Intelligent furnace tube leakage diagnosis method based on big data AI

Abstract

The invention relates to the technical field of furnace tube intelligent diagnosis, in particular to a big data AI-based furnace tube leakage intelligent diagnosis method, which comprises the following steps: collecting boiler operation data in real time through a multi-source sensor, including temperature, pressure, vibration, sound wave and flue gas component data; performing time sequence alignment, abnormal value elimination and standardized preprocessing on the collected original data; inputting the preprocessed data into a pre-trained deep neural network model, wherein the deep neural network model is obtained through comparison training of historical normal data and leakage accident data; generating a real-time diagnosis result and an early warning level based on the leakage probability value and the feature contribution degree analysis output by the model; through multi-source sensing, AI deep analysis, dynamic early warning and predictive maintenance closed loop, early, accurate and automatic diagnosis of boiler tube leakage is realized, and the safety and operation and maintenance efficiency of industrial equipment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent diagnosis of furnace pipes, and in particular to an intelligent diagnosis method for furnace pipe leakage based on big data AI. BACKGROUND

[0002] Furnace pipe leakage diagnosis refers to the process of judging and analyzing whether the furnace pipe of a boiler leaks and the location, cause and other conditions of the leakage. In the operation of a boiler, the furnace pipe covers all the heating surfaces of the boiler, such as the water wall, the superheater, the economizer and the reheater, which bear the pressure of the working medium and the action of some chemical components inside, and bear the high temperature, corrosion and wear environment outside, and are prone to failure and leakage problems. In the prior art, a single parameter threshold alarm is relied on, and the false alarm rate is as high as 30-40%, and manual inspection cannot achieve 24-hour continuous monitoring, and the existing sound wave detection technology is not sensitive to early micro-leakage, lacks multi-parameter collaborative analysis capability, and the monitoring effect is limited. SUMMARY

[0003] The present application provides an intelligent diagnosis method for furnace pipe leakage based on big data AI, which is beneficial to improve the alarm accuracy, reduce the trouble of manual monitoring, and ensure the monitoring effect.

[0004] In order to solve the problems in the prior art, the present application discloses an intelligent diagnosis method for furnace pipe leakage based on big data AI, which comprises: Step S1, collecting boiler operation data in real time through a multi-source sensor, including temperature, pressure, vibration, sound wave and flue gas composition data; Step S2, performing time sequence alignment, outlier elimination and standardization preprocessing on the collected raw data; Step S3, inputting the preprocessed data into a pre-trained deep neural network model, wherein the deep neural network model is obtained by comparing and training historical normal data and leakage accident data; Step S4, generating real-time diagnosis results and warning levels based on the leakage probability value and feature contribution degree analysis output by the model; Step S5, when the leakage probability exceeds the threshold, an alarm triggering mechanism is adopted to automatically trigger an alarm and locate the position of the leaking pipe section.

[0005] Further, the multi-source sensor comprises: an infrared thermal imager array for monitoring the temperature field distribution of the outer wall of the furnace pipe; a pressure sensor for monitoring the pressure condition inside the furnace pipe; a vibration sensor for recording the mechanical vibration spectrum of the pipe wall; a high-frequency sound wave sensor for capturing the leakage characteristic sound print in the frequency band of 20kHz-100kHz; and a laser gas analyzer for detecting the concentration change of CO, H2 and other characteristic gases.

[0006] Furthermore, the deep neural network model adopts a multimodal fusion architecture, including: a bidirectional LSTM branch for processing time series data; and a CNN branch for analyzing spatial temperature distribution.

[0007] Furthermore, a 1D convolutional neural network branch is used for voiceprint feature extraction; a joint learning layer is used to fuse the features of each branch through the attention mechanism.

[0008] Furthermore, the training process of the model includes: using transfer learning technology to pre-train based on a public industrial equipment failure dataset; expanding small sample leakage scenario data through an adversarial generative network; and using a Focal Loss function to solve the imbalance problem of the number of normal / leakage samples.

[0009] Furthermore, the feature contribution analysis includes: calculating the SHAP value to quantify the influence of each input feature on the leakage judgment; generating an interpretable heat map to mark the key sensor readings that cause the alarm; and dynamically displaying the three-dimensional visualization interface of the health status of each pipe segment.

[0010] Furthermore, it also includes: establishing a leakage evolution prediction module to predict the leakage risk in the next 6-72 hours based on current data trends; combining equipment maintenance records to recommend the optimal maintenance time window; and automatically generating a diagnostic report containing root cause analysis of the fault.

[0011] Furthermore, the alarm triggering mechanism includes: a multi-level warning strategy, when the probability value is in the range of 60%-80%, a yellow warning is triggered, 80%-95% triggers an orange warning, and above 95% triggers a red emergency response; the DCS system is automatically linked to execute load reduction or shutdown protection procedures.

[0012] Furthermore, edge computing nodes are used for real-time data preprocessing and rapid reasoning; the cloud platform big data center stores full life cycle operation data; the digital twin module synchronously simulates the operating status of the physical boiler; and the mobile terminal APP receives push alarm information.

[0013] Furthermore, the adaptive learning module regularly uses newly collected data to incrementally train the model and dynamically updates the model parameters to adapt to equipment aging or changes in operating conditions; the safety redundancy mechanism automatically switches to the backup sensor when any sensor fails, and uses a data reconstruction algorithm based on Kalman filtering to ensure diagnostic continuity; the blockchain evidence storage unit stores key alarm events and diagnostic results on the chain to ensure that the operation log cannot be tampered with and is traceable.

[0014] Compared with the prior art, the present invention achieves the following beneficial effects: Through the deployment of multi-source sensors composed of infrared thermal imagers, pressure sensors, vibration sensors, high-frequency acoustic wave sensors, and laser gas analyzers, covering multi-dimensional data monitoring such as temperature, pressure, vibration, acoustic fingerprint, and gas composition, combined with time alignment algorithm and improved DBSCAN outlier elimination, ensuring data comprehensiveness and reliability, and the pre-training model of ResNet-LSTM hybrid architecture combined with multi-modal attention mechanism can capture spatial temperature anomalies and time sequence dynamic characteristics at the same time, and the leakage probability prediction accuracy is improved by more than 30%, through dynamic early warning mechanism and multi-level threshold response, effectively balancing production loss and maintenance cost, therefore, through multi-source perception, AI deep analysis, dynamic early warning, and predictive maintenance closed loop, the technology realizes early, accurate, and automated diagnosis of boiler tube leakage, significantly improving the safety and operation efficiency of industrial equipment. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The diagnostic method steps of the present application are shown in the figure; Figure 2 The multi-source sensor composition framework of the diagnostic method of the present application is shown in the figure; Figure 3 The diagnostic system flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0016] An intelligent boiler tube leakage diagnosis method based on big data AI: Step S1, real-time acquisition of boiler operation data by multi-source sensors, including temperature, pressure, vibration, acoustic wave, and flue gas composition data; Step S2, time alignment, outlier elimination, and standardization preprocessing of the collected raw data; Step S3, input the preprocessed data into the pre-trained deep neural network model, which is obtained by comparing and training historical normal data and leakage accident data; Step S4, based on the leakage probability value and feature contribution analysis output by the model, generate real-time diagnosis results and warning levels; Step S5, when the leakage probability exceeds the threshold, use the alarm triggering mechanism to automatically trigger the alarm and locate the leakage pipe segment position.

[0017] The intelligent diagnosis method realizes early leakage detection through the above five-stage closed loop, and in the specific implementation process: A high-precision sensor network is deployed in each pipe section of the boiler. The temperature measurement range is set at 0-1200℃±0.5%, the pressure measurement is set at 0-25MPa±0.1%FS, the vibration measurement is set at a bandwidth of 0.5-10kHz, and the acoustic measurement sensitivity is set at -40dB±2dB to capture 20kHz-100kHz soundprints. A time alignment algorithm is used to solve the difference in multi-sensor sampling frequencies (such as temperature 1Hz / acoustic 100kHz). An improved DBSCAN clustering algorithm is used to eliminate outliers, and Z-score normalization is used to normalize data of different dimensions. First, when performing deep inference, the pre-trained model uses a ResNet-LSTM hybrid architecture. The input layer receives standardized data in a 60-second time window, extracts spatial features through a 12-layer residual network, and a bidirectional LSTM layer (128 hidden units) captures temporal dependencies. Finally, the Sigmoid output layer generates a leakage probability of 0-1. When making a decision output, a dynamic threshold is set (default 0.85). When the probability exceeds the limit, the alarm circuit is activated. At the same time, the Grad-CAM algorithm is used to generate a feature heat map to locate abnormal sensors with a contribution rate exceeding 15%. The alarm signal is transmitted to the DCS system via the OPC UA protocol, triggering the preset interlock protection logic (such as opening the backup pipeline and starting spray cooling).

[0018] In some embodiments, the multi-source sensor includes: an infrared thermal imager array for monitoring the temperature field distribution of the outer wall of the furnace tube; a pressure sensor for monitoring the pressure condition inside the furnace tube; a vibration sensor for recording the mechanical vibration spectrum of the tube wall; a high-frequency acoustic wave sensor for capturing the characteristic soundprint of leakage in the frequency band of 20kHz-100kHz; a laser gas analyzer for detecting changes in the concentration of characteristic gases such as CO and H2 In this embodiment, a high-precision sensor network is deployed in each pipe section of the boiler. The infrared thermal imager array adopts a K-type thermocouple array (measuring range 0-1200℃±0.5%) and a 640×512 resolution uncooled focal plane detector (working band 8-14μm). The outer wall of the furnace tube is scanned every 5 seconds. The local hot spot is identified by the temperature gradient algorithm (ΔT>30℃ / m is judged as abnormal) to monitor the temperature field distribution of the outer wall of the furnace tube. The pressure sensor adopts a piezoelectric transmitter (0-25MPa±0.1%FS) to monitor the pressure condition in the furnace tube. The vibration sensor adopts a MEMS triaxial accelerometer (bandwidth 0.5-10kHz) to record the mechanical vibration spectrum of the tube wall. The high-frequency acoustic wave sensor adopts a high-temperature resistant piezoelectric ceramic probe based on the resonance acoustic spectrum principle and is configured with 40k A Hz bandpass filter uses wavelet packet decomposition (db10 wavelet basis) to extract burst features (duration <1ms and energy exceeding the baseline by 3σ) from acoustic emission signals to capture characteristic soundprints of leaks in the 20kHz-100kHz frequency band. A laser gas analyzer is used to detect concentration changes in characteristic gases such as CO and H2. It uses TDLAS technology to scan the CO absorption line (1579.74nm) and the H2 absorption line (2121.8nm). Therefore, by deploying multimodal sensors such as infrared thermal imagers, high-frequency acoustic wave sensors, and laser gas analyzers, covering multi-dimensional data such as temperature, pressure, vibration, soundprints, and gas composition, combined with a time series alignment algorithm (to address sampling rate differences between 1Hz and 100kHz) and improved DBSCAN outlier removal, data comprehensiveness and reliability are ensured.

[0019] In some embodiments, the deep neural network model adopts a multimodal fusion architecture, including: a bidirectional LSTM branch for processing time series data; a CNN branch for analyzing spatial temperature distribution; a 1D convolutional neural network branch for voiceprint feature extraction; and a joint learning layer that fuses the features of each branch through an attention mechanism.

[0020] In this embodiment, a deep neural network model is used to preprocess data using a multimodal fusion architecture, a time series alignment algorithm is used to solve the sampling frequency differences of multiple sensors (such as temperature 1Hz / sound wave 100kHz), an improved DBSCAN clustering algorithm is used to remove outliers, and data of different dimensions are normalized by Z-score standardization. Then, a deep inference engine is used to process time series data using a ResNet-LSTM hybrid architecture. The input layer receives standardized data of a 60-second time window, and spatial features are extracted through a 12-layer residual network. The bidirectional LSTM layer (128 hidden units) captures temporal dependencies, and finally a 0-1 leakage probability is generated by the Sigmoid output layer. Then, a 2D-CNN branch is used to analyze the spatial features of the infrared thermal map (ResNet18 architecture) to analyze the spatial temperature, and a 1D-CNN branch is used to extract the sound wave spectrum features (5 layers of convolution, kernel size=7), realizes voiceprint feature extraction, and fuses the joint learning layer of each branch feature through the multi-head attention mechanism. Therefore, the pre-trained model based on the ResNet-LSTM hybrid architecture (12-layer residual network + bidirectional LSTM) combined with the multimodal attention mechanism (CNN processing heat map, 1D-CNN analysis of voiceprint) can simultaneously capture spatial temperature anomalies and temporal dynamic features, and the accuracy of leakage probability prediction is improved by more than 30%.

[0021] In some embodiments, the training process of the model includes: using transfer learning technology to pre-train based on a public industrial equipment failure data set; expanding small sample leakage scenario data through an adversarial generative network; and using a FocalLoss function to solve the problem of imbalance in the number of normal / leakage samples.

[0022] In this embodiment, a CNN feature extractor is pre-trained on the public PHM2014 bearing dataset. The convolution kernel weights are retained and then fine-tuned with a boiler-specific classifier. Data augmentation is then performed. Wasserstein GAN is used to generate synthetic leakage data. The generator inputs a 100-dimensional noise vector and outputs simulated acoustic-temperature joint features through five layers of transposed convolution, covering leakage patterns under different pipe diameters and pressure conditions. Loss function optimization is then performed, with Focal Loss set to α=0.75 and γ=2. A 3x weight is applied to minority class samples. Training is performed for 200 epochs with the Adam optimizer (lr=1e-4), focusing on optimizing the classification boundaries of the minority class (leakage samples). Therefore, the Focal Loss function is used to enhance training of small sample leakage data to address the sample imbalance problem, significantly enhancing the model's generalization ability for rare leakage scenarios.

[0023] In some embodiments, a leakage evolution prediction module is established to predict the leakage risk in the next 6-72 hours based on the current data trend, and recommend the optimal maintenance time window in combination with the equipment maintenance record: a diagnostic report containing the root cause analysis of the fault is automatically generated, and the alarm triggering mechanism includes: a multi-level early warning strategy, when the probability value is in the range of 60%-80%, a yellow warning is triggered, 80%-95% triggers an orange warning, and above 95% triggers a red emergency response; the DCS system is automatically linked to execute a load reduction or shutdown protection program.

[0024] In this embodiment, feature contribution analysis and SHAP value calculation are performed: the Monte Carlo sampling approximation algorithm is used to quantify the contribution of each sensor to the current prediction (for example, the 87th frequency band of the acoustic spectrum contributes +23%). 3D visualization is performed: the digital twin of the boiler is rendered using the Unity3D engine, and pipe sections with a leakage probability greater than 95% are highlighted in red (with a positioning accuracy of ±15 cm). Then, a dynamic early warning mechanism and multi-level threshold response are implemented: Root Cause Analysis Report: Automatically correlates historical maintenance records, recommends the most likely failure modes (such as high-temperature creep and stress corrosion), and performs predictive maintenance. It uses a time convolutional network (TCN) to predict leakage evolution trends. It inputs 72 hours of historical data and outputs a risk curve for the next 6-72 hours (with a 95% confidence interval). It then dynamically optimizes maintenance plans based on reinforcement learning to balance production losses and maintenance costs.

[0025] In some embodiments, edge computing nodes are used for real-time data preprocessing and rapid reasoning; cloud platform big data centers store full life cycle operation data; digital twin modules synchronously simulate the operating status of physical boilers; and mobile terminal apps receive push alarm information.

[0026] In this embodiment, edge-cloud collaborative computing is finally performed as follows: Edge computing nodes use NVIDIA Jetson AGX Xavier (32TOPS computing power) to deploy lightweight models (latency <15ms after TensorRT optimization); The cloud platform big data center is built on a Kubernetes-based big data lake cluster. The time series database uses InfluxDB (with an 80% compression rate), storing more than 10 years of full data. Spark ML is used for processing and supports Spark offline analysis. The digital twin module synchronizes digital twins and uses ANSYS Twin Builder to build a 1:1 simulation model. The high-fidelity simulation model is updated every 5 minutes, integrating physics engines such as FEM stress analysis and CFD flow field simulation. Model recalibration is triggered when the deviation exceeds 5%. Mobile terminal APP, mobile terminal function supports AR inspection navigation, maintenance personnel superimposes the health status of the pipe section (green / yellow / red) through the mobile phone camera, supports alarm confirmation (fingerprint + GPS location double-factor authentication), real-time viewing of the three-dimensional section view of the pipe section, and finally pushes the alarm information containing the on-site disposal guide.

[0027] In some embodiments, the adaptive learning module periodically performs incremental training on the model with newly collected data, dynamically updates the model parameters to adapt to equipment aging or working condition changes; the safety redundancy mechanism automatically switches to the backup sensor when any sensor fails, and uses a Kalman filter-based data reconstruction algorithm to ensure continuous diagnosis; the blockchain storage unit stores key alarm events and diagnosis results on the chain to ensure that the operation log cannot be tampered with and is traceable.

[0028] In this embodiment, the adaptive learning module performs incremental training and dynamic updating. The system periodically collects new data (such as vibration, temperature, current, etc.) of the equipment running, uses these data to perform incremental training (rather than training from scratch) on the existing diagnostic model (such as a deep learning or machine learning model), dynamically adjusts the model parameters, adapts to aging and working condition changes, and through continuous learning, the model can capture the performance degradation trend of the equipment or environmental changes (such as load fluctuations), avoiding the decline in diagnostic accuracy due to long-term use; The safety redundancy mechanism automatically switches to the backup sensor when the system detects that a sensor has failed (such as signal loss or exceeding the reasonable range), ensuring uninterrupted data collection. The data reconstruction algorithm uses Kalman filtering (an optimal estimation algorithm) to reconstruct missing or abnormal data, and then predicts the current missing value by combining historical data through the physical correlation between sensors (such as multiple measuring points of the same device), maintaining the continuity of the diagnosis process; The blockchain storage unit writes alarm events (such as device anomaly warning), diagnosis results (such as fault type), and operation logs (such as maintenance records) into the blockchain network, and then uses the distributed ledger and hash encryption technology of the blockchain to ensure that the data cannot be modified once it is on the chain. Through the timestamp and chain transaction ID, the complete timeline and responsible person operation of the event can be traced back, suitable for compliance audit or fault review, collaborative workflow, and real-time diagnosis. The system collects data through multiple sensors, ensures data integrity through the redundancy mechanism, inputs the data into the diagnostic model, dynamically optimizes it, and then the adaptive module updates the model with new data periodically to maintain diagnostic accuracy. Finally, a safe record is made, and when the diagnosis result triggers an alarm, the blockchain unit automatically stores the evidence to form a trusted record. Therefore, through redundancy and filtering algorithms, the impact of hardware failure is reduced, the model is self-adapting to extend the effective life cycle of the system, and the blockchain meets the stringent requirements of the industrial field for data authenticity.

Claims

1. A furnace tube leakage intelligent diagnosis method based on big data AI, a furnace tube leakage intelligent diagnosis method based on big data AI, characterized in that, include: Step S1, collecting boiler operation data in real time through multi-source sensors, including temperature, pressure, vibration, sound wave and flue gas composition data; Step S2, performing time series alignment, outlier removal and standardization preprocessing on the collected raw data; Step S3: inputting the pre-processed data into a pre-trained deep neural network model, wherein the deep neural network model is trained by comparing historical normal data with leakage accident data; Step S4: Based on the leakage probability value and feature contribution analysis output by the model, generate real-time diagnosis results and warning levels; Step S5: When the leakage probability exceeds a threshold, an alarm triggering mechanism is used to automatically trigger an alarm and locate the leaking pipe section.

2. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 1 is characterized in that: The multi-source sensor comprises: Infrared thermal imager array, used to monitor the temperature field distribution on the outer wall of the furnace tube; Pressure sensor, used to monitor the pressure condition inside the furnace tube; Vibration sensor, used to record the mechanical vibration spectrum of the pipe wall; High-frequency acoustic wave sensor, used to capture leakage characteristic sound patterns in the 20kHz-100kHz frequency band; Laser gas analyzer, used to detect changes in concentration of characteristic gases such as CO and H2.

3. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 2 is characterized in that: The deep neural network model adopts a multimodal fusion architecture, including: Bidirectional LSTM branch for processing time series data; CNN branch for analyzing spatial temperature distribution.

4. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 3 is characterized in that: The deep neural network model adopts a multimodal fusion architecture and also includes: 1D convolutional neural network branch for voiceprint feature extraction; A joint learning layer that fuses features from each branch through an attention mechanism.

5. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 1, characterized in that: The training process of the model includes: Use transfer learning technology and pre-train based on a public industrial equipment failure dataset; Expanding small sample leakage scenario data through adversarial generative networks; The Focal Loss function is used to solve the problem of imbalance in the number of normal / leakage samples.

6. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 1, characterized in that: The feature contribution analysis includes: Calculate the SHAP value to quantify the influence of each input feature on the leakage judgment; Generate interpretable heat maps that annotate key sensor readings that led to alarms; A three-dimensional visualization interface that dynamically displays the health status of each pipe segment.

7. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 1, characterized in that: Also includes: Establish a leakage evolution prediction module to predict leakage risks in the next 6-72 hours based on current data trends; Recommend the optimal maintenance time window based on equipment maintenance records; Automatically generate diagnostic reports including root cause analysis.

8. The intelligent diagnosis method for furnace tube leakage based on big data AI according to claim 1, characterized in that: The alarm triggering mechanism includes: Multi-level warning strategy: when the probability value is between 60% and 80%, a yellow warning is triggered; when it is between 80% and 95%, an orange warning is triggered; and when it is above 95%, a red emergency response is triggered; Automatically link the DCS system to execute load reduction or shutdown protection procedures.

9. An intelligent diagnostic system for implementing the method according to any one of claims 1 to 7, characterized in that: include: Edge computing nodes, used for real-time data preprocessing and fast reasoning; Cloud platform big data center, storing full life cycle operation data; Digital twin module, synchronously simulates the operating status of the physical boiler; Mobile terminal APP, receive push alarm information.

10. The intelligent diagnostic system according to claim 9, characterized in that: Also includes: Adaptive learning module, which regularly uses newly collected data to incrementally train the model and dynamically updates model parameters to adapt to equipment aging or changes in operating conditions; Safety redundancy mechanism: automatically switches to the backup sensor when any sensor fails, and uses a data reconstruction algorithm based on Kalman filtering to ensure diagnostic continuity; The blockchain evidence storage unit stores key alarm events and diagnostic results on-chain, ensuring that operation logs cannot be tampered with and are traceable.

Images