IoT fire-fighting spraying pipe network intelligent micro-leakage alarm method and system based on time period pressure drop and pressure fluctuation analysis

Abstract

The invention discloses an IoT fire-fighting spraying pipe network intelligent micro-leakage alarm method and system based on time period pressure drop and pressure fluctuation analysis. According to the method, a multi-node IoT pressure sensor is deployed to collect pipe network pressure data in real time, and a pressure drop-fluctuation joint detection model is constructed by adopting frequency domain analysis combining dynamic time period division and pressure drop rate calculation with wavelet transform. Multi-modal feature fusion is realized by using a Transform-LSTM-GNN hybrid model, graded alarm is triggered through a dynamic threshold value, and a leakage point is positioned based on a pipe network topology and a pressure gradient. The system comprises a data acquisition module, a feature analysis module, an AI model training module, an alarm decision module, a maintenance scheme generation module and the like, and a'monitoring-analysis-alarm-maintenance-optimization 'closed loop is formed. Time period dynamic division is realized through a DTW-BIRCH algorithm, and the model is optimized through a multi-task loss function joint; topological feature fusion is enhanced through a graph attention network; and the detection precision is continuously improved through an incremental learning mechanism, and early-stage accurate identification of tiny leakage can be realized.

Description

Technical Field

[0001] This invention relates to the field of fire sprinkler network monitoring technology, and in particular to an IoT-based intelligent micro-leakage alarm method and system for fire sprinkler networks based on time-period pressure drop and pressure fluctuation analysis, applicable to micro-leakage monitoring of fire water-filled networks such as wet sprinkler systems and fire hydrant systems. Background Technology

[0002] In recent years, with the acceleration of urbanization, the safety and reliability of fire sprinkler systems, as important fire protection infrastructure, have received increasing attention. However, during long-term operation, these systems are prone to micro-leakage due to material aging, external damage, or installation defects. If not detected in time, this can lead to serious safety accidents. Traditional leak detection methods mainly rely on manual inspections or fixed threshold alarms, which suffer from problems such as delayed response, high false alarm rates, and insufficient location accuracy.

[0003] Currently, while pressure monitoring-based leak detection technology has some applications, it faces the following technical bottlenecks: Insufficient data utilization: Existing methods mostly rely on a single pressure index and fail to effectively combine the correlation between time-domain pressure drop and frequency-domain fluctuation characteristics; Poor model generalization: Traditional machine learning models are difficult to adapt to complex pipeline topologies and dynamic operating condition changes; Rigid alarm mechanisms: Fixed threshold alarms cannot distinguish between environmental interference and actual leaks, leading to false alarms and missed alarms; Lack of operation and maintenance closed loop: Detection, location, and maintenance are disconnected, lacking intelligent decision support.

[0004] While some studies have attempted to incorporate AI algorithms, these have largely been limited to single tasks (such as leak classification) and have failed to address core issues such as multimodal feature fusion, model interpretability, and end-to-end closed-loop optimization. Therefore, there is an urgent need for a micro-leak detection method that integrates multi-source data, possesses adaptive capabilities, and supports intelligent decision-making throughout the entire process to improve the safety and maintenance of fire protection pipe networks. Summary of the Invention

[0005] This invention proposes an intelligent micro-leakage alarm method and system for IoT fire sprinkler networks based on time-period pressure drop and pressure fluctuation analysis. By dynamically analyzing water pressure change characteristics, it achieves accurate monitoring of minute leaks.

[0006] In a first aspect, embodiments of this application provide an intelligent micro-leakage alarm method for IoT fire sprinkler networks based on time-period pressure drop and pressure fluctuation analysis, the method comprising:

[0007] S1. By deploying IoT pressure sensors at key nodes of the fire sprinkler network, real-time pressure data of the network is collected and transmitted to the cloud or edge computing platform.

[0008] S2. Divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the pressure drop rate for each time period.

[0009] S3. Use wavelet transform to perform frequency domain analysis on pressure data, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation.

[0010] S4. A large model is used to train the pressure data under historical normal and leakage conditions to construct a pressure drop-fluctuation joint detection model;

[0011] S5. Input the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, trigger an intelligent alarm and locate potential leak points through the visualization platform.

[0012] S6. Based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, intelligently analyze the location and cause of leakage, and generate a maintenance plan.

[0013] S7. Continuously optimize AI model parameters based on feedback data and link with pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

[0014] Optionally, in one implementation of the first aspect of the present invention, the acquisition in S1 includes:

[0015] Three levels of pressure monitoring nodes are set up at the main pipeline, the junction of branch pipes and the end of the pipeline network. Each node is equipped with an IoT pressure sensor with self-calibration function.

[0016] The sensor uses a dual-mode communication protocol of LoRaWAN and 5G to collect pressure data in real time at a sampling frequency of 1Hz;

[0017] The raw data is preprocessed through edge computing nodes, including: using sliding window filtering to eliminate transient interference; achieving multi-node data synchronization through timestamp alignment; and performing Mahalanobis distance detection and interpolation repair for outliers.

[0018] The preprocessed data is compressed and packaged, and then uploaded to the cloud data lake via the MQTT protocol.

[0019] A spatiotemporal database is established in the cloud, and pipeline topology tags are attached to the pressure data to form a structured spatiotemporal sequence dataset.

[0020] Optionally, in one implementation of the first aspect of the present invention, step S2, dividing the collected pressure data into preset time periods, calculating the average pressure value within each time period, comparing it with historical reference pressure, and calculating the pressure drop rate for each time period, includes:

[0021] S2.1. The DTW-BIRCH hybrid clustering algorithm is used to dynamically divide historical pressure data into time periods, including: measuring data similarity through dynamic time warping, constructing a time period feature tree, and generating the optimal time period division scheme by combining water use patterns.

[0022] S2.2 Perform multi-level feature extraction within each time period, including: calculating the arithmetic mean stress value and standard deviation of the base layer, the weighted average stress value of the topology layer, and the stress change gradient of the trend layer;

[0023] S2.3. Generate a dynamic baseline pressure curve based on the Transformer-TSA model. This model processes historical pressure data, environmental parameters and maintenance records through a time attention mechanism and outputs a dynamic baseline that considers multiple factors.

[0024] S2.4. An adaptive pressure drop rate calculation model is adopted, which integrates the base pressure drop rate, stability correction factor and topology correction factor, and dynamically adjusts the weight of each factor through an online learning mechanism.

[0025] S2.5 Establish a closed-loop feedback optimization system to evaluate data quality, computational reliability, and anomalies, and feed the evaluation results back to the time period division and benchmark generation modules in real time.

[0026] Optionally, in one implementation of the first aspect of the present invention, step S3, performing frequency domain analysis on the pressure data using wavelet transform to extract high-frequency fluctuation features and calculating the pressure fluctuation intensity by combining the time domain standard deviation, includes:

[0027] S3.1. An adaptive wavelet basis selection algorithm is used to match the optimal wavelet basis function, and the pressure signal is decomposed into three levels to extract high-frequency detail coefficients and low-frequency approximation coefficients, and a wavelet energy spectrum feature matrix is ​​established.

[0028] S3.2. Through the time-frequency domain fusion analysis engine, the pressure standard deviation and peak-to-peak value in the time domain are comprehensively calculated, as well as the high-frequency energy ratio, characteristic frequency band energy ratio and wavelet entropy in the frequency domain, and the attention mechanism is used to dynamically weight the features.

[0029] S3.3 Construct a volatility intensity index, where the weight coefficients are obtained through training with historical data, and establish a dynamic threshold adjustment mechanism;

[0030] S3.4 Set up a closed-loop optimization system to continuously optimize the analysis parameters through feature quality assessment and online learning modules, and feed the results back to the time period segmentation module.

[0031] Optionally, in one implementation of the first aspect of the present invention, step S4, training a large model on historical pressure data under normal and leakage conditions to construct a pressure drop-fluctuation joint detection model, includes:

[0032] S4.1 Constructing a multimodal feature fusion and hybrid model architecture: After aligning the time-period pressure drop features, pressure fluctuation features and pipeline topology features in time and space, a joint feature representation is generated through a feature cross-network;

[0033] An attention mechanism is used to dynamically weight the importance of each feature dimension and input it into a Transformer-LSTM-GNN hybrid model. The Transformer layer processes the aligned long-term pressure change pattern, the LSTM layer extracts the local temporal characteristics of the joint features, and the graph neural network layer parses the pipeline topology relationship.

[0034] S4.2 Establish an interpretable incremental learning system: Integrate a SHAP value interpretation module at the output of the hybrid model to generate a feature contribution heatmap in real time; establish a rule extraction engine based on heatmap analysis to output interpretable judgment logic; design an online update pipeline that includes sample importance weighting, and achieve continuous model optimization without forgetting through elastic weight solidification technology;

[0035] S4.3 Forming an end-to-end joint optimization closed loop: Design a multi-task loss function that integrates voltage drop detection and fluctuation analysis to drive the collaborative training of the hybrid model; feed back the performance indicators output by the model to the front-end acquisition system to dynamically adjust the sensor sampling strategy; trigger an automatic rollback mechanism through the model performance monitoring module to ensure system stability.

[0036] Optionally, in one implementation of the first aspect of the present invention, the multimodal feature fusion and hybrid model architecture is implemented through the following interrelated technical solutions:

[0037] Constructing a feature interaction fusion network: First, the voltage drop feature F is calculated through a feature cross-layer. drop With fluctuation characteristics F wave Interactive items: in, σ represents the outer product operation, used to generate higher-order feature combinations; W1 represents the learnable weight matrix; and σ represents the activation function, used to introduce nonlinearity.

[0038] Combine the interaction results with the topological feature F topo The secondary fusion is performed using the following formula:

[0039] F fused =LayerNorm(F cross1 +GAT(F topo ,F cross1 )),

[0040] Among them, F fused This indicates that the interaction result F will be... cross1 With topological feature F topoBy fusing GAT with a graph attention network, importance weights between network nodes are dynamically allocated, and LayerNorm represents layer normalization.

[0041] Stabilize the training process;

[0042] Achieving dynamic feature optimization: The feature importance weight α is calculated using a multilayer perceptron (MLP), with the following formula:

[0043] α = Softmax(MLP([F drop ||F wave ||F topo ])),

[0044] Where || represents the concatenation operation, and Softmax represents normalization to a probability distribution to ensure that the sum of the weights is 1;

[0045] The final weighted feature representation is as follows: Based on the calculated weights, the pressure drop feature, fluctuation feature, and topological feature are weighted and summed, as shown in the formula:

[0046] F final =α drop ·F drop +α wave ·F wave +α topo ·F topo ,

[0047] Among them, F final This indicates that the weighted summation yields the final fusion feature, α. drop α wave α topo These represent the weights corresponding to the pressure drop characteristic, fluctuation characteristic, and topological characteristic, respectively, given by α = Softmax(MLP([F drop ||F wave ||F topo The generation of each modality reflects its importance to the current task.

[0048] Deploying a hybrid processing architecture: Transformer layer processing flow:

[0049] F trans =TransformerEncoder(F final ,positional_ending(x,y,z)),

[0050] Specifically, the Transformer model is used to process sequence data, and the Transformer layer processes the fused features F. finalTo capture long-term pressure change patterns, positional_ending(x,y,z) represents injecting the spatial coordinates (x,y,z) of the pipeline network and preserving the topological location information;

[0051] LSTM layer processing flow: F lstm =BiLSTM(F trans ,t win It uses a bidirectional long short-term memory network to process temporal information, and a bidirectional LSTM to extract local temporal features, with a window length of t. win Control the scope of short-term dependence;

[0052] Graph network layer processing flow: F gnn =GraphsomorphismNetwork(F lstm ,edge_attr), utilize graph isomorphic networks to process graph-structured data, F gnn This indicates that the graph isomorphic network handles the pipe network topology relationship, and edge_attr contains pipe segment attributes;

[0053] Implement joint optimization: The total loss function includes the task loss L task Reconstruction loss L recon and orthogonal constraint L orth The formula is:

[0054] L total =L task +λL recon +μL orth ,

[0055] Among them, the task loss L task Used to supervise the model in completing the main task of leak detection, and to reconstruct the loss L. recon Orthogonal constraint L is used to ensure the reversibility of feature fusion. orth To force different modal features to remain independent and avoid redundancy, λ and μ represent the weights of the hyperparameter balancing loss term.

[0056] Optionally, in one implementation of the first aspect of the present invention, step S5 involves inputting the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds a preset threshold, an intelligent alarm is triggered, and potential leakage points are located through a visualization platform, including:

[0057] Data cleaning is performed on the input real-time pressure drop rate and fluctuation intensity, including outlier removal and smoothing. The pre-processed data is spatiotemporally aligned with the pipeline topology information, and the current operating condition feature vector is extracted, including time period features, environmental parameters and equipment status.

[0058] Input the feature vector into the trained AI model, and output the leakage probability value;

[0059] Multi-level alarm judgment is performed by combining a dynamic threshold library, including early warning, confirmed alarm and emergency alarm;

[0060] A delayed confirmation mechanism is activated for critical alarm states to avoid false alarms caused by momentary interference;

[0061] The potential leakage area was calculated based on pressure gradient changes and a pipeline hydraulic model.

[0062] Mark potential leak points and confidence levels on a 3D digital twin model of the pipeline network;

[0063] Generate a visual report that includes location results, risk level, and scope of impact;

[0064] Automatically push alarm information to the mobile terminals of relevant personnel;

[0065] Record the operator's response and the on-site confirmation results;

[0066] The verified alarm cases are fed back to the model training system for incremental learning.

[0067] Optionally, in one implementation of the first aspect of the present invention, step S6, based on the output of the AI ​​model, combines historical leakage data and pipeline topology to intelligently analyze the location and cause of leakage and generate a maintenance plan, including:

[0068] Integrate the leakage probability distribution output by the AI ​​model with the pipeline topology data, associate it with similar operating condition characteristics in the historical leakage case library, and overlay equipment operation logs and maintenance record data;

[0069] The possible leakage area is calculated based on the pressure gradient propagation model, the correlation of pipeline nodes is analyzed using graph neural network, the most likely leakage point and confidence level are determined by Bayesian inference, and the potential causes of leakage are analyzed by combining the material aging model.

[0070] The system automatically retrieves the pipe section design parameters based on the leak location, assesses the leak's impact range and risk level, matches the maintenance resource database to generate alternative solutions, and calculates the cost, time consumption, and expected effects of each solution.

[0071] Generates reports containing location results, cause analysis, and maintenance recommendations; marks leak points and affected areas in the 3D pipeline network model; provides a maintenance solution comparison interface and recommendations; and supports maintenance process tracking and result feedback.

[0072] Secondly, embodiments of this application provide an IoT fire sprinkler network intelligent micro-leakage alarm system based on time-period pressure drop and pressure fluctuation analysis, applied to the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in the first aspect, the system comprising:

[0073] The data acquisition module is used to collect real-time pipeline pressure data through IoT pressure sensors deployed at key nodes of the fire sprinkler network and transmit it to the cloud or edge computing platform.

[0074] The time period division and pressure drop calculation module is used to divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the pressure drop rate of the time period.

[0075] The pressure fluctuation analysis module is used to perform frequency domain analysis on pressure data using wavelet transform, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation.

[0076] The AI ​​model training module is used to train a large model on historical pressure data under normal and leakage conditions to build a pressure drop-fluctuation joint detection model.

[0077] The micro-leak detection and alarm module is used to input real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, an intelligent alarm is triggered, and potential leak points are located through a visualization platform.

[0078] The Leakage Location and Repair Decision Module is used to intelligently analyze the location and cause of leakage based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, and generate a repair plan.

[0079] The dynamic optimization and equipment control module is used to continuously optimize the AI ​​model parameters based on feedback data and link with the pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

[0080] Thirdly, embodiments of this application provide an electronic device, including:

[0081] processor;

[0082] Memory used to store processor-executable instructions;

[0083] The processor is configured to implement the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in the first aspect when executing the instructions.

[0084] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in the first aspect.

[0085] This invention discloses an intelligent micro-leak alarm method and system for IoT fire sprinkler networks based on time-period pressure drop and pressure fluctuation analysis. The method deploys multi-node IoT pressure sensors to collect network pressure data in real time, employing dynamic time-period division and pressure drop rate calculation combined with wavelet transform frequency domain analysis to construct a joint pressure drop-fluctuation detection model. A Transformer-LSTM-GNN hybrid model is used to achieve multi-modal feature fusion, triggering graded alarms through dynamic thresholds, and locating leak points based on network topology and pressure gradients. The system includes modules for data acquisition, feature analysis, AI model training, alarm decision-making, and maintenance plan generation, forming a closed loop of "monitoring-analysis-alarm-maintenance-optimization." Its innovations are reflected in: 1) dynamic time-period division using the DTW-BIRCH algorithm; 2) a multi-task loss function joint optimization model; 3) graph attention network-enhanced topology feature fusion; and 4) an incremental learning mechanism to continuously improve detection accuracy. This method can achieve early and accurate identification of micro-leakage, with a location error within ±3 meters and a false alarm rate of less than 2%, significantly improving the efficiency of network safety operation and maintenance. Attached Figure Description

[0086] Figure 1 This is a schematic diagram of the process of an IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis, provided as an embodiment of this application.

[0087] Figure 2 The present application provides a process for constructing a pressure drop-fluctuation joint detection model according to an embodiment of this application.

[0088] Figure 3 This is a diagram of a multimodal feature fusion and hybrid model architecture provided in an embodiment of this application.

[0089] Figure 4 This application provides an embodiment of an IoT-based intelligent micro-leakage alarm system architecture for fire sprinkler networks, based on time-period pressure drop and pressure fluctuation analysis.

[0090] Figure 5 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0091] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0092] It should be noted that, in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0093] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features specified as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0094] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0095] Example 1

[0096] Figure 1 This is a schematic flowchart illustrating an embodiment of the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis. Figure 1 As shown, an intelligent micro-leakage alarm method for IoT fire sprinkler networks based on time-period pressure drop and pressure fluctuation analysis includes:

[0097] S1. By deploying IoT pressure sensors at key nodes of the fire sprinkler network, real-time pressure data of the network is collected and transmitted to the cloud or edge computing platform.

[0098] Specifically, in this embodiment, the data collection in S1 includes:

[0099] Three levels of pressure monitoring nodes are set up along the main pipeline, at the junctions of branch pipes, and at the end of the pipeline network. Each node is equipped with an IoT pressure sensor with self-calibration function. The IoT sensor is used to collect key parameters such as water pressure and flow rate in real time.

[0100] The sensor employs a dual-mode communication protocol combining LoRaWAN and 5G, acquiring pressure data in real time at a sampling frequency of 1Hz. LoRaWAN is a low-power wide-area network (LPWA) technology suitable for low-power, wide-area IoT applications.

[0101] Edge computing nodes perform preprocessing on raw data, including: using sliding window filtering to eliminate transient interference; achieving multi-node data synchronization through timestamp alignment; and performing Mahalanobis distance detection and interpolation repair for outliers. Edge computing is used in IoT systems to reduce the burden on the cloud and improve real-time performance.

[0102] The preprocessed data is compressed and packaged, then uploaded to the cloud data lake via the MQTT protocol. A spatiotemporal database is established in the cloud, and pipeline topology tags are attached to the pressure data to form a structured spatiotemporal sequence dataset. The MQTT protocol is widely used for IoT data transmission.

[0103] S2. Divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the pressure drop rate for each time period.

[0104] Specifically, in this embodiment, step S2 involves dividing the collected pressure data into preset time periods, calculating the average pressure value within each time period, comparing it with historical baseline pressure, and calculating the pressure drop rate for each time period, including:

[0105] S2.1 A DTW-BIRCH hybrid clustering algorithm is used to dynamically divide historical stress data into time periods. This includes: measuring data similarity through dynamic time warping, constructing a time period feature tree, and generating the optimal time period division scheme by combining water usage patterns. DTW (Dynamic Time Warping) is an algorithm for time series analysis used to measure the similarity between two time series, and is widely used in speech recognition, time series pattern recognition, and other fields. BIRCH (Balanced Iterative Reducing and Clustering) is a clustering algorithm used to efficiently process large-scale datasets. The hybrid method combining DTW and BIRCH can be used for time series clustering. Similarity is measured using DTW, and the time period division is generated by combining it with the clustering algorithm.

[0106] Specifically, the time period identification module based on pipeline network operation characteristics adopts an improved DTW-BIRCH hybrid clustering algorithm, which is implemented through the following steps: a) Dynamic Time Warping (DTW) similarity measurement of historical pressure data streams; b) Construction of time period feature trees using hierarchical clustering (BIRCH); c) Automatic generation of optimal time period division scheme by combining water use pattern recognition; The real-time monitoring module continuously tracks the pressure change trend, and triggers the time period re-division mechanism when the detected change in water use pattern exceeds the threshold.

[0107] S2.2. Perform multi-level feature extraction within each time period, including: calculating the arithmetic mean and standard deviation of the base layer, the weighted average pressure value of the topology layer, and the pressure change gradient of the trend layer. Through feature extraction at different levels, the dynamic changes of the pressure data can be described more comprehensively.

[0108] Specifically, the multi-level pressure drop feature extraction includes the following steps: Within each defined time period, a three-level feature extraction architecture is used to calculate: a) Base layer: calculate the arithmetic mean pressure value (P_mean) and standard deviation (P_std); b) Topology layer: calculate the weighted average pressure value (P_weighted) based on the weight coefficients of the pipeline nodes; c) Trend layer: calculate the pressure change gradient (P_gradient) using the sliding window technique; the feature extraction process is updated synchronously with the time period division results in real time.

[0109] S2.3. A dynamic baseline stress curve is generated based on the Transformer-TSA model. This model processes historical stress data, environmental parameters, and maintenance records through a time attention mechanism, outputting a dynamic baseline that considers multiple factors. The Transformer model is widely used in natural language processing and time series analysis. Its self-attention mechanism can handle long-series data and is suitable for processing time series data. TSA (Time Series Analysis) is a common method for time series analysis, and combining it with the Transformer model can improve time series forecasting and modeling capabilities.

[0110] Specifically, the dynamic baseline stress generation engine includes: constructing a time series prediction model based on an attention mechanism (Transformer-TSA), and generating a dynamic baseline through the following steps: a) inputting historical stress data, environmental parameters, and maintenance records; b) capturing long-term dependencies through a time attention layer; c) outputting a dynamic baseline stress curve that considers multiple factors; the baseline generation module shares feature extraction results with the time period segmentation system.

[0111] S2.4. An adaptive pressure drop rate calculation model is adopted, which integrates the base pressure drop rate, stability correction factor, and topology correction factor, and dynamically adjusts the weights of each factor through an online learning mechanism. Through this online learning mechanism, the model can adjust parameters in real time to adapt to data changes.

[0112] Specifically, the intelligent pressure drop rate calculation framework is designed as follows: an adaptive pressure drop rate calculation model is designed, consisting of: a) base pressure drop rate, b) stability correction factor, c) topology correction factor, and d) final pressure drop rate; the weights of each correction factor are dynamically adjusted through an online learning mechanism.

[0113] S2.5 Establish a closed-loop feedback optimization system to evaluate data quality, computational reliability, and anomalies, and feed the evaluation results back to the time period segmentation and benchmark generation modules in real time. Through the feedback mechanism, the system can continuously optimize the model and parameters, thereby improving system performance.

[0114] Specifically, the closed-loop feedback optimization system may include: establishing a computation result quality assessment module: data quality detection: assessing the integrity and reliability of the original data; computation credibility scoring: based on feature consistency and historical consistency; anomaly detection: identifying potential deviations in the computation process; and real-time feedback of the assessment results to the time period division and benchmark generation module to form a closed-loop optimization.

[0115] S3. Use wavelet transform to perform frequency domain analysis on the pressure data, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation.

[0116] Specifically, in this embodiment, step S3, performing frequency domain analysis on the pressure data using wavelet transform to extract high-frequency fluctuation features and calculating the pressure fluctuation intensity by combining the time domain standard deviation, includes:

[0117] S3.1. An adaptive wavelet basis selection algorithm is used to match the optimal wavelet basis function, and the pressure signal is decomposed into three levels to extract high-frequency detail coefficients and low-frequency approximation coefficients, and a wavelet energy spectrum feature matrix is ​​established; the signal is decomposed into multiple scales through wavelet basis functions to extract features of different frequency components.

[0118] S3.2 Through a time-frequency domain fusion analysis engine, the system comprehensively calculates the pressure standard deviation and peak-to-peak value in the time domain, as well as the high-frequency energy proportion, characteristic frequency band energy ratio, and wavelet entropy in the frequency domain, and employs an attention mechanism for dynamic feature weighting. This combines time-domain and frequency-domain analysis methods, integrating features extracted through wavelet transform with traditional time-domain statistics (such as standard deviation) to enhance feature expressiveness. The introduction of the attention mechanism helps to dynamically adjust feature weights to adapt to signal changes under different operating conditions.

[0119] S3.3 Construct a fluctuation intensity index, where the weight coefficients are obtained through training on historical data, and establish a dynamic threshold adjustment mechanism. Using machine learning and adaptive algorithms, a model is trained on historical data to determine the weights of features, and the threshold is dynamically adjusted to adapt to signal changes under different operating conditions. This mechanism helps improve the system's adaptability and robustness.

[0120] S3.4 A closed-loop optimization system is set up to continuously optimize the analysis parameters through feature quality assessment and online learning modules, and the results are fed back to the time period segmentation module. By utilizing closed-loop control and online learning technologies, and through continuous optimization of analysis parameters and feedback mechanisms, the system's performance and adaptability are improved.

[0121] S4. A large model is used to train pressure data under historical normal and leakage conditions to construct a pressure drop-fluctuation joint detection model.

[0122] Specifically, Figure 2 This document describes a process for constructing a joint detection model for pressure drop and fluctuation, as provided in one embodiment of this application. Figure 2 As shown, in this embodiment, step S4 involves training a large model on historical pressure data under normal and leakage conditions to construct a pressure drop-fluctuation joint detection model, including:

[0123] S4.1 Constructing a multimodal feature fusion and hybrid model architecture: After aligning the time-period pressure drop features, pressure fluctuation features and pipeline topology features in time and space, a joint feature representation is generated through a feature cross-network;

[0124] An attention mechanism is employed to dynamically weight the importance of each feature dimension, which is then input into a hybrid Transformer-LSTM-GNN model. The Transformer layer handles aligned long-term stress variation patterns, the LSTM layer extracts local temporal characteristics of joint features, and the graph neural network layer parses network topology relationships. Multimodal data fusion integrates data from different sources, improving the model's expressive power. The Transformer layer handles long-term dependencies, the LSTM layer captures local temporal characteristics, and the graph neural network layer parses complex topological relationships. This hybrid model architecture effectively handles complex data structures and task requirements.

[0125] Specifically, Figure 3 A diagram of a multimodal feature fusion and hybrid model architecture provided in an embodiment of this application is shown below. Figure 3 As shown, the multimodal feature fusion and hybrid model architecture is implemented through the following interrelated technical solutions:

[0126] Constructing a feature interaction fusion network: First, the voltage drop feature F is calculated through a feature cross-layer. drop With fluctuation characteristics F wave Interactive items: in, σ represents the outer product operation, used to generate higher-order feature combinations; W1 represents the learnable weight matrix; and σ represents the activation function, used to introduce nonlinearity.

[0127] Combine the interaction results with the topological feature F topo The secondary fusion is performed using the following formula:

[0128] F fused =LayerNorm(F cross1+GAT(F topo ,F cross1 )),

[0129] Among them, F fused This indicates that the interaction result F will be... cross1 With topological feature F topo By fusing GAT with a graph attention network, importance weights between network nodes are dynamically allocated, and LayerNorm represents layer normalization.

[0130] Stabilize the training process;

[0131] Achieving dynamic feature optimization: The feature importance weight α is calculated using a multilayer perceptron (MLP), with the following formula:

[0132] α = Softmax(MLP([F drop ||F wave ||F topo ])),

[0133] Where || represents the concatenation operation, and Softmax represents normalization to a probability distribution to ensure that the sum of the weights is 1;

[0134] The final weighted feature representation is as follows: Based on the calculated weights, the pressure drop feature, fluctuation feature, and topological feature are weighted and summed, as shown in the formula:

[0135] F final =α drop ·F drop +α wave ·F wave +α topo ·F topo ,

[0136] Among them, F final This indicates that the weighted summation yields the final fusion feature, α. drop α wave α topo These represent the weights corresponding to the pressure drop characteristic, fluctuation characteristic, and topological characteristic, respectively, given by α = Softmax(MLP([F drop ||F wave ||F topo The generation of each modality reflects its importance to the current task.

[0137] Deploying a hybrid processing architecture: Transformer layer processing flow:

[0138] F trans =TransformerEncoder(F final ,positional_ending(x,y,z)),

[0139] Specifically, the Transformer model is used to process sequence data, and the Transformer layer processes the fused features F. final To capture long-term pressure change patterns, positional_ending(x,y,z) represents injecting the spatial coordinates (x,y,z) of the pipeline network and preserving the topological location information;

[0140] LSTM layer processing flow: F lstm =BiLSTM(F trans ,t win It uses a bidirectional long short-term memory network to process temporal information, and a bidirectional LSTM to extract local temporal features, with a window length of t. win Control the scope of short-term dependence;

[0141] Graph network layer processing flow: F gnn =GraphsomorphismNetwork(F lstm ,edge_attr), utilize graph isomorphic networks to process graph-structured data, F gnn This indicates that the graph isomorphic network handles the pipe network topology relationship, and edge_attr contains pipe segment attributes;

[0142] Implement joint optimization: The total loss function includes the task loss L task Reconstruction loss L recon and orthogonal constraint L orth The formula is:

[0143] L total =L task +λL recon +μL orth ,

[0144] Among them, the task loss L task Used to supervise the model in completing the main task of leak detection, and to reconstruct the loss L. recon Orthogonal constraint L is used to ensure the reversibility of feature fusion. orth To force different modal features to remain independent and avoid redundancy, λ and μ represent the weights of the hyperparameter balancing loss term.

[0145] S4.2 Establish an interpretable incremental learning system: Integrate a SHAP value interpretation module at the output of the hybrid model to generate feature contribution heatmaps in real time; establish a rule extraction engine based on heatmap analysis to output interpretable judgment logic; design an online update pipeline that includes sample importance weighting, and use elastic weight solidification technology to ensure continuous model optimization without forgetting old knowledge. The SHAP value interpretation module provides interpretability for model decisions, helping users understand the model's judgment logic. The rule extraction engine can extract interpretable rules from the model output, improving the model's transparency and credibility. The elastic weight solidification technology ensures that the model does not forget old knowledge during continuous optimization, improving the model's stability and adaptability.

[0146] S4.3 Forming an end-to-end joint optimization closed loop: Design a multi-task loss function that integrates voltage drop detection and fluctuation analysis to drive the collaborative training of the hybrid model; feed back the performance metrics output by the model to the front-end acquisition system to dynamically adjust the sensor sampling strategy; trigger an automatic rollback mechanism through the model performance monitoring module to ensure system stability. The multi-task loss function can optimize multiple tasks simultaneously, improving the overall performance of the model. The performance metric feedback mechanism can dynamically adjust system parameters, improving the system's adaptability and stability. The automatic rollback mechanism can ensure that the system can recover quickly when anomalies occur, improving the system's reliability and stability.

[0147] The S4 component utilizes techniques such as multimodal data fusion, hybrid model architecture, incremental learning, and end-to-end optimization to construct a highly efficient joint detection model for voltage drop and fluctuation. This model can not only handle complex data structures and task requirements but also improve its performance and reliability through interpretability and continuous optimization mechanisms.

[0148] S5. Input the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, trigger an intelligent alarm and locate potential leak points through the visualization platform.

[0149] Specifically, in this embodiment, step S5 involves inputting the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds a preset threshold, an intelligent alarm is triggered, and potential leak points are located through a visualization platform, including:

[0150] The system performs data cleaning on the input real-time pressure drop rate and fluctuation intensity, including outlier removal and smoothing. The pre-processed data is then spatiotemporally aligned with the pipeline network topology information to extract current operating condition feature vectors, including time-period characteristics, environmental parameters, and equipment status. These feature vectors are input into a trained AI model, which outputs a leakage probability value. A multi-level alarm judgment is performed using a dynamic threshold library, including early warning, confirmed alarm, and emergency alarm. A delayed confirmation mechanism is activated for critical alarm states to avoid false alarms caused by momentary interference. Potential leakage areas are calculated based on pressure gradient changes and the pipeline network hydraulic model. Potential leakage points and their confidence levels are marked on a 3D pipeline network digital twin model. A visual report containing location results, risk level, and impact range is generated. Alarm information is automatically pushed to the mobile terminals of relevant personnel. Operator response and on-site confirmation results are recorded. Validated alarm cases are fed back to the model training system for incremental learning.

[0151] Specifically, data cleaning and preprocessing include: outlier detection (e.g., using statistical or machine learning methods) and smoothing (e.g., moving average, wavelet transform) of real-time pressure drop rate and fluctuation intensity data to remove noise and interference. The preprocessed data is then spatiotemporally aligned with the pipeline network topology information to extract current operating condition feature vectors, including time-period characteristics, environmental parameters, and equipment status.

[0152] Model inference and alarm judgment include: inputting the extracted feature vectors into a trained AI model (such as a deep learning model, a hybrid model, etc.) and outputting a leakage probability value. Multi-level alarm judgment (early warning, confirmed alarm, emergency alarm) is performed using a dynamic threshold library, and a delayed confirmation mechanism is activated to avoid false alarms.

[0153] Location and visualization include: Leakage area calculation: Calculating potential leak areas based on pressure gradient changes and a pipeline hydraulic model. 3D visualization: Marking potential leak points and confidence levels on a 3D digital twin model of the pipeline network and generating a visualization report. Alarms and feedback include: Alarm push and response: Automatically pushing alarm information to the mobile terminals of relevant personnel and recording operator responses and on-site confirmation results. Model optimization: Feeding validated alarm cases back to the model training system for incremental learning, continuously optimizing model performance.

[0154] By integrating data processing, machine learning, digital twins, and visualization technologies, the entire process from data preprocessing to model inference, alarm judgment, location analysis, and feedback aims to achieve efficient and accurate leak detection and response.

[0155] S6. Based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, intelligently analyze the location and cause of leakage, and generate maintenance plans.

[0156] Specifically, in this embodiment, step S6, based on the output of the AI ​​model, combines historical leakage data and pipeline topology to intelligently analyze the location and cause of leakage, and generate a maintenance plan, including:

[0157] The system integrates the leakage probability distribution output by the AI ​​model with pipeline topology data, correlates similar operating conditions in the historical leakage case database, and overlays equipment operation logs and maintenance records.

[0158] The potential leakage area is calculated based on the pressure gradient propagation model. The correlation between pipeline nodes is analyzed using graph neural networks. The most likely leakage point and confidence level are determined through Bayesian inference. The potential causes of leakage are analyzed in combination with the material aging model.

[0159] The system automatically retrieves the pipe section design parameters based on the leak location, assesses the scope and risk level of the leak's impact, matches the maintenance resource database to generate alternative solutions, and calculates the cost, time, and expected effects of each solution.

[0160] Generates reports containing location results, cause analysis, and maintenance recommendations; marks leak points and affected areas in the 3D pipeline network model; provides a maintenance solution comparison interface and recommendations; and supports maintenance process tracking and result feedback.

[0161] S7. Continuously optimize AI model parameters based on feedback data and link with pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

[0162] Specifically, in terms of model optimization, various methods can be used to adjust and optimize parameters. For example, random search and Bayesian optimization can be used to find the optimal model parameters. Furthermore, regularization techniques (such as L1 and L2 regularization) can effectively prevent overfitting and improve the model's generalization ability. Meanwhile, data preprocessing (such as data cleaning and feature engineering) is also an important means of improving model performance.

[0163] Regarding feedback mechanisms, establishing a closed-loop feedback system is crucial for continuous model optimization. By collecting feedback from users or maintenance personnel and regularly optimizing the model and rules, model performance can be continuously improved. For example, A / B testing can be used to compare the effects of different model versions and select the optimal solution.

[0164] In terms of system integration, the output of the AI ​​model can be linked with the pipeline valve control system to achieve automatic isolation of leaking areas and remote dispatch of maintenance resources. For example, after the AI ​​model detects an anomaly, the system can automatically trigger valve closure, isolate the leaking area, and dispatch maintenance resources for handling. By combining model optimization, feedback mechanisms, and system integration, a more efficient and intelligent leak detection and handling process can be achieved.

[0165] Example 2

[0166] like Figure 4 As shown in the figure, this application provides an architecture diagram of an IoT fire sprinkler network intelligent micro-leak alarm system based on time-period pressure drop and pressure fluctuation analysis. It is applied to the IoT fire sprinkler network intelligent micro-leak alarm system based on time-period pressure drop and pressure fluctuation analysis as described in Embodiment 1. It includes a data acquisition module 11, a time period division and pressure drop calculation module 12, a pressure fluctuation analysis module 13, an AI model training module 14, a micro-leak detection and alarm module 15, a leak location and maintenance decision module 16, and a dynamic optimization and equipment control module 17.

[0167] The data acquisition module 11 is used to collect network pressure data in real time through IoT pressure sensors deployed at key nodes of the fire sprinkler network and transmit it to the cloud or edge computing platform.

[0168] The time period division and pressure drop calculation module 12 is used to divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the time period pressure drop rate.

[0169] The pressure fluctuation analysis module 13 is used to perform frequency domain analysis on pressure data using wavelet transform, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation.

[0170] AI model training module 14 is used to train a large model on historical pressure data under normal and leakage conditions to build a pressure drop-fluctuation joint detection model.

[0171] The micro-leak detection and alarm module 15 is used to input the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, an intelligent alarm is triggered, and potential leak points are located through the visualization platform.

[0172] The Leakage Location and Repair Decision Module 16 is used to intelligently analyze the location and cause of leakage based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, and generate a repair plan.

[0173] The dynamic optimization and equipment control module 17 is used to continuously optimize the AI ​​model parameters based on feedback data and link with the pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

[0174] Figure 5 This is an electronic device provided in one embodiment of this application. For example... Figure 5 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0175] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0176] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0177] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these devices is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (Flash ROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0178] It should be noted that a portion of the electronic device described in the above embodiments can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0179] It should be noted that the term "computer" as used here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, "computer-readable recording media" refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage devices such as hard drives built into a computer.

[0180] Furthermore, a "computer-readable recording medium" can include: a medium that dynamically stores a program for a short period of time, such as a communication line used when transmitting a program via a network such as the Internet or a communication line such as a telephone line; or a medium that stores a program for a fixed period of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining with programs already recorded in the computer.

[0181] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (device group) composed of multiple devices. Each device constituting the device group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a device group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0182] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A smart micro-leakage alarm method for IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis, characterized in that, The method includes: S1. By deploying IoT pressure sensors at key nodes of the fire sprinkler network, real-time pressure data of the network is collected and transmitted to the cloud or edge computing platform. S2. Divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the pressure drop rate for each time period. S3. Use wavelet transform to perform frequency domain analysis on pressure data, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation. S4. A large model is used to train the pressure data under historical normal and leakage conditions to construct a pressure drop-fluctuation joint detection model; S5. Input the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, trigger an intelligent alarm and locate potential leak points through the visualization platform. S6. Based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, intelligently analyze the location and cause of leakage, and generate a maintenance plan. S7. Continuously optimize AI model parameters based on feedback data and link with pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

2. The method for intelligent micro-leakage alarm of IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis as described in claim 1, characterized in that, The data acquisition in S1 includes: Three levels of pressure monitoring nodes are set up at the main pipeline, the junction of branch pipes and the end of the pipeline network. Each node is equipped with an IoT pressure sensor with self-calibration function. The sensor uses a dual-mode communication protocol of LoRaWAN and 5G to collect pressure data in real time at a sampling frequency of 1Hz; The raw data is preprocessed through edge computing nodes, including: using sliding window filtering to eliminate transient interference; achieving multi-node data synchronization through timestamp alignment; and performing Mahalanobis distance detection and interpolation repair for outliers. The preprocessed data is compressed and packaged, and then uploaded to the cloud data lake via the MQTT protocol. A spatiotemporal database is established in the cloud, and pipeline topology tags are attached to the pressure data to form a structured spatiotemporal sequence dataset.

3. The method for intelligent micro-leakage alarm of IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis as described in claim 1, characterized in that, Step S2 involves dividing the collected pressure data into preset time periods, calculating the average pressure value within each time period, comparing it with historical benchmark pressure, and calculating the pressure drop rate for each time period, including: S2.

1. The DTW-BIRCH hybrid clustering algorithm is used to dynamically divide historical pressure data into time periods, including: measuring data similarity through dynamic time warping, constructing a time period feature tree, and generating the optimal time period division scheme by combining water use patterns. S2.2 Perform multi-level feature extraction within each time period, including: calculating the arithmetic mean stress value and standard deviation of the base layer, the weighted average stress value of the topology layer, and the stress change gradient of the trend layer; S2.

3. Generate a dynamic baseline pressure curve based on the Transformer-TSA model. This model processes historical pressure data, environmental parameters and maintenance records through a time attention mechanism and outputs a dynamic baseline that considers multiple factors. S2.

4. An adaptive pressure drop rate calculation model is adopted, which integrates the base pressure drop rate, stability correction factor and topology correction factor, and dynamically adjusts the weight of each factor through an online learning mechanism. S2.5 Establish a closed-loop feedback optimization system to evaluate data quality, computational reliability, and anomalies, and feed the evaluation results back to the time period division and benchmark generation modules in real time.

4. The IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in claim 2, characterized in that, S3, using wavelet transform to perform frequency domain analysis on the pressure data, extracting high-frequency fluctuation features, and combining the time domain standard deviation to calculate the pressure fluctuation intensity, includes: S3.

1. An adaptive wavelet basis selection algorithm is used to match the optimal wavelet basis function, and the pressure signal is decomposed into three levels to extract high-frequency detail coefficients and low-frequency approximation coefficients, and a wavelet energy spectrum feature matrix is ​​established. S3.

2. Through the time-frequency domain fusion analysis engine, the pressure standard deviation and peak-to-peak value in the time domain are comprehensively calculated, as well as the high-frequency energy ratio, characteristic frequency band energy ratio and wavelet entropy in the frequency domain, and the attention mechanism is used to dynamically weight the features. S3.3 Construct a volatility intensity index, where the weight coefficients are obtained through training with historical data, and establish a dynamic threshold adjustment mechanism; S3.4 Set up a closed-loop optimization system to continuously optimize the analysis parameters through feature quality assessment and online learning modules, and feed the results back to the time period segmentation module.

5. The method for intelligent micro-leakage alarm of IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis as described in claim 2, characterized in that, S4 involves training a large model using historical pressure data under normal and leakage conditions to construct a pressure drop-fluctuation joint detection model, including: S4.1 Constructing a multimodal feature fusion and hybrid model architecture: After aligning the time-period pressure drop features, pressure fluctuation features and pipeline topology features in time and space, a joint feature representation is generated through a feature cross-network; An attention mechanism is used to dynamically weight the importance of each feature dimension and input it into a Transformer-LSTM-GNN hybrid model. The Transformer layer processes the aligned long-term pressure change pattern, the LSTM layer extracts the local temporal characteristics of the joint features, and the graph neural network layer parses the pipeline topology relationship. S4.2 Establish an interpretable incremental learning system: Integrate a SHAP value interpretation module at the output of the hybrid model to generate a feature contribution heatmap in real time; establish a rule extraction engine based on heatmap analysis to output interpretable judgment logic; design an online update pipeline that includes sample importance weighting, and achieve continuous model optimization without forgetting through elastic weight solidification technology; S4.3 Forming an end-to-end joint optimization closed loop: Design a multi-task loss function that integrates voltage drop detection and fluctuation analysis to drive the collaborative training of the hybrid model; feed back the performance indicators output by the model to the front-end acquisition system to dynamically adjust the sensor sampling strategy; trigger an automatic rollback mechanism through the model performance monitoring module to ensure system stability.

6. The method for intelligent micro-leakage alarm of IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis as described in claim 5, characterized in that, The multimodal feature fusion and hybrid model architecture is implemented through the following interrelated technical solutions: Constructing a feature interaction fusion network: First, the voltage drop feature F is calculated through a feature cross-layer. drop With fluctuation characteristics F wave Interactive items: in, σ represents the outer product operation, used to generate higher-order feature combinations; W1 represents the learnable weight matrix; and σ represents the activation function, used to introduce nonlinearity. Combine the interaction results with the topological feature F topo The secondary fusion is performed using the following formula: F fused =LayerNorm(F cross1 +GAT(F topo ,F cross1 )), Among them, F fused This indicates that the interaction result F will be... cross1 With topological feature F topo Fusion through graph attention network GAT dynamically assigns importance weights between network nodes, while LayerNorm represents layer normalization to stabilize the training process. Achieving dynamic feature optimization: The feature importance weight α is calculated using a multilayer perceptron (MLP), with the following formula: α=Softmax(MLP([F drop ||F wave ||F topo ])), Where || represents the concatenation operation, and Softmax represents normalization to a probability distribution to ensure that the sum of the weights is 1; The final weighted feature representation is as follows: Based on the calculated weights, the pressure drop feature, fluctuation feature, and topological feature are weighted and summed, as shown in the formula: F final =a drop ·F drop +a wave ·F wave +a topo ·F topo , Among them, F final This indicates that the weighted summation yields the final fusion feature, α. drop α wave α topo These represent the weights corresponding to the pressure drop characteristic, fluctuation characteristic, and topological characteristic, respectively, by α = Softmax(MLP([F drop ||F wave ||F topo The generation of each modality reflects its importance to the current task. Deploying a hybrid processing architecture: Transformer layer processing flow: F trans =TransformerEncoder(F final ,positional_ending(x,y,z)), Specifically, the Transformer model is used to process sequence data, and the Transformer layer processes the fused features F. final To capture long-term pressure change patterns, positional_ending(x,y,z) represents injecting the spatial coordinates (x,y,z) of the pipeline network and preserving the topological location information; LSTM layer processing flow: F lstm =BiLSTM(F trans ,t win It uses a bidirectional long short-term memory network to process temporal information, and a bidirectional LSTM to extract local temporal features, with a window length of t. win Control the scope of short-term dependence; Graph network layer processing flow: F gnn =GraphsomorphismNetwork(F lstm ,edge_attr), utilize graph isomorphic networks to process graph-structured data, F gnn This indicates that the graph isomorphic network handles the pipe network topology relationship, and edge_attr contains pipe segment attributes; Implement joint optimization: The total loss function includes the task loss L task Reconstruction loss L recon and orthogonal constraint L orth The formula is: L total =L task +λL recon +μL orth , Among them, the task loss L task Used to supervise the model in completing the main task of leak detection, and to reconstruct the loss L. recon Orthogonal constraint L is used to ensure the reversibility of feature fusion. orth To force different modal features to remain independent and avoid redundancy, λ and μ represent the weights of the hyperparameter balancing loss term.

7. The IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis according to claim 1, characterized in that, S5 involves inputting the real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds a preset threshold, an intelligent alarm is triggered, and potential leak points are located through a visualization platform, including: Data cleaning is performed on the input real-time pressure drop rate and fluctuation intensity, including outlier removal and smoothing. The pre-processed data is spatiotemporally aligned with the pipeline topology information, and the current operating condition feature vector is extracted, including time period features, environmental parameters and equipment status. Input the feature vector into the trained AI model, and output the leakage probability value; Multi-level alarm judgment is performed by combining a dynamic threshold library, including early warning, confirmed alarm and emergency alarm; A delayed confirmation mechanism is activated for critical alarm states to avoid false alarms caused by momentary interference; The potential leakage area was calculated based on pressure gradient changes and a pipeline hydraulic model. Mark potential leak points and confidence levels on a 3D digital twin model of the pipeline network; Generate a visual report that includes location results, risk level, and scope of impact; Automatically push alarm information to the mobile terminals of relevant personnel; Record the operator's response and the on-site confirmation results; The verified alarm cases are fed back to the model training system for incremental learning.

8. The method for intelligent micro-leakage alarm of IoT fire sprinkler network based on time-period pressure drop and pressure fluctuation analysis as described in claim 5, characterized in that, S6, based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, intelligently analyzes the location and cause of leakage, and generates a maintenance plan, including: Integrate the leakage probability distribution output by the AI ​​model with the pipeline topology data, associate it with similar operating condition characteristics in the historical leakage case library, and overlay equipment operation logs and maintenance record data; The possible leakage area is calculated based on the pressure gradient propagation model, the correlation of pipeline nodes is analyzed using graph neural network, the most likely leakage point and confidence level are determined by Bayesian inference, and the potential causes of leakage are analyzed by combining the material aging model. The system automatically retrieves the pipe section design parameters based on the leak location, assesses the leak's impact range and risk level, matches the maintenance resource database to generate alternative solutions, and calculates the cost, time consumption, and expected effects of each solution. Generates reports containing location results, cause analysis, and maintenance recommendations; marks leak points and affected areas in the 3D pipeline network model; provides a maintenance solution comparison interface and recommendations; and supports maintenance process tracking and result feedback.

9. An IoT fire sprinkler network intelligent micro-leakage alarm system based on time-period pressure drop and pressure fluctuation analysis, applied to the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is used to collect real-time pipeline pressure data through IoT pressure sensors deployed at key nodes of the fire sprinkler network and transmit it to the cloud or edge computing platform. The time period segmentation and pressure drop calculation module is used to divide the collected pressure data into preset time periods, calculate the average pressure value in each time period, compare it with the historical benchmark pressure, and calculate the pressure drop rate of the time period. The pressure fluctuation analysis module is used to perform frequency domain analysis on pressure data using wavelet transform, extract high-frequency fluctuation features, and calculate the pressure fluctuation intensity by combining the time domain standard deviation. The AI ​​model training module is used to train a large model on historical pressure data under normal and leakage conditions to build a pressure drop-fluctuation joint detection model. The micro-leak detection and alarm module is used to input real-time pressure drop rate and fluctuation intensity into the AI ​​model. If the output probability exceeds the preset threshold, an intelligent alarm is triggered, and potential leak points are located through a visualization platform. The Leakage Location and Repair Decision Module is used to intelligently analyze the location and cause of leakage based on the output of the AI ​​model, combined with historical leakage data and pipeline topology, and generate a repair plan. The dynamic optimization and equipment control module is used to continuously optimize the AI ​​model parameters based on feedback data and link with the pipeline valve control system to achieve automatic isolation of leaking areas and remote scheduling of maintenance resources.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that instructs the device to execute the IoT fire sprinkler network intelligent micro-leakage alarm method based on time-period pressure drop and pressure fluctuation analysis as described in any one of claims 1 to 8.

Images