CN-122024436-A - Pressure regulating well surrounding rock detection method and early warning system based on intelligent early warning

CN122024436ACN 122024436 ACN122024436 ACN 122024436ACN-122024436-A

Abstract

The invention discloses a pressure regulating well surrounding rock detection method and an early warning system based on intelligent early warning, and relates to the technical field of safety monitoring of hydraulic and hydroelectric engineering. The method comprises the steps of dynamically arranging a multi-dimensional sensing unit group, constructing an improved CNN-LSTM model integrated with an attention mechanism and an Adam optimization algorithm by combining wavelet denoising, kalman filtering and systematic preprocessing of temperature and osmotic pressure coupling correction, realizing accurate extraction of space-time characteristics of multiple indexes, adopting a hierarchical analysis method to establish a stability grading evaluation system, generating an optimal emergency scheme by the fusion mechanism, realizing dynamic updating of the model and the evaluation system by means of incremental learning, and constructing six mutually-coordinated modules consisting of multi-dimensional data acquisition, preprocessing, intelligent analysis, stability evaluation, emergency treatment and model self-updating to cooperatively form a pressure regulating well surrounding rock detection and early warning system. The invention has wide monitoring coverage, small positioning error and quick response time, and provides an efficient and accurate intelligent monitoring and early warning solution for the safety of the pressure regulating well surrounding rock.

Inventors

SHAN JIANGUO
Qin Kunhe
LI HAIBIN
ZHAO JIANBIN
WANG KANG
JIANG XIAOWEI
YANG SHAOFEI
GOU WU
ZHU MINXIA
LI XUAN
LI XINXIN

Assignees

中国水电建设集团十五工程局有限公司

Dates

Publication Date: 20260512
Application Date: 20251210

Claims (9)

1. The utility model provides a surge shaft surrounding rock detection method based on intelligent early warning which is characterized in that the method comprises the following steps: Constructing a multidimensional data acquisition system, arranging sensing unit groups at different depths and different orientations of the surrounding rock of the pressure regulating well, and preprocessing multisource raw data acquired by the sensing unit groups; based on the preprocessed multi-source data, training a CNN-LSTM intelligent analysis model fused by an improved convolutional neural network and a long-term and short-term memory network through an adaptive moment estimation Adam optimization algorithm and an attention mechanism; setting evaluation indexes according to the strain change rate, the microseismic event frequency, the osmotic pressure change amplitude and the surface temperature abnormal area output by the intelligent analysis model, calculating the comprehensive score of the stability of the surrounding rock through hierarchical analysis and weighted summation, and constructing a classification evaluation system of the stability of the surrounding rock; based on multisource data acquired in real time, the CNN-LSTM intelligent analysis model performs analysis calculation once at intervals, outputs comprehensive surrounding rock stability scores and corresponding early warning grades, automatically triggers an audible and visual alarm device when the scores are too low, simultaneously starts an emergency response plan recommendation module, and generates an emergency treatment scheme according to the early warning grades, surrounding rock specific parameters and historical treatment cases; And (3) comparing the actual surrounding rock state detection result with the intelligent analysis model prediction result at regular intervals, calculating a prediction error, starting a model parameter updating program when the error is overlarge, adjusting the CNN-LSTM intelligent analysis model through an incremental learning algorithm, and updating a surrounding rock stability grading evaluation system at the same time, so that the model prediction accuracy and the evaluation system applicability are ensured to be continuously optimized along with the change of the actual working condition.
2. The method according to claim 1, wherein the arrangement of the sensor unit group is specifically that the diameter of the pressure regulating well is set as Number of microseismic sensor layout The method meets the following conditions: And when In the time-course of which the first and second contact surfaces, Wherein Indicating the depth of embedding the osmotic pressure sensor Modulus of elasticity with surrounding rock The method meets the following conditions: Wherein 3 and 40 are respectively preset depth range constants.
3. The method according to claim 1, wherein the preprocessing specifically comprises removing the environmental interference noise in the microseismic signal by a wavelet threshold denoising algorithm, wherein a threshold calculation formula of wavelet threshold denoising is as follows: Wherein the method comprises the steps of Is the wavelet coefficient of the microseismic signal, The number of the sampling points for the signals; the fiber bragg grating strain data is smoothed by a Kalman filtering algorithm, and a state updating formula is as follows: Wherein, the Is the first The time-filtered data is then used to determine, In order for the filter gain to be a function of, In order to predict the error covariance of the error, In order to observe the matrix, In order to observe the covariance of the noise, Dynamically adjusting according to the fluctuation amplitude of the strain data; And integrating multi-source features by adopting a channel splicing mode to form a model input tensor to provide a multi-dimensional input basis for the space feature extraction of the CNN module.
4. The method of claim 1, wherein the constructing of the CNN-LSTM intelligent analysis model specifically includes inputting a model input tensor into a CNN spatial feature extraction layer, outputting a 2-dimensional feature map, inputting the feature flattened feature map into an LSTM module, adding a attention layer between the LSTM output layer and a full connection layer, and dynamically calculating feature weights according to the formula: Wherein, the For the dynamic feature weights to be used, Is the first The feature vector of the class data, As a matrix of weights, the weight matrix, Is a bias term; And constructing a full-connection output layer through two layers of full connection, and calculating the error between a model predicted value and an actual monitoring value by using a mean square error MSE loss function, wherein the formula is as follows: Wherein, the Is the first Sample number The actual value of the individual index is calculated, Ensuring the prediction precision of the model on each evaluation index for the corresponding prediction value; And updating parameters of the model in a test stage by an Adam optimization algorithm, wherein the formula is as follows: , Wherein the method comprises the steps of 、 Respectively, a first-order and a second-order momentum estimation, 、 The momentum attenuation coefficients are respectively given to the two, In order to loss-function gradients, In order for the rate of learning to be high, For the number of iterations, And after passing the test set test, obtaining the improved CNN-LSTM intelligent analysis model.
5. The method of claim 1, wherein the step of calculating the comprehensive surrounding rock stability score by analytic hierarchy process and weighted summation comprises determining a matrix in the step of determining the evaluation index weight by analytic hierarchy process Meets the consistency test and the consistency proportion Wherein, the consistency index , , As weight vector, random consistency index Is the corresponding value of the 4-order matrix; target weight vector Wherein For the rate of change of strain weight, Is the frequency weight of the microseismic event, For the weight of the amplitude of the osmotic pressure variation, Area weight is the abnormal area of the surface temperature.
6. The method according to claim 4, wherein the comprehensive score of the stability of the surrounding rock is in particular: calculating a strain rate of change score The formula is: wherein In order to be able to achieve a rate of change of the actual strain, A strain limit threshold, 100 is a total score, 5 is a limit threshold; Calculating the frequency score of microseismic events The formula is: wherein For the actual microseismic frequency, Is a microseismic limit threshold; Calculating the osmotic pressure change amplitude score The formula is: wherein For the actual amplitude of the osmotic pressure change, Is the osmotic pressure limit threshold; calculating area score of abnormal temperature area of surface The formula is: wherein Is the area of the abnormal temperature region, The surface area of surrounding rock of the pressure regulating well is; scoring by strain rate of change Frequency score of microseismic events Osmotic pressure variation amplitude score Area score for surface temperature anomaly region Calculating a comprehensive score for surrounding rock stability The formula is: Wherein the method comprises the steps of Is the first The score of each evaluation index is calculated, Is the first The weight of each evaluation index.
7. The method according to claim 1, wherein the generating of the contingency treatment plan is specifically generating the contingency treatment plan by a case-wise reasoning and rule-wise reasoning fusion mechanism, a plan priority The calculation formula of (2) is as follows: Wherein, the For the purpose of case-wise reasoning about weights, For the degree of matching of the historical cases, For the degree of matching of the rule base, For the current stability score to be a score, 、 Weight for case reasoning Is a range constant of (2); setting a matching degree threshold Among the generated sets of schemes, when the scheme with the highest priority meets And the implementation cost of the scheme And expected reinforcing effect The ratio of (2) satisfies In this case, the scheme is adopted, wherein, 、 The cost and effect of each scheme are respectively.
8. The method according to claim 1, wherein the model parameter updating program is configured to calculate the prediction error by a root mean square error RMSE indicator, and the formula is: Wherein, the In order to predict the error of the signal, In order to compare the number of samples, Is the first The result of the actual detection is that, Is the predicted value corresponding to the model when When the model parameter updating program is started, the incremental learning algorithm fixes CNN layer parameters, only the weights of the LSTM layer and the full-connection layer are updated, and the parameter adjustment amplitude meets the following conditions: Wherein, the The amplitude is adjusted for the parameter(s), Is a learning rate coefficient.
9. The method for realizing any one of claims 1-7, comprising a multidimensional data acquisition module, a data preprocessing module, a CNN-LSTM intelligent analysis module, a stability evaluation module, an emergency treatment module and a model self-updating module, wherein each module is in bidirectional communication with an edge computing gateway through an industrial Ethernet, specifically: the multi-dimensional data acquisition module comprises a sensing unit group, a data acquisition unit and a synchronous control unit, wherein the sensing unit group comprises a fiber bragg grating strain sensor, a microseismic sensor, a osmoticum sensor and an infrared thermal imaging probe which are distributed according to different depths and different directions; the data preprocessing module receives the multidimensional data output by the multidimensional data acquisition module, preprocesses the data through the wavelet threshold denoising unit, the Kalman filtering module unit, the temperature compensation unit and the characteristic fusion unit, integrates the preprocessed data into a model input tensor and outputs the model input tensor to the CNN-LSTM intelligent analysis module; the CNN-LSTM intelligent analysis module comprises a model training unit and a real-time reasoning unit, wherein the model training unit receives historical monitoring data to perform model training, and automatically outputs a model weight file after training is completed; The stability evaluation module comprises a weight calculation unit, a comprehensive scoring unit and an audible and visual alarm unit, wherein the weight calculation unit automatically calculates and stores the weight of each evaluation index through a hierarchical analysis method; The emergency treatment module comprises a case library, a rule library and a scheme generation unit, wherein the case library stores historical emergency treatment cases, the rule library is internally provided with rules formed by industry specifications and expert experiences, the scheme generation unit calculates scheme priorities, generates a plurality of sets of emergency treatment schemes and screens out optimal schemes. The model self-updating module comprises an error calculation unit, an increment learning unit and an evaluation system updating unit, wherein the error calculation unit calculates model prediction errors, actual detection data is regularly imported to be compared with model prediction values, self-updating is triggered when conditions are met, the increment learning module fixes CNN layer parameters through a parameter adjustment algorithm and only updates LSTM layer weights and full-connection layer weights, the evaluation system updating module synchronously updates a judgment matrix and a grade threshold value, and after updating, the judgment matrix and the grade threshold value automatically pass consistency test, and the new weights and the threshold value are synchronized to the stability evaluation and early warning module.

Description

Pressure regulating well surrounding rock detection method and early warning system based on intelligent early warning Technical Field The invention relates to the technical field of safety monitoring of water conservancy and hydropower engineering, in particular to a pressure regulating well surrounding rock detection method and an early warning system based on intelligent early warning. Background The pressure regulating well is used as a core underground structure for connecting a pressure pipeline and a factory building in water conservancy and hydropower engineering, and the stability of surrounding rock is directly related to the safe operation of the whole hydropower junction. In the construction and long-term service process of the pressure regulating well, surrounding rock is subjected to the coupling effects of multiple factors such as excavation unloading, osmotic pressure change, temperature fluctuation and the like, the oversaturation precursors such as strain concentration, micro-vibration activity aggravation, osmotic pressure sudden rise and the like are easy to appear, and if the monitoring is not timely or the early warning is inaccurate, serious safety accidents such as surrounding rock collapse, well wall cracking and the like can be possibly caused. Therefore, the development of an efficient and accurate surrounding rock detection and early warning technology becomes a key research direction in the field of hydraulic and hydroelectric engineering safety. Along with the development of sensing technology and artificial intelligence, the surrounding rock monitoring field has emerged a lot of technical schemes for fusing multi-source data and deep learning, and the technologies realize the improvement of monitoring precision in a specific scene, but aiming at the monitoring requirement of a special structure of a pressure regulating well, the existing method still has the following remarkable defects. The Chinese patent (publication No. CN 202310867542.3) discloses a method for dividing the grades of surrounding rock according to the elastic wave velocity, wherein the method only adopts a single elastic wave velocity index to evaluate the surrounding rock, does not adopt a analytic hierarchy process to scientifically distribute multi-index weights, ignores the coupling influence of osmotic pressure and temperature, leads to the deviation of an evaluation result and actual stability to be 2-3 grades, and has single stability evaluation. Chinese patent (publication No. CN 110472729B) discloses a rock burst state prediction method based on comprehensive CNN-LSTM, which realizes rock burst risk assessment by fusing microseismic and stress data, but does not consider the influence of pressure regulating well diameter difference on monitoring coverage, and the distribution of a sensing unit lacks scene suitability and has insufficient data acquisition effectiveness. Chinese patent (publication No. CN2024131358924. X) discloses a TBM rock machine parameter prediction method adopting a CNN-LSTM-Adaboost model, which improves stratum adaptability by fusing tunneling parameters, focuses on a TBM tunneling scene, does not design a special sensing layout rule aiming at the coupling characteristics of the osmotic pressure and the temperature of the pressure regulating well surrounding rock, and the model does not allocate weights aiming at the space-time characteristics of surrounding rock data, so that the space positioning characteristics of microseismic events and the time evolution rule of strain cannot be captured efficiently, and the model structure is not optimized enough. In conclusion, the prior pressure regulating well surrounding rock monitoring related technical scheme lacks scene suitability, has weak generalization capability and insufficient practicability, has poor synergism on system integration, and is difficult to realize accurate monitoring and high-efficiency early warning of surrounding rock stability. Therefore, an intelligent early warning detection method which has scene suitability, model suitability and decision practicability is developed, and the intelligent early warning detection method has important practical significance for guaranteeing the safety of water conservancy and hydropower engineering. Disclosure of Invention Based on the technical problems, the application discloses a method for detecting the surrounding rock of a pressure regulating well based on intelligent early warning, which comprises the following steps: Constructing a multidimensional data acquisition system, arranging sensing unit groups at different depths and different orientations of the surrounding rock of the pressure regulating well, and preprocessing multisource raw data acquired by the sensing unit groups; based on the preprocessed multi-source data, training a CNN-LSTM intelligent analysis model fused by an improved convolutional neural network and a long-term and short-term memory network th