Self-monitoring karst treatment construction method

Abstract

The invention relates to the technical field of karst treatment construction, in particular to a self-monitoring karst treatment construction method which comprises the following steps that S1, holes are formed in the ground; s2, a self-sensing FRP rib is put down along the open hole; s3, the self-sensing FRP rib is connected with a strain demodulator; s4, the karst cave and the open hole are subjected to filling treatment; and S5, processing effect detection: S6, carrying out self-monitoring, sensing a data basis provided by the FRP rib, extracting and mining information values through multi-scale features, diagnosing a global health state through graph convolution, carrying out reinforcement learning and optimizing an early warning decision, and capturing a mutation behavior through GRU. The system realizes a sensing-diagnosis-early warning-optimization closed loop through tight coupling of a data stream and an algorithm stream, constructs a self-sensing, self-learning and self-perfecting intelligent karst cave monitoring system, and lays a methodological foundation for intelligent geotechnical engineering.

Description

Technical Field

[0001] The invention relates to the technical field of karst treatment construction, and more specifically to a self-monitoring karst treatment construction method. Background Art

[0002] In recent years, infrastructure construction in my country's karst areas has been accelerating, and cave management faces new opportunities and challenges. As a special geological structure, caves have complex internal structures and unclear evolution mechanisms, which pose major safety hazards to engineering construction. Traditional cave management mainly uses reinforcement measures such as grouting, but lacks effective monitoring methods, making it difficult to fully grasp the management effect and long-term stability.

[0003] At present, cave management monitoring mainly adopts a combination of manual inspection and buried sensors. Manual inspection relies on the experience of engineers, is highly subjective, inefficient, and difficult to achieve full-time and full-area coverage; although buried sensors can provide objective data, they have high deployment costs and high damage risks, and data analysis mainly relies on simple threshold comparisons, lacking intelligent diagnosis and early warning capabilities. These problems seriously restrict the scientific and refined level of cave management.

[0004] In terms of sensors, resistance strain sensors represented by vibrating steel string meters are widely used in cave monitoring. Steel string meters indirectly reflect the strain changes of the surrounding rock mass by measuring the vibration frequency of the steel string, and then judge the stability of the cave. However, steel string meters are easily disturbed by environmental factors such as temperature, and their long-term stability and reliability are insufficient. Moreover, steel string meters are point sensors, which are difficult to fully reflect the spatial distribution characteristics of cave deformation. In addition, steel string meters are prone to failure problems such as steel string breakage and rust, and have high maintenance costs. Fiber Bragg grating sensing technology has attracted attention in the field of geotechnical engineering monitoring in recent years. Compared with steel string meters, fiber Bragg gratings have the advantages of strong anti-interference ability, good durability, high sensitivity, and can be connected in series to achieve the effect of distributed monitoring. However, existing studies mainly attach fiber Bragg gratings to the surface of steel bars or anchor rods, and the monitoring performance is limited by the base material, making it difficult to maximize the potential of fiber grating sensing. Moreover, the fiber Bragg grating installation is complex, prone to damage, and insufficient in practicality.

[0005] In terms of data analysis, existing methods mainly use statistical indicators, such as mean and variance, to manually set thresholds for abnormal judgment and early warning. However, cave deformation data has complex characteristics such as nonlinearity and non-stationarity, and traditional indicators often fail. Moreover, the subjective setting of thresholds lacks theoretical basis, has poor robustness, and is prone to problems such as missed alarms and false alarms, and the reliability of early warnings is not high. In addition, a large number of studies have shown that cave deformation is closely related to surrounding environmental factors and hydrogeological conditions. However, existing monitoring schemes often view single parameters in isolation and lack a comprehensive analysis of multi-source influencing factors. There is also a lack of effective mining and utilization of the massive monitoring data obtained, and the value of big data has not yet been fully utilized.

[0006] In summary, innovative breakthroughs are urgently needed in cave management and monitoring. Summary of the invention

[0007] The present invention is aimed at the above technical problems and proposes a self-monitoring karst treatment construction method, which integrates perception, transmission, analysis and early warning, improves the quality and analysis level of monitoring data, provides reliable safety diagnosis and risk early warning for cave treatment projects, and promotes the transition of geotechnical engineering monitoring from experience to intelligence. This is exactly the necessity of the present invention.

[0008] The present invention provides a self-monitoring karst treatment construction method, comprising the following steps:

[0009] Step S1: Drilling a hole in the ground;

[0010] Step S2: lowering the self-sensing FRP reinforcement along the opening;

[0011] Step S3: the self-sensing FRP tendon is connected to a strain demodulator;

[0012] Step S4: filling the cave and the opening;

[0013] Step S5: Processing effect detection, including:

[0014] S5.1) Multi-scale feature adaptive extraction: The multi-parameter monitoring sequence data X={x1, x2,…, x n} Perform wavelet transform and obtain the wavelet coefficient matrix W = [W1, W2, ..., W n ], where x i represents the time series of the ith monitoring parameter, W i The multi-scale time-frequency domain characteristics representing the i-th monitoring parameter perform singular value decomposition on the wavelet coefficient matrix W W = U × S × V T , take the left singular vector U corresponding to the first k singular values k =[u1,u2,…,u k ] as a multi-scale feature method;

[0015] S5.2) Graph Convolutional Neural Network Cave Health Diagnosis: Based on the multi-scale feature U k Construct cave health status diagram G = (V, E, A) n Where V is the node set, E is the edge set, and A is the adjacency matrix through the graph convolution operation H (l+1) =σ(D -1 / 2 ×A×D -1 / 2 ×H (l) ×W (l) ) Update node features, where H (l) is the first layer node feature matrix, W(l) is the convolution kernel parameter matrix, D is the degree matrix, and σ is the activation function. The node embedding feature is input into the classifier Y = Softmax (f (H (L) )) performs health diagnosis, where γ is the diagnosis result and f is the classifier;

[0016] Step S6: Self-monitoring.

[0017] Preferably, the diameter of the hole opened in step S1 is 0.11-0.2 m, and the depth is 10-50 m.

[0018] Preferably, the self-sensing FRP rib in step S2 has a diameter of 10 mm, an embedded fiber grating sensor, a measurement range of ±2000 με, and a sensitivity of 1.2 pm / με.

[0019] Preferably, the strain demodulator in step S3 has 16 channels, a sampling frequency of 100 Hz, and a resolution of 0.1 pm.

[0020] Preferably, the cave filling material in step S4 is a modified cement-based grouting material, and the 28-day unconfined compressive strength is not less than 30 MPa.

[0021] Preferably, in step S5.1, the wavelet transform uses Morlet wavelet, and the decomposition scale is 3 to 5 layers.

[0022] Preferably, in step S5.2, the graph convolutional neural network adopts 2 layers of convolution, the convolution kernel size is 3, the activation function is ReLU; and the classifier f is a 2-layer fully connected layer.

[0023] Preferably, step S6 comprises:

[0024] S6.1) Threshold adaptive optimization based on deep reinforcement learning: The threshold setting is modeled as a Markov decision process (S, A, P, R), where S is the state space, A is the action space, P is the state transition probability, R is the reward function, and the policy network outputs the Q-value function Q(s t ,a t )=f(s t ,a t ; θ), minimize the time difference error L(θ) = E[(r t +γ×max a Q(s t+1 ,a;θ)-Q(s t ,a t ;θ)) 2 ] Update the policy network parameters Θ, input state s t To monitor the multi-scale features of sequence X;

[0025] 6.2) Cave mutation behavior capture: An anomaly detection model based on the gated recurrent unit GRU is used to reset the gate r t , update gate z t and the hidden state h t Iteration, monitoring value x at time t t Encoding, when the reconstruction error e t =||x t -f(h t )||2 is greater than the threshold value, it is determined as a mutation abnormality, where f is the reconstruction function

[0026] Preferably, in step S6.1, the state space S is represented as a continuous high-dimensional vector, which is extracted from the monitoring sequence X through the encoder LS TM; the action space A is a discrete value, corresponding to the threshold adjustment; the reward function R is designed based on the diagnosis accuracy and warning timeliness; the policy network adopts an LSTM encoder and an MLP decoder.

[0027] Preferably, the reset gate, update gate and hidden state recursive formula of the gated recurrent unit GRU in step S6.2 are:

[0028] r t =σ(W r ×[h t-1 ,x t ])

[0029] z t =σ(W z ×[h t-1 ,x t ])

[0030] n t =tanh(W n ×[r t ×h t-1 ,x t ])

[0031] h t =(1-z t )×n t +z t ×h t-1

[0032] Among them, W r ,W z ,W n is the gating parameter matrix, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function; the reconstruction function f adopts a 2-layer MLP.

[0033] The present invention has the following beneficial effects:

[0034] 1. First, the buried self-sensing FRP bars give caves awareness. Traditional cave management often lacks effective internal monitoring methods, while the present invention uses the strain sensing characteristics of FRP bars to achieve real-time perception of cave deformation, laying a data foundation for subsequent diagnosis and early warning. At the same time, FRP bars also have a reinforcing and toughening effect, and together with the cave filling material, they construct a composite reinforcement system with built-in intelligent sensing, high strength and low deformation.

[0035] Secondly, multi-scale time-frequency domain feature extraction significantly improves the expression ability of monitoring information. The deformation behavior of caves has complex characteristics such as multiple time scales and non-stationarity, which are difficult to describe with traditional time domain indicators. The present invention innovatively introduces wavelet transform and singular value decomposition to adaptively extract key features, overcome the limitations of artificial features, and provide information enhancement for subsequent intelligent diagnosis and early warning.

[0036] Furthermore, graph convolutional neural networks enable global assessment of cave health status. Existing methods often view data from a single monitoring point in isolation, ignoring spatial correlation information. This paper introduces graph convolution into cave health diagnosis for the first time, models the complex topological relationship between monitoring parameters through graph structure, learns node embedding representation, and comprehensively analyzes the global status. The seamless integration of graph convolution and feature extraction maximizes the diagnostic potential of monitoring data.

[0037] In addition, the deep reinforcement learning strategy realizes the adaptive optimization of the warning threshold. Traditional threshold setting often relies on experience and lacks robustness. The present invention models the threshold adjustment as a Markov decision process and avoids the limitations of human experience through self-learning of the strategy network. The LSTM encoder extracts the evolution characteristics of the monitoring sequence and learns the optimal threshold strategy end-to-end to ensure the adaptability and accuracy of the warning.

[0038] Finally, the GRU network is endowed with the ability to intelligently capture cave mutations. Cave mutations are a major hidden danger to engineering safety, and traditional methods are difficult to achieve effective early warning. The present invention uses GRU to perform high-dimensional dynamic modeling of the monitoring sequence to sensitively capture abnormal mutations in the cave state. The gating mechanism enables GRU to adaptively forget or remember the evolution laws of different time scales, ensuring the timeliness and accuracy of mutation detection.

[0039] In summary, the present invention deeply integrates cave management with artificial intelligence, and multiple innovative links are organically connected and synergistic. Self-sensing FRP bars provide a data basis, multi-scale feature extraction mines information value, graph convolution diagnoses the global health status, reinforcement learning optimizes early warning decisions, and GRU captures mutation behavior. Through the tight coupling of data flow and algorithm flow, the system realizes the closed loop of perception-diagnosis-early warning-optimization, constructs a self-perceiving, self-learning, and self-improving intelligent cave monitoring system, and lays a methodological foundation for smart geotechnical engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the self-monitoring karst treatment construction method of the present invention. DETAILED DESCRIPTION

[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the self-monitoring karst treatment construction method proposed by the present invention, its specific implementation method, structure, characteristics and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0042] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0043] The specific scheme of the self-monitoring karst treatment construction method provided by the present invention is described in detail below with reference to the accompanying drawings.

[0044] See also Figure 1 The present invention provides a self-monitoring karst treatment construction method, which realizes intelligent diagnosis and long-term stability monitoring and early warning of the cave treatment effect by burying self-sensing FRP bars during the cave treatment process. The whole process includes 6 key steps:

[0045] Step S1: Drill a hole on the ground; the diameter of the hole is 0.11-0.2m and the depth is 10-50m.

[0046] Step S2: lowering the self-sensing FRP reinforcement along the opening; this step can be implemented in a variety of ways, one of which is to tie a counterweight block to one end of the self-sensing FRP reinforcement, and place the self-sensing FRP reinforcement at a preset position in the cave through the gravity of the counterweight block.

[0047] Step S3: the self-sensing FRP tendon is connected to a strain demodulator;

[0048] Step S4: filling the cave and the opening;

[0049] Step S5: Processing effect detection;

[0050] Step S6: Self-monitoring.

[0051] The present invention uses a self-sensing FRP bar with a built-in fiber grating sensor, with a diameter of 10mm, a measurement range of ±2000με, and a sensitivity of 1.2pm / με. Distributed monitoring of cave deformation is achieved. The grid size is reasonably configured according to the span and shape of the cave, generally 1 to 3m.

[0052] The strain demodulator uses a 16-channel strain demodulator to collect data from the buried FRP grating sensor. The sampling frequency is set to 100Hz and the resolution reaches 0.1pm to meet the needs of dynamic deformation monitoring. The strain demodulator is connected to the monitoring server via wired or wireless means to achieve real-time data transmission and storage.

[0053] For filling treatment of caves and openings, modified cement-based grouting materials are used to fill and reinforce the caves and openings, and the self-sensing FRP bars in the holes are fixed. The modified cement-based grouting materials are purchased from Xi'an Wuhe New Materials Technology Group Co., Ltd. The grouting materials form a dense limestone-cement complex in the cave through a special grouting pipe, which improves the overall strength and stability of the cave. The 28-day unconfined compressive strength of the grouting material is not less than 30MPa to ensure long-term performance.

[0054] Step S5: Treatment effect detection. Wavelet transform, singular value decomposition and graph convolutional neural network are integrated to perform intelligent diagnosis on the health status of the treated cave.

[0055] S5.1) Multi-scale feature adaptive extraction of multi-parameter monitoring sequence data X = {x1, x2, ..., x n} Perform wavelet transform to extract multi-scale time-frequency domain features. Use Morlet wavelet, decompose the scale into 3 to 5 layers, and obtain the wavelet coefficient matrix W = [W1, W2, ..., W n ], where W i Represents the multi-scale characteristics of the i-th monitoring parameter. Further singular value decomposition of W is performed: W = U × S × V T , select the left singular vector U corresponding to the first k singular values k =[u1,u2,…,u k ] as the key feature subset. The k value is determined by the cumulative contribution rate of singular values, and generally the cumulative contribution rate is taken to be more than 90%.

[0056] S5.2) Graph convolutional neural network for cave health diagnosis.

[0057] The extracted multi-scale features U k As nodes, a cave health status graph G = (V, E, A) is constructed. V is a node set corresponding to the monitoring parameters; E is an edge set, indicating the correlation between parameters; A is an adjacency matrix, describing the topological structure between nodes. Node information is aggregated through graph convolution operations to update node features: H (l+1) =σ(D -1 / 2 ×A×D -1 / 2 ×H (l) ×W (l)) function. Two convolution layers are stacked, the convolution kernel size is 3, and the activation function is ReLU. The node embedding features obtained by convolution are input into the two-layer fully connected classifier, and the Softmax function is used to output the cave health status diagnosis result Y. The diagnosis results can be divided into three categories: "qualified", "re-inspection" and "reinforcement", which provide a basis for processing quality acceptance and subsequent decision-making.

[0058] Step S6: Self-monitoring. Based on deep reinforcement learning, the warning threshold is optimized to achieve intelligent warning of cave stability; the gated recurrent unit network GRU is used to capture the cave mutation behavior in real time.

[0059] S6.1) Threshold Adaptive Optimization Based on Deep Reinforcement Learning

[0060] The warning threshold setting is regarded as a strategy learning problem. The optimal threshold strategy is learned through continuous trial and error and environmental interaction. The state space S is a continuous high-dimensional vector, and the LSTM encoder is used to extract state features from the monitoring sequence X. The action space A is a set of discrete threshold adjustment actions, such as increasing by 10%, decreasing by 596, etc. The reward function R is designed according to the accuracy and timeliness of the warning. Accurate warnings are given positive rewards, and missed warnings and false alarms are punished. The transition probability P obeys the random strategy, and the exploration or exploitation action is selected through the e-greedy method. The policy network outputs the action value function Q(s t ,a t )=f(s t ,a t ; θ), the LSTM encoder is used to extract state features, and the MLP decoder outputs the Q value of each action. By minimizing the temporal difference error L(θ) = E[(r t +γ×max a Q(s t+1 ,a;θ)-Q(s t ,a t ;θ)) 2 ] Update the policy network parameters θ, γ as discount factors. The network is optimized through self-supervised learning and selects the threshold action that maximizes the long-term reward in a given state.

[0061] S6.2) Capturing the mutation behavior of caves. The gated recurrent unit (GRU) is used to sensitively capture the mutation behavior of monitoring parameters. For the monitoring value x at time t t , GRU resets the gate r t , update gate z t and the hidden state h t Iterative calculation of , extracting its time-dependent features. Reset gate r t Control the hidden state h at the previous moment t-1 Information reset:

[0062] r t=σ(W r ×[h t-1 ,x t ])

[0063] Update gate z t Controls the degree of preservation of the previous hidden state:

[0064] z t =σ(W z ×[h t-1 ,x t ])

[0065] Candidate hidden state n t Combining the current input and the historical information after reset:

[0066] n t =tanh(W n ×[r t ×h t-1 ,x t ])

[0067] The final hidden state h t From the previous state h t -1 and candidate state n t Press Update Gate t Weighted to get:

[0068] h t =(1-z t )×n t +z t ×h t-1

[0069] The monitoring sequence is encoded by GRU, and the reconstruction error e t =||x t -f(h t )||2 When it exceeds the given threshold, it is determined that a mutation event has occurred. The reconstruction function f uses a 2-layer MLP (multi-layer perceptron) and the threshold is updated in real time by the reinforcement learning strategy in S7.1. When a mutation is determined to occur, the system automatically issues an early warning and triggers the emergency response process.

[0070] In terms of hardware structure, the self-monitoring karst treatment construction system of the present invention consists of a field monitoring unit, a data transmission unit, an edge computing unit and a cloud service unit.

[0071] The on-site monitoring unit includes a self-sensing FRP bar and a strain demodulator. The self-sensing FRP bar uses a composite bar with a diameter of 10mm and an embedded fiber grating sensor, with a measurement range of ±2000με and a sensitivity of 1.2pm / με. The strain demodulator uses a high-precision acquisition device with 16 channels, a sampling frequency of 100Hz, and a resolution of 0.1pm, which is responsible for the real-time acquisition and digital conversion of FRP bar strain data.

[0072] The data transmission unit uses an industrial-grade 4G or 5G wireless router to support two-way communication between the edge computing unit and the field monitoring unit. The wireless router is equipped with a high-gain antenna with a transmission distance of up to 1km, ensuring the real-time and reliability of data transmission.

[0073] The edge computing unit uses an industrial computer as the hardware platform, with a built-in Intel i7 processor and 16GB memory, and is equipped with a Linux operating system. The industrial computer is responsible for receiving on-site monitoring data in real time and running intelligent algorithms such as graph convolutional neural networks and GRU to achieve cave health diagnosis and mutation warning. The industrial computer is equipped with a 512GB solid-state hard drive, which can locally store more than 30 days of monitoring data and diagnostic results.

[0074] The cloud service unit uses a high-performance server cluster equipped with 2-way Intel Xeon Gold 6248 processors, 256GB memory and 4 NVIDIA Tesla V100 graphics cards, providing powerful computing and storage capabilities. The server cluster deploys a deep reinforcement learning platform to train and update threshold optimization strategies. At the same time, the server also undertakes tasks such as remote data backup, historical data analysis and visualization, providing big data support for engineering decision-making. The cloud service unit and the edge computing unit are interconnected through a dedicated optical fiber network, ensuring the security and efficiency of data transmission.

[0075] The present invention also integrates virtual instrument technology, and builds self-monitoring virtual dashboards on the edge computing unit and cloud service unit respectively. The dashboard is developed using LabVIEW, which displays monitoring data, health diagnosis results, mutation warning information, etc. in real time and interacts with human-computer interaction, providing an intuitive and friendly human-computer interface for engineering managers.

[0076] In addition, the present invention also provides a mobile APP, which supports managers to remotely access monitoring data and diagnostic results through portable devices such as smart phones and tablets, realizing mobile office and real-time control. The mobile APP and the cloud service unit use an encrypted communication protocol to ensure data security.

[0077] To summarize, the present invention constructs a three-level intelligent monitoring architecture of end-edge-cloud, and integrates information technology such as virtual instruments and mobile Internet to form a complete hardware and software system, realizing the automation and intelligence of the entire process from on-site perception, data analysis to information presentation, providing efficient and reliable intelligent monitoring guarantee for cave management projects.

[0078] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A self-monitoring karst treatment construction method, characterized in that: The following steps are involved: Step S1: Drilling a hole in the ground; Step S2: lowering the self-sensing FRP reinforcement along the opening; Step S3: the self-sensing FRP tendon is connected to a strain demodulator; Step S4: filling the cave and the opening; Step S5: Processing effect detection, including: S5.1) Multi-scale feature adaptive extraction: The multi-parameter monitoring sequence data X={x1, x2,…, x n } Perform wavelet transform and obtain the wavelet coefficient matrix W = [W1, W2, ..., W n ], where x i represents the time series of the i-th monitoring parameter, W i The multi-scale time-frequency domain characteristics representing the i-th monitoring parameter perform singular value decomposition on the wavelet coefficient matrix W W = U × S × V T , take the left singular vector U corresponding to the first k singular values k =[u1,u2,…,u k ] as a multi-scale feature method; S5.2) Graph Convolutional Neural Network Cave Health Diagnosis: Based on the multi-scale feature U k Construct cave health status diagram G = (V, E, A) n Where V is the node set, E is the edge set, and A is the adjacency matrix through the graph convolution operation H (l+1) =σ(D -1 / 2 ×A×D -1 / 2 ×H (l) ×W (l) ) Update node features, where H (l) is the first layer node feature matrix, W (l) is the convolution kernel parameter matrix, D is the degree matrix, and σ is the activation function. The node embedding feature is input into the classifier Y = Softmax (f (H (L) )) performs health diagnosis, where γ is the diagnosis result and f is the classifier; Step S6: Self-monitoring.

2. The self-monitoring karst treatment construction method according to claim 1, characterized in that: The diameter of the hole opened in step S1 is 0.11-0.2 m, and the depth is 10-50 m.

3. The self-monitoring karst treatment construction method according to claim 1, characterized in that: The self-sensing FRP rib in step S2 has a diameter of 10 mm, an embedded fiber grating sensor, a measurement range of ±2000 με, and a sensitivity of 1.2 pm / με.

4. The self-monitoring karst treatment construction method according to claim 1, characterized in that: In step S3, the strain demodulator has 16 channels, a sampling frequency of 100 Hz, and a resolution of 0.1 pm.

5. The self-monitoring karst treatment construction method according to claim 1, characterized in that: In step S4, the cave filling material is a modified cement-based grouting material, and the 28-day unconfined compressive strength is not less than 30 MPa.

6. The self-monitoring karst treatment construction method according to claim 1, characterized in that: In the step S5.1, the wavelet transform uses Morlet wavelet, and the decomposition scale is 3 to 5 layers.

7. The self-monitoring karst treatment construction method according to claim 1, characterized in that: In step S5.2, the graph convolutional neural network adopts 2 layers of convolution, the convolution kernel size is 3, and the activation function is ReLU; the classifier f is a 2-layer fully connected layer.

8. The self-monitoring karst treatment construction method according to claim 1, characterized in that: The step S6 comprises: S6.1) Threshold adaptive optimization based on deep reinforcement learning: The threshold setting is modeled as a Markov decision process (S, A, P, R), where S is the state space, A is the action space, P is the state transition probability, R is the reward function, and the policy network outputs the Q-value function Q(s t ,a t )=f(s t ,a t ; θ), minimize the time difference error L(θ) = E[(r t +γ×max a Q(s t+1 ,a;θ)-Q(s t ,a t ;θ) 2 ] Update the policy network parameters Θ, input state s t To monitor the multi-scale features of sequence X; 6.2) Capturing cave mutation behavior: An anomaly detection model based on the gated recurrent unit (GRU) is constructed by resetting the gate r t , update gate z t and the hidden state h t Iteration, monitoring value x at time t t Encoding, when the reconstruction error e t =||x t -f(h t )||2 is greater than the threshold value, it is determined as a mutation abnormality, where f is the reconstruction function.

9. The self-monitoring karst treatment construction method according to claim 8, characterized in that: In step S6.1, the state space S is represented as a continuous high-dimensional vector, which is extracted from the monitoring sequence X through the encoder LSTM; the action space A is a discrete value, corresponding to the threshold adjustment; the reward function R is designed based on the diagnosis accuracy and warning timeliness; the policy network adopts the LSTM encoder and the MLP decoder.

10. The self-monitoring karst treatment construction method according to claim 9, characterized in that: The reset gate, update gate and hidden state recursive formula of the gated recurrent unit GRU in step S6.2 are: r t =σ(W r ×[h t-1 ,x t ]) z t =σ(W z ×[h t-1 ,x t ]) n t =tanh(W n ×[r t ×h t-1 ,x t ]) h t =(1-z t )×n t +z t ×h t-1 Among them, W r ,W z ,W n is the gating parameter matrix, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function; the reconstruction function f adopts a 2-layer MLP.

Images