A method for constructing a multi-source monitoring and safety early warning system for canal projects
By constructing a multi-source monitoring and safety early warning system for canal projects, and employing a CNN-LSTM fusion model and a Bayesian network, efficient fusion of multi-source data and accurate feature mining were achieved. This solved the problems of incomplete data coverage and false alarms/missed alarms in traditional canal project monitoring and early warning systems, and improved the timeliness of early warnings and the efficiency of emergency response.
Patent Information
- Application Number
- CN202611266201.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional canal engineering monitoring and early warning systems rely on a single type of sensor, resulting in incomplete data coverage, inability to identify multi-source coupling risks, serious false alarms and missed alarms, lack of emergency response support, and difficulty in adapting to complex and ever-changing service environments.
A multi-source monitoring and safety early warning system for canal projects was constructed. Through data acquisition, preprocessing, feature extraction and fusion, health assessment and dynamic early warning, a CNN-LSTM fusion model, Bayesian network and dynamic threshold mechanism were adopted to achieve multi-source data fusion and accurate feature mining, and generate targeted disposal suggestions.
It has achieved efficient integration and in-depth utilization of multi-source data, accurately identified risk characteristics, improved the timeliness and reliability of early warning, reduced false alarms and missed alarms, improved emergency response efficiency, and realized closed-loop management of "early warning-response".
Smart Images

Figure CN123117779A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of canal engineering safety monitoring technology, specifically to a method for constructing multi-source monitoring and safety early warning systems for canal engineering. Background Technology
[0002] As a core hub for national strategic transportation infrastructure and water resource allocation, the long-term stable operation of canal projects, including waterways, dikes, and locks, is directly related to regional economic development, smooth material transportation, and the safety of life and property of people along the canals. However, canal projects are generally characterized by large spans, long service periods, and complex geological conditions, and are constantly operating in a dynamically changing environment, facing multiple risks and challenges. On the one hand, extreme meteorological and hydrological events such as torrential rains, strong winds, and sudden changes in water levels can easily lead to safety hazards such as abnormal seepage, sudden displacement, and crack expansion in the engineering structures. On the other hand, human factors such as ship traffic loads and maintenance disturbances further exacerbate the coupled risks of multiple disasters (piping, dike breaches, structural failures, etc.), posing stringent requirements for ensuring the safety of the projects.
[0003] Currently, traditional safety management models for canal projects are no longer adequate for the complex needs of risk prevention and control, mainly due to the following prominent problems: Existing monitoring relies heavily on localized data collection from single-type sensors, focusing only on individual physical indicators such as deformation and seepage, neglecting the value of unstructured data such as inspection records and historical disaster cases. This results in incomplete data coverage and insufficient correlation, failing to fully reflect the multi-source coupling characteristics of engineering risks and hindering early risk identification. Multi-source monitoring data suffers from format heterogeneity, spatiotemporal misalignment, and outlier interference, lacking a systematic preprocessing mechanism, making data quality difficult to guarantee. Furthermore, traditional feature extraction methods often employ simple statistical analysis or single algorithms, failing to effectively capture spatial correlation patterns and long-term evolutionary characteristics in monitoring data, leading to low accuracy in risk feature identification and a high likelihood of missed or false alarms. Existing early warning systems often use fixed thresholds based on engineering design specifications, failing to consider the dynamic changes in engineering risks under different operating conditions such as flood season, dry season, and extreme weather, resulting in missed or false alarms in early warning results, making them unsuitable for complex and ever-changing service environments. Traditional early warning systems only output simple risk levels, lacking accurate identification of key risk sources and systematic analysis of risk transmission paths. Furthermore, they lack a standardized emergency response knowledge base, making it difficult for managers to quickly formulate targeted response measures. This results in delayed emergency response and an inability to achieve closed-loop management of "early warning-response".
[0004] Against this backdrop, to overcome the limitations of traditional monitoring and early warning models, there is an urgent need to construct a safety early warning system capable of integrating multi-source data, accurately mining risk characteristics, dynamically adapting to changes in operating conditions, and providing full-process decision support. This system, developed through technological innovation, aims to achieve early identification, accurate early warning, and efficient handling of safety risks in canal engineering, providing strong support for the long-term stable operation of the project. This is the core motivation and practical significance of this system's development. Among existing patents, patent CN12032338A provides a multi-source fusion-based intelligent monitoring method and system for multi-dimensional diseases of waterway bank protection, and patent CN120909202A provides an intelligent monitoring system and method for water conservancy projects. While these provide relevant ideas, they lack specific system support for addressing the problems of single-dimensional monitoring data, inaccurate feature extraction, poor early warning adaptability, and lack of decision support in canal monitoring. Summary of the Invention
[0005] The purpose of this invention is to provide a method for constructing a multi-source monitoring and safety early warning system for canal projects, which realizes multi-source data fusion, precise feature mining, dynamic early warning evaluation and key risk handling suggestions for canal projects, improves the timeliness, accuracy and reliability of safety early warning for canal projects, and provides technical support for ensuring the safety of canal projects in service.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A method for constructing a multi-source monitoring and safety early warning system for canal projects, the system comprising a data acquisition layer, a data preprocessing layer, a feature extraction and fusion layer, a health evaluation and early warning layer and an output layer connected in series, wherein the output layer and the feature extraction and fusion layer realize bidirectional data interaction for model parameter optimization feedback;
[0008] The construction method includes the following steps:
[0009] S1, Multi-source data acquisition: A multi-dimensional monitoring system is constructed through the data acquisition layer to collect multi-source data of the core structures of the canal project. The multi-source data includes real-time sensor monitoring data, inspection record data, and historical disaster case data. The sensors include displacement gauges, strain gauges, piezometers, water level gauges, and anemometers, which are used to monitor deformation, stress and strain, seepage, and hydrological and meteorological indicators, respectively. The inspection record data includes text, image, and video data.
[0010] S2, Data Preprocessing: The multi-source data obtained in step S1 is sequentially subjected to outlier removal, spatiotemporal alignment, and format standardization through the data preprocessing layer to obtain standardized data; the outlier removal adopts a combination of the 3σ criterion and the isolated forest algorithm.
[0011] S3, Multi-source data feature extraction and fusion: The standardized data obtained in step S2 is input into the CNN-LSTM fusion model through the feature extraction and fusion layer for feature extraction and fusion, resulting in a multi-source data fusion feature vector. The CNN-LSTM fusion model includes a parallel CNN branch and an LSTM branch. The CNN branch is used to extract spatial features, and it introduces an SE attention mechanism to adaptively weight the extracted spatial features. The LSTM branch is used to extract temporal features, and it introduces residual connections to add the output of the previous LSTM layer to the input of the next LSTM layer. The CNN-LSTM fusion model also includes a feature fusion layer, which concatenates the spatial features output by the CNN branch with the temporal features output by the LSTM branch to output a multi-source data fusion feature vector.
[0012] S4, Health Assessment and Dynamic Early Warning: Through the health assessment and early warning layer, based on the fusion feature vector obtained in step S3, a combined weight evaluation system and a dynamic threshold mechanism are established to quantify the health status level of the engineering structure and determine the early warning level. The combined weight evaluation system is constructed by combining the Analytic Hierarchy Process (AHP) and the entropy weight method. The dynamic threshold mechanism uses the sliding window method to achieve adaptive updating of the early warning threshold based on the statistical characteristics of historical health score data.
[0013] S5, Risk Identification and Decision Output: Through the output layer, based on the evaluation results of step S4, a risk transmission model is constructed using a Bayesian network to identify key risk sources and output graded early warning information and targeted handling suggestions.
[0014] Furthermore, the real-time monitoring data from the sensors mentioned in step S1 is transmitted via a 5G / BeiDou communication module; the inspection record data is entered via a mobile app or automatically imported via drone inspection; and the historical disaster case data is stored in a relational database, supporting data backtracking and related queries.
[0015] Furthermore, the specific process of data preprocessing in step S2 is as follows:
[0016] S21, Outlier Removal and Imputation: First, outliers exceeding the mean ± 3 standard deviations are removed using the 3σ criterion. Then, nonlinear outliers are identified using the Isolation Forest algorithm. For missing data after outlier removal, linear interpolation is used to imputate missing data when the missing rate is less than 5%, and LSTM prediction is used to imputate missing data when the missing rate is 5% to 20%.
[0017] S22, Spatiotemporal Alignment: Based on GPS timestamps, the sampling frequency of different sensors is unified to 10 minutes / time; the inspection records are matched with the inspection time and geographical location to achieve time synchronization and spatial correlation of multi-source data;
[0018] S23, Format standardization: Numerical data is normalized to the [0,1] interval using min-max normalization; textual data is converted into vector form using a BERT pre-trained model; image data is normalized after feature extraction using CNN.
[0019] Furthermore, the specific structure of the CNN-LSTM fusion model described in step S3 is as follows:
[0020] S31, CNN branch: The input is single-index standardized time-series data, which is sequentially passed through three one-dimensional convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer. Finally, the extracted spatial features are adaptively weighted through the SE attention mechanism. The kernel sizes of the three one-dimensional convolutional layers are 3×1, 5×1, and 7×1, and the number of kernels are 32, 64, and 128, respectively. The kernel size of the max pooling layer is 2×1.
[0021] S32, LSTM branch: The spatial features of all indicators after processing by the CNN branch are concatenated along the time dimension and input into two LSTM layers; the number of hidden units in the two LSTM layers are 256 and 128, respectively; a dropout layer is added between the LSTM layers with a dropout probability of 0.3; at the same time, a residual connection is introduced between the LSTM layers to add the output of the previous LSTM layer to the input of the next LSTM layer;
[0022] S33, Feature Fusion: The temporal features output from the LSTM branch are concatenated with the spatial features output from the CNN branch and input into a fully connected layer with 64 neurons. After passing through the ReLU activation function, the output is a multi-source data fusion feature vector with a dimension of 32.
[0023] Furthermore, the specific implementation of the combined weight evaluation system and dynamic threshold mechanism described in step S4 is as follows:
[0024] S41, Combined Weight Calculation: Establish an engineering health evaluation index system, including four primary indicators (deformation index, stress-strain index, seepage index, and hydro-meteorological index) and twelve secondary indicators; construct a judgment matrix based on relevant engineering standards using the Analytic Hierarchy Process (AHP), calculate subjective weights, and perform a consistency test (CR < 0.1); calculate the information entropy of each indicator using the entropy weight method, and determine the objective weights based on the entropy values; use a linear weighting method to integrate subjective and objective weights, with weight coefficients of 0.4 and 0.6 respectively, to obtain the combined weight W = [w1, w2, ..., w12];
[0025] S42, Health Status Quantification: The fused feature vector is mapped to scores for each secondary indicator, with a score range of 0-100. A comprehensive health status score is calculated through weighted summation. ,in For the first scores of each indicator, and divides the health status into 5 levels according to the S value (excellent: S≥90, good: 80≤S<90, average: 60≤S<80, poor: 40≤S<60, dangerous: S<40);
[0026] S43, dynamic threshold update: adopting a sliding window method with a window size of 30 days, calculating the mean μ and standard deviation σ of the scores within the window based on the comprehensive health score data in the historical 30 days, and determining the dynamic early warning thresholds: the blue early warning threshold T1=μ-σ, the yellow early warning threshold T2=μ-2σ, the orange early warning threshold T3=μ-2.5σ, and the red early warning threshold T4=μ-3σ; the window data is automatically updated and the thresholds are recalculated every early morning to realize adaptive adjustment of the thresholds.
[0027] Further, the specific implementation of the risk identification and decision output in step S5 is as follows:
[0028] S51, key risk source identification: constructing a risk conduction model based on a Bayesian network, taking the abnormality of each monitoring indicator as a parent node and the engineering structure failure as a child node, training the network through historical disaster case data to obtain a conditional probability table, adopting posterior probability inference to identify the key risk source that has the greatest impact on the health status, and determining the indicators with a posterior probability ≥0.7 as key risk sources;
[0029] S52, graded early warning information output: outputting an early warning level according to the comparison result between the comprehensive health score S and the dynamic thresholds: S≥T1 is no warning, T2≤S<T1 is blue warning, T3≤S<T2 is yellow warning, T4≤S<T3 is orange warning, and S<T4 is red warning; the early warning information includes the early warning level, triggering indicator, current health score, risk source location and development trend;
[0030] S53, disposal suggestion generation: calling a preset disposal suggestion knowledge base based on the key risk source type and the early warning level to generate targeted disposal suggestions.
[0031] Further, the construction method further includes step S6 model optimization iteration, in which the parameters of the improved CNN-LSTM fusion model are reversely updated by the gradient descent method through comparison between the real-time early warning results and the actual engineering status, wherein the learning rate is 0.001 and the number of iterations is 100 rounds, and the combined weight evaluation system and the disposal suggestion knowledge base are updated regularly to improve the long-term applicability of the model.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] (1) Multi-source data fusion and preprocessing effect: Integrating multiple types of data such as real-time sensor data, inspection record data and historical disaster case data, outlier removal, spatiotemporal alignment and format standardization are performed sequentially through the data preprocessing layer to achieve effective fusion and in-depth utilization of multi-source heterogeneous data. Among them, outlier removal adopts a combination of 3σ criterion and isolated forest algorithm to take into account the identification of linear and nonlinear anomalies; missing data is filled in according to the missing rate (linear interpolation is used for missing rate less than 5%, and LSTM prediction is used for missing rate of 5%~20%); spatiotemporal alignment is based on GPS timestamp and the sampling frequency is uniformly set at 10 minutes / time to realize the three-dimensional association of "time-space-data"; format standardization adopts min-max normalization, BERT pre-trained model vectorization and CNN feature extraction to ensure high-quality input of multi-source data.
[0034] (2) Deep learning feature extraction effect: The improved CNN-LSTM fusion model introduces SE attention mechanism and residual connection, which significantly improves the accuracy of spatial feature extraction and the ability to capture temporal dependencies. The CNN branch uses SE attention mechanism to adaptively weight the extracted spatial features, enhance the feature weights of key indicators such as seepage and displacement, and suppress noise interference; the LSTM branch adds the output of the previous LSTM layer to the input of the next LSTM layer through residual connection to solve the gradient vanishing problem in long sequence training, and uses dropout layer (probability 0.3) to prevent overfitting. After the temporal features and spatial features are concatenated, they are compressed into a 32-dimensional fusion feature vector through a fully connected layer, which can accurately identify the coupling risk of seepage anomalies and displacement abrupt changes.
[0035] (3) Health assessment and dynamic early warning effect: The AHP-entropy weight combination weight evaluation system takes into account both expert experience and objective data laws. The analytic hierarchy process (AHP) is used to calculate subjective weights and conduct consistency tests (CR<0.1). The entropy weight method calculates objective weights based on the information entropy of monitoring data. The linear weighting method is used to integrate the weight coefficients of 0.4 (subjective) and 0.6 (objective), so that the health assessment results are both scientific and engineering applicable. The dynamic threshold mechanism adopts the 30-day sliding window method. The early warning threshold is adaptively updated based on the mean μ and standard deviation σ of historical health score data (T1=μ-σ, T2=μ-2σ, T3=μ-2.5σ, T4=μ-3σ). It is automatically updated every morning. Compared with the fixed threshold, the adaptability to different working conditions such as flood season, dry season and extreme weather is improved by 40%, effectively reducing the phenomenon of "missed reports" or "false reports".
[0036] (4) Risk Identification and Decision Support Effectiveness: A risk transmission model was constructed based on Bayesian networks, with 12 secondary indicator anomalies as parent nodes and engineering structural failures as child nodes. Key risk sources were identified through posterior probability inference (posterior probability ≥ 0.7), realizing quantitative analysis of the transmission path from "single indicator anomaly" to "multi-hazard coupled risk". Based on the type of key risk source and the warning level, a preset disposal suggestion knowledge base was invoked to generate targeted disposal suggestions. Compared with the traditional method of only outputting the warning level, the emergency response efficiency was significantly improved, realizing "early warning-disposal" closed-loop management.
[0037] (5) Long-term applicability of the model: Through bidirectional data interaction between the output layer and the feature extraction fusion layer, the gradient descent method is used to update the parameters of the CNN-LSTM fusion model in reverse. At the same time, the combined weight evaluation system and disposal suggestion knowledge base are updated regularly to ensure that the model is continuously optimized as the engineering operation status evolves, thereby improving long-term applicability. Attached Figure Description
[0038] Figure 1 This is a diagram illustrating the overall architecture of the multi-source monitoring and safety early warning system for the canal project.
[0039] Figure 2 This is a schematic diagram of the CNN-LSTM fusion model structure;
[0040] Figure 3 This is a schematic diagram of the combined weight evaluation and dynamic early warning process;
[0041] Figure 4 Flowchart for model application implementation examples. Detailed Implementation
[0042] The invention will be further described in detail below with reference to a specific canal embankment project.
[0043] like Figure 1 As shown, the canal engineering multi-source monitoring and safety early warning system of the present invention includes a data acquisition layer, a data preprocessing layer, a feature extraction and fusion layer, a health evaluation and early warning layer and an output layer connected in series. The output layer and the feature extraction and fusion layer realize bidirectional data interaction for model parameter optimization feedback.
[0044] Example 1: Model Deployment and Data Acquisition
[0045] A model was deployed on a section (10km long) of the canal revetment. This section is a silty clay embankment that has historically experienced piping disasters.
[0046] S1, Multi-source data acquisition:
[0047] A multi-dimensional monitoring system is constructed through the data acquisition layer to collect multi-source data of the core structures of the canal project, including real-time sensor monitoring data, inspection record data, and historical disaster case data.
[0048] Real-time sensor monitoring data: One set of piezometers (20 sets in total) is deployed every 500m on the water-facing slope of the dike, one set of displacement gauges (10 sets in total) is deployed every 1000m on the top of the dike, strain gauges (5 sets) are deployed near the lock, and water level gauges (3 sets) and anemometers (2 sets) are deployed along the line. The sensors are connected to the monitoring platform through 5G / BeiDou communication modules to achieve data transmission every 10 minutes, ensuring the stability and timeliness of data transmission.
[0049] Inspection data recording: A mobile inspection app has been developed, allowing inspectors to input problem descriptions (text) and on-site photos / videos. Simultaneously, drones equipped with high-definition cameras and infrared thermal imagers are deployed to conduct aerial inspections of dikes and waterways. Inspection data is automatically linked to geographical location and timestamps for structured storage. Inspectors conduct daily inspections using the mobile app and weekly drone aerial inspections. Inspection data is uploaded to the platform in real time. If even minor cracks are found in the dike, photos and location information of the cracks are uploaded simultaneously.
[0050] Historical disaster case data: Establish a relational database to store historical disaster cases (such as dike breaches, piping, and lock failures) along the canal project, engineering design parameters, maintenance records, and monitoring data over the years. The database supports data backtracking and related queries.
[0051] Example 2: Data Preprocessing and Feature Extraction
[0052] S2: Data Preprocessing
[0053] The multi-source data obtained in Example 1 is processed sequentially by the data preprocessing layer to remove outliers, perform spatiotemporal alignment and format standardization, and obtain standardized data.
[0054] S21, Outlier Removal and Imputation: In daily monitoring data, a sudden change in the piezometer reading (1.2 MPa, far exceeding the normal range of 0.1-0.3 MPa) was observed. First, outliers exceeding the mean ± 3 standard deviations were removed using the 3σ criterion. Then, the Isolation Forest algorithm was used to identify nonlinear anomalies. For missing data after outlier removal, linear interpolation was used for imputation when the missing rate was below 5%. In this embodiment, LSTM prediction was used for imputation (missing rate approximately 8%, falling within the 5%~20% range).
[0055] S22, Spatiotemporal Alignment: Based on GPS timestamps, the sampling frequency of different sensors is unified to 10 minutes / time; for inspection records, the GPS location of the inspectors or the coordinates of drone aerial photography are matched with the spatial coordinates of the engineering structures, and the inspection crack data of a certain section of the embankment is associated with the displacement and seepage data of the area, realizing the three-dimensional association of "time-space-data".
[0056] S23, Format Standardization: Establish unified data coding standards. Numerical data (such as displacement and stress) is normalized to the [0,1] interval using min-max normalization; textual data (such as inspection records) is converted into 768-dimensional vectors using a BERT pre-trained model; image data (such as drone aerial images) is standardized after feature extraction using CNN. All data is standardized into tensor form (dimension: 144×32, 144 being the daily time step and 32 being the number of features).
[0057] S3, Feature Extraction and Fusion of Multi-Source Data:
[0058] The standardized data obtained in step S2 is input into the CNN-LSTM fusion model through the feature extraction and fusion layer to obtain a multi-source data fusion feature vector.
[0059] S31, CNN Branch Spatial Feature Extraction: For standardized time-series data of a single indicator (such as 10-minute / time data from a displacement meter), the data is sequentially processed through three one-dimensional convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer. The kernel sizes of the three one-dimensional convolutional layers are 3×1, 5×1, and 7×1, respectively, with 32, 64, and 128 kernels. The max pooling layer has a 2×1 kernel size. An SE attention mechanism (squeeze-activation module) is introduced. Through squeezing operations (global average pooling) and activation operations (weighted by fully connected layers), the extracted spatial features are adaptively weighted to enhance the feature weights of key indicators such as seepage and displacement, and to suppress the influence of noisy data.
[0060] S32, LSTM Branch Temporal Feature Extraction: The spatial features of all indicators after processing by the CNN branch are concatenated along the time dimension (forming a matrix of "time step × number of features") and input into two LSTM layers. The number of hidden units in the two LSTM layers are 256 and 128, respectively. The first LSTM layer (256 hidden units) extracts local temporal dependencies (such as the trend of indicator changes within one day), and the second LSTM layer (128 hidden units) extracts long-term temporal dependencies (such as the trend evolution within one month). Dropout layers are added between the LSTM layers with a dropout probability of 0.3; at the same time, residual connections are introduced between the LSTM layers, adding the output of the previous LSTM layer to the input of the next LSTM layer to solve the gradient vanishing problem in long sequence training.
[0061] S33, Feature Fusion Output: The temporal features (128-dimensional) output from the LSTM branch and the spatial features (128-dimensional) output from the CNN branch are concatenated and input into a fully connected layer (64 neurons). After passing through the ReLU activation function, a 32-dimensional fused feature vector is output. This vector simultaneously contains spatial correlation and temporal evolution information from multiple data sources. Among them, the feature vector of segment K12+300 shows that the seepage and crack feature weights account for 45%.
[0062] Example 3: Health Assessment and Early Warning Output
[0063] S4, Health Assessment and Dynamic Early Warning:
[0064] Through the health assessment and early warning layer, based on the fusion feature vector obtained in step S3, a combined weight evaluation system and dynamic threshold mechanism are established to quantify the health status level of the engineering structure and determine the early warning level.
[0065] S41, Combined Weight Calculation: An engineering health evaluation index system is established, selecting 12 secondary indicators from four primary dimensions: "deformation safety, stress rationality, seepage stability, and environmental adaptability," including levee crest settlement, horizontal displacement, maximum stress, seepage velocity, water level rise, and wind speed, forming a hierarchical evaluation index system. The AHP method is used to construct pairwise judgment matrices (e.g., comparing the importance of "seepage velocity" and "levee crest settlement"). Subjective weights are calculated using the sum-product method and a consistency test is performed (CR=0.08<0.1, meeting the requirements). The entropy weight method calculates the information entropy of each indicator based on monitoring data from the past 30 days; the seepage velocity has the lowest entropy value (0.32) and the highest weight (0.18). A linear weighting method is used to integrate subjective and objective weights, with weight coefficients of 0.4 and 0.6 respectively, resulting in a combined weight W=[w1,w2,...,w12]. For example: the subjective weight of the seepage velocity determined by the AHP method expert scoring is 0.14, the objective weight is calculated by the entropy weight method to be 0.18, and the combined weight is 0.16; the combined weight of the levee crest settlement is 0.15.
[0066] S42, Health Status Quantification: The fused feature vector is mapped to scores for each secondary indicator, with a score range of 0-100. A comprehensive health status score is calculated through weighted summation. ,in For the first Each indicator is scored. The weighted sum of the scores of each indicator in the K12+300 segment is S=68, which corresponds to a "poor" health status (general: 60≤S<80).
[0067] S43, Dynamic threshold update: Adopt the sliding window method with a window size of 30 days. Based on the comprehensive health score data within the latest 30 days, calculate the mean μ and standard deviation σ of the scores in the window to determine the dynamic warning thresholds. For the sliding window from July 1 to 15, μ=78 and σ=6, so the blue warning threshold T1=μ-σ=72, the yellow warning threshold T2=μ-2σ=66, the orange warning threshold T3=μ-2.5σ=63, and the red warning threshold T4=μ-3σ=60. A score S=68 triggers a blue warning (T2=66≤68<T1=72). The window data is automatically updated and the thresholds are recalculated every early morning to realize adaptive threshold adjustment. Compared with fixed thresholds, the adaptability of dynamic thresholds to different working conditions such as flood seasons and dry seasons is increased by 40%.
[0068] S5, Risk identification and decision output:
[0069] Through the output layer, based on the evaluation result of step S4, a Bayesian network is used to construct a risk propagation model, identify key risk sources, and output graded warning information and targeted disposal suggestions.
[0070] S51, Key risk source identification: Construct a Bayesian network, take the anomalies of 12 secondary indicators as parent nodes and "engineering failure" as the child node, and train the network using the data of 100 historical disaster cases to obtain the conditional probability table. Adopt posterior probability inference to identify the key risk source that has the greatest impact on the health state (an indicator with a posterior probability ≥0.7 is determined as a key risk source). The Bayesian network inference shows that the posterior probability of seepage velocity anomaly is 0.85. Combined with the crack inspection data, the key risk source is determined as "levee crack propagation caused by abnormal seepage in the K12+300 section".
[0071] S52: Graded warning information output. The warning level is output according to the comparison result between the comprehensive health score S and the dynamic warning thresholds: S≥T1 means no warning, T2≤S<T1 means blue warning, T3≤S<T2 means yellow warning, T4≤S<T3 means orange warning, and S<T4 means red warning. In this embodiment, S=68 points, T2=66≤68<T1=72, which triggers a blue warning. The warning information includes: warning level, triggering indicator, current health score, risk source location and development trend. The warning information is pushed to management personnel through the warning platform: "Blue warning: Abnormal seepage velocity (0.8m / d) in the K12+300 section of the levee, current health score is 68 points, and the risk is expected to escalate within 12 hours", and the warning level and risk location are displayed on the LED screen at the engineering site at the same time.
[0072] S53, Disposal Recommendation Generation: Based on the key risk source type and warning level, the system calls the preset disposal recommendation knowledge base, matches the "abnormal seepage + blue warning" scenario, and generates targeted recommendations: ① Adjust the frequency of piezometer monitoring to 5 minutes / time; ② Inspectors check the crack expansion once every 2 hours; ③ Transport 200㎡ of geomembrane and 500 sandbags to the site; ④ Notify the maritime department to suspend ship navigation in the K12-K13 section to reduce the impact of load.
[0073] Example 4: Model Validation and Optimization
[0074] Verification of early warning effectiveness: On-site investigation revealed small-scale piping, which was promptly treated with geomembrane for seepage prevention, and no disaster occurred, verifying that the early warning accuracy rate was 100%; compared with traditional early warning methods, this method provides early warning and sufficient time for response.
[0075] S6, Model Optimization Iteration:
[0076] By comparing real-time early warning results with actual engineering conditions, the parameters of the improved CNN-LSTM fusion model are updated in reverse using the gradient descent method (learning rate 0.001, 100 iterations), resulting in a 5% improvement in seepage feature extraction accuracy. Simultaneously, the combined weight evaluation system and disposal suggestion knowledge base are updated regularly (quarterly), and suggestions for "relevant emergency response procedures" are added based on disposal experience to enhance the model's long-term applicability.
[0077] Two-way data interaction: The output layer and the feature extraction and fusion layer achieve two-way data interaction for model parameter optimization feedback. At the end of each quarter, case data is collected, and model parameters are updated using gradient descent. The updated parameters are then fed back to the feature extraction and fusion layer through a feedback channel, enabling continuous model optimization.
Claims
1. A method for constructing a multi-source monitoring and safety early warning system for canal projects, characterized in that, The system comprises a data acquisition layer, a data preprocessing layer, a feature extraction and fusion layer, a health evaluation and early warning layer, and an output layer connected in sequence. The output layer and the feature extraction and fusion layer achieve bidirectional data interaction for model parameter optimization feedback. The construction method includes the following steps: S1, Multi-source data acquisition: A multi-dimensional monitoring system is constructed through the data acquisition layer to collect multi-source data of the core structures of the canal project. The multi-source data includes real-time sensor monitoring data, inspection record data, and historical disaster case data. The sensors include displacement gauges, strain gauges, piezometers, water level gauges, and anemometers, which are used to monitor deformation, stress and strain, seepage, and hydrological and meteorological indicators, respectively. The inspection record data includes text, image, and video data. S2, Data Preprocessing: The multi-source data obtained in step S1 is sequentially subjected to outlier removal, spatiotemporal alignment, and format standardization through the data preprocessing layer to obtain standardized data; the outlier removal adopts a combination of the 3σ criterion and the isolated forest algorithm. S3, Multi-source data feature extraction and fusion: The standardized data obtained in step S2 is input into the CNN-LSTM fusion model through the feature extraction and fusion layer for feature extraction and fusion, resulting in a multi-source data fusion feature vector. The CNN-LSTM fusion model includes a parallel CNN branch and an LSTM branch. The CNN branch is used to extract spatial features, and it introduces an SE attention mechanism to adaptively weight the extracted spatial features. The LSTM branch is used to extract temporal features, and it introduces residual connections to add the output of the previous LSTM layer to the input of the next LSTM layer. The CNN-LSTM fusion model also includes a feature fusion layer, which concatenates the spatial features output by the CNN branch with the temporal features output by the LSTM branch to output a multi-source data fusion feature vector. S4, Health Assessment and Dynamic Early Warning: Through the health assessment and early warning layer, based on the fusion feature vector obtained in step S3, a combined weight evaluation system and a dynamic threshold mechanism are established to quantify the health status level of the engineering structure and determine the early warning level. The combined weight evaluation system is constructed by combining the Analytic Hierarchy Process (AHP) and the entropy weight method. The dynamic threshold mechanism uses the sliding window method to achieve adaptive updating of the early warning threshold based on the statistical characteristics of historical health score data. S5, Risk Identification and Decision Output: Through the output layer, based on the evaluation results of step S4, a risk transmission model is constructed using a Bayesian network to identify key risk sources and output graded early warning information and targeted handling suggestions.
2. The method for constructing a multi-source monitoring and safety early warning system for canal engineering according to claim 1, characterized in that, The real-time monitoring data from the sensors in step S1 is transmitted via a 5G / BeiDou communication module; the inspection record data is entered via a mobile app or automatically imported via drone inspection; and the historical disaster case data is stored in a relational database, supporting data backtracking and related queries.
3. The method for constructing a multi-source monitoring and safety early warning system for canal engineering according to claim 1, characterized in that, The specific process of data preprocessing in step S2 is as follows: S21, Outlier Removal and Imputation: First, outliers exceeding the mean ± 3 standard deviations are removed using the 3σ criterion. Then, nonlinear outliers are identified using the Isolation Forest algorithm. For missing data after outlier removal, linear interpolation is used to imputate missing data when the missing rate is less than 5%, and LSTM prediction is used to imputate missing data when the missing rate is 5% to 20%. S22, Spatiotemporal Alignment: Based on GPS timestamps, the sampling frequency of different sensors is unified to 10 minutes / time; the inspection records are matched with the inspection time and geographical location to achieve time synchronization and spatial correlation of multi-source data; S23, Format standardization: Numerical data is normalized to the [0,1] interval using min-max normalization; textual data is converted into vector form using a BERT pre-trained model; image data is normalized after feature extraction using CNN.
4. The method for constructing a multi-source monitoring and safety early warning system for canal engineering according to claim 1, characterized in that, The specific structure of the CNN-LSTM fusion model described in step S3 is as follows: S31, CNN branch: The input is single-index standardized time-series data, which is sequentially passed through three one-dimensional convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer. Finally, the extracted spatial features are adaptively weighted through the SE attention mechanism. The kernel sizes of the three one-dimensional convolutional layers are 3×1, 5×1, and 7×1, and the number of kernels are 32, 64, and 128, respectively. The kernel size of the max pooling layer is 2×1. S32, LSTM branch: The spatial features of all indicators after processing by the CNN branch are concatenated along the time dimension and input into two LSTM layers; the number of hidden units in the two LSTM layers are 256 and 128, respectively; a dropout layer is added between the LSTM layers with a dropout probability of 0.3; at the same time, a residual connection is introduced between the LSTM layers to add the output of the previous LSTM layer to the input of the next LSTM layer; S33, Feature Fusion: The temporal features output from the LSTM branch are concatenated with the spatial features output from the CNN branch and input into a fully connected layer with 64 neurons. After passing through the ReLU activation function, the output is a multi-source data fusion feature vector with a dimension of 32.
5. The method for constructing a multi-source monitoring and safety early warning system for canal engineering according to claim 1, characterized in that, The specific implementation of the combined weight evaluation system and dynamic threshold mechanism described in step S4 is as follows: S41, Combined Weight Calculation: Establish an engineering health evaluation index system, including four primary indicators (deformation index, stress-strain index, seepage index, and hydro-meteorological index) and twelve secondary indicators; construct a judgment matrix based on relevant engineering standards using the Analytic Hierarchy Process (AHP), calculate subjective weights, and perform a consistency test (CR < 0.1); calculate the information entropy of each indicator using the entropy weight method, and determine the objective weights based on the entropy values; use a linear weighting method to integrate subjective and objective weights, with weight coefficients of 0.4 and 0.6 respectively, to obtain the combined weight W = [w1, w2, ..., w12]; S42, Health Status Quantification: The fused feature vector is mapped to scores for each secondary indicator, with a score range of 0-100. A comprehensive health status score is calculated through weighted summation. ,in For the first The health status is divided into 5 levels based on the S-value (Excellent: S≥90, Good: 80≤S<90, Average: 60≤S<80, Poor: 40≤S<60, Dangerous: S<40). S43, dynamic threshold updating: adopting a sliding window method with a window size of 30 days, calculating the mean μ and standard deviation σ of the scores within the window based on the historical comprehensive health score data of the past 30 days, and determining the dynamic warning thresholds: blue warning threshold T1=μ-σ, yellow warning threshold T2=μ-2σ, orange warning threshold T3=μ-2.5σ, red warning threshold T4=μ-3σ; the window data is automatically updated and the thresholds are recalculated at wee hours every day to realize adaptive threshold adjustment.
6. The method for constructing a multi-source monitoring and safety early warning system for canal engineering according to claim 1, characterized in that, The specific implementation of risk identification and decision output described in step S5 is as follows: S51, key risk source identification: constructing a risk conduction model based on a Bayesian network, taking the abnormality of each monitoring indicator as a parent node and engineering structure failure as a child node, training the network through historical disaster case data to obtain a conditional probability table, identifying key risk sources by posterior probability inference, and determining indicators with a posterior probability ≥0.7 as key risk sources; S52, graded warning information output: outputting the warning level according to the comparison result between the comprehensive health score S and the dynamic thresholds: S≥T1 means no warning, T2≤S<T1 means blue warning, T3≤S<T2 means yellow warning, T4≤S<T3 means orange warning, S<T4 means red warning; the warning information includes the warning level, triggering indicator, current health score, risk source location and development trend; S53, disposal suggestion generation: invoking a preset disposal suggestion knowledge base based on the key risk source type and warning level to generate targeted disposal suggestions.
7. A method for constructing a multi-source monitoring and safety early warning system for canal engineering according to any one of claims 1-6, characterized in that, The method further includes step S6 model optimization and iteration: by comparing the real-time warning result with the actual engineering state, the gradient descent method is used to reversely update and improve the parameters of the CNN-LSTM fusion model, wherein the learning rate is 0.001 and the number of iterations is 100 rounds, and the combined weight evaluation system and the disposal suggestion knowledge base are updated regularly at the same time to improve the long-term applicability of the model.
Citation Information
Patent Citations
Water conservancy project intelligent monitoring system and method
CN120909202A
