sequences of each monitoring source. The network weights of the first continuous time hidden state trajectory generation network, the second continuous time hidden state trajectory generation network, the third continuous time hidden state trajectory generation network, and the initial dynamic interactive coordination module are jointly adjusted by minimizing the training loss metric. After the training loss metric converges, the initial dynamic interaction coordination module that has been trained is used as the dynamic interaction coordination module. The rock deformation hidden state trajectory sequence output by the first continuous time hidden state trajectory generation network, the support stress hidden state trajectory sequence output by the second continuous time hidden state trajectory generation network, and the microseismic signal hidden state trajectory sequence output by the third continuous time hidden state trajectory generation network are input into the dynamic interaction coordination module for temporal coupling interaction processing.

3. The method of claim 2, wherein, The process of inputting the hidden state trajectory sequence of the surrounding rock deformation, the hidden state trajectory sequence of the support stress, and the hidden state trajectory sequence of the microseismic signal into the initial dynamic interactive coordination module, and performing multi-round time-series interactive coding processing, includes: The first interactive coding layer of the initial dynamic interactive coordination module performs time alignment on the input hidden state trajectory sequence of surrounding rock deformation, hidden state trajectory sequence of support stress, and hidden state trajectory sequence of microseismic signal to generate time-synchronized hidden state trajectory sequences of surrounding rock deformation, support stress, and microseismic signal. The second interactive coding layer of the initial dynamic interactive coordination module calculates the first unidirectional response intensity of the time-synchronized surrounding rock deformation latent state trajectory sequence to the time-synchronized microseismic signal latent state trajectory sequence, and updates the time-synchronized microseismic signal latent state trajectory sequence accordingly. The third interactive coding layer of the initial dynamic interactive coordination module is used to calculate the second unidirectional response intensity of the updated time-synchronized microseismic signal hidden state trajectory sequence to the time-synchronized support stress hidden state trajectory sequence, and the time-synchronized support stress hidden state trajectory sequence is updated accordingly. The third unidirectional response intensity of the updated time-synchronized support stress latent state trajectory sequence to the time-synchronized surrounding rock deformation latent state trajectory sequence is calculated through the fourth interactive coding layer of the initial dynamic interactive coordination module, and the time-synchronized surrounding rock deformation latent state trajectory sequence is updated accordingly. The time-synchronized hidden state trajectory sequence of surrounding rock deformation, the time-synchronized hidden state trajectory sequence of support stress, and the time-synchronized hidden state trajectory sequence of microseismic signal obtained in this round of updates are residually connected with the trajectory sequence corresponding to the previous round to generate the hidden state trajectory sequence of surrounding rock deformation with cumulative interaction effect, the hidden state trajectory sequence of support stress with cumulative interaction effect, and the hidden state trajectory sequence of microseismic signal with cumulative interaction effect. If the current interaction round does not reach the preset round limit, the trajectory sequences of the cumulative interaction effect are used as the input for the next round and the above interaction coding layer processing is repeated. If the limit is reached, the trajectory sequences of the final cumulative interaction effect are output as the initial interaction coordination trajectory sequence set.

4. The method of claim 2, wherein, The step of constructing a training loss metric based on the initial interactive coordination trajectory sequence set, wherein the training loss metric is a weighted sum of the reconstruction difference metric generated from each monitoring sequence after dimensional unification processing and the interaction consistency metric among the hidden state trajectory sequences of each monitoring source, and the network weights of the first continuous-time hidden state trajectory generation network, the second continuous-time hidden state trajectory generation network, the third continuous-time hidden state trajectory generation network, and the initial dynamic interactive coordination module are jointly adjusted by minimizing the training loss metric, including: The hidden state trajectory sequences of surrounding rock deformation, hidden state trajectory sequences of support stress, and hidden state trajectory sequences of microseismic signals in the initial interactive coordination trajectory sequence set are respectively input into the corresponding generation and reconstruction decoder to generate the reconstructed surrounding rock deformation monitoring sequence, the reconstructed support stress monitoring sequence, and the reconstructed microseismic signal monitoring sequence. The reconstruction difference of surrounding rock deformation is calculated based on the difference between the reconstructed surrounding rock deformation monitoring sequence and the original surrounding rock deformation monitoring sequence. The reconstruction difference of support stress is calculated based on the difference between the reconstructed support stress monitoring sequence and the original support stress monitoring sequence. The reconstruction difference of microseismic signal is calculated based on the difference between the reconstructed microseismic signal monitoring sequence and the original microseismic signal monitoring sequence. The reconstruction difference of surrounding rock deformation, the reconstruction difference of support stress, and the reconstruction difference of microseismic signal are all dimensionless. After eliminating the dimensional differences of each monitored physical quantity, the three reconstruction difference quantities are weighted and summed to obtain the generated reconstruction difference metric. At the end of each round of interactive coding, the hidden state change trends of the time-synchronized surrounding rock deformation hidden state trajectory sequence, the time-synchronized support stress hidden state trajectory sequence, and the time-synchronized microseismic signal hidden state trajectory sequence at continuous time are extracted, and an interactive consistency metric is constructed based on the degree of synchronization between the hidden state change trends. The training loss metric is obtained by multiplying the generated reconstruction difference metric and the interaction consistency metric by a preset interaction consistency constraint weight and then adding them together. The training loss metric is calculated via backpropagation, with respect to the adjustment gradient of the weights of each network in the first continuous-time hidden state trajectory generation network, the second continuous-time hidden state trajectory generation network, the third continuous-time hidden state trajectory generation network, and the initial dynamic interaction coordination module. Based on the adjusted gradient, update the weights of each network, and iteratively perform generation reconstruction, difference metric calculation and weight update until the training loss metric stabilizes.

5. The method of claim 1, wherein, The dynamic interactive coordination module is invoked to perform time-series coupling interactive processing on the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signals. This captures the dynamic interactive response between the hidden state trajectories from different sources within a continuous time frame, generating the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signals, including: The hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal are input into the trajectory time parameter unification layer of the dynamic interactive coordination module. The timestamps and time intervals of each hidden state trajectory sequence are extracted to generate a three-source hidden state trajectory time alignment sequence with a unified time reference. The time-aligned sequence of the three-source hidden state trajectory is input into the bidirectional cross-interactive coding layer of the dynamic interactive coordination module. The instantaneous response vector of the hidden state of surrounding rock deformation to the hidden state of support stress is calculated through the first direction coding branch, and the instantaneous response vector of the hidden state of support stress to the hidden state of surrounding rock deformation is calculated through the second direction coding branch. After splicing, a bidirectional interactive response vector of surrounding rock deformation and support stress is generated. The instantaneous response vector of the latent state of surrounding rock deformation to the latent state of microseismic signal is calculated by the third direction coding branch, and the instantaneous response vector of the latent state of microseismic signal to the latent state of surrounding rock deformation is calculated by the fourth direction coding branch. After splicing, a bidirectional interactive response vector of surrounding rock deformation and microseismic signal is generated. The instantaneous response vector of the hidden state of support stress to the hidden state of microseismic signal is calculated by the fifth direction coding branch, and the instantaneous response vector of the hidden state of microseismic signal to the hidden state of support stress is calculated by the sixth direction coding branch. After splicing, a bidirectional interactive response vector of support stress and microseismic signal is generated. The bidirectional interactive response vectors of the surrounding rock deformation and support stress, the bidirectional interactive response vectors of the surrounding rock deformation and microseismic signals, and the bidirectional interactive response vectors of the support stress and microseismic signals are input into the interactive response fusion layer. The full-source interactive response comprehensive vector is generated by splicing the vectors in the channel dimension. The interactive correction vectors that match the dimensions of each hidden state trajectory sequence are generated through a fully connected mapping network. The modified component of the corresponding hidden state trajectory sequence of surrounding rock deformation in the interactive correction vector is added to the original hidden state trajectory sequence of surrounding rock deformation at each time step to obtain the interactively coordinated hidden state trajectory sequence of surrounding rock deformation. The modified component of the corresponding hidden state trajectory sequence of support stress is added to the original hidden state trajectory sequence of support stress at each time step to obtain the interactively coordinated hidden state trajectory sequence of support stress. The modified component of the corresponding hidden state trajectory sequence of microseismic signal is added to the original hidden state trajectory sequence of microseismic signal at each time step to obtain the interactively coordinated hidden state trajectory sequence of microseismic signal.

6. The method of claim 5, wherein, The step of inputting the time-aligned sequence of the three-source hidden state trajectories into the bidirectional cross-interactive coding layer of the dynamic interaction coordination module includes: Obtain the surrounding rock deformation hidden state vector, support stress hidden state vector and microseismic signal hidden state vector for each time point in the three-source hidden state trajectory time alignment sequence. Generate a surrounding rock deformation query vector and a surrounding rock deformation key vector from the surrounding rock deformation hidden state vector through the first linear projection layer. Generate a support stress query vector and a support stress key vector from the support stress hidden state vector through the second linear projection layer. The interaction attention weights of the surrounding rock deformation on the support stress are generated by the inner product of the surrounding rock deformation query vector and the support stress key vector. The support stress hidden state vector is passed through the third linear projection layer to obtain the support stress value vector. The support stress value vector is weighted and aggregated by the interaction attention weights to obtain the support stress interaction response vector received by the surrounding rock deformation as the output of the first direction encoding branch. The interaction attention weight of support stress on surrounding rock deformation is generated by the inner product of the support stress query vector and the surrounding rock deformation key vector. The surrounding rock deformation value vector is obtained by passing the hidden state vector of surrounding rock deformation through the fourth linear projection layer. The surrounding rock deformation value vector is weighted and aggregated by the interaction attention weight to obtain the surrounding rock deformation interaction response vector received by support stress as the output of the second direction encoding branch. The interaction response vector of the support stress received by the surrounding rock deformation and the interaction response vector of the surrounding rock deformation received by the support stress are spliced ​​along the channel dimension, and after being filtered by the first gated fusion unit, a bidirectional interaction response vector of the surrounding rock deformation and the support stress is generated. Based on the same processing structure as the first and second directional coding branches, a bidirectional interactive response vector of surrounding rock deformation and microseismic signal is generated through the third and fourth directional coding branches, and a bidirectional interactive response vector of support stress and microseismic signal is generated through the fifth and sixth directional coding branches.

7. The method of claim 5, wherein, The step of adding the correction component of the corresponding hidden state trajectory sequence of surrounding rock deformation in the interactive correction vector to the original hidden state trajectory sequence of surrounding rock deformation time-by-time to obtain the interactively coordinated hidden state trajectory sequence of surrounding rock deformation; the step of adding the correction component of the corresponding hidden state trajectory sequence of support stress to the original hidden state trajectory sequence of support stress time-by-time to obtain the interactively coordinated hidden state trajectory sequence of support stress; and the step of adding the correction component of the corresponding hidden state trajectory sequence of microseismic signal to the original hidden state trajectory sequence of microseismic signal time-by-time to obtain the interactively coordinated hidden state trajectory sequence of microseismic signal, further includes: The interconnected and coordinated hidden state trajectory sequence of surrounding rock deformation, the interconnected and coordinated hidden state trajectory sequence of support stress, and the interconnected and coordinated hidden state trajectory sequence of microseismic signal are input into the trajectory smoothing layer for temporal convolution smoothing to obtain the smoothed hidden state trajectory sequence of surrounding rock deformation, the smoothed hidden state trajectory sequence of support stress, and the smoothed hidden state trajectory sequence of microseismic signal. The smoothed hidden state trajectory sequences are input into the trajectory consistency verification layer. The similarity of the hidden state vectors of any two source hidden state trajectory sequences at the same time is calculated to obtain the consistency scores of the surrounding rock deformation and support stress trajectory, the consistency scores of the surrounding rock deformation and microseismic signal trajectory, and the consistency scores of the support stress and microseismic signal trajectory. Determine whether the consistency scores of the three trajectories all fall within a preset reasonable consistency range. If any score exceeds the preset reasonable consistency range, generate a trajectory correction feedback signal based on the degree of exceedance. The trajectory correction feedback signal is sent back to the interactive response fusion layer, the mapping weights of the fully connected mapping network are adjusted to regenerate the interactive correction vector, and the time-by-time addition, trajectory smoothing and consistency verification are repeatedly performed until all trajectory consistency scores fall into the preset reasonable consistency range. After the verification is passed, the smoothed surrounding rock deformation hidden state trajectory sequence is taken as the final interactively coordinated surrounding rock deformation hidden state trajectory sequence, the smoothed support stress hidden state trajectory sequence is taken as the final interactively coordinated support stress hidden state trajectory sequence, and the smoothed microseismic signal hidden state trajectory sequence is taken as the final interactively coordinated microseismic signal hidden state trajectory sequence. The final interconnected and coordinated hidden state trajectory sequence of surrounding rock deformation, the final interconnected and coordinated hidden state trajectory sequence of support stress, and the final interconnected and coordinated hidden state trajectory sequence of microseismic signal are output to the processing step of synchronously extracting hidden state vectors at the risk query time.

8. The method of claim 1, wherein, At the time of risk query, the hidden state vectors of the interactively coordinated surrounding rock deformation hidden state trajectory sequence, the interactively coordinated support stress hidden state trajectory sequence, and the interactively coordinated microseismic signal hidden state trajectory sequence are simultaneously extracted to obtain the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector. Then, an asynchronous feature synchronous fusion processing is performed on the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector through a multi-head cross-attention fusion mechanism to generate a multi-source fused hidden state representation vector, including: Obtain the timestamp value of the risk query time specified externally or input by the user, and obtain the hidden state vector of the surrounding rock deformation hidden state in the interactively coordinated hidden state trajectory sequence based on the timestamp value. Obtain the hidden state vector of the surrounding rock deformation hidden state in the interactively coordinated hidden state trajectory sequence. Obtain the hidden state vector of the support stress hidden state in the interactively coordinated hidden state trajectory sequence. Obtain the hidden state vector of the microseismic signal hidden state in the interactively coordinated hidden state trajectory sequence. The hidden state vectors of surrounding rock deformation, support stress, and microseismic signal are stacked into a three-source hidden state vector stacking matrix. This matrix is ​​then input into the first attention head of the multi-head cross-attention fusion mechanism. The first attention head uses the hidden state vector of surrounding rock deformation as the query, the hidden state vector of support stress as the key, and the hidden state vector of microseismic signal as the value to perform attention aggregation and generate the first attention head output vector. The stacked matrix of the three hidden state vectors is input into the second attention head. The second attention head uses the hidden state vector of the support stress as the query, the hidden state vector of the surrounding rock deformation as the key, and the hidden state vector of the microseismic signal as the value to perform attention aggregation and generate the output vector of the second attention head. The stacked matrix of the three hidden state vectors is input into the third attention head. The third attention head uses the hidden state vector of the microseismic signal as the query, the hidden state vector of the surrounding rock deformation as the key, and the hidden state vector of the support stress as the value to perform attention aggregation and generate the output vector of the third attention head. The first attention head output vector, the second attention head output vector, and the third attention head output vector are concatenated along the channel dimension to obtain a multi-head cross-attention concatenation vector. Then, the multi-head cross-attention concatenation vector is dimensionality reduced and fused through a linear projection fusion layer to generate a multi-source fusion hidden state representation vector.

9. A tunnel construction early warning device, characterized in that, include: The data acquisition module is used to acquire asynchronous monitoring data streams collected by multi-source monitoring equipment in the tunnel construction area. The asynchronous monitoring data streams include surrounding rock deformation monitoring sequences, support stress monitoring sequences, and microseismic signal monitoring sequences. Each monitoring sequence has an independent and non-equally spaced timestamp. The trajectory mapping module is used to construct a continuous-time latent state trajectory generation network for the surrounding rock deformation monitoring sequence, the support stress monitoring sequence and the microseismic signal monitoring sequence, respectively, and to map the discrete observation values ​​of each monitoring sequence to the continuous posterior latent state trajectory of the corresponding monitoring source, so as to obtain the surrounding rock deformation latent state trajectory sequence, the support stress latent state trajectory sequence and the microseismic signal latent state trajectory sequence. The sequence generation module is used to call the dynamic interaction coordination module to perform time-series coupling interaction processing on the hidden state trajectory sequence of surrounding rock deformation, the hidden state trajectory sequence of support stress, and the hidden state trajectory sequence of microseismic signal. It captures the dynamic interaction response between the hidden state trajectories of different sources in a continuous time frame and generates the interactively coordinated hidden state trajectory sequence of surrounding rock deformation, the interactively coordinated hidden state trajectory sequence of support stress, and the interactively coordinated hidden state trajectory sequence of microseismic signal. The cross-fusion module is used to obtain the risk query time, and simultaneously extract the hidden state vectors of the interactively coordinated surrounding rock deformation hidden state trajectory sequence, the interactively coordinated support stress hidden state trajectory sequence, and the interactively coordinated microseismic signal hidden state trajectory sequence at the risk query time, so as to obtain the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector. Then, the surrounding rock deformation hidden state vector, the support stress hidden state vector, and the microseismic signal hidden state vector are asynchronously and synchronously fused through a multi-head cross-attention fusion mechanism to generate a multi-source fused hidden state representation vector. The instruction generation module is used to simultaneously output the risk category prediction distribution, risk severity prediction value, and risk occurrence time window prediction interval through a multi-task prediction network based on the multi-source fusion hidden state representation vector, and generate tunnel construction early warning instructions according to the risk category prediction distribution, the risk severity prediction value, and the risk occurrence time window prediction interval.

10. A computer system, characterized by include: processor; And a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 8.

Images