Water conservancy project communication network node quality evaluation method and system

Abstract

The invention relates to the technical field of water conservancy project communication network intelligentization, in particular to a water conservancy project communication network node quality evaluation method and system, and the method comprises the following steps: a five-dimensional index system construction step: building a five-dimensional evaluation system; a dynamic weight calculation step: calculating the dynamic weight of the five-dimensional index system; an expert weight calculation step: calculating an expert time sequence weight of the five-dimensional index system; a weight fusion calculation step: a fusion weight hybrid model evaluation step of calculating a dynamic weight and an expert time sequence weight of the same index through a credibility weighting formula; carrying out weighted summation on each network node of the hydraulic engineering communication network through a gating fusion module of the hybrid model, and outputting a comprehensive score; and a decision early warning step: performing decision assessment early warning according to the comprehensive score. According to the method, multi-dimensional data can be fused, the weight can be dynamically adjusted, and intelligent evaluation is performed in combination with expert experience, so that the safety and the operation efficiency of a water conservancy system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent water conservancy project communication networks, and in particular to a water conservancy project communication network node quality assessment method and system. Background Art

[0002] As core infrastructure for water resource scheduling and flood prevention and disaster reduction, the quality of water conservancy transmission networks (such as pumping stations, gates, sensor terminals, and communication hubs) directly impacts the safety, reliability, and operational efficiency of water conservancy systems. With the widespread adoption of IoT technology, a large number of sensors (for pressure, flow, water quality, etc.) have been deployed at water conservancy nodes, generating over 10,000 high-frequency data points daily. Furthermore, the complexity of network structures (such as chain, ring, and star hybrid topologies), the diversity of services (such as video surveillance, real-time scheduling, and equipment control), and the demand for intelligent capabilities (such as AI algorithm integration and 4G / 5G upgrades) have placed higher demands on node evaluation.

[0003] Existing technologies struggle to address the complexity and dynamic nature of water transmission networks due to incomplete indicator systems, static weighting mechanisms, insufficient real-time processing capabilities, and a lack of human-machine collaboration. This results in low assessment accuracy, delayed response, and high operation and maintenance costs. An intelligent assessment method that integrates multi-dimensional data, dynamically adjusts weights, and incorporates expert experience is urgently needed to improve the safety and operational efficiency of water systems. Summary of the Invention

[0004] One of the objectives of the present invention is to provide a method for evaluating the quality of communication network nodes in water conservancy projects, which can integrate multi-dimensional data, dynamically adjust weights and perform intelligent evaluation based on expert experience to improve the safety and operational efficiency of water conservancy systems.

[0005] In order to achieve the above object, a method for evaluating the quality of nodes in a water conservancy project communication network is provided, comprising the following steps:

[0006] Steps for constructing a five-dimensional indicator system: Establish a five-dimensional evaluation system that includes network structure characteristics, functional performance, security and reliability, economic cost, and intelligent optimization potential as first-level indicators; the first-level indicators include several second-level indicators;

[0007] Dynamic weight calculation steps: The first-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights; the second-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights;

[0008] Expert weight calculation steps: The first-level indicators of the five-dimensional indicator system are respectively generated through the hierarchical analysis method to generate their respective expert time-series weights; the second-level indicators of the five-dimensional indicator system are respectively generated through the hierarchical analysis method to generate their respective expert time-series weights;

[0009] Weight fusion calculation steps: Calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data;

[0010] Hybrid model evaluation steps: Obtain an LSTM-GNN-Expert hybrid model, which includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module; input multidimensional time series data, network structure data, and expert unstructured annotations, and output a time series feature vector, a network structure feature vector, and an expert correction vector respectively, and input them together into the gated fusion module. The gated fusion module performs a weighted summation on each network node of the water conservancy project communication network and outputs a comprehensive score;

[0011] Decision-making and early warning steps: Make decision-making assessment and early warning based on the comprehensive score.

[0012] Furthermore, the steps of generating the dynamic weight in the dynamic weight calculation step are as follows:

[0013] Generate the basic weight of each dynamic weight of the first-level indicator or the second-level indicator through the intelligent agent;

[0014] Generate the optimized weights of the dynamic weights of the primary indicators or secondary indicators through the Q-learning intelligent agent optimization algorithm;

[0015] Adjust the time decay coefficient of the weight through the time decay algorithm;

[0016] The multi-head self-attention mechanism is used to capture the coupling relationship between indicators and obtain the attention correction weight;

[0017] The dynamic weight is obtained according to the basic weight, optimization weight, time decay coefficient and attention correction weight.

[0018] Furthermore, the generation step of the expert time series weight in the expert weight calculation step is as follows:

[0019] Based on the hierarchical analysis method, the working condition correction coefficient of the weight is obtained according to the working condition label to obtain the initial weight;

[0020] The expert calibration text is converted into a weight correction vector through BERT, and the initial weight is updated to obtain the expert time series weight;

[0021] Adjust the time decay coefficient of the initial weight through the time decay algorithm And update the expert timing weight;

[0022] Update expert time series weights smoothly based on historical calibration data through Markov chain;

[0023] Set the weight ratio of the expert's time series weight. The weight ratio formula is:

[0024] is the time decay coefficient.

[0025] Furthermore, the credibility weighting formula in the weight fusion calculation step is as follows:

[0026]

[0027] Data credibility = 1-percentage of abnormal data; expert activity is the quarterly questionnaire completion rate;

[0028] in, is the fusion weight, α is the dynamic weight ratio, and the ratio range is (0.6-0.8). is the dynamic weight, is the expert temporal weight, α expert The proportion range is (0.2-0.4).

[0029] Furthermore, the gated fusion module includes a splicing layer, a weight calculation layer, and a weighted summation layer;

[0030] Splicing layer: used to splice the time series feature vector and the structural feature vector to form a joint feature h containing time series dynamics and network structure information LG , represents the joint hidden state of the LSTM layer and the GNN layer;

[0031] Weight calculation layer: by splicing the joint hidden state h of the LSTM layer and the GNN layer LG Combined with the expert semantic hidden state, the fusion weight σ is generated through full connection and sigmoid activation function to achieve soft-gated fusion of dynamic scoring and expert scoring;

[0032] Weighted summation layer: Generates a comprehensive score through a gated fusion formula; the gated fusion formula is as follows:

[0033] Score=σ(W g ·[h LG ;h erpert ])·LGScore+(1-σ)·ExpertScore

[0034] Among them, W g represents the weight matrix of gated fusion; h expertRepresents the processing result of the expert-annotated text from the BERT semantic layer; LGScore: represents the preliminary score generated by the LSTM-GNN model in the LSTM-GNN-Expert hybrid model based on time series data and network structure features; ExpertScore: represents the score generated based on expert semantic knowledge.

[0035] Furthermore, the first-level indicators include network structure characteristics, functional performance, security and reliability, economic cost and intelligent optimization potential; the network structure characteristics include chain topology position, ring structure participation, star centrality, structural vulnerability, number of cross-substructure connections, and path uniqueness dependence, a total of 6 second-level indicators; functional performance includes data transmission volume, bandwidth redundancy, real-time requirements, and multi-service compatibility, a total of 4 second-level indicators; security and reliability include anti-attack capability, redundancy, fault recovery time, and historical failure rate, a total of 4 second-level indicators; economic cost includes construction cost, operation and maintenance cost, and upgrade cost, a total of 3 second-level indicators; intelligent optimization potential includes data support capability, algorithm adaptability, and scalability, a total of 3 second-level indicators.

[0036] Furthermore, it also includes:

[0037] Edge computing steps: Deploy edge computing nodes in cities and prefectures to collect and preprocess data obtained by the sensor layer in real time to obtain several first-level and second-level indicators.

[0038] Furthermore, in the step of generating the optimized weights of the dynamic weights of the primary indicators or the secondary indicators by the Q-learning agent optimization algorithm, the step of obtaining the optimized weights is as follows:

[0039] Obtain real-time indicator data for primary or secondary indicators and obtain optimized weights through reward functions;

[0040] Reward function = |prediction score - expert quarterly calibration value| -1

[0041] Prediction score: machine score calculated by the agent in real time, expert quarterly calibration value: manual benchmark score by domain experts.

[0042] A second object of the present invention is to provide a water conservancy project communication network node quality assessment system, comprising:

[0043] Sensor layer: used to collect sensor data of the water conservancy project communication network through a number of sensors;

[0044] Edge computing module: used to deploy edge computing nodes in cities and prefectures to collect and pre-process sensor data in real time;

[0045] Core evaluation module: includes a five-dimensional indicator system construction unit, a dynamic weight calculation unit, an expert weight calibration unit, a weight fusion engine, and a hybrid model reasoning unit;

[0046] Five-dimensional indicator system construction unit: used to construct a five-dimensional indicator system based on preprocessed sensor data;

[0047] Dynamic weight calculation unit: used to generate dynamic weights for the first-level indicators of the five-dimensional indicator system through the corresponding intelligent agents; and to generate dynamic weights for the second-level indicators of the five-dimensional indicator system through the corresponding intelligent agents;

[0048] Expert weight calibration unit: used to generate respective expert time series weights for the first-level indicators of the five-dimensional indicator system through the analytic hierarchy process; and to generate respective expert time series weights for the second-level indicators of the five-dimensional indicator system through the analytic hierarchy process;

[0049] Weight fusion engine: used to calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data;

[0050] Hybrid model inference unit: used to obtain an LSTM-GNN-Expert hybrid model, which includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module. The hybrid model inputs multidimensional time series data, network structure data, and expert unstructured annotations, and outputs a time series feature vector, a network structure feature vector, and an expert correction vector, which are input into the gated fusion module. The gated fusion module performs a weighted summation on each network node of the water conservancy project communication network and outputs a comprehensive score.

[0051] Early warning decision module: Execute four-level early warning based on the comprehensive score.

[0052] Furthermore, the expert collaboration platform is used to provide a scenario-based weight calibration interface, enabling experts to complete the quarterly questionnaire through the WeChat mini-program, and convert the expert unstructured annotations in the quarterly questionnaire into weight correction vectors through the built-in BERT semantic encoder. The weight correction vectors are used by the expert weight calibration unit to generate expert time-series weights for each indicator, and are used to obtain expert correction vectors in the hybrid model inference unit.

[0053] Principles and advantages:

[0054] 1. This solution integrates multi-dimensional data, covering 32 quantitative indicators, and incorporates network structure vulnerability and upgrade costs into the assessment to address the complexity of water conservancy transmission networks. This addresses the one-sidedness of existing technical indicators and facilitates the accuracy of subsequent water conservancy project communication network node quality assessments. By fusing dynamic weights with expert time-series weights to obtain fused weights, a combined analysis of model evaluation and expert experience is achieved, facilitating intelligent and accurate assessment of water conservancy project communication network node quality. Furthermore, multi-dimensional time-series data, network structure data, and expert experience are input into the LSTM-GNN-Expert hybrid model to quantitatively score each network node in the water conservancy project communication network. This results in a high-precision, adaptive quantitative assessment of network node quality to support the safe operation and maintenance and intelligent decision-making of water conservancy systems, addressing the low assessment accuracy, delayed response, and high operation and maintenance costs of existing technologies.

[0055] 2. This solution has carried out a three-dimensional evolution of the analytic hierarchy process (AHP) into "time series + semantics + data";

[0056] (1) Time series dynamics: Through quarterly calibration + time decay, the expert weights are adapted to seasonal and diurnal changes; it breaks through the static weight limitations of traditional AHP and achieves time series adaptation of weights through time decay and multi-period data integration.

[0057] (2) Semantic quantification: BERT converts qualitative descriptions such as expert-annotated text into computable weight corrections, reducing subjective bias;

[0058] (3) Data-driven: The weights are smoothed based on historical calibration data through a Markov chain to avoid drastic fluctuations caused by human intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flowchart of a method for evaluating the quality of nodes in a water conservancy project communication network according to an embodiment of the present invention;

[0060] Figure 2 This is a logic block diagram of a water conservancy project communication network node quality assessment system according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the correlation between one-dimensional indicators and two-dimensional indicators in the five-dimensional evaluation system;

[0062] Figure 4 Schematic diagram of the structure for fusing dynamic weights with expert temporal weights;

[0063] Figure 5 Schematic diagram of the structure of the LSTM-GNN-Expert hybrid evaluation model. DETAILED DESCRIPTION

[0064] The following is further described in detail through specific implementation methods:

[0065] Example

[0066] A method for evaluating the quality of water conservancy project communication network nodes is basically as follows: Figure 1 As shown, the following steps are included:

[0067] S1. Edge computing steps: deploy edge computing nodes in prefecture-level cities to collect and pre-process data obtained by the sensor layer in real time to obtain several primary and secondary indicators.

[0068] S101, sensor deployment and data collection;

[0069] 1. Collect network structure data: Obtain the physical location of nodes through Wi-Fi / LoRa / 4G / 5G positioning modules, build a network topology map in real time, and update six structural indicators such as chain betweenness and ring participation every 5 minutes (error <±20 meters).

[0070] 2. Collect functional performance data: Deploy multi-mode industrial routers to collect data transmission volume (GB / h) and bandwidth redundancy (calculated in real time as (peak bandwidth - real-time bandwidth) / peak bandwidth). Use a high-precision timer to measure real-time latency (jitter within a 1-minute window ≤ 5ms).

[0071] 3. Collect security and reliability data: Use the GAN algorithm to generate attack traffic to test anti-attack capabilities (average recovery time ≤ 10 seconds), and record the fault recovery time through the system (calculated by the 95th percentile).

[0072] 4. Collect economic cost data: Connect to the financial system to obtain construction costs (equipment procurement + deployment costs) and operation and maintenance costs (monthly labor and spare parts costs, depreciated at an annual interest rate of 4.5%). Use interface testing tools to measure upgrade costs (renovation time <8 hours / time).

[0073] 5. Collect intelligent potential data: Calculate the average daily effective data percentage (≥90% after removing noise) and obtain the algorithm adaptability F1 score through offline testing (≥85% under new working conditions).

[0074] S101, edge computing preprocessing;

[0075] 1. Use city-level edge nodes to perform spatiotemporal alignment of second-level high-frequency data (10Hz):

[0076] 1.1 Timing data: Align the clocks of different sensors using an improved DTW algorithm, with an error of <±1 second;

[0077] 1.2. Structural data: Matched with physical parameter timestamps to generate a standardized feature vector (32 dimensions) with a 5-minute granularity.

[0078] 2. Anomaly detection: Use the isolation forest algorithm to filter noisy data, with single-node processing delay <50ms and anomaly data identification accuracy ≥98%.

[0079] S2. Steps for constructing a five-dimensional indicator system: Establish a five-dimensional evaluation system that includes network structure characteristics, functional performance, security and reliability, economic cost, and intelligent optimization potential as first-level indicators; the first-level indicators include several second-level indicators; such as Figure 3 shown.

[0080] The first-level indicators include network structure characteristics W1, functional performance W2, security and reliability W3, economic cost W4 and intelligent optimization potential W5; the network structure characteristics include chain topology position, ring structure participation, star centrality, structural vulnerability, number of cross-substructure connections, and path uniqueness dependence, a total of 6 second-level indicators; functional performance includes data transmission volume, bandwidth redundancy, real-time requirements, and multi-service compatibility, a total of 4 second-level indicators; security and reliability include anti-attack capability, redundancy, fault recovery time, and historical failure rate, a total of 4 second-level indicators; economic cost includes construction cost, operation and maintenance cost, and upgrade cost, a total of 3 second-level indicators; intelligent optimization potential includes data support capability, algorithm adaptability, and scalability, a total of 3 second-level indicators.

[0081] S3, dynamic weight calculation step: the first-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights; the second-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights;

[0082] The steps for generating the dynamic weight in the dynamic weight calculation step are as follows:

[0083] S301. Generate the basic weight of the dynamic weight of each first-level indicator or second-level indicator through the intelligent agent; in this embodiment, the five intelligent agents correspond to five one-dimensional indicators respectively, and the input of each intelligent agent is the real-time characteristics of the current dimension (6-3 items ranging) and historical scores, and the output is the dynamic weight of the dimension (0-1).

[0084] S302: Generate optimized weights for the dynamic weights of the primary or secondary indicators using a Q-learning agent optimization algorithm. In the above steps, the steps for obtaining the optimized weights are as follows:

[0085] S3021. Obtain real-time indicator data of the primary indicator or the secondary indicator, and obtain the optimized weight through the reward function;

[0086] Reward function = |prediction score - expert quarterly calibration value| -1

[0087] Prediction score: machine score calculated by the agent in real time, expert quarterly calibration value: manual benchmark score by domain experts.

[0088] S303. Adjust the time decay coefficient of the weight using a time decay algorithm. The time decay coefficient formula is as follows:

[0089] λ t =0.95 24-t , t is the tth hour of the day, 0≤t<24;

[0090] S304. Capture the coupling relationship between indicators through the multi-head self-attention mechanism to obtain the attention correction weight; Multi-head self-attention mechanism: 8 attention heads capture indicator coupling. For example, when the structural vulnerability is greater than 0.7, the safety reliability weight is automatically increased by 25% (obtained through training with historical fault data).

[0091] S305: Obtain dynamic weights based on the basic weights, optimized weights, time decay coefficients, and attention correction weights. An example is as follows:

[0092] Network structure (first-level indicator W1)

[0093] Secondary indicators: chain betweenness, ring participation, node degree, connectivity redundancy, average path length, and vulnerability index. Dynamic weight generation logic:

[0094] Q-learning agent optimization:

[0095] Input: Real-time structural data (e.g., ring participation = 0.8, vulnerability index = 0.7).

[0096] Reward trigger: When the fragility index is greater than 0.6 and the connectivity redundancy is less than 0.5, the prediction score decreases, driving the agent to increase the fragility index weight (e.g., from 0.25 to 0.35).

[0097] Time decay adjustment:

[0098] Night (t = 22): Chain betweenness weight × 0.9 (λt = 0.95 2 =0.9025), the structural impact is weakened due to the low flow at night.

[0099] Attention coupling (association security reliability W3):

[0100] If the ring participation is greater than 0.7 (W1 anomaly), the attention mechanism increases the "redundancy" weight under safety and reliability by +0.15, and the ring participation weight under the network structure by +0.2.

[0101] Example weight table (ring hub node, daytime t=10):

[0102]

[0103] S4, expert weight calculation step: the first-level indicators of the five-dimensional indicator system are respectively generated by the hierarchical analysis method to generate their respective expert time-series weights; the second-level indicators of the five-dimensional indicator system are respectively generated by the hierarchical analysis method to generate their respective expert time-series weights;

[0104] The steps for generating the expert time series weight in the expert weight calculation step are as follows:

[0105] S401, based on the analytic hierarchy process, obtain the working condition correction coefficient of the weight according to the working condition label to obtain the initial weight;

[0106] The AHP method is a quarterly scenario-based AHP method, which constructs a 24×24 judgment matrix based on the quarterly scenario-based questionnaire;

[0107] For example: Working condition label: irrigation peak + equipment aging;

[0108] Qualitative description: The structural redundancy is insufficient, and it is recommended to increase the safety and economic weights;

[0109] Weight formula: δ is the working condition correction coefficient;

[0110] Initial weights (traditional AHP): W1 = 0.2, W3 = 0.25, W4 = 0.2.

[0111] S402: Convert the expert calibration text into a weight correction vector through BERT, and update the initial weight to obtain the expert time series weight; for example:

[0112] Input: Expert calibration text (such as "The equipment is seriously aging and the economic cost weight needs to be increased"), which is obtained through expert unstructured annotation. Expert unstructured annotation refers to the subjective judgment information provided by experts through open text (such as free description, suggestions, scoring reasons, etc.) during the evaluation process, which is different from the structured scoring based on a fixed scale in traditional AHP. The source scenarios are: 1. Quarterly scenario questionnaire: open questions are designed in each evaluation cycle (such as quarterly), and experts are invited to fill in specific reasons for the importance of indicators. 2. Expert interviews / meeting minutes: text records collected through semi-structured interviews, or non-quantitative opinions put forward by experts in review meetings.

[0113] Processing flow:

[0114] i. BERT refers specifically to the BERT-base model, which is used to extract semantic features and generate 768-dimensional vectors;

[0115] ii. The fully connected layer is mapped to a five-dimensional weight correction vector (ΔW1-ΔW5) in the range [-0.1, 0.1];

[0116] Example: BERT output: ΔW1 = +0.05 (structure), ΔW3 = +0.1 (safety), ΔW4 = +0.08 (economy);

[0117] Corrected weights: W1=0.25, W3=0.35, W4=0.28.

[0118] S403: Adjust the time decay coefficient of the initial weight through the time decay algorithm And update the expert timing weight;

[0119] (τ is the last calibration quarter, T is the current quarter);

[0120] The last calibration was 2024Q1 (τ=1), Weight decay 2%: W3 = 0.35 × 0.98 = 0.343, W4 = 0.28 × 0.98 = 0.274.

[0121] S404, updating the expert time series weights by smoothing through the Markov chain and based on historical calibration data;

[0122] State definition: expert weight vector S t =[ω1,ω2,ω3,ω4,ω5] (sum=1).

[0123] Transition probability: P(S t+1 |S t )=exp(-γ·|prediction score t -Expert calibration values t |), γ=0.1, the smaller the error, the higher the probability that the weight remains in its original state.

[0124] Iterative logic: After quarterly calibration, the transfer matrix is updated based on 2.8 million historical data points (including 12 expert interventions) to avoid drastic fluctuations in weights.

[0125] In the above example, the expert time series weights of the one-dimensional indicator [ω1,ω2,ω3,ω4,ω5] are [0.25,0.15,0.343,0.274,0.183].

[0126] Steps S402-S404 are essentially weight range corrections, and the fluctuation range of a single indicator weight is ≤±0.15.

[0127] S405. Set the weight ratio of the expert's time series weight. The weight ratio formula is:

[0128] is the time decay coefficient.

[0129] During the day (sufficient data, ): Expert time series weight ratio ≤ 0.4-0.2×0.8=0.24; that is, the dynamic weight is 1-0.24=0.76;

[0130] Nighttime (data sparse, ) : Expert time series weight ratio ≥ 0.4 - 0.2 × 0.95 = 0.21. That is, the dynamic weight is 1 - 0.21 = 0.79.

[0131] S4, weight fusion calculation step: calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data; such as Figure 4 shown.

[0132] The credibility weighting formula in the weight fusion calculation step is as follows:

[0133]

[0134] Data credibility = 1-percentage of abnormal data; expert activity is the quarterly questionnaire completion rate;

[0135] in, is the fusion weight, α is the dynamic weight ratio, and the ratio range is (0.6-0.8). is the dynamic weight, is the expert temporal weight, α expert The proportion range is (0.2-0.4).

[0136] Expert timing weight Refers to the weight generated by the subjective judgment of experts and dynamically adjusted over time (such as quarterly and annually), reflecting the changes in the experts' cognition of the importance of indicators at different time points (the proportion range is 0.2-0.4, which is different from the dynamic weight driven by dynamic data). complement each other).

[0137] S5. Hybrid model evaluation step: obtaining an LSTM-GNN-Expert hybrid model, wherein the hybrid model includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module; inputting multidimensional time series data, network structure data, and expert unstructured annotations, outputting a time series feature vector, a network structure feature vector, and an expert correction vector, respectively, and inputting them together into the gated fusion module, performing a weighted summation on each network node of the water conservancy project communication network through the gated fusion module, and outputting a comprehensive score;

[0138] like Figure 5 As shown:

[0139] LSTM layer: 2 layers of bidirectional LSTM units, 256 units per layer, processing the past 120 hours of time series data and outputting a time series feature vector h_LSTM (dimension = 256). The five-dimensional evaluation system of the water conservancy project communication network and the fusion weights of each indicator are organized as part of the multidimensional time series data;

[0140] The GNN layer uses the GraphSAGE algorithm to capture network structural characteristics by aggregating neighbor node features (sample count = 7) and outputs a structural feature vector h_GNN (dimension = 256). The LSTM layer processes time series data from the past 120 hours and supports high-frequency data input in seconds (sampling frequency ≥ 10 Hz).

[0141] BERT semantic layer: Input expert-annotated text, obtain a 5-dimensional weight correction vector h_expert (dimension = 5, value range [-0.15, 0.15]), and finally output the expert correction vector.

[0142] The gated fusion module includes a splicing layer, a weight calculation layer and a weighted summation layer;

[0143] Splicing layer: used to splice the time series feature vector and the structural feature vector to form a joint feature h containing time series dynamics and network structure information LG , h LG =[h LSTM ;h GNN ] dimension 512, representing the joint hidden state of the LSTM layer and the GNN layer;

[0144] Weight calculation layer: by splicing the joint hidden state h of the LSTM layer and the GNN layer LG Combined with the expert semantic hidden state, the fusion weight σ is generated through full connection and sigmoid activation function to achieve soft-gated fusion of dynamic scoring and expert scoring;

[0145]

[0146] Weighted summation layer: Generates a comprehensive score through a gated fusion formula; the gated fusion formula is as follows:

[0147] Score=σ(W g ·[h LG ;h erpert ])·LGScore+(1-σ)·ExpertScore

[0148] Among them, W g represents the weight matrix of gated fusion; h expertRepresents the processing result of expert-annotated text from the BERT semantic layer; LGScore: represents the preliminary score generated by the LSTM-GNN model in the LSTM-GNN-Expert hybrid model based on time series data and network structure features, that is, the dynamic score; ExpertScore: represents the score generated based on expert semantic knowledge, that is, the expert score.

[0149] The dynamic score LGScore is the LSTM-GNN output (0-100), and the expert score ExpertScore is the expert experience score (initial value = 80, quarterly calibration update).

[0150] S6. Decision-making and early warning step: Decision-making assessment and early warning are performed based on the comprehensive score. In this embodiment, a four-level early warning system (green ≥ 90, yellow 70-89, orange 50-69, and red < 50) is implemented based on the comprehensive score. When the dynamic score deviates from the expert score by > 20%, the expert intervention mechanism is triggered. A review request is automatically pushed to the WeChat mini-program, and the expert response time is < 30 minutes, forming a collaborative process of "model initial judgment - expert calibration - closed-loop optimization."

[0151] A water conservancy project communication network node quality assessment system, which uses the above method, includes:

[0152] Sensor layer: used to collect sensor data of the water conservancy project communication network through a number of sensors;

[0153] Edge computing module: used to deploy edge computing nodes in cities and prefectures to collect and pre-process sensor data in real time;

[0154] Core evaluation module: A data center built on a self-built cloud cluster, including a five-dimensional indicator system construction unit, a dynamic weight calculation unit, an expert weight calibration unit, a weight fusion engine, and a hybrid model inference unit;

[0155] Five-dimensional indicator system construction unit: used to construct a five-dimensional indicator system based on preprocessed sensor data;

[0156] Dynamic weight calculation unit: used to generate dynamic weights for the first-level indicators of the five-dimensional indicator system through the corresponding intelligent agents; and to generate dynamic weights for the second-level indicators of the five-dimensional indicator system through the corresponding intelligent agents;

[0157] Expert weight calibration unit: used to generate respective expert time series weights for the first-level indicators of the five-dimensional indicator system through the analytic hierarchy process; and to generate respective expert time series weights for the second-level indicators of the five-dimensional indicator system through the analytic hierarchy process;

[0158] Weight fusion engine: used to calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data;

[0159] Hybrid model inference unit: used to obtain an LSTM-GNN-Expert hybrid model, which includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module. The hybrid model inputs multidimensional time series data, network structure data, and expert unstructured annotations, and outputs a time series feature vector, a network structure feature vector, and an expert correction vector, which are input into the gated fusion module. The gated fusion module performs a weighted summation on each network node of the water conservancy project communication network and outputs a comprehensive score.

[0160] LSTM layer: 2 layers of bidirectional LSTM, 256 units per layer, process the past 120 hours of time series data and output the time series feature vector h_LSTM (dimension = 256). The five-dimensional evaluation system of the water conservancy project communication network and the fusion weights of each indicator are organized as part of the multidimensional time series data;

[0161] The GNN layer uses the GraphSAGE algorithm to capture network structural characteristics by aggregating neighbor node features (sample count = 7) and outputs a structural feature vector h_GNN (dimension = 256). The LSTM layer processes time series data from the past 120 hours and supports high-frequency data input in seconds (sampling frequency ≥ 10 Hz).

[0162] BERT semantic layer: Input expert-annotated text, obtain a 5-dimensional weight correction vector h_expert (dimension = 5, value range [-0.15, 0.15]), and finally output the expert correction vector.

[0163] The gated fusion module includes a splicing layer, a weight calculation layer and a weighted summation layer;

[0164] Splicing layer: used to splice the time series feature vector and the structural feature vector to form a joint feature h containing time series dynamics and network structure information LG , h LG =[h LSTM ;h GNN ] dimension 512, representing the joint hidden state of the LSTM layer and the GNN layer;

[0165] Weight calculation layer: by splicing the joint hidden state h of the LSTM layer and the GNN layer LG Combined with the expert semantic hidden state, the fusion weight σ is generated through full connection and sigmoid activation function to achieve soft-gated fusion of dynamic scoring and expert scoring;

[0166]

[0167] Weighted summation layer: Generates a comprehensive score through a gated fusion formula; the gated fusion formula is as follows:

[0168] Score=σ(W g ·[h LG ;h erpert ])·LGScore+(1-σ)·ExpertScore

[0169] Among them, W g represents the weight matrix of gated fusion; h expert Represents the processing result of expert-annotated text from the BERT semantic layer; LGScore: represents the preliminary score generated by the LSTM-GNN model in the LSTM-GNN-Expert hybrid model based on time series data and network structure features, that is, the dynamic score; ExpertScore: represents the score generated based on expert semantic knowledge, that is, the expert score.

[0170] The dynamic score LGScore is the LSTM-GNN output (0-100), and the expert score ExpertScore is the expert experience score (initial value = 80, quarterly calibration update).

[0171] Expert Collaboration Platform: This platform provides a scenario-based weight calibration interface, enabling experts to complete quarterly questionnaires through the WeChat mini-program. The platform also uses the built-in BERT semantic encoder to convert unstructured annotations in the quarterly questionnaires into weight correction vectors. These weight correction vectors are used by the expert weight calibration unit to generate expert time-series weights for each indicator, and are also used to obtain expert correction vectors in the hybrid model inference unit.

[0172] Early warning decision module: Execute four levels of early warning based on the comprehensive score. In this embodiment, four levels of early warning are implemented based on the comprehensive score (green ≥ 90, yellow 70-89, orange 50-69, red <50). The comprehensive score is <50 and the expert marks it as "high risk". When the dynamic score deviates from the expert score by >20%, the expert intervention mechanism is triggered. The WeChat mini-program review request is automatically pushed, and the expert response time is <30 minutes, forming a collaborative process of "model initial judgment-expert calibration-closed-loop optimization".

[0173] The above is only an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme will not be described in detail here. Those of ordinary skill in the art are aware of all common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Those of ordinary skill in the art can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for those of ordinary skill in the art to implement this application. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the structure of the present invention. These should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for evaluating the quality of nodes in a water conservancy project communication network, characterized in that: The following steps are involved: Steps for constructing a five-dimensional indicator system: Establish a five-dimensional evaluation system that includes network structure characteristics, functional performance, security and reliability, economic cost, and intelligent optimization potential as first-level indicators; the first-level indicators include several second-level indicators; Dynamic weight calculation steps: The first-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights; the second-level indicators of the five-dimensional indicator system are respectively generated through the corresponding intelligent agents to generate their respective dynamic weights; Expert weight calculation steps: The first-level indicators of the five-dimensional indicator system are respectively generated through the hierarchical analysis method to generate their respective expert time-series weights; the second-level indicators of the five-dimensional indicator system are respectively generated through the hierarchical analysis method to generate their respective expert time-series weights; Weight fusion calculation steps: Calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data; Hybrid model evaluation step: Obtain an LSTM-GNN-Expert hybrid model, which includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module; The system inputs multi-dimensional time series data, network structure data, and expert unstructured annotations, and outputs a time series feature vector, a network structure feature vector, and an expert correction vector. These are then input into a gated fusion module, which then performs a weighted summation of the network nodes of the water conservancy project communication network and outputs a comprehensive score. Decision-making and early warning steps: Make decision-making assessment and early warning based on the comprehensive score.

2. A water conservancy project communication network node quality assessment method according to claim 1, characterized in that: The steps for generating the dynamic weight in the dynamic weight calculation step are as follows: Generate the basic weight of each dynamic weight of the first-level indicator or the second-level indicator through the intelligent agent; Generate the optimized weights of the dynamic weights of the primary indicators or secondary indicators through the Q-learning intelligent agent optimization algorithm; Adjust the time decay coefficient of the weight through the time decay algorithm; The multi-head self-attention mechanism is used to capture the coupling relationship between indicators and obtain the attention correction weight; The dynamic weight is obtained according to the basic weight, optimization weight, time decay coefficient and attention correction weight.

3. A water conservancy project communication network node quality assessment method according to claim 2, characterized in that: The steps for generating the expert time series weight in the expert weight calculation step are as follows: Based on the hierarchical analysis method, the working condition correction coefficient of the weight is obtained according to the working condition label to obtain the initial weight; The expert calibration text is converted into a weight correction vector through BERT, and the initial weight is updated to obtain the expert time series weight; Adjust the time decay coefficient of the initial weight through the time decay algorithm And update the expert timing weight; Update expert time series weights smoothly based on historical calibration data through Markov chain; Set the weight ratio of the expert's time series weight. The weight ratio formula is: is the time decay coefficient.

4. A water conservancy project communication network node quality assessment method according to claim 3, characterized in that: The credibility weighting formula in the weight fusion calculation step is as follows: Data credibility = 1-percentage of abnormal data; expert activity is the quarterly questionnaire completion rate; in, is the fusion weight, α is the dynamic weight ratio, and the ratio range is (0.6-0.8). is the dynamic weight, is the expert temporal weight, α expert The proportion range is (0.2-0.4).

5. A water conservancy project communication network node quality assessment method according to claim 4, characterized in that: The gated fusion module includes a splicing layer, a weight calculation layer and a weighted summation layer; Splicing layer: used to splice the time series feature vector and the structural feature vector to form a joint feature h containing time series dynamics and network structure information LG , represents the joint hidden state of the LSTM layer and the GNN layer; Weight calculation layer: by splicing the joint hidden state h of the LSTM layer and the GNN layer LG Combined with the expert semantic hidden state, the fusion weight σ is generated through full connection and sigmoid activation function to achieve soft-gated fusion of dynamic scoring and expert scoring; Weighted summation layer: Generates a comprehensive score through a gated fusion formula; the gated fusion formula is as follows: Score=σ(W g ·[h LG ;h erpert ])·LGScore+(1-σ)·ExpertScore Among them, W g represents the weight matrix of gated fusion; h expert Represents the processing result of the expert-annotated text from the BERT semantic layer; LGScore: represents the preliminary score generated by the LSTM-GNN model in the LSTM-GNN-Expert hybrid model based on time series data and network structure features; ExpertScore: represents the score generated based on expert semantic knowledge.

6. A water conservancy project communication network node quality assessment method according to claim 1, characterized in that: The first-level indicators include network structure characteristics, functional performance, security and reliability, economic cost and intelligent optimization potential; the network structure characteristics include chain topology position, ring structure participation, star centrality, structural vulnerability, number of cross-substructure connections, and path uniqueness dependence, a total of 6 second-level indicators; functional performance includes data transmission volume, bandwidth redundancy, real-time requirements, and multi-service compatibility, a total of 4 second-level indicators; security and reliability include anti-attack capability, redundancy, fault recovery time, and historical failure rate, a total of 4 second-level indicators; economic cost includes construction cost, operation and maintenance cost, and upgrade cost, a total of 3 second-level indicators; intelligent optimization potential includes data support capability, algorithm adaptability, and scalability, a total of 3 second-level indicators.

7. A water conservancy project communication network node quality assessment method according to claim 6, characterized in that: Also includes: Edge computing steps: Deploy edge computing nodes in cities and prefectures to collect and preprocess data obtained by the sensor layer in real time to obtain several first-level and second-level indicators.

8. A water conservancy project communication network node quality assessment method according to claim 2, characterized in that: In the step of generating the optimized weights of the dynamic weights of the primary indicators or the secondary indicators by the Q-learning agent optimization algorithm, the step of obtaining the optimized weights is as follows: Obtain real-time indicator data for primary or secondary indicators and obtain optimized weights through reward functions; Reward function = |prediction score - expert quarterly calibration value| - 1 Prediction score: machine score calculated by the agent in real time, expert quarterly calibration value: manual benchmark score by domain experts.

9. A water conservancy project communication network node quality assessment system, characterized in that: A method for evaluating the quality of a water conservancy project communication network node according to any one of claims 1 to 8 is used, comprising: Sensor layer: used to collect sensor data of the water conservancy project communication network through a number of sensors; Edge computing module: used to deploy edge computing nodes in cities and prefectures to collect and pre-process sensor data in real time; Core evaluation module: includes a five-dimensional indicator system construction unit, a dynamic weight calculation unit, an expert weight calibration unit, a weight fusion engine, and a hybrid model reasoning unit; Five-dimensional indicator system construction unit: used to construct a five-dimensional indicator system based on preprocessed sensor data; Dynamic weight calculation unit: used to generate dynamic weights for the first-level indicators of the five-dimensional indicator system through the corresponding intelligent agents; and to generate dynamic weights for the second-level indicators of the five-dimensional indicator system through the corresponding intelligent agents; Expert weight calibration unit: used to generate respective expert time series weights for the first-level indicators of the five-dimensional indicator system through the analytic hierarchy process; and to generate respective expert time series weights for the second-level indicators of the five-dimensional indicator system through the analytic hierarchy process; Weight fusion engine: used to calculate the fusion weight of the dynamic weight of the same indicator and the expert time series weight through the credibility weighting formula; organize the five-dimensional evaluation system of the water conservancy project communication network and the fusion weight of each indicator into multi-dimensional time series data; Hybrid model inference unit: used to obtain an LSTM-GNN-Expert hybrid model, which includes an LSTM layer for processing multidimensional time series data, a GNN layer for processing network structure data of the water conservancy project communication network, a BERT semantic layer for processing expert-annotated text, and a gated fusion module. The hybrid model inputs multidimensional time series data, network structure data, and expert unstructured annotations, and outputs a time series feature vector, a network structure feature vector, and an expert correction vector, which are input into the gated fusion module. The gated fusion module performs a weighted summation on each network node of the water conservancy project communication network and outputs a comprehensive score. Early warning decision module: Execute four-level early warning based on the comprehensive score.

10. A water conservancy project communication network node quality assessment system according to claim 9, characterized in that: Also includes: Expert Collaboration Platform: Used to provide a scenario-based weight calibration interface, enabling experts to complete quarterly questionnaires through the WeChat mini-program, and convert the expert unstructured annotations in the quarterly questionnaire into weight correction vectors through the built-in BERT semantic encoder. The weight correction vectors are used by the expert weight calibration unit to generate expert time-series weights for each indicator, and are used to obtain expert correction vectors in the hybrid model inference unit.

Images