Boiler state abnormity monitoring system based on artificial intelligence

Abstract

The invention discloses a boiler state abnormity monitoring system based on artificial intelligence, and particularly relates to the field of abnormity monitoring, comprising a multi-modal sensor data acquisition module, an edge computing node module, a migration enhanced abnormity detection module and a fault tracing module; according to the invention, a boiler multi-mode sensor network is constructed, vibration, temperature, pressure and other data acquisition modules are integrated, adaptive filtering and working condition classification are realized through edge computing nodes, and a data enhancement strategy is dynamically adjusted; the anomaly detection module fuses a physical mechanism model and a twin neural network, generates a multi-dimensional deviation index, and starts fault tracing after a dynamic threshold triggers an alarm; probability propagation is carried out based on a Bayesian knowledge graph, fault trajectories are simulated in combination with an adversarial generative network, root cause probability ranking and maintenance priority scores are generated, real-time support is provided for operation and maintenance decisions, and intelligent diagnosis and dynamic optimization of the boiler state are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormality monitoring, and more particularly to an artificial intelligence-based boiler state abnormality monitoring system. Background Art

[0002] In the field of industrial production, boilers, as key thermal power equipment, are widely used in many industries such as electricity, chemical industry, and metallurgy. Their safe and stable operation is crucial to ensuring production continuity, improving energy utilization efficiency, and avoiding safety accidents. However, in actual operation, boilers face complex operating conditions, variable fuel characteristics, and long-term mechanical wear and tear, and are prone to various abnormal conditions such as dust accumulation, coking, leakage, and component aging. If these abnormalities are not discovered and handled in a timely manner, they will not only lead to reduced boiler operating efficiency and energy waste, but may also cause serious safety accidents, resulting in significant casualties and property losses.

[0003] Traditional boiler status monitoring methods mainly rely on manual inspections and simple instrument monitoring. They have problems such as limited monitoring range, untimely data collection, and highly subjective abnormality judgment. They are unable to meet the high requirements of modern industrial production for safe and efficient boiler operation. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a boiler state abnormality monitoring system based on artificial intelligence to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Multimodal sensor data acquisition module: including vibration sensor, temperature sensor group, pressure sensor, voiceprint acquisition unit and infrared thermal imager, which respectively collect the mechanical vibration spectrum, three-dimensional temperature field distribution data, steam pressure waveform, voiceprint characteristics and thermal radiation image of the boiler body in real time;

[0007] Edge computing node module: includes an adaptive filtering unit and an operating condition classifier, which are used to perform operating condition perception and noise reduction on raw sensor data and generate dynamic data enhancement strategies based on the boiler's current load rate and combustion mode;

[0008] Transfer-enhanced anomaly detection module: This module integrates a pre-trained physical mechanism simulation model with an online updated twin neural network. By comparing the virtual sensor output of the digital twin with the difference between the measured data, it generates multi-dimensional deviation indicators.

[0009] The fault tracing module constructs a Bayesian knowledge graph for the boiler, with nodes containing equipment, faults, and parameters. When an anomaly occurs, probability propagation is used to generate the maximum possible fault chain, combined with data verification and adversarial generation network simulation trajectory.

[0010] Preferably, in the multimodal sensor data acquisition module, the three-dimensional vibration sensor network uses a MEMS accelerometer array to capture the 0-5kHz wide-band mechanical vibration spectrum at a sampling frequency of 10kHz, and 32-dimensional characteristic parameters such as blade passing frequency and shaft torsional vibration can be extracted through FFT analysis;

[0011] The temperature sensor group consists of 48 PT100 platinum resistors forming a three-layer ring array. Combined with the thermocouple compensation algorithm, it can achieve three-dimensional temperature field reconstruction in the boiler combustion chamber within the range of 0-1200℃ with a spatial resolution of 5cm.

[0012] The digital pressure sensor uses STRAIN-GAGE technology to monitor the dynamic pressure waveform of 0-25MPa in the steam pipeline with a response time of 1 millisecond and integrates a peak hold function;

[0013] The voiceprint collection unit includes a 4-directional microphone array, which uses beamforming technology to locate noise sources within a 120dB dynamic range and can extract characteristic frequency bands such as combustion oscillations.

[0014] The infrared thermal imager is equipped with a 640×512 vanadium oxide detector, which outputs 14-bit thermal radiation images at a 30Hz frame rate in the 8-14 micron band, and can realize hot spot detection in conjunction with temperature field data.

[0015] Preferably, in the edge computing node module, the adaptive filtering unit includes a three-stage processing pipeline: first, a dynamic noise eliminator based on the LMS algorithm performs a 50-800Hz bandpass filter on the vibration signal to eliminate mechanical transmission chain interference; secondly, an improved Kalman filter is used to smooth the temperature field data in the time and space domain, effectively suppressing temperature measurement jumps while maintaining an accuracy of ±2°C; finally, the voiceprint signal is processed by a wavelet threshold denoising module to retain the key frequency band characteristics of 200-5000Hz;

[0016] The operating condition classifier deploys a lightweight deep neural network. The input layer receives preprocessed multimodal feature vectors and uses a three-layer convolutional and two-layer fully connected network to identify eight typical boiler operating conditions in real time, keeping classification latency within 15ms.

[0017] The system dynamically adjusts data enhancement strategies for different operating conditions. Under high-load conditions, it initiates time-domain stretching of vibration signals to generate simulation data that includes rotor imbalance characteristics. Under low-NOx combustion mode, it uses a temperature field completion algorithm based on a generative adversarial network to effectively expand the sample size in the sparse temperature measurement area. All processed data is timestamped and labeled with operating conditions and uploaded to the cloud platform via the OPC UA protocol. A 10-minute sliding window cache is also retained locally to support iterative updates of the real-time health assessment algorithm at the edge.

[0018] The kurtosis-entropy joint index of each sensor signal is calculated by sliding the time window, and the decision tree splitting threshold is dynamically adjusted. The calculation method of the kurtosis-entropy joint index is as follows:

[0019]

[0020] Where J is the combined index of kurtosis and entropy, b is the fourth-order moment kurtosis of the vibration signal, H(p) is the Shannon entropy of the temperature field distribution, and b is the max Expressed as the maximum value of the vibration signal kurtosis, H max Expressed as the maximum value of Shannon entropy;

[0021] This indicator is used to dynamically adjust the splitting threshold of the decision tree, so that stricter splitting conditions are adopted under high-load conditions (J>3.0) and the splitting restrictions are relaxed under transient conditions (J<1.5) to improve the model's adaptability to non-stationary processes.

[0022] When an unknown operating mode is detected (i.e., the prediction confidence of all decision trees is lower than the threshold δ = 0.65), the semi-supervised clustering unit is activated. This unit is based on the improved DBSCAN algorithm and uses the Mahalanobis distance of multimodal data as the similarity measure. After the clustering results generate temporary operating condition labels, the incremental learning mechanism is triggered: first, the c decision trees with the highest confidence are selected through integrated weighted voting, and then the model parameters are updated using the online bagging method. At the same time, a forgetting factor is introduced to prevent overfitting of historical data. After each incremental learning, the system will re-evaluate the feature importance and dynamically adjust the feature selection strategy.

[0023] Preferably, the migration-enhanced anomaly detection module adopts a dual-channel hybrid architecture to deeply integrate the physical mechanism model with the data-driven approach. The physical mechanism simulation model is constructed based on the thermodynamic principles of boilers, and numerical simulation calculations are used to generate benchmark data containing key parameters such as temperature field and pressure distribution, forming a virtual sensor dataset for the digital twin.

[0024] The simulation model outputs spatiotemporal evolution data of key parameters, including temperature fields and pressure distribution, forming a virtual sensor dataset for the digital twin. The twin neural network, operating in parallel, uses an online incremental learning mechanism. Its network structure includes a feature extraction branch that uses multi-scale residual units to process measured sensor data.

[0025] The prior probability P(abn) is adaptively adjusted according to the runtime, and the likelihood function P(V|abn) is modeled using a mixed Gaussian distribution. When the posterior probability exceeds the dynamic threshold, an abnormal alarm is triggered and the fault tracing and analysis process is started.

[0026] Preferably, in the fault tracing module, the Bayesian knowledge graph of the boiler status is constructed using a triple structure, which includes three types of nodes: equipment components, fault modes and monitoring parameters, and defines the conditional probability through the edge relationship. The edge condition is defined according to the conditional probability: equipment → fault, P(F j |C i ): Component C i Causes fault F j The prior probability of comes from historical maintenance records;

[0027] Fault → Parameter, P(S d |F j ): Fault F j Leading to monitoring parameter S d The probability of anomalies comes from mechanistic model simulation of equipment components such as burners, economizers, and steam drums, failure modes such as dust accumulation, corrosion, and leakage, and monitoring modes including vibration spectrum, temperature gradient, and pressure fluctuation;

[0028] When the anomaly detection unit triggers an alarm, it initializes observation evidence and performs probability propagation calculations along the graph edge. It uses the Loopy Belief Propagation algorithm to calculate the posterior probability of the faulty node and reversely infers the probability of device failure, ultimately generating the maximum possible fault chain.

[0029] The adversarial generative network is used to simulate the fault development trajectory. The generator inputs the fault chain and current data and outputs virtual sensor data for a period of time in the future. The discriminator evaluates the authenticity of the generated data to predict the evolution trend and potential impact of the fault. The specific method for calculating the future virtual sensor data is as follows:

[0030]

[0031] in, represents the generated future virtual sensor data, G represents the generator, and Y 1:t It represents the temperature sensor data in the past t minutes, and N represents the fault chain;

[0032] During the diagnostic report generation phase, the report outputs root cause probability rankings, dynamic fault propagation simulations, and maintenance priority scores. The scores comprehensively consider the probability, severity, and urgency of the fault, providing decision support for operations and maintenance personnel.

[0033] Technical effects and advantages of the present invention:

[0034] The present invention integrates multimodal sensors to collect boiler operation data in real time, performs adaptive filtering and operating condition classification through edge computing nodes, and then uses digital twins and twin neural networks for comparative analysis to generate multi-scale deviation indicators. Combined with the Bayesian knowledge graph, fault tracing is achieved, and GAN is used to predict fault evolution. The present invention has real-time monitoring and rapid response capabilities, and can issue early warning signals in time at the early stage of anomalies, thereby gaining valuable processing time for maintenance personnel and avoiding accidents. Through edge computing and dynamic data enhancement strategies, the system can quickly adapt to different operating conditions to ensure that anomalies can be detected in time under various operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a module schematic diagram of the present invention.

[0036] Figure 2 Schematic diagram of the anomaly detection process of the present invention. DETAILED DESCRIPTION

[0037] 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.

[0038] See also Figure 1 As shown, the present invention provides an artificial intelligence-based boiler state abnormality monitoring system, including a multimodal sensor data acquisition module, an edge computing node module, a migration-enhanced abnormality detection module, and a fault tracing module:

[0039] Multimodal sensor data acquisition module: including vibration sensor, temperature sensor group, pressure sensor, voiceprint acquisition unit and infrared thermal imager, which respectively collect the mechanical vibration spectrum, three-dimensional temperature field distribution data, steam pressure waveform, voiceprint characteristics and thermal radiation image of the boiler body in real time;

[0040] In the multimodal sensor data acquisition module, the three-dimensional vibration sensor network uses a MEMS accelerometer array to capture the 0-5kHz wide-band mechanical vibration spectrum at a 10kHz sampling frequency. Through FFT analysis, 32-dimensional characteristic parameters such as blade passing frequency and shaft torsional vibration can be extracted.

[0041] The temperature sensor group consists of 48 PT100 platinum resistors forming a three-layer ring array. Combined with the thermocouple compensation algorithm, it can achieve three-dimensional temperature field reconstruction in the boiler combustion chamber within the range of 0-1200℃ with a spatial resolution of 5cm.

[0042] The digital pressure sensor uses STRAIN-GAGE technology to monitor the dynamic pressure waveform of 0-25MPa in the steam pipeline with a response time of 1 millisecond and integrates a peak hold function;

[0043] The voiceprint collection unit includes a 4-directional microphone array, which uses beamforming technology to locate noise sources within a 120dB dynamic range and can extract characteristic frequency bands such as combustion oscillations.

[0044] The infrared thermal imager is equipped with a 640×512 vanadium oxide detector, which outputs 14-bit thermal radiation images at a 30Hz frame rate in the 8-14 micron band, and can realize hot spot detection in conjunction with temperature field data.

[0045] Edge computing node module: includes an adaptive filtering unit and an operating condition classifier, which are used to perform operating condition perception and noise reduction on raw sensor data and generate dynamic data enhancement strategies based on the boiler's current load rate and combustion mode;

[0046] In the edge computing node module, the adaptive filtering unit includes a three-stage processing pipeline: first, a dynamic noise eliminator based on the LMS algorithm performs a 50-800Hz bandpass filter on the vibration signal to eliminate mechanical transmission chain interference; second, an improved Kalman filter is used to smooth the temperature field data in the time and space domain, effectively suppressing temperature jumps while maintaining an accuracy of ±2°C; finally, the voiceprint signal is processed through a wavelet threshold denoising module to retain the key frequency band characteristics of 200-5000Hz;

[0047] The operating condition classifier deploys a lightweight deep neural network. The input layer receives preprocessed multimodal feature vectors and uses a three-layer convolutional and two-layer fully connected network to identify eight typical boiler operating conditions in real time, keeping classification latency within 15ms.

[0048] The system dynamically adjusts data enhancement strategies for different operating conditions. Under high-load conditions, it initiates time-domain stretching of vibration signals to generate simulation data that includes rotor imbalance characteristics. Under low-NOx combustion mode, it uses a temperature field completion algorithm based on a generative adversarial network to effectively expand the sample size in the sparse temperature measurement area. All processed data is timestamped and labeled with operating conditions and uploaded to the cloud platform via the OPC UA protocol. A 10-minute sliding window cache is also retained locally to support iterative updates of the real-time health assessment algorithm at the edge.

[0049] The working condition classifier adopts the improved random forest algorithm:

[0050] Construct an operating condition characteristic vector including combustion efficiency, load change gradient, and fuel calorific value compensation coefficient;

[0051] The calculation method of combustion efficiency is as follows:

[0052]

[0053] Where η represents the combustion efficiency, Q 实际 Expressed as the actual heat released, Q 理论 Expressed as the theoretically released heat, T i It is the temperature of the ith partition measured by the infrared thermal imager, A i It is represented as the area weight corresponding to the i-th partition, α is the heat loss compensation coefficient, a is the fuel flow rate, and LHV is the lower heating value of the fuel, which represents the heat that can be released when the fuel per unit mass or volume is completely burned;

[0054] The load change gradient uses a sliding time window to calculate the first-order differential entropy value of the pressure signal. The calculation method is as follows:

[0055]

[0056] Among them, H represents the first-order difference entropy value, p k It is expressed as the probability of the k-th pressure change value, and M is expressed as the total number of pressure change values ​​in the sliding time window;

[0057]

[0058] Among them, p k Expressed as the probability of the kth pressure change value, Expressed as the kth pressure change value, ΔP t Expressed as the sum of the absolute values ​​of all pressure changes;

[0059] The fuel calorific value compensation coefficient is dynamically corrected by the voiceprint main frequency offset. The calculation method of the fuel calorific value compensation coefficient is as follows:

[0060]

[0061] Among them, β represents the fuel calorific value compensation coefficient, γ represents the correction coefficient, and f 实际 It is expressed as the actual measured value of the main frequency of the voiceprint, f 基准 It is expressed as the base value of the voiceprint main frequency;

[0062] The kurtosis-entropy joint index of each sensor signal is calculated by sliding the time window, and the decision tree splitting threshold is dynamically adjusted. The calculation method of the kurtosis-entropy joint index is as follows:

[0063]

[0064] Where J is the combined index of kurtosis and entropy, b is the fourth-order moment kurtosis of the vibration signal, H(p) is the Shannon entropy of the temperature field distribution, and b is the max Expressed as the maximum value of the vibration signal kurtosis, H max Expressed as the maximum value of Shannon entropy;

[0065] This indicator is used to dynamically adjust the splitting threshold of the decision tree, so that stricter splitting conditions are adopted under high-load conditions (J>3.0) and the splitting restrictions are relaxed under transient conditions (J<1.5) to improve the model's adaptability to non-stationary processes.

[0066] When an unknown operating mode is detected (i.e., the prediction confidence of all decision trees is lower than the threshold δ = 0.65), the semi-supervised clustering unit is activated. This unit is based on the improved DBSCAN algorithm and uses the Mahalanobis distance of multimodal data as the similarity measure. After the clustering results generate temporary operating condition labels, the incremental learning mechanism is triggered: first, the c decision trees with the highest confidence are selected through integrated weighted voting, and then the model parameters are updated using the online bagging method. At the same time, a forgetting factor is introduced to prevent overfitting of historical data. After each incremental learning, the system will re-evaluate the feature importance and dynamically adjust the feature selection strategy.

[0067] Transfer-enhanced anomaly detection module: This module integrates a pre-trained physical mechanism simulation model with an online updated twin neural network. By comparing the virtual sensor output of the digital twin with the difference between the measured data, it generates multi-dimensional deviation indicators.

[0068] The migration-enhanced anomaly detection module uses a dual-channel hybrid architecture to deeply integrate the physical mechanism model with the data-driven approach. The physical mechanism simulation model is built based on the thermodynamic principles of boilers. Numerical simulations generate benchmark data containing key parameters such as temperature fields and pressure distribution, forming a virtual sensor dataset for the digital twin. The model discretizes the three-dimensional heat transfer equation using the finite volume method:

[0069]

[0070] Where ρ represents density, c represents specific heat capacity, Indicates the rate of change of heat capacity per unit volume, Q comb Expressed as the heat released by fuel combustion, The partial derivative of temperature with time represents the rate of change of temperature with time, where R represents temperature and t represents time; Heat conduction term, q is the thermal conductivity, is the temperature gradient, Expressed as a divergence operator;

[0071] The simulation model outputs spatiotemporal evolution data of key parameters such as temperature field and pressure distribution, forming a virtual sensor dataset for the digital twin. The twin neural network, working in parallel, uses an online incremental learning mechanism. Its network structure includes a feature extraction branch that uses multi-scale residual units to process the measured sensor data:

[0072] f(B t )=ResBlock n (…ResBlock1(B t ))

[0073] Among them, f(B t ) is represented as a deep learning model based on residual network, which is used to extract the input data B t Extract features and predict faults; B t Represented as real-time input data, ResBlock n It is represented as a deep network composed of n residual blocks, which is used to process complex nonlinear relationships. t ) is represented as the first residual block in the network, responsible for preliminary feature extraction;

[0074] The relationship measurement branch calculates the feature space distance between the virtual data and the measured data. The calculation method is as follows:

[0075] e t =g φ (f(B t ),f(D t ))

[0076] Among them, e t It is expressed as the feature space distance at time t, which represents the distance measure between virtual data and measured data in the feature space; g φ Denote as a function representing the metric branch, f(D t ) is expressed as the virtual data D t Perform feature extraction;

[0077] e t The formula describes the time t, through the feature extraction function f (*) to measure the data B t and dummy data D t Mapped to the feature space, and then measured by the relationship function g φ Calculate the feature space distance e between themt ;

[0078] The calculation of the deviation index introduces a multi-scale analysis framework and uses a sliding window to calculate statistical features in the time dimension, including instantaneous deviation, trend deviation, and periodic deviation. The periodic deviation is obtained by extracting the phase difference of the dominant frequency component through FFT.

[0079] The calculation method of instantaneous deviation is as follows:

[0080]

[0081] Among them, W inst Expressed as instantaneous deviation, e t Expressed as the feature space distance at time t, μ e Expressed as the mean of the historical feature space distance, σ e Expressed as the standard deviation of the feature space distance;

[0082] The calculation method of trend deviation is as follows:

[0083] W trend =LinReg(e t-o ,…,e t )

[0084] Among them, W trend It is expressed as trend deviation, LinReg is expressed as linear regression function, which is used to fit historical data; t-o Expressed as the feature space distance at time to;

[0085] In the spatial dimension, a propagation model based on a graph structure is constructed. Nodes represent monitoring points, and edge weights reflect the strength of physical associations. The calculation method for abnormal propagation values ​​is as follows:

[0086]

[0087] Among them, L d is represented by the abnormal propagation value of node d, K(d) is represented by the neighbor set of node d, and W h Expressed as the deviation of node h, w dh Expressed as the edge weight between nodes d and h, x d Expressed as the spatial coordinate of node d, x h It is represented as the spatial coordinate of node h, and l is represented as the attenuation value, which is used to control the attenuation speed of the abnormal propagation value;

[0088]

[0089] Among them, P(abn|V) is represented as the posterior probability, which is the probability that the boiler is in an abnormal state under the observed data V. P(V|abn) is represented as the likelihood function, which is the probability of observing data V in an abnormal state. P(abn) is represented as the prior probability, which is the initial probability that the boiler is in an abnormal state. P(V|w) is represented as the likelihood function, which is the probability of observing data V in state w. P(w) is represented as the initial probability that the boiler is in state w.

[0090] The prior probability P(abn) is adaptively adjusted based on the runtime, and the likelihood function P(V|abn) is modeled using a mixed Gaussian distribution. When the posterior probability exceeds the dynamic threshold, an abnormality alarm is triggered, and the fault tracing and analysis process is initiated.

[0091] The dynamic threshold is calculated as follows:

[0092] τ t =μ t +3σ t

[0093] Among them, τ t Expressed as dynamic threshold, μ t Expressed as the historical mean at time t, σ t Expressed as historical standard deviation.

[0094] For example, digital twin benchmark data is generated based on boiler design parameters and CFD simulation, and a superheater tube wall temperature distribution model under normal operating conditions is established (350±5°C on side A and 348±5°C on side B). Fault characteristic modes such as dust accumulation and coking are also pre-stored. When the infrared thermal imager detects that the tube wall temperature on side A abnormally rises to 360°C, and the vibration sensor captures a 15% increase in energy in the 1-2kHz frequency band, the twin neural network immediately initiates multi-dimensional analysis: wavelet packet decomposition is used to identify the abnormal frequency band of the vibration signal, asymmetric convolution is used to process the thermal imaging data to discover local temperature gradient distortion, and the dynamic time warping algorithm shows that the temperature rise rate deviates from the benchmark by 20%. The energy difference coefficient of the vibration signal in the 1.5kHz frequency band is equal to 1.8 (the threshold is 1.3), and the time domain trend deviation (W trend =2.3) and other multi-scale indicators, and calculated an abnormal probability of 89% through Bayesian reasoning, far exceeding the dynamic threshold of 80%, immediately triggering an abnormality alarm and initiating the fault tracing and analysis process.

[0095] The fault tracing module builds a Bayesian knowledge graph for boilers, with nodes including equipment, faults, and parameters. When an anomaly occurs, probability propagation is used to generate the maximum possible fault chain, combining data verification with adversarial generative network simulation trajectories.

[0096] In the fault tracing module, the Bayesian knowledge graph of boiler status is constructed using a triple structure, which includes three types of nodes: equipment components, fault modes, and monitoring parameters, and defines conditional probabilities through edge relationships;

[0097] According to the definition of conditional probability: equipment → failure, P(F j |C i ): Component C i Causes fault F j The prior probability of comes from historical maintenance records;

[0098] Fault → Parameter, P(S d |F j ): Fault F j Leading to monitoring parameter S d The probability of anomaly, comes from the mechanism model simulation;

[0099] For example: economizer (P = 0.15) → tube wall corrosion (P = 0.9) → feedwater flow deviation;

[0100] Equipment components such as burners, economizers, and steam drums, failure modes such as dust accumulation, corrosion, and leakage, monitoring modes including vibration spectrum, temperature gradient, and pressure fluctuation;

[0101] When the anomaly detection unit triggers an alarm, it initializes observation evidence and performs probability propagation calculations along the graph edge. It uses the Loopy Belief Propagation algorithm to calculate the posterior probability of the faulty node and reversely infers the probability of device failure, ultimately generating the maximum possible fault chain.

[0102] Activate the relevant monitoring parameter nodes and initialize the observation evidence: E={S1=s1,…,S m =s m}

[0103] Among them, E represents the set of observation evidence, which represents the currently activated monitoring parameter nodes and their observation values; S1, S2, ..., S m Represented as monitoring parameter nodes, s1, s2, ..., s m It is expressed as an observation value, which represents the current state measurement value of the corresponding monitoring parameter node;

[0104] Probability propagation along the edge: Calculate the posterior probability of the faulty node. The calculation method is as follows:

[0105]

[0106] Among them, P(F j |E) represents fault F j The posterior probability reflects the matching degree between the fault and the observed evidence; P(F j) is represented by fault F j The prior probability, P(S m =s m |F j ) is represented by fault F j Leading to monitoring parameter S m The probability of anomaly, E is represented by the set of observation evidence;

[0107]

[0108] Among them, P(C i |E) is represented as component C i The posterior probability of component C is given the observed evidence E. i The probability of causing a failure; P(F j |E) represents fault F j The posterior probability, P(C i |F j )Component C i Causes fault F j probability;

[0109]

[0110] Among them, N * It is represented as the maximum possible fault chain, Z is the fault chain, (u, v) is the edge in the fault chain, indicating the causal relationship, and P(v|u) is the conditional probability, indicating the probability of v occurring when u occurs.

[0111] Dynamic hypothesis testing is performed in combination with real-time operating data. First, the expected parameter change pattern of the fault hypothesis is constructed, and the degree of matching between it and the measured data is calculated using the dynamic time warping (DTW) algorithm to screen out candidate faults that meet the current operating conditions. The specific calculation method is as follows:

[0112] D DTW =DTW(Y hyp ,Y real )

[0113] Among them, D DTW It is represented by the dynamic time warping distance, DTW represents the dynamic time warping algorithm, which is used to calculate the nonlinear alignment distance between two time series; Y hyp Expressed as the theoretical predicted value of the monitoring parameter under the fault mode, Y real It represents the actual recorded value of the monitoring parameter when the equipment is running;

[0114] When the calculated D DTW When it is greater than the threshold, the current hypothesis is rejected;

[0115] The adversarial generative network is used to simulate the fault development trajectory. The generator inputs the fault chain and current data and outputs virtual sensor data for a period of time in the future. The discriminator evaluates the authenticity of the generated data to predict the evolution trend and potential impact of the fault. The specific method for calculating the future virtual sensor data is as follows:

[0116]

[0117] in, represents the generated future virtual sensor data, G represents the generator, and Y 1:t It represents the temperature sensor data in the past t minutes, and N represents the fault chain;

[0118] During the diagnostic report generation phase, the report outputs root cause probability rankings, dynamic fault propagation simulations, and maintenance priority scores. The scores comprehensively consider the probability, severity, and urgency of the fault, providing decision support for operations and maintenance personnel.

[0119] For example, when detecting drum water level fluctuations and abnormal furnace negative pressure, BKG infers the maximum probability failure chain (probability 68%) of "feedwater pump failure → water level control failure → unstable combustion." GAN further simulation shows that this failure may trigger a safety interlock after 6 minutes, and then pushes a suggestion to "switch to the backup pump." After confirmation by the operation and maintenance personnel, the conditional probability weights of the relevant nodes are automatically adjusted to achieve dynamic evolution of the knowledge base.

[0120] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A boiler status abnormality monitoring system based on artificial intelligence, characterized in that: include: Multimodal sensor data acquisition module: including vibration sensor, temperature sensor group, pressure sensor, voiceprint acquisition unit and infrared thermal imager, which respectively collect the mechanical vibration spectrum, three-dimensional temperature field distribution data, steam pressure waveform, voiceprint characteristics and thermal radiation image of the boiler body in real time; Edge computing node module: includes an adaptive filtering unit and an operating condition classifier, which are used to perform operating condition perception and noise reduction on raw sensor data and generate dynamic data enhancement strategies based on the boiler's current load rate and combustion mode; Transfer-enhanced anomaly detection module: This module integrates a pre-trained physical mechanism simulation model with an online updated twin neural network. By comparing the virtual sensor output of the digital twin with the difference between the measured data, it generates multi-dimensional deviation indicators. The fault tracing module constructs a Bayesian knowledge graph for the boiler, with nodes containing equipment, faults, and parameters. When an anomaly occurs, probability propagation is used to generate the maximum possible fault chain, combined with data verification and adversarial generation network simulation trajectory.

2. The artificial intelligence-based boiler status abnormality monitoring system according to claim 1, characterized in that: In the edge computing node module, the adaptive filtering unit includes a three-stage processing pipeline: a dynamic noise eliminator based on the LMS algorithm performs 50-800Hz bandpass filtering on the vibration signal; an improved Kalman filter is then used to smooth the temperature field data in the spatiotemporal domain; and finally, a wavelet threshold denoising module is used to process the voiceprint signal, retaining the key frequency band characteristics of 200-5000Hz. The operating condition classifier deploys a lightweight deep neural network. The input layer receives the preprocessed multimodal feature vector and uses a 3-layer convolution and 2-layer fully connected network to identify 8 typical operating conditions of the boiler in real time. The classification delay is controlled within 15ms.

3. The artificial intelligence-based boiler status abnormality monitoring system according to claim 2, characterized in that: The working condition classifier adopts the improved random forest algorithm: Construct an operating condition characteristic vector including combustion efficiency, load change gradient, and fuel calorific value compensation coefficient; The calculation method of combustion efficiency is as follows: Where η represents the combustion efficiency, Q 实际 Expressed as the actual heat released, Q 理论 Expressed as the theoretically released heat, T i It is the temperature of the ith partition measured by the infrared thermal imager, A i It is represented as the area weight corresponding to the i-th partition, α is the heat loss compensation coefficient, a is the fuel flow rate, and LHV is the lower heating value of the fuel, which represents the heat that can be released when the fuel per unit mass or volume is completely burned; The load change gradient uses a sliding time window to calculate the first-order differential entropy value of the pressure signal. The calculation method is as follows: Among them, H represents the first-order difference entropy value, p k It is expressed as the probability of the k-th pressure change value, and M is expressed as the total number of pressure change values ​​in the sliding time window; The fuel calorific value compensation coefficient is dynamically corrected by the voiceprint main frequency offset. The calculation method of the fuel calorific value compensation coefficient is as follows: Among them, β represents the fuel calorific value compensation coefficient, γ represents the correction coefficient, and f 实际 It is expressed as the actual measured value of the main frequency of the voiceprint, f 基准 Indicates the base value of the voiceprint's main frequency.

4. The artificial intelligence-based boiler status abnormality monitoring system according to claim 1, characterized in that: The migration-enhanced anomaly detection module adopts a dual-channel hybrid architecture to deeply integrate the physical mechanism model with the data-driven approach. The physical mechanism simulation model is built based on the thermodynamic principles of boilers. Through numerical simulation calculations, benchmark data containing key parameters such as temperature field and pressure distribution are generated to form a virtual sensor dataset for the digital twin. The model discretizes the three-dimensional heat transfer equation using the finite volume method: Where ρ represents density, c represents specific heat capacity, Indicates the rate of change of heat capacity per unit volume, Q comb Expressed as the heat released by fuel combustion, The partial derivative of temperature with time represents the rate of change of temperature with time, where R represents temperature and t represents time; Heat conduction term, q is the thermal conductivity, is the temperature gradient, Expressed as a divergence operator; The simulation model outputs spatiotemporal evolution data of key parameters such as temperature field and pressure distribution, forming a virtual sensor data set of the digital twin; the twin neural network working in parallel adopts an online incremental learning mechanism, and its network structure includes a feature extraction branch, which uses multi-scale residual units to process the measured sensor data.

5. The artificial intelligence-based boiler status abnormality monitoring system according to claim 4, characterized in that: In the spatial dimension, a propagation model based on a graph structure is constructed. Nodes represent monitoring points, and edge weights reflect the strength of physical associations. The calculation method for abnormal propagation values ​​is as follows: Among them, L d is represented by the abnormal propagation value of node d, K(d) is represented by the neighbor set of node d, and W h Expressed as the deviation of node h, w dh Expressed as the edge weight between nodes d and h, x d Expressed as the spatial coordinate of node d, x h It is represented as the spatial coordinate of node h, and l is represented as the attenuation value, which is used to control the attenuation speed of the abnormal propagation value; Among them, P(abn|V) is represented as the posterior probability, which is the probability that the boiler is in an abnormal state under the observed data V. P(V|abn) is represented as the likelihood function, which is the probability of observing data V in an abnormal state. P(abn) is represented as the prior probability, which is the initial probability that the boiler is in an abnormal state. P(V|w) is represented as the likelihood function, which is the probability of observing data V in state w. P(w) is represented as the initial probability that the boiler is in state w. The prior probability P(abn) is adaptively adjusted according to the runtime, and the likelihood function P(V|abn) is modeled using a mixed Gaussian distribution. When the posterior probability exceeds the dynamic threshold, an abnormal alarm is triggered and the fault tracing and analysis process is started.

6. The artificial intelligence-based boiler status abnormality monitoring system according to claim 1, characterized in that: In the fault tracing module, the Bayesian knowledge graph of boiler status is constructed using a triple structure, which includes three types of nodes: equipment components, fault modes, and monitoring parameters, and defines conditional probabilities through edge relationships; When the anomaly detection unit triggers an alarm, it initializes the observation evidence and performs probability propagation calculation along the edge of the graph. The Loopy Belief Propagation algorithm is used to calculate the posterior probability of the faulty node, and the probability of device failure is reversely inferred, ultimately generating the maximum possible fault chain.

7. The artificial intelligence-based boiler status abnormality monitoring system according to claim 6, characterized in that: Perform dynamic hypothesis testing based on real-time operating data. First, construct the expected parameter change pattern of the fault hypothesis and calculate its matching degree with the measured data through the dynamic time warping algorithm to screen out candidate faults that meet the current operating conditions. The adversarial generative network is used to simulate the fault development trajectory. The generator inputs the fault chain and current data and outputs virtual sensor data for a period of time in the future. The discriminator evaluates the authenticity of the generated data and predicts the evolution trend of the fault. The specific method for calculating virtual sensor data in the future is: in, represents the generated future virtual sensor data, G represents the generator, and Y 1:t It represents the temperature sensor data in the past t minutes, and N represents the fault chain.

Images