Intelligent fire-fighting equipment fault prediction method

Abstract

The invention discloses an intelligent fire-fighting equipment fault prediction method, and belongs to the technical field of intelligent maintenance of fire-fighting equipment. According to the method, pressure, electrical waveform and environmental data are dynamically acquired through a multi-source sensor, spatial-temporal features are extracted in parallel by adopting a bidirectional LSTM and a multi-scale convolutional network, and feature fusion is realized by utilizing a cross-modal attention mechanism; outputting a fault probability based on a gradient boosting decision tree model, and predicting the remaining service life in combination with Bayesian regression; the firmware upgrading authority is controlled through a health degree dynamic scoring system, and the upgrading safety is guaranteed through AES-256-CBC encryption and SM3 signature verification; and generating an optimal maintenance path based on an improved genetic algorithm, and scheduling an engine to respond to a sudden fault in real time. Tests show that the fault early warning accuracy rate reaches 93.1%, the maintenance cost is reduced by 45%, the upgrade failure blocking rate is 100%, and the intelligent level and reliability of operation and maintenance of the fire-fighting equipment are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent maintenance of fire-fighting equipment, and in particular to a fault prediction method for intelligent fire-fighting equipment. Background Art

[0002] With the widespread adoption of smart firefighting systems in high-rise buildings, industrial parks, and other scenarios, numerous firefighting terminals are prone to failures. Traditional firefighting equipment maintenance suffers from the following technical flaws: Detection mechanisms are lagging: Relying on a single sensor threshold alarm (e.g., pressure exceeding the limit), it fails to identify complex faults (e.g., the coupling effect of abnormal pressure and electrical short circuits), resulting in an underreporting rate exceeding 30%. However, some existing technical improvement solutions still have limitations. Consequently, a smart firefighting equipment fault prediction method is currently lacking to address these issues. Summary of the Invention

[0003] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a method for predicting failures of intelligent fire fighting equipment.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] A method for predicting failures of intelligent firefighting equipment comprises the following steps:

[0006] S1. Dynamically collect pressure, electrical waveform, and environmental data through multi-source sensors. The pressure sensor sampling rate is dynamically adjusted according to the pressure change rate.

[0007] S2. A bidirectional recurrent neural network is used to extract the long-term temporal features of pressure data, and a multi-scale convolutional network is used to capture the local mutation patterns of the electrical waveform.

[0008] S3. Use the cross-modal attention mechanism to dynamically weight the temporal features and waveform features to generate a comprehensive spatiotemporal feature matrix;

[0009] S4. Output the fault type and probability based on the gradient boosting decision tree model, and combine it with Bayesian regression to predict the remaining service life of the equipment;

[0010] S5. When the device health score falls below the threshold, the firmware upgrade is blocked and AES-256-CBC encryption and SM3 signature verification are used to ensure upgrade security.

[0011] S6. Generate the optimal maintenance path based on the dynamic programming algorithm and respond to sudden failures through the real-time scheduling engine.

[0012] As a further optimization of the present technical solution, the dynamic adjustment strategy of the pressure sensor sampling rate in step S1 is: a 50 Hz sampling rate is used when the pressure is stable, and it is automatically increased to 100 Hz when the pressure change rate exceeds 2 MPa / s.

[0013] As a further optimization of the present technical solution, the number of hidden units of the bidirectional recurrent neural network in step S2 is 32, and the time step is set to 60 sampling points.

[0014] As a further optimization of the present technical solution, the weight distribution of the cross-modal attention mechanism in step S3 satisfies the following rules: when burrs and distortions appear in the electrical waveform, the waveform feature weight is increased to 0.75-0.85; when the pressure drop exceeds the threshold, the timing feature weight accounts for more than 60%.

[0015] As a further optimization of this technical solution, the equipment health score in step S5 is calculated as follows: the failure probability weight accounts for 70%, and the remaining service life accounts for 30%; when the health score is lower than 0.6, the equipment self-check program is triggered.

[0016] As a further optimization of the present technical solution, the device self-test program includes memory check and verification, sensor reference value comparison, and communication module loop test.

[0017] As a further optimization of this technical solution, the constraints of the dynamic programming algorithm in step S6 include: emergency failure equipment response time ≤ 2 hours; key node equipment maintenance interval ≥ 14 days; the maximum number of devices maintained per day does not exceed 50.

[0018] As a further optimization of this technical solution, the real-time scheduling engine performs the following operations: fire alarm failures are automatically upgraded to the highest priority; sudden failures trigger online re-planning of maintenance paths, and the re-planning takes ≤30 seconds.

[0019] As a further optimization of this technical solution, it also includes: deploying edge computing nodes to perform real-time prediction, and configuring NPU coprocessors to make the inference speed ≥100 frames / second.

[0020] As a further optimization of this technical solution, the electrical waveform analysis includes 3rd / 5th / 7th harmonic detection, and the sampling accuracy is not less than 12 bits.

[0021] Beneficial effects: The present invention provides a method for predicting faults of intelligent fire-fighting equipment. This device achieves the following core advantages through technological innovation: 1. Significantly improved fault warning capabilities: Multimodal data fusion enables the accuracy of complex fault identification to reach 93.1% (68.7% for traditional methods), and the false alarm rate is reduced to 3.1%. The remaining service life (RUL) prediction error is ≤9.1 hours (>24 hours for traditional methods). 2. Optimization of upgrade security and efficiency: The dynamic health assessment mechanism achieves a 100% upgrade failure blocking rate and shortens the upgrade time by 65%. Differential upgrade technology reduces firmware transmission volume by 85%, and the rollback mechanism ensures a 100% system abnormality recovery rate. 3. Efficient utilization of maintenance resources: The dynamic programming algorithm reduces the maintenance path cost by 41%, and the response time to sudden faults is shortened to 1.8 hours (8.5 hours for traditional methods). AR-assisted inspections and blockchain evidence storage technology increase maintenance efficiency by 53% and reduce manual verification workload by 70%. 4. Enhanced system reliability: The edge-cloud collaborative architecture (inference latency <10ms) ensures real-time performance, and dual redundant communications (LoRaWAN + NB-IoT) ensure a data transmission integrity rate of ≥99.99%. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 : Schematic diagram of the system architecture block diagram (four-level architecture);

[0023] Figure 2 : Schematic diagram of feature fusion flow chart;

[0024] Figure 3 : Parallel processing branches;

[0025] Figure 4 : Schematic diagram of the dynamic programming algorithm flow chart. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0027] The present invention provides a technical solution: a method for predicting faults of intelligent fire-fighting equipment, comprising the following steps:

[0028] S1. Dynamically collect pressure, electrical waveform, and environmental data through multi-source sensors. The pressure sensor sampling rate is dynamically adjusted according to the pressure change rate.

[0029] S2. A bidirectional recurrent neural network is used to extract the long-term temporal features of pressure data, and a multi-scale convolutional network is used to capture the local mutation patterns of the electrical waveform.

[0030] S3. Use the cross-modal attention mechanism to dynamically weight the temporal features and waveform features to generate a comprehensive spatiotemporal feature matrix;

[0031] S4. Output the fault type and probability based on the gradient boosting decision tree model, and combine it with Bayesian regression to predict the remaining service life of the equipment;

[0032] S5. When the device health score falls below the threshold, the firmware upgrade is blocked and AES-256-CBC encryption and SM3 signature verification are used to ensure upgrade security.

[0033] S6. Generate the optimal maintenance path based on the dynamic programming algorithm and respond to sudden failures through the real-time scheduling engine.

[0034] In specific implementation, the dynamic adjustment strategy of the pressure sensor sampling rate in step S1 is: adopting a 50 Hz sampling rate when the pressure is stable, and automatically increasing to 100 Hz when the pressure change rate exceeds 2 MPa / s.

[0035] In a specific implementation, the number of hidden units of the bidirectional recurrent neural network in step S2 is 32, and the time step is set to 60 sampling points.

[0036] In specific implementation, the weight distribution of the cross-modal attention mechanism in step S3 satisfies the following rules: when burrs and distortions appear in the electrical waveform, the waveform feature weight is increased to 0.75-0.85; when the pressure drop exceeds the threshold, the timing feature weight accounts for more than 60%.

[0037] In specific implementation, the device health score in step S5 is calculated as follows: the failure probability weight accounts for 70%, and the remaining service life accounts for 30%; when the health score is lower than 0.6, the device self-check program is triggered.

[0038] In specific implementation, the device self-test program includes memory check and verification, sensor reference value comparison, and communication module loop test.

[0039] In specific implementation, the constraints of the dynamic programming algorithm in step S6 include: emergency failure equipment response time ≤ 2 hours; key node equipment maintenance interval ≥ 14 days; the maximum number of devices maintained per day does not exceed 50.

[0040] In specific implementation, the real-time scheduling engine performs the following operations: fire alarm failure is automatically upgraded to the highest priority; sudden failure triggers online re-planning of the maintenance path, and the re-planning takes ≤30 seconds.

[0041] The specific implementation also includes: deploying edge computing nodes to perform real-time predictions, and configuring NPU coprocessors to make the inference speed ≥100 frames per second.

[0042] In a specific implementation, the electrical waveform analysis includes 3rd / 5th / 7th harmonic detection, and the sampling accuracy is not less than 12 bits.

[0043] In specific implementation, when applied to a gas fire extinguishing system, when the pressure drop rate of the pressure vessel exceeds 0.1MPa / h, an aging warning of the sealing ring is automatically triggered and a spare parts procurement list is generated.

[0044] Corresponding to Example 4, the innovative point of protecting pressure monitoring and linking with the supply chain.

[0045] In specific implementation, during the maintenance of the fire smoke exhaust system, AR equipment is used to display fault location information and maintenance instructions in real time.

[0046] In specific implementation, during the upgrade of the fire alarm system, when the false alarm rate exceeds 5%, it will automatically roll back to the previous stable version and trigger system self-check.

[0047] like Figure 1 The system architecture shown includes the following core modules:

[0048] 1. Intelligent Perception Layer

[0049] Pressure monitoring unit: uses piezoelectric sensor (range 0-15MPa, accuracy ±0.3%), dynamically adjusts sampling rate (50-100Hz);

[0050] Electrical analysis unit: capture current / voltage waveforms, identify 3rd / 5th / 7th harmonic components, sampling rate 1MHz;

[0051] Environmental sensing unit: monitors the temperature and humidity around the device (-20℃~80℃), with a data update interval of 1 second.

[0052] 2. Feature Fusion Layer

[0053] Time series processing module: a bidirectional recurrent neural network extracts long-term dependencies of pressure data, with 32 hidden units;

[0054] Waveform analysis module: Multi-scale convolutional network (3×1 dilated convolution kernel, dilation rate 1 / 2 / 4) captures the sudden change characteristics of electrical waveforms;

[0055] Attention weighting module: Dynamically allocates pressure and electrical feature weights, and increases the distorted waveform weight to 0.8.

[0056] 3. Security Control Layer

[0057] Health assessment: Comprehensive failure probability (gradient boosting decision tree output) and remaining life (Bayesian regression prediction), with a threshold set to 0.6;

[0058] Encryption upgrade protocol: AES-256-CBC encrypted transmission, SM3 signature verification, key generation based on device MAC address hash.

[0059] 4. Maintenance Optimization Layer

[0060] Dynamic planning engine: Generates the most cost-effective path based on device topology, supporting real-time calculation for up to 50 nodes;

[0061] Emergency dispatch module: Sudden failures trigger path re-planning, with response delay less than 30 seconds.

[0062] Technical Effects

[0063] Tested by the National Fire Electronic Product Quality Inspection Center (Report No. NFTC-2024-0218):

[0064] index The present invention Traditional methods Improvement Fault warning accuracy 93.1％ 68.7％ +24.4％ Average maintenance response time 2.1h 8.5h -75％ Upgrade failure blocking rate 100％ 0％ +100％ DETAILED DESCRIPTION

[0065] Example 1: Fire pump system monitoring

[0066] Data collection

[0067] When the pressure sensor detects a change rate greater than 2 MPa / s, the sampling rate switches from 50 Hz to 100 Hz; 256 points are collected per cycle of the current waveform to identify harmonic distortion greater than 5%.

[0068] Model training

[0069] Data set: 12 months of historical data (8,000 groups of normal data, 3,000 groups of 12 types of fault data);

[0070] Training parameters: batch size 32, early stopping patience value 15, learning rate exponential decay (decay coefficient 0.95).

[0071] Upgrade Verification

[0072] Health threshold: Upgrade is blocked when failure probability is greater than 85% and remaining life is less than 72 hours; Signature verification: SM3 hash check error tolerance is ±0.01%.

[0073] Example 2: Emergency Lighting System Maintenance

[0074] Path optimization

[0075] Dynamic programming constraints: The upper limit for daily maintenance is 50 units, and the path cost weight is 70%; emergency fault handling: A fire alarm failure triggers the maintenance priority to be the highest.

[0076] Effect verification

[0077] 50-node scenario: path cost reduced by 41%, working hours compressed by 35%; sudden fault response: average processing time shortened from 4.2 hours to 1.8 hours.

[0078] Example 3: Intelligent Fire Alarm System Monitoring

[0079] Application scenario: High-rise office building fire alarm network

[0080] Technical Implementation

[0081] Data collection:

[0082] Smoke detector: monitors smoke concentration (0-10% obs / m), sampling rate 1Hz; Sound and light alarm: collects sound pressure level data (30-120dB), and identifies abnormal audio features;

[0083] Feature fusion:

[0084] Time series features: Bidirectional LSTM analysis of smoke concentration trends (time step 120);

[0085] Spatial features: CNN identifies fault noise in the acoustic spectrum (frequency band 50-5000 Hz);

[0086] Attention weight distribution: When smoke suddenly increases, the weight of the smoke feature is increased to 0.9;

[0087] Security upgrades:

[0088] Health threshold: H = 0.65 (α = 0.6, β = 0.4); Upgrade package fragmentation: Each 512KB data block is independently encrypted, and transmission failures are automatically retried ≤ 3 times; Maintenance optimization:

[0089] Dynamic planning constraints: maintenance time window for each floor ≤ 15 minutes; sudden fault response: trigger system self-check when false alarm rate exceeds 5%;

[0090] Technical results: The false alarm recognition rate has been increased to 98.5% (compared to 82% for traditional methods); the system self-check time has been reduced from 20 minutes to 3 minutes;

[0091] Example 4: Gas Fire Extinguishing System Pressure Vessel Monitoring

[0092] Application scenario: Heptafluoropropane fire extinguishing device in data center

[0093] Technical implementation:

[0094] Data collection:

[0095] Pressure vessel: monitors pressure value (2.5-4.2MPa), with dynamic adjustment of sampling rate (10Hz when stable, 50Hz when leaking); Valve status: records the number of solenoid valve actions and response time;

[0096] Failure prediction:

[0097] Feature Engineering: Extracting pressure fluctuation variance (window length 60 seconds); Prediction Model: XGBoost classifier identifying seal aging patterns (accuracy 94.2%)

[0098] Security upgrades:

[0099] Encryption protocol: Quantum key distribution (QKD) is used to transmit firmware; Health blocking: Upgrade is prohibited when the container pressure drop rate is greater than 0.1MPa / h;

[0100] Maintenance strategy:

[0101] Route optimization: prioritizes maintenance of the top 10% of containers with the highest pressure drop; spare parts scheduling: pre-delivers seals based on container life prediction;

[0102] Technical effect: Leakage warning is provided 72 hours before failure occurs; spare parts inventory costs are reduced by 37%;

[0103] Example 5: Fire smoke exhaust system fan monitoring

[0104] Application scenario: underground garage smoke exhaust system

[0105] Technical implementation:

[0106] Multi-source data: Fan vibration: triaxial accelerometer (range ±50g); exhaust efficiency: CO concentration gradient monitoring (ppm / s); electrical parameters: motor current harmonic analysis (alarm when THD>8%);

[0107] Model optimization: Transfer learning: Fine-tuning the water pump system model (freezing the first three layers of the network); Anomaly detection: Isolation Forest identifies early bearing wear;

[0108] Dynamic upgrade: Differential upgrade: only transmit firmware difference blocks (compression rate 85%); Rollback mechanism: automatically restore the old version within 24 hours after the upgrade;

[0109] Maintenance verification: AR-assisted inspection: Maintenance personnel receive fault location data through AR glasses; blockchain evidence storage: Maintenance records are stored on the chain with tamper-proof timestamps;

[0110] Technical effect: Fan fault false detection rate <0.5%; average maintenance time shortened to 45 minutes.

[0111] Technical Effect Summary Table

[0112]

[0113] Working Principle: This invention realizes fault prediction and maintenance optimization of intelligent firefighting equipment based on multimodal data fusion and closed-loop control mechanism. The specific working process is as follows:

[0114] Data Perception and Dynamic Collection: Pressure sensors, electrical analyzers, and environmental sensors collect real-time data on equipment operating status. The pressure sensor dynamically adjusts the sampling rate based on the rate of pressure change (50Hz for steady-state conditions and 100Hz for sudden changes) to ensure timely capture of abnormal signals. Electrical waveform data undergoes high-precision sampling (≥1MHz) and harmonic analysis (3rd, 5th, and 7th harmonic detection) to identify circuit anomalies.

[0115] Spatiotemporal Feature Fusion: Temporal feature extraction: A bidirectional LSTM network (32 hidden units) analyzes long-term dependencies in pressure data and captures slowly changing fault modes (such as pipeline leaks). Spatial feature extraction: A multi-scale dilated convolutional network (with dilation rates of 1 / 2 / 4) identifies local distortion features in electrical waveforms (such as voltage sags and harmonic glitches). Cross-modal attention: Dynamically weights pressure and electrical features (for example, a 60% weight for sudden pressure changes) to generate a 128-dimensional comprehensive feature vector.

[0116] Fault prediction and safety control: The gradient boosted decision tree (GBDT) model outputs fault type and probability based on fused features, and the Bayesian regression model predicts the remaining useful life (RUL). A health scoring system (H = 0.7P_f + 0.3RUL) dynamically assesses device status. When H < 0.6, firmware upgrades are blocked and self-test procedures (memory checksum and sensor baseline calibration) are triggered.

[0117] Dynamic optimization of maintenance paths: Based on device topology and fault priorities, an improved genetic algorithm (population size 200, adaptive crossover rate 0.7) is used to generate cost-optimized paths. The real-time scheduling engine responds to sudden faults (such as fire alarm anomalies), and replanning takes less than 30 seconds.

[0118] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for predicting faults of intelligent fire-fighting equipment, characterized in that: The following steps are involved: S1. Dynamically collect pressure, electrical waveform, and environmental data through multi-source sensors. The pressure sensor sampling rate is dynamically adjusted according to the pressure change rate. S2. A bidirectional recurrent neural network is used to extract the long-term temporal features of pressure data, and a multi-scale convolutional network is used to capture the local mutation patterns of the electrical waveform. S3. Use the cross-modal attention mechanism to dynamically weight the temporal features and waveform features to generate a comprehensive spatiotemporal feature matrix; S4. Output the fault type and probability based on the gradient boosting decision tree model, and combine it with Bayesian regression to predict the remaining service life of the equipment; S5. When the device health score falls below the threshold, the firmware upgrade is blocked and AES-256-CBC encryption and SM3 signature verification are used to ensure upgrade security. S6. Generate the optimal maintenance path based on the dynamic programming algorithm and respond to sudden failures through the real-time scheduling engine.

2. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The dynamic adjustment strategy of the pressure sensor sampling rate in step S1 is: a 50 Hz sampling rate is used when the pressure is stable, and it is automatically increased to 100 Hz when the pressure change rate exceeds 2 MPa / s.

3. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The number of hidden units of the bidirectional recurrent neural network in step S2 is 32, and the time step is set to 60 sampling points.

4. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The weight distribution of the cross-modal attention mechanism in step S3 satisfies the following rules: when burrs and distortions appear in the electrical waveform, the waveform feature weight increases to 0.75-0.85; when the pressure drop exceeds the threshold, the timing feature weight accounts for more than 60%.

5. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The device health score in step S5 is calculated as follows: the failure probability weight accounts for 70%, and the remaining service life accounts for 30%; when the health score is lower than 0.6, the device self-check program is triggered.

6. The intelligent firefighting equipment fault prediction method according to claim 5, characterized in that: The equipment self-test procedure includes memory check and verification, sensor reference value comparison, and communication module loop test.

7. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The constraints of the dynamic programming algorithm in step S6 include: emergency failure equipment response time ≤ 2 hours; key node equipment maintenance interval ≥ 14 days; the maximum number of devices maintained per day does not exceed 50.

8. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The real-time scheduling engine performs the following operations: fire alarm failures are automatically upgraded to the highest priority; sudden failures trigger online re-planning of maintenance paths, and the re-planning takes ≤30 seconds.

9. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: Also includes: Deploy edge computing nodes to perform real-time predictions and configure NPU coprocessors to achieve an inference speed of ≥100 frames per second.

10. The intelligent firefighting equipment fault prediction method according to claim 1, characterized in that: The electrical waveform analysis includes 3rd / 5th / 7th harmonic detection, and the sampling accuracy is not less than 12 bits.

Images