Reservoir fusion analysis method and system based on multiple sensors

Abstract

The invention discloses a reservoir fusion analysis method and system based on multiple sensors, belongs to the technical field of artificial intelligence, and aims to solve the technical problem of how to realize accurate sensing and comprehensive evaluation of a reservoir operation state and timely discover potential safety hazards. Comprising the following steps: collecting reservoir monitoring data, meteorological data and human activity data as original time sequence data, and constructing a data set based on the original time sequence data and a corresponding reservoir safety level; performing data preprocessing on the data set; screening out features with high mutual information and large information difference as target features, and constructing a sample set based on the target features and corresponding reservoir safety levels; constructing a reservoir monitoring model based on an RNN network, a C NN network, an attention mechanism and a conditional probability distribution layer; and inputting the preprocessed to-be-measured time series data into the trained reservoir monitoring model, and predicting and outputting the reservoir safety level and the corresponding probability value of the target reservoir through the reservoir monitoring model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a reservoir fusion analysis method and system based on multiple sensors. BACKGROUND

[0002] In recent years, with the rapid development of social economy and population growth, the importance of water resources has become increasingly prominent. As a key infrastructure for water resources regulation and utilization, reservoirs have been expanding in size and are responsible for multiple important tasks such as flood control, irrigation, water supply, and power generation. Reservoirs face many complex and severe challenges during operation.

[0003] The environment of the reservoir is complex and variable, and is significantly affected by various natural factors such as rainfall, geological activity, and climate change. For example, extreme rainfall events can cause the water level of the reservoir to rise rapidly, increasing the flood control pressure on the dam; geological structure movements can cause stability problems of the dam and surrounding mountains, threatening the safety of the reservoir. The facilities of the reservoir gradually age during long-term operation, and some early-stage constructed reservoirs have different degrees of wear, corrosion, and functional degradation phenomena in dam structure, flood discharge facilities, monitoring equipment, etc. due to long service life.

[0004] In the past period of time, safety accidents such as dam leakage, piping, landslides, and even dam breaches have occurred in reservoirs across the country, causing not only huge economic losses but also serious threats to the lives and safety of residents in downstream areas, and also having immeasurable damage to the ecological environment. Especially during critical periods such as the flood season, the safe operation of the reservoir is crucial to the overall situation.

[0005] To address these problems, various monitoring systems have been introduced by local reservoir management departments, which collect key parameters in real time during the operation of the reservoir by installing water level sensors, seepage sensors, stress and strain sensors, etc. However, these monitoring systems currently operate independently, with scattered data and lack of effective fusion analysis means, making it difficult to comprehensively and accurately assess the operating state and potential risks of the reservoir.

[0006] How to accurately perceive and comprehensively evaluate the operating state of the reservoir and timely discover potential safety hazards is a technical problem that needs to be solved. SUMMARY

[0007] The technical task of the present application is to provide a reservoir fusion analysis method and system based on multiple sensors to solve the technical problem of how to accurately perceive and comprehensively evaluate the operating state of the reservoir and timely discover potential safety hazards.

[0008] In a first aspect, the present application provides a reservoir fusion analysis method based on multiple sensors, comprising the following steps:

[0009] Dataset construction: collect reservoir monitoring data, meteorological data and human activity data as original time series data, and construct a dataset based on the original time series data and the corresponding reservoir safety level;

[0010] Data preprocessing: data preprocessing is performed on the dataset, and the original time series data is converted in format, outliers are processed, and missing values are filled through data preprocessing to obtain a preprocessed dataset;

[0011] Feature selection: for the preprocessed dataset, the sample data in the dataset is used as feature data affecting the reservoir operation state, and feature selection is performed based on the mutual information between the feature data and the reservoir safety level and the information difference between the features. The features with high mutual information and large information difference are selected as target features, and a sample set is constructed based on the target features and the corresponding reservoir safety level;

[0012] Model construction: a reservoir monitoring model is constructed based on RNN network, CNN network, attention mechanism and conditional probability distribution layer, and a dynamic regularization strategy is introduced. The reservoir monitoring model is trained based on the sample set to obtain a trained reservoir monitoring model. The reservoir monitoring model is used to extract features from input time series data through RNN network and CNN network, weight the features output by the RNN network based on the attention mechanism, and predict and analyze the reservoir operation state through the conditional probability distribution layer based on the features output by the CNN network and the weighted features output by the attention mechanism, output the reservoir safety level and the corresponding probability value;

[0013] Reservoir monitoring: for the target reservoir to be monitored, collect reservoir monitoring data, meteorological data and human activity data as test time series data, perform data preprocessing on the test time series data, convert the test time series data in format, process outliers and fill missing values through data preprocessing, input the preprocessed test time series data into the trained reservoir monitoring model, and output the reservoir safety level and the corresponding probability value of the target reservoir through the reservoir monitoring model.

[0014] Preferably, data preprocessing includes the following steps:

[0015] Data format conversion: reshape the data to adapt to the analysis modeling;

[0016] Outlier processing: for values that exceed the sensor range or are not reasonable, instantaneous mutation values, data is selected through z_score, and for continuous constants, moving average method or linear interpolation method is used for processing;

[0017] Missing value processing: missing data is filled by cubic spline interpolation method.

[0018] As a preferred method, feature selection includes the following steps:

[0019] L100, calculate the mutual information between characteristic data and reservoir safety level, mutual information I(X) i The formula for calculating Y is expressed as:

[0020]

[0021] Where Y represents the safety level of the reservoir, X i Let p(x) represent the i-th feature. i Let p(x, y) represent the joint probability distribution. i p(y) and p(y) are marginal probability distributions;

[0022] L200. Calculate the information difference between features. The information difference between features is the redundancy between features. Feature X i With feature X j The information difference R(X) between (j≠i) i The calculation formula is expressed as follows:

[0023]

[0024] Where |S| represents the total number of features;

[0025] L300. Quantify the information difference by calculating the difference between mutual information and redundancy, and select the feature with the largest difference. The calculation formula is as follows:

[0026] Score(X i )=I(X i ;Y)-R(X i ),

[0027] Alternatively, calculate the ratio of mutual information to redundancy, using the following formula:

[0028]

[0029] Where ε is the local minimum value to avoid division by zero;

[0030] L400, Feature selection: Initialize the feature set S with all candidate features, and iteratively select... Add the selection results to the selected feature set S. m Remove the selected features from the initial feature set S, and repeat steps L100-L400 until the preset number of features or score threshold is reached.

[0031] Preferably, the reservoir monitoring model includes an RNN network module, a CNN network module, an attention mechanism module, a splicing layer, a fully connected layer, and a conditional probability distribution layer.

[0032] The RNN network module takes time series data as input, extracts long-term dependence features of the time series data based on a bidirectional LSTM network model, and outputs a feature vector;

[0033] The CNN network module takes time series data as input, extracts local features of the time series data based on a convolutional layer and a max pooling layer, and outputs a feature vector;

[0034] The attention mechanism module is connected with the RNN network module, and is used for attention weighting of the feature vector output by the RNN network module, and outputs a weighted feature vector;

[0035] The concatenation layer is used for feature fusion of the output of the CNN network module and the output of the attention mechanism module, and outputs a fused feature;

[0036] The fully connected layer is connected with the concatenation layer, and is used for linear mapping of the fused feature and output to the conditional probability distribution layer;

[0037] The conditional probability distribution layer is used for receiving the feature vector output by the fully connected layer, performing reservoir operation state prediction analysis based on the feature vector through a Softmax function, and outputting a reservoir safety level and a corresponding probability value;

[0038] Wherein, when the reservoir monitoring model is trained based on the sample set, the target features in the sample set are subjected to Z-score standardization processing to obtain standardized target features, and the standardized target features are divided into fixed-length time series samples based on a predefined sliding window, and a final sample set is constructed based on the fixed-length time series samples and corresponding reservoir safety levels, and the reservoir monitoring model is trained based on the final sample set, and during the training process, the weight matrix of the fully connected layer is subjected to dynamic Dropout and L2 regularization constraint based on a dynamic regularization strategy to prevent overfitting;

[0039] Correspondingly, the total loss L of the reservoir monitoring model total is expressed as:

[0040] L total =L CE +λ·L MLE +γ·L L2 ,

[0041] Wherein, λ and γ are hyperparameters, L L2 represents a regularization term;

[0042] L CE represents a cross-entropy loss, and the calculation formula is as follows:

[0043]

[0044] y i is one-hot encoding of true label, p i is the probability output by the reservoir monitoring model;

[0045] L MLE represents the maximum likelihood loss, and the calculation formula is as follows:

[0046] L MLE =-logP(X|θ),

[0047] wherein X represents the input target feature, and θ is the parameter of the reservoir monitoring model.

[0048] In a second aspect, the present application discloses a reservoir fusion analysis system based on multiple sensors, which comprises a data set construction module, a data preprocessing module, a feature screening module, a model construction module, and a reservoir monitoring module.

[0049] The data set construction module is configured to perform the following operation: collecting reservoir monitoring data, meteorological data, and human activity data as original time series data, and constructing a data set based on the original time series data and corresponding reservoir safety levels.

[0050] The data preprocessing module is configured to perform the following operation: data preprocessing on the data set, format conversion, outlier processing, and missing value filling on the original time series data through data preprocessing, and obtaining a preprocessed data set.

[0051] The feature screening module is configured to perform the following operation: for the preprocessed data set, the sample data in the data set is used as feature data affecting the reservoir operation state, and the feature screening is performed based on the mutual information between the feature data and the reservoir safety level and the information difference between the features, and the features with high mutual information and large information difference are screened out as target features, and a sample set is constructed based on the target features and corresponding reservoir safety levels.

[0052] The model construction module is configured to perform the following operation: constructing a reservoir monitoring model based on an RNN network, a CNN network, an attention mechanism, and a conditional probability distribution layer, introducing a dynamic regularization strategy, and performing model training on the reservoir monitoring model based on the sample set, to obtain a trained reservoir monitoring model, wherein the reservoir monitoring model is configured to perform feature extraction on the input time series data through the RNN network and the CNN network, weight the features output by the RNN network based on the attention mechanism, and perform reservoir operation state prediction analysis through the conditional probability distribution layer based on the features output by the CNN network and the weighted features output by the attention mechanism, to output the reservoir safety level and the corresponding probability value.

[0053] The reservoir monitoring module performs the following operations: For the target reservoir to be monitored, it collects reservoir monitoring data, meteorological data, and human activity data as the time series data to be measured. It performs data preprocessing on the time series data to be measured, including format conversion, outlier handling, and missing value filling. The preprocessed time series data is then input into the trained reservoir monitoring model. The reservoir monitoring model then predicts and outputs the reservoir safety level of the target reservoir and the corresponding probability value.

[0054] Preferably, the data preprocessing module is used to perform the following operations:

[0055] Data format conversion: Reshaping the data to adapt it for analysis and modeling.

[0056] Outlier handling: For values ​​that exceed the sensor's range, are illogical, or are sudden changes, data is filtered using z_score. For continuous constant values, moving average or linear interpolation methods are used.

[0057] Missing value handling: Missing data is filled in using cubic spline interpolation.

[0058] Preferably, the feature filtering module is used to perform the following operations:

[0059] L100, calculate the mutual information between characteristic data and reservoir safety level, mutual information I(X) i The formula for calculating Y is expressed as:

[0060]

[0061] Where Y represents the safety level of the reservoir, X i Let p(x) represent the i-th feature. i Let p(x, y) represent the joint probability distribution. i p(y) and p(y) are marginal probability distributions;

[0062] L200. Calculate the information difference between features. The information difference between features is the redundancy between features. Feature X i With feature X j The information difference R(X) between (j≠i) i The calculation formula is expressed as follows:

[0063]

[0064] Where |S| represents the total number of features;

[0065] L300. Quantify the information difference by calculating the difference between mutual information and redundancy, and select the feature with the largest difference. The calculation formula is as follows:

[0066] Score(X i )=I(X i ;Y)-R(X i ),

[0067] Alternatively, the ratio of mutual information and redundancy is calculated, and the calculation formula is as follows:

[0068]

[0069] Wherein, ε is a minimum value, to avoid division by zero;

[0070] L400, screening features, initializing the feature set S as all candidate features, iteratively selecting The selection result is added to the selected feature set S m The selected features are removed from the initialized feature set S, and steps L100-L400 are repeated until a preset feature number or score threshold is reached.

[0071] As preferred, the reservoir monitoring model comprises an RNN network module, a CNN network module, an attention mechanism module, a concatenation layer, a full connection layer, and a conditional probability distribution layer;

[0072] The RNN network module takes time series data as input, and is used to extract long-term dependence features of the time series data based on a bidirectional LSTM network model, and outputs a feature vector;

[0073] The CNN network module takes time series data as input, and is used to extract local features of the time series data based on a convolution layer and a max pooling layer, and outputs a feature vector;

[0074] The attention mechanism module is connected with the RNN network module, and is used to perform attention weighting on the feature vector output by the RNN network module, and outputs a weighted feature vector;

[0075] The concatenation layer is used to fuse the output of the CNN network module and the output of the attention mechanism module, and outputs the fused features;

[0076] The full connection layer is connected with the concatenation layer, and is used to perform linear mapping on the fused features, and outputs to the conditional probability distribution layer;

[0077] The conditional probability distribution layer is used to receive the feature vector output by the full connection layer, and perform reservoir operation state prediction analysis based on the feature vector through a Softmax function, and output the reservoir safety level and the corresponding probability value;

[0078] Wherein, when the reservoir monitoring model is trained based on the sample set, the target features in the sample set are subjected to Z-score standardization processing to obtain the standardized target features, and the standardized target features are divided into time series samples of fixed length based on a predefined sliding window, a final sample set is constructed based on the time series samples of fixed length and corresponding reservoir safety levels, the reservoir monitoring model is trained based on the final sample set, and in the training process, the weight matrix of the full connection layer is subjected to dynamic Dropout and L2 regularization constraint based on a dynamic regularization strategy to prevent overfitting.

[0079] Correspondingly, the total loss L of the reservoir monitoring model total is expressed as:

[0080] L total = L CE + λ·L MLE + γ·L L2 ,

[0081] Wherein, λ and γ are hyperparameters, L L2 represents a regularization term.

[0082] L CE represents a cross-entropy loss, and the calculation formula is as follows:

[0083]

[0084] y i is the one-hot encoding of the true label, and p i is the probability output by the reservoir monitoring model.

[0085] L MLE represents a maximum likelihood loss, and the calculation formula is as follows:

[0086] L MLE = -logP(X|θ),

[0087] Wherein, X represents the input target features, and θ is the parameter of the reservoir monitoring model.

[0088] The reservoir fusion analysis method and system based on multiple sensors have the following advantages:

[0089] 1. Improve the prediction accuracy of reservoir operation parameters: by fusing multiple sensor data, using advanced hybrid neural network models, fully considering the spatio-temporal correlation and complex features between different sensor data, effectively solving the problem of insufficient prediction accuracy of reservoir operation parameters, and more accurately predicting the water level trend, seepage rate and other key parameters, providing reliable basis for scientific scheduling of reservoirs;

[0090] 2. Enhancing the accuracy of reservoir risk event identification and prediction: With the help of multi-sensor data fusion analysis, combined with attention mechanism and feature optimization algorithm, the potential rules behind the data are deeply mined, the identification and prediction accuracy of reservoir risk events are improved, the false alarm situation is significantly reduced, and the management personnel can more accurately master the reservoir safety situation;

[0091] 3. Optimizing model lightweight and early warning timeliness: To ensure the timeliness of early warning, the model is pruned and optimized to remove redundant parameters and structures, improving the lightweight level of the model. This enables the model to maintain a high prediction analysis frequency even with limited computing resources, allowing it to quickly process real-time monitoring data and issue timely warning signals, giving the reservoir management department more emergency response time and effectively reducing the loss caused by reservoir risk events. BRIEF DESCRIPTION OF DRAWINGS

[0092] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0093] The present application will be further described below in conjunction with the drawings.

[0094] Figure 1 A flowchart of a reservoir fusion analysis method based on multiple sensors according to Embodiment 1. DETAILED DESCRIPTION

[0095] The present application will be further described below in conjunction with the drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it. However, the embodiments are not limiting to the present application, and the technical features in the embodiments and the embodiments can be combined with each other without conflict.

[0096] The present application provides a reservoir fusion analysis method and system based on multiple sensors, which is used to solve the technical problem of how to realize accurate perception and comprehensive evaluation of reservoir operation state and timely discover potential safety hazards.

[0097] Embodiment 1:

[0098] The present application provides a reservoir fusion analysis method based on multiple sensors, which includes five steps of data set construction, data preprocessing, feature selection, model construction and reservoir monitoring.

[0099] Step S100 dataset construction: collect reservoir monitoring data, meteorological data and human activity data as original time series data, and construct a dataset based on the original time series data and the corresponding reservoir safety level.

[0100] In this embodiment, the monitoring data of 50 key monitoring points of the reservoir (including water level monitoring points, seepage monitoring points, stress and strain monitoring points, etc.) in the past three years is collected, covering parameters such as water level, seepage, stress and strain; meteorological data in the past three years, such as rainfall, air temperature, wind speed, etc.; topographic data of the reservoir basin, surrounding human activity data (such as reservoir surrounding engineering construction situation, agricultural irrigation water situation, etc.); design drawings, historical maintenance records and other materials of the reservoir dam are collected.

[0101] Step S200 data preprocessing: data preprocessing is performed on the dataset, and the original time series data is converted in format, abnormal value processing and missing value filling through data preprocessing to obtain a preprocessed dataset.

[0102] As a specific implementation of data preprocessing, this step includes the following operations:

[0103] (1) Data format conversion: remodeling the data, and adapting the data to the analysis modeling through remodeling;

[0104] (2) Abnormal value processing: for values exceeding the sensor range or not reasonable, instantaneous mutation values, data filtering is performed through z_score, and for continuous constant values, moving average method or linear interpolation method is used for processing;

[0105] (3) Missing value processing: missing data is filled by cubic spline interpolation method.

[0106] Step S300 feature selection: for the preprocessed dataset, the sample data in the dataset is used as feature data affecting the reservoir operation state, and feature selection is performed based on the mutual information between the feature data and the reservoir safety level and the information difference between the features, and the features with high mutual information and large information difference are selected as target features, and a sample set is constructed based on the target features and the corresponding reservoir safety level.

[0107] As a specific implementation of feature selection, this step includes the following operations:

[0108] L100, calculate the mutual information between the feature data and the reservoir safety level, and the calculation formula of the mutual information I(X i ;Y) is represented as:

[0109]

[0110] Where Y represents the reservoir safety level, X i represents the i-th feature, and p(xi y) represents a joint probability distribution, p(x i ) and p(y) are marginal probability distributions;

[0111] L200, calculate the information difference between features, the information difference between features is the redundancy between features, the feature X i and the feature X j (j≠i) information difference R(X i ) calculation formula is:

[0112]

[0113] Where |S| represents the total number of features;

[0114] L300, quantize the information difference, calculate the difference value of mutual information and redundancy, select the feature with the largest difference value, and the calculation formula is as follows:

[0115] Score(X i )=I(X i ;Y)-R(X i ),

[0116] Or, calculate the ratio of mutual information and redundancy, and the calculation formula is as follows:

[0117]

[0118] Where ε is a minimum value to avoid division by zero;

[0119] L400, screen features, initialize the feature set S as all candidate features, and iteratively select The selected results are added to the selected feature set S m Remove the selected features from the initialized feature set S, repeat steps L100-L400 until the preset feature quantity or score threshold is reached.

[0120] Meteorological observation data, surrounding human activity data and other features are directly affected by the reservoir operation state, which directly affects the reservoir safety analysis results. Considering the redundancy caused by the complex mapping relationship between features, the maximum correlation and minimum redundancy algorithm based on mutual information is used to select features. According to the algorithm, the mutual information between the features and the reservoir operation state (such as dam safety index, water level change trend, etc.) is calculated, and the information difference between each feature and the remaining features is calculated, and the most valuable features for reservoir safety analysis are selected.

[0121] Step S400 model construction: based on the RNN network, the CNN network, the attention mechanism and the conditional probability distribution layer, a reservoir monitoring model is constructed, and a dynamic regularization strategy is introduced, the reservoir monitoring model is trained based on the sample set, and a trained reservoir monitoring model is obtained, wherein the reservoir monitoring model is used for feature extraction on input time series data through the RNN network and the CNN network, the features output by the RNN network are weighted based on the attention mechanism, and the features output by the CNN network and the weighted features output by the attention mechanism are used for reservoir operation state prediction analysis through the conditional probability distribution layer, and the reservoir safety level and the corresponding probability value are output.

[0122] In this embodiment, the reservoir monitoring model includes an RNN network module, a CNN network module, an attention mechanism module, a splicing layer, a full connection layer and a conditional probability distribution layer.

[0123] The RNN network module takes time series data as input, and is used for extracting long-term dependence features of time series data based on a bidirectional LSTM network model, and outputs a feature vector.

[0124] As a specific implementation of the RNN network module, the structure selection is: a bidirectional LSTM (BiLSTM) is used as the basic structure to solve the gradient disappearance problem of traditional RNN and capture forward and backward time series dependence. Input: 3D data (sample number x time step x feature number), such as time series of reservoir water level, seepage, and meteorological data. Output: hidden state of each time step (dimension hidden_size), and the final output is the hidden state of the last time step (or the context vector weighted by the attention mechanism). Function: extract long-term dependence features of time series data, such as the correlation between water level change trend and rainfall.

[0125] The CNN network module takes time series data as input, and is used for local feature extraction of time series data based on convolution layer and max pooling layer, and outputs a feature vector.

[0126] As a specific implementation of the CNN network module, the structure is: 1D-CNN layer (convolution kernel size = 3, step = 1, channel number = 64) + MaxPooling layer (pooling size = 2). Input: share the original 3D data with RNN, and independently process to extract local time series patterns (such as peak features of sudden rainfall). Output: feature map (dimension sample number x time step / 2 x 64), which is spliced with RNN output after flattening. Function: capture local time series features, and complement global dependence of RNN.

[0127] The attention mechanism module is connected with the RNN network module, and is used for attention weighting on the feature vector output by the RNN network module, and outputs a weighted feature vector.

[0128] As a specific implementation of the attention focusing module, its structure is: self-attention mechanism (Scaled Dot-Product Attention), the query vector (Q), the key vector (K), and the value vector (V) are all from the RNN hidden state. Input: the hidden state of all time steps of the RNN (H = [h1, h2,..., h_T]). Output: the weighted context vector c = Σ (α_i * h_i), where α_i is the attention weight (calculated by Softmax). Function: dynamically focusing on key time steps (such as extreme weather event corresponding to water level mutation point).

[0129] The concatenation layer is used to fuse the output of the CNN network module and the output of the attention mechanism module, and output the fused features.

[0130] The full connection layer is connected with the concatenation layer, which is used to linearly map the fused features and output to the conditional probability distribution layer.

[0131] The conditional probability distribution layer is used to receive the feature vector output by the full connection layer, and based on the feature vector, the reservoir operation state prediction analysis is carried out by the Softmax function, and the reservoir safety level and the corresponding probability value are output.

[0132] As a specific implementation of the conditional probability layer, its structure is: Softmax output layer, output dimension = number of categories (such as dam safety level: safe / warning / danger). Input: the concatenation of the context vector c output by the attention module and the CNN feature. Output: conditional probability distribution P (Y|X), indicating the probability of each safety level given the input feature. Function: quantify uncertainty, support risk decision (such as triggering inspection when high probability warning).

[0133] Wherein, when training the reservoir monitoring model based on the sample set, the target features in the sample set are subjected to Z-score standardization processing to obtain standardized target features, and the standardized target features are divided into fixed-length time series samples based on a predefined sliding window, and the final sample set is constructed based on the fixed-length time series samples and the corresponding reservoir safety level, and the reservoir monitoring model is trained based on the final sample set. During the training process, the weight matrix of the full connection layer is subjected to dynamic Dropout and L2 regularization constraint based on the dynamic regularization strategy to prevent overfitting;

[0134] Correspondingly, the total loss L of the reservoir monitoring model total is represented as:

[0135] L total = L CE + λ·L MLE + γ·L L2 ,

[0136] where λ and γ are hyperparameters, L L2 denotes the regularization term;

[0137] L CE denotes the cross-entropy loss, calculated as follows:

[0138]

[0139] y i is the one-hot encoding of the true label, p i is the probability output by the reservoir monitoring model;

[0140] L MLE denotes the maximum likelihood loss, calculated as follows:

[0141] L MLE = -log P(X| θ),

[0142] where X denotes the input target features, and θ are the parameters of the reservoir monitoring model.

[0143] As a specific implementation of model training, it involves data preprocessing, forward propagation, loss function, and backpropagation and optimization.

[0144] Data preprocessing involves standardization and sliding window. Standardization: Z-score standardization of water level, seepage, and other features. Sliding window: dividing continuous monitoring data into fixed-length (e.g., 24 hours) time series samples.

[0145] For forward propagation, the CNN branch: original data → 1D-CNN → MaxPooling → Flatten → feature vector f_cnn; RNN branch: original data → BiLSTM → hidden state sequence H → attention mechanism → context vector c; fusion: concatenate f_cnn and c → fully connected layer → Softmax output P(Y|X).

[0146] The loss function includes cross-entropy loss and likelihood function estimation loss. Cross-entropy loss measures the difference between predicted probability distribution and true label. Likelihood function estimation loss: based on maximum likelihood estimation (MLE), optimize model parameters to maximize the joint probability of observed data.

[0147] During backpropagation and optimization, the optimizer used is Adam (learning rate = 0.001, β1 = 0.9, β1 = 0.999). Gradient clipping: prevent RNN gradient explosion (threshold = 1.0). Dynamic adjustment: after every 10 training rounds, if the validation loss does not decrease, the learning rate is reduced to 0.1 times.

[0148] In this embodiment, the conditional probability distribution application scenario is: directly modeling P(Y|X) at the output layer, supporting risk quantification (such as calculating the early warning probability threshold), and providing interpretability, for example, “the dam danger probability is 15% under the current rainfall”.

[0149] The likelihood function estimation application scenario is: optimizing model parameters in the training process to maximize the probability of observed data (such as historical water level records), enhancing the model's fitting ability to data distribution (such as adapting to different seasonal water level patterns), and combining with cross-entropy loss to improve classification robustness (for example, reducing false positive rate).

[0150] The model is a multi-modal fusion that combines the local features of CNN and the global dependence of RNN to improve feature expression capability, and through attention mechanism and dynamic regularization, it adapts to data distribution changes (such as sudden floods), conditional probability output supports decision threshold optimization (such as adjusting the inspection frequency according to the early warning probability), and the likelihood function estimation provides model confidence to assist engineers in diagnosing abnormalities (such as low likelihood value samples marked as potential risk points).

[0151] Step S500 reservoir monitoring: For the target reservoir to be monitored, collect reservoir monitoring data, meteorological data, and human activity data as time series data to be measured, perform data preprocessing on the time series data to be measured, format conversion, outlier processing, and missing value filling on the time series data to be measured through data preprocessing, input the preprocessed time series data to be measured into the trained reservoir monitoring model, and output the reservoir safety level and corresponding probability value of the target reservoir through the reservoir monitoring model.

[0152] The method of this embodiment integrates multiple types of sensors, including water level sensors, rainfall sensors, seepage sensors, stress and strain sensors, etc. These sensors are distributed in key locations such as reservoirs, dams, and surrounding areas, and collect various data during the operation of the reservoir. The collected data is characterized by multiple sources, heterogeneous, massive, etc. Advanced data fusion algorithms are used to preprocess, feature extraction, and correlation analysis of different sensor data, eliminate redundancy and conflicts between data, mine hidden association information behind the data, and form a unified and accurate reservoir operation state information model.

[0153] Based on this information model, the technology can realize multiple key functions. It can monitor the reservoir water level, rainfall, seepage, dam stress and strain, and other key parameters in real time, and when the parameters exceed the preset safety threshold, it can timely issue an early warning signal to provide decision-making basis for reservoir management personnel and effectively prevent safety accidents. Through deep mining and analysis of historical data, the future water level change trend of the reservoir and the dam structure health condition can be predicted to provide scientific guidance for the operation, maintenance and repair of the reservoir.

[0154] Embodiment 2:

[0155] The application discloses a reservoir fusion analysis system based on multiple sensors, which comprises a data set construction module, a data preprocessing module, a feature screening module, a model construction module and a reservoir monitoring module.

[0156] The data set construction module is used for collecting reservoir monitoring data, meteorological data and human activity data as original time series data, and constructing a data set based on the original time series data and corresponding reservoir safety levels.

[0157] In the embodiment, monitoring data of 50 key monitoring points (including water level monitoring points, seepage monitoring points, stress and strain monitoring points and the like) of the reservoir in the past three years are collected, covering parameters such as water level, seepage and stress and strain; meteorological data in the past three years such as rainfall, air temperature and wind speed are collected; terrain data of the reservoir basin, surrounding human activity data (such as reservoir surrounding engineering construction conditions, agricultural irrigation water conditions and the like) are collected; design drawings, historical maintenance records and the like of the reservoir dam are collected.

[0158] The data preprocessing module is used for performing data preprocessing on the data set, and performing format conversion, abnormal value processing and missing value filling on the original time series data through the data preprocessing to obtain a preprocessed data set.

[0159] As a specific implementation of the data preprocessing module, the module is used for performing the following operations:

[0160] (1) Data format conversion: the data is subjected to reshaping processing, and the data is adapted to analysis modeling through the reshaping processing;

[0161] (2) Abnormal value processing: for values beyond the sensor range or unreasonable, instantaneous mutation values, data screening is performed through z_score, and for continuous constants, a moving average method or a linear interpolation method is adopted for processing;

[0162] (3) Missing value processing: missing data is filled through a cubic spline interpolation method.

[0163] The feature screening module is used for performing the following operations: for the preprocessed data set, sample data in the data set are taken as feature data affecting the reservoir operation state, feature screening is performed based on mutual information between the feature data and the reservoir safety levels and information differences between the features, features with high mutual information and large information differences are screened as target features, and a sample set is constructed based on the target features and corresponding reservoir safety levels.

[0164] As a specific implementation of the feature screening module, the module is used for performing the following operations:

[0165] L100, mutual information between the feature data and the reservoir safety levels is calculated, mutual information I (X iThe calculation formula of Y is:

[0166]

[0167] wherein Y represents the reservoir safety level, X i represents the i-th feature, p(x i ,y) represents the joint probability distribution, p(x i ) and p(y) are the marginal probability distributions;

[0168] L200, calculate the information difference between features, the information difference between features is the redundancy between features, the feature X i and the feature X j (j≠i) The calculation formula of the information difference R(X i ) is:

[0169]

[0170] wherein |S| represents the total number of features;

[0171] L300, quantize the information difference, calculate the difference value of mutual information and redundancy, select the feature with the maximum difference value, and the calculation formula is as follows:

[0172] Score(X i )=I(X i ;Y)-R(X i ),

[0173] Alternatively, calculate the ratio of mutual information and redundancy, and the calculation formula is as follows:

[0174]

[0175] wherein ε is a minimum value, which avoids division by zero;

[0176] L400, screen features, initialize the feature set S as all candidate features, and iteratively select the selected features are added to the selected feature set S m , the selected features are removed from the initialized feature set S, and steps L100-L400 are repeated until a preset number of features or a score threshold is reached.

[0177] Meteorological observation data, surrounding human activity data and other features are directly affected by the reservoir operation state, which directly affects the reservoir safety analysis result. Considering the redundancy caused by the complex mapping relationship between features, the maximum correlation and minimum redundancy algorithm based on mutual information is used to select the features. According to the algorithm, the mutual information between the features and the reservoir operation state (such as dam safety index, water level change trend, etc.) and the information difference between each feature and the remaining features are calculated, and the features most valuable to the reservoir safety analysis are selected.

[0178] The model construction module is configured to construct a reservoir monitoring model based on an RNN network, a CNN network, an attention mechanism, and a conditional probability distribution layer, introduce a dynamic regularization strategy, and train the reservoir monitoring model based on a sample set to obtain a trained reservoir monitoring model. The reservoir monitoring model is configured to extract features of input time series data through the RNN network and the CNN network, weight the features output by the RNN network based on the attention mechanism, and perform reservoir operation state prediction analysis through the conditional probability distribution layer based on the features output by the CNN network and the weighted features output by the attention mechanism, and output a reservoir safety level and a corresponding probability value.

[0179] In this embodiment, the reservoir monitoring model includes an RNN network module, a CNN network module, an attention mechanism module, a concatenation layer, a full connection layer, and a conditional probability distribution layer.

[0180] The RNN network module takes time series data as input and is configured to extract long-term dependence features of the time series data based on a bidirectional LSTM network model and output a feature vector.

[0181] As a specific implementation of the RNN network module, the structure is selected as follows: a bidirectional LSTM (BiLSTM) is used as the basic structure to solve the gradient disappearance problem of the traditional RNN and capture forward and backward time series dependencies. Input: 3D data (number of samples x time step x number of features), such as time series of reservoir water level, seepage, and meteorological data. Output: hidden state of each time step (dimension hidden_size), and the final output is the hidden state of the last time step of the sequence (or the context vector weighted by the attention mechanism). Function: extract long-term dependence features of time series data, such as the correlation between water level change trend and rainfall.

[0182] The CNN network module takes time series data as input and is configured to extract local features of the time series data based on a convolution layer and a max pooling layer and output a feature vector.

[0183] As a specific implementation of the CNN network module, the structure is as follows: 1D-CNN layer (convolution kernel size = 3, step = 1, number of channels = 64) + MaxPooling layer (pooling size = 2). Input: share the original 3D data with RNN and independently process to extract local time series patterns (such as peak features of sudden rainfall). Output: feature map (dimension sample number x time step / 2 x 64), which is concatenated with the RNN output after flattening. Function: capture local time series features and complement global dependencies of RNN.

[0184] The attention mechanism module is connected with the RNN network module, and is configured to perform attention weighting on a feature vector output by the RNN network module, and output a weighted feature vector.

[0185] As a specific implementation of the attention focusing module, the structure is: self-attention mechanism (Scaled Dot-Product Attention), the query vector (Q), the key vector (K), and the value vector (V) are all from the RNN hidden state. Input: the hidden state of all time steps of the RNN (H = [h1, h2,..., h_T]). Output: the weighted context vector c = Σ (α_i * h_i), where α_i is the attention weight (calculated by Softmax). Function: dynamically focusing on key time steps (such as extreme weather event corresponding to water level mutation point).

[0186] The concatenation layer is configured to perform feature fusion on the output of the CNN network module and the output of the attention mechanism module, and output a fused feature.

[0187] The fully connected layer is connected with the concatenation layer, and is configured to perform linear mapping on the fused feature and output to the conditional probability distribution layer.

[0188] The conditional probability distribution layer is configured to receive the feature vector output by the fully connected layer, and perform reservoir operation state prediction analysis based on the feature vector by using a Softmax function, and output a reservoir safety level and a corresponding probability value.

[0189] As a specific implementation of the conditional probability layer, the structure is: a Softmax output layer, the output dimension = the number of categories (such as dam safety level: safe / warning / danger). Input: the concatenation of the context vector c output by the attention module and the CNN feature. Output: conditional probability distribution P (Y|X), indicating the probability of each safety level given the input feature. Function: quantify uncertainty, support risk decision (such as triggering inspection when high probability warning).

[0190] In the model training of the reservoir monitoring model based on the sample set, the target feature in the sample set is subjected to Z-score standardization processing to obtain a standardized target feature, and the standardized target feature is divided into time series samples of a fixed length based on a predefined sliding window. Based on the time series samples of the fixed length and the corresponding reservoir safety level, a final sample set is constructed, and the reservoir monitoring model is trained based on the final sample set. During the training process, the weight matrix of the fully connected layer is subjected to dynamic Dropout and L2 regularization constraint based on a dynamic regularization strategy, so as to prevent overfitting.

[0191] Correspondingly, the total loss L of the reservoir monitoring model total is expressed as:

[0192] L total = L CE + λ·L MLE + γ·L L2 ,

[0193] where λ and γ are hyperparameters, L L2 represents the regularization term;

[0194] L CE represents the cross-entropy loss, calculated as follows:

[0195]

[0196] y i is the one-hot encoding of the true label, p i is the probability output by the reservoir monitoring model;

[0197] L MLE represents the maximum likelihood loss, calculated as follows:

[0198] L MLE = -log P(X|θ),

[0199] where X represents the input target features, and θ are the parameters of the reservoir monitoring model.

[0200] As a specific implementation of model training, it involves data preprocessing, forward propagation, loss function, and backpropagation and optimization.

[0201] Data preprocessing involves standardization and sliding window. Standardization: Z-score normalization of water level, seepage, and other features. Sliding window: dividing continuous monitoring data into fixed-length (e.g., 24 hours) time series samples.

[0202] For forward propagation, the CNN branch: original data → 1D-CNN → MaxPooling → Flatten → feature vector f_cnn; RNN branch: original data → BiLSTM → hidden state sequence H → attention mechanism → context vector c; fusion: concatenate f_cnn and c → fully connected layer → Softmax output P(Y|X).

[0203] The loss function includes cross-entropy loss and likelihood function estimation loss. Cross-entropy loss measures the difference between predicted probability distribution and true label. Likelihood function estimation loss: based on maximum likelihood estimation (MLE), optimize model parameters to maximize the joint probability of observed data.

[0204] The optimizer used in back propagation and optimization is: Adam (learning rate = 0.001, beta1 = 0.9, beta1 = 0.999). Gradient clipping: prevent RNN gradient explosion (threshold = 1.0), dynamic adjustment: after every 10 training rounds, if the validation loss does not decrease, the learning rate is reduced to 0.1 times.

[0205] In this embodiment, the conditional probability distribution application scenario is: directly modeling P(Y|X) in the output layer, supporting risk quantification (such as calculating the warning probability threshold), providing interpretability, for example, "under the current rainfall, the dam danger probability is 15%".

[0206] The likelihood function estimation application scenario is: optimizing the model parameters in the training process to maximize the probability of observed data (such as historical water level records), enhancing the model's fitting ability to the data distribution (such as adapting to different seasonal water level patterns), and combining with cross-entropy loss to improve classification robustness (for example, reduce false positive rate).

[0207] The model is a multi-modal fusion that combines the local features of CNN and the global dependence of RNN to improve feature expression capability, and through attention mechanism and dynamic regularization, it adapts to data distribution changes (such as sudden floods), conditional probability output supports decision threshold optimization (such as adjusting the inspection frequency according to the warning probability), and the likelihood function estimation provides model confidence to assist engineers in diagnosing abnormalities (such as low likelihood value samples are marked as potential risk points).

[0208] The reservoir monitoring module is configured to perform the following operations: for a target reservoir to be monitored, collecting reservoir monitoring data, meteorological data, and human activity data as to-be-tested time series data, performing data preprocessing on the to-be-tested time series data, performing format conversion, outlier processing, and missing value filling on the to-be-tested time series data through the data preprocessing, inputting the preprocessed to-be-tested time series data into the trained reservoir monitoring model, and predicting and outputting the reservoir safety level and the corresponding probability value of the target reservoir through the reservoir monitoring model.

[0209] The system of the embodiment can perform the method disclosed in embodiment 1 to realize reservoir monitoring.

[0210] The multi-sensor-based reservoir fusion analysis method and system provided by the present application are described in detail above, and the principles and implementation modes of the present application are described by applying specific examples; the above embodiment is only used to help understand the method and core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed; in view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A multi-sensor based reservoir fusion analysis method, characterized by, The method comprises the following steps: Data set construction: collecting reservoir monitoring data, meteorological data and human activity data as original time series data, and constructing a data set based on the original time series data and the corresponding reservoir safety level; Data preprocessing: data preprocessing is performed on the data set, and the original time series data is converted in format, abnormal value processing and missing value filling through data preprocessing to obtain a preprocessed data set; Feature selection: for the preprocessed data set, the sample data in the data set is used as feature data affecting the reservoir operation state, and feature selection is performed based on the mutual information between the feature data and the reservoir safety level and the information difference between the features, and features with high mutual information and large information difference are selected as target features, and a sample set is constructed based on the target features and the corresponding reservoir safety level; Model construction: a reservoir monitoring model is constructed based on an RNN network, a CNN network, an attention mechanism and a conditional probability distribution layer, a dynamic regularization strategy is introduced, and the reservoir monitoring model is trained based on the sample set to obtain a trained reservoir monitoring model, wherein the reservoir monitoring model is used for feature extraction on input time series data through an RNN network and a CNN network, weighting of features output by the RNN network based on an attention mechanism, and reservoir operation state prediction and analysis based on the features output by the CNN network and the weighted features output by the attention mechanism through a conditional probability distribution layer, and output of a reservoir safety level and a corresponding probability value; Reservoir monitoring: for a target reservoir to be monitored, reservoir monitoring data, meteorological data and human activity data are collected as test time series data, data preprocessing is performed on the test time series data, the test time series data is converted in format, abnormal value processing and missing value filling are performed through data preprocessing, the preprocessed test time series data is input into the trained reservoir monitoring model, and the reservoir safety level and the corresponding probability value of the target reservoir are output by the reservoir monitoring model.

2. The multi-sensor based reservoir fusion analysis method of claim 1, wherein, Data preprocessing comprises the following steps: Data format conversion: remodeling the data to adapt to the analysis modeling; Abnormal value processing: for values exceeding the sensor range or being unreasonable, instantaneous mutation values, data is screened through z_score, and for continuous constants, moving average method or linear interpolation method is used for processing; Missing value processing: missing data is filled by cubic spline interpolation method.

3. The multi-sensor based reservoir fusion analysis method of claim 1, wherein, Feature selection Comprises the following steps: L100, mutual information between the calculated feature data and the reservoir safety level, the calculation formula of mutual information I(X i ; Y) is represented as: where Y represents the reservoir safety level, X i represents the i-th feature, p(x i ,y) represents the joint probability distribution, p(x i ) and p(y) are the marginal probability distributions; L200, calculate the information difference between features, the information difference between features is the redundancy between features, feature X i and feature X j (j≠i) between information difference R(X i ) The formula is: Wherein, |S| represents the total number of features; L300, the information difference is quantified, the difference value of mutual information and redundancy is calculated, the feature with the largest difference value is selected, and the calculation formula is as follows: Score(X i ) = I(X i ; Y) - R(X i ), Or, the ratio of mutual information and redundancy is calculated, and the calculation formula is as follows: Wherein, ε is a minimum value, to avoid division by zero; L400, select a feature from the candidate feature set S, and add it to the selected feature set S L300, add the selected result to the selected feature set S m L200, remove the selected feature from the initialized feature set S, repeat steps L100-L400 until a pre-set number or score threshold is reached.

4. The multi-sensor based reservoir fusion analysis method of claim 1, wherein, The reservoir monitoring model comprises an RNN network module, a CNN network module, an attention mechanism module, a splicing layer, a full connection layer and a conditional probability distribution layer; The RNN network module takes time series data as input, and is used for extracting long-term dependence features of the time series data based on a bidirectional LSTM network model, and outputs a feature vector; The CNN network module takes time series data as input, and is configured to perform local feature extraction on the time series data based on a convolutional layer and a max-pooling layer, and output a feature vector; The attention mechanism module is connected to the RNN network module, and is configured to perform attention weighting on the feature vector output by the RNN network module, and output a weighted feature vector; The concatenation layer is configured to perform feature fusion on the output of the CNN network module and the output of the attention mechanism module, and output a fused feature; The fully connected layer is connected to the concatenation layer, and is configured to perform linear mapping on the fused feature, and output to a conditional probability distribution layer; The conditional probability distribution layer is configured to receive the feature vector output by the fully connected layer, and perform reservoir operation state prediction analysis based on the feature vector by using a Softmax function, and output a reservoir safety level and a corresponding probability value; In the model training of the reservoir monitoring model based on the sample set, the target features in the sample set are subjected to Z-score standardization processing to obtain standardized target features, and the standardized target features are divided into time series samples of a fixed length based on a predefined sliding window, and a final sample set is constructed based on the time series samples of the fixed length and corresponding reservoir safety levels, and the reservoir monitoring model is trained based on the final sample set, and in the training process, the weight matrix of the fully connected layer is subjected to dynamic Dropout and L2 regularization constraint based on a dynamic regularization strategy to prevent overfitting. Correspondingly, the total loss L of the reservoir monitoring model total is represented as: L total = L CE + λ · L MLE + γ · L L2 , where λ and γ are hyperparameters, L L2 denotes the regularization term; L CE represents the cross-entropy loss, and the calculation formula is as follows: y i is a one-hot encoding of the true label, p i is the probability output by the reservoir monitoring model; L MLE represents the maximum likelihood loss, and the calculation formula is as follows: L MLE = -log P(X | θ), In the formula, X represents the input target feature, and θ represents the parameter of the reservoir monitoring model.

5. A multi-sensor based reservoir fusion analysis system, characterized by, The reservoir monitoring system comprises a data set construction module, a data preprocessing module, a feature screening module, a model construction module, and a reservoir monitoring module. The data set construction module is configured to perform the following operations: collecting reservoir monitoring data, meteorological data, and human activity data as original time series data, and constructing a data set based on the original time series data and corresponding reservoir safety levels; The data preprocessing module is configured to perform the following operations: data preprocessing on the data set, and format conversion, outlier processing, and missing value filling on the original time series data through data preprocessing to obtain a preprocessed data set; The feature screening module is configured to perform the following operations: for the preprocessed data set, the sample data in the data set are taken as feature data affecting the reservoir operation state, and feature screening is performed based on the mutual information between the feature data and the reservoir safety level and the information difference between the features, and features with high mutual information and large information difference are selected as target features, and a sample set is constructed based on the target features and corresponding reservoir safety levels; The reservoir monitoring module is configured to perform the following operations: receiving the sample set, and performing reservoir operation state prediction analysis on the sample set based on the reservoir monitoring model to output a reservoir safety level and a corresponding probability value. The model construction module is configured to construct a reservoir monitoring model based on an RNN network, a CNN network, an attention mechanism, and a conditional probability distribution layer, introduce a dynamic regularization strategy, and train the reservoir monitoring model based on a sample set to obtain a trained reservoir monitoring model, wherein the reservoir monitoring model is configured to extract features of input time series data through the RNN network and the CNN network, weight features output by the RNN network based on the attention mechanism, and perform reservoir operation state prediction analysis through the conditional probability distribution layer based on features output by the CNN network and the weighted features output by the attention mechanism, and output a reservoir safety level and a corresponding probability value. The reservoir monitoring module is configured to collect reservoir monitoring data, meteorological data, and human activity data as to-be-tested time series data for a target reservoir to be monitored, perform data preprocessing on the to-be-tested time series data, convert the format of the to-be-tested time series data, process outliers, and fill in missing values, input the preprocessed to-be-tested time series data into the trained reservoir monitoring model, and output a reservoir safety level and a corresponding probability value of the target reservoir through the reservoir monitoring model.

6. The multi-sensor based reservoir fusion analysis system of claim 5, wherein, The data preprocessing module is configured to perform the following operations: Data format conversion: remodeling the data to adapt to the analysis modeling; Outlier processing: for values that exceed the sensor range or are not reasonable, instantaneous mutation values, data is filtered through z_score, and for continuous constants, a moving average method or linear interpolation method is used for processing; Missing value processing: missing data is filled in through a cubic spline interpolation method.

7. The multi-sensor based reservoir fusion analysis system of claim 5, wherein, The feature screening module is configured to perform the following operations: L100, mutual information between the calculated feature data and the reservoir safety level, the calculation formula of mutual information I(X i ; Y) is represented as: where Y represents the reservoir safety level, X i represents the ith feature, p(x i ,y) represents the joint probability distribution, p(x i ) and p(y) are the marginal probability distributions; L200, calculate the information difference between features, the information difference between features is the redundancy between features, feature X i and feature X j (j≠i) between the information difference R(X i ) The formula is: Wherein, |S| represents the total number of features; L300, quantifying the information difference, calculating the difference value of mutual information and redundancy, selecting the feature with the largest difference value, and the calculation formula is as follows: Score(X i ) = I(X i ; Y) - R(X i ), Or, calculating the ratio of mutual information and redundancy, and the calculation formula is as follows: Wherein, ε is a minimum value to avoid division by zero; L400, select a feature from the candidate feature set S, and add the selected feature to the selected feature set S L300, add the selected result to the selected feature set S m L200, remove the selected feature from the initialized feature set S, and repeat steps L100-L400 until a pre-set number or score threshold of features is reached.

8. The multi-sensor based reservoir fusion analysis system of claim 5, wherein, The reservoir monitoring model comprises an RNN network module, a CNN network module, an attention mechanism module, a concatenation layer, a full connection layer, and a conditional probability distribution layer; The RNN network module takes time series data as input, and is configured to extract long-term dependence features of the time series data based on a bidirectional LSTM network model and output a feature vector; The CNN network module takes time series data as input, and is configured to extract local features of the time series data based on a convolution layer and a max pooling layer and output a feature vector; The attention mechanism module is connected with the RNN network module and is configured to perform attention weighting on the feature vector output by the RNN network module and output a weighted feature vector; The concatenation layer is configured to fuse the output of the CNN network module and the output of the attention mechanism module and output fused features; The full connection layer is connected with the concatenation layer and is configured to perform linear mapping on the fused features and output to the conditional probability distribution layer. The conditional probability distribution layer is configured to receive the feature vector of the full connection layer output, perform reservoir operation state prediction analysis based on the feature vector through a Softmax function, and output a reservoir safety level and a corresponding probability value; In the model training of the reservoir monitoring model based on the sample set, the target features in the sample set are subjected to Z-score standardization processing to obtain standardized target features, the standardized target features are divided into time series samples of a fixed length based on a predefined sliding window, a final sample set is constructed based on the time series samples of the fixed length and corresponding reservoir safety levels, and the reservoir monitoring model is trained based on the final sample set; in the training process, the weight matrix of the full connection layer is subjected to dynamic Dropout and L2 regularization constraint based on a dynamic regularization strategy to prevent overfitting. Correspondingly, the total loss L of the reservoir monitoring model total is represented as: L total = L CE + λ · L MLE + γ · L L2 , where λ and γ are hyperparameters, L L2 denotes the regularization term; L CE represents the cross-entropy loss, and the calculation formula is as follows: y i is a one-hot encoding of the true label, p i is the probability output by the reservoir monitoring model; L MLE represents the maximum likelihood loss, and the calculation formula is as follows: L MLE = -log P(X | 0), In the formula, X represents the input target feature, and θ is a parameter of the reservoir monitoring model.

Images