A combustor prognostics and health management system

Abstract

The application relates to the technical field of equipment state monitoring, and specifically discloses a combustor fault prediction and health management system which comprises nine core modules of multi-modal sensor fusion, sensor verification and recovery, intelligent feature extraction, digital twin simulation, health state evaluation, residual useful life prediction, risk perception decision, uncertainty quantification and management, and adaptive control; through a multi-source data fusion method combined with a physical model and data driving, comprehensive monitoring and accurate prediction of the state of the combustor are realized; risk perception decision is made through distributed reinforcement learning, maintenance strategies are verified through a digital twin platform, and closed-loop optimization is realized in combination with adaptive control. Through an adaptive weighting algorithm of the multi-modal sensor fusion module, the application improves the comprehensiveness of state perception; through a double-path network of the intelligent feature extraction module, multi-dimensional feature collaborative analysis can be realized.

Description

Technical Field

[0001] This invention relates to the field of equipment monitoring technology, specifically to a burner fault prediction and health management system. Background Technology

[0002] As key equipment in industries such as energy, petrochemicals, and metallurgy, the operating status of burners directly affects production safety, energy efficiency, and environmental performance. Traditional burner maintenance mainly adopts a periodic inspection and reactive repair model, which suffers from insufficient or excessive maintenance and fails to achieve accurate condition perception and predictive maintenance.

[0003] Existing burner monitoring systems lack effective multi-source heterogeneous data fusion mechanisms; sensor data with different sampling rates are difficult to synchronize in time, resulting in insufficient data feature extraction and incomplete condition monitoring; furthermore, existing burner monitoring systems lack effective sensor fault detection and data recovery mechanisms, leading to a significant decrease in system monitoring capabilities when sensors malfunction. Existing health assessment methods rely too heavily on empirical thresholds, resulting in insufficient accuracy in remaining life prediction. Summary of the Invention

[0004] Therefore, the present invention provides a burner fault prediction and health management system to solve the problems in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A burner failure prediction and health management system, comprising:

[0007] The multimodal sensor fusion module collects raw data from multiple types of sensors, processes it uniformly, and outputs the processed raw data to the intelligent feature extraction module and the sensor verification and recovery module.

[0008] The sensor verification and recovery module analyzes the processed raw data and outputs a quality report;

[0009] The intelligent feature extraction module extracts feature vectors from the processed raw data that can comprehensively reflect the health status of the burner.

[0010] The digital twin simulation module constructs a multi-physics coupled virtual model of the burner, providing a simulation environment for condition assessment, life prediction, and decision verification.

[0011] The health status assessment module receives the extracted feature vectors, and then quantifies the health status of the burner and its key components based on hierarchical Bayesian networks and fuzzy comprehensive evaluation methods, and outputs the health status assessment results.

[0012] The remaining useful life prediction module receives the extracted feature vectors and outputs the prediction results of the generated remaining useful life probability distribution; and adapts to the degradation characteristics of different components through multi-timescale modeling.

[0013] The risk perception and decision-making module receives health status assessment results and remaining life prediction results, and generates the optimal maintenance decision based on distributed reinforcement learning and multi-objective optimization algorithms, taking into account the uncertainty of prediction.

[0014] The uncertainty quantification and management module quantifies uncertainty in data, models, and forecasts throughout the entire process, providing a comprehensive uncertainty measurement for risk perception and decision-making.

[0015] The adaptive control module translates maintenance decisions into precise control actions; the execution results are fed back to the system through the multimodal sensor fusion module.

[0016] Furthermore: the fusion process of the multimodal sensor fusion module is as follows:

[0017] Suppose there are n sensors for the same physical quantity, and the measurement value of the i-th sensor at time k is... The true value is x(k); its calculation process is divided into timestamp synchronization, adaptive weighted fusion and Kalman filter noise reduction;

[0018] (1) Timestamp synchronization:

[0019] Different sensor sampling rates cause time point misalignment. Let the target time point be t, and the two most recent sampling points of sensor i be... and ;

[0020] The synchronization value at time t is obtained through linear interpolation:

[0021] ;

[0022] This process ensures that all sensor data is aligned on a unified timeline;

[0023] (2) Adaptive weighted fusion:

[0024] Weight Dynamic updates are based on historical sensor errors. The variance estimate of the i-th sensor within window W is defined as:

[0025] ;

[0026] in, It is the fusion estimate from the previous moment;

[0027] Weights are inversely proportional to variance:

[0028] ;

[0029] Calculation of fusion estimate:

[0030] ;

[0031] Among them, weight normalization ensures Online variance updates enable the system to adapt to sensor degradation.

[0032] (3) Kalman filtering noise reduction:

[0033] After fusion As the observation input, a discrete Kalman filter is applied;

[0034] Then its state equation is:

[0035] ;

[0036] The observation equation is:

[0037] ;

[0038] Perform filtering recursion; then the prediction step is:

[0039] ;

[0040] ;

[0041] The update steps are as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] Q and R are calibrated using historical data, and the output is filtered. The fused data after denoising;

[0046] The calculation formulas are summarized as follows:

[0047] ;

[0048] Weighted fusion:

[0049] ;

[0050] The Kalman filter calculation formula is:

[0051] ;

[0052] ;

[0053] ;

[0054] in, These are the sensor values ​​after time synchronization; , Indicates the adjacent time points of sensor i; This represents the variance estimate of sensor i within window W; For sensor i, the adaptive weights; These are the estimated values ​​of the physical quantities after fusion; This is the true state vector; Here, A is the observation vector; H is the state transition matrix; Q is the observation matrix; and R is the process noise variance. The state estimation error covariance; is the Kalman gain; I is the identity matrix.

[0055] Furthermore, the specific verification process for the sensor verification and recovery module is as follows:

[0056] Let the sensor reading sequence be ;

[0057] (1) Physical rationality check is to set the readings of each sensor within the physical rationality range and implement the first layer of verification;

[0058] Define the physical boundaries of the sensor [L] i U i ];L i U i These represent the lower and upper physical limits of sensor i, respectively;

[0059] The fault determination rule is: if or If so, it is marked as a potential fault;

[0060] (2) Signal consistency analysis is performed by cluster analysis to detect the consistency between redundant sensors and to implement the second layer of verification;

[0061] For redundant sensor groups, the pairwise Euclidean distance is calculated using the following formula:

[0062] ;

[0063] Then k-means clustering is applied to... Grouping: If a sensor is more than δ away from most other sensors, it is considered inconsistent; the threshold is set based on historical data.

[0064] Therefore, the sensor fault score is:

[0065] ;

[0066] in, For indicator functions; when If this occurs, a second-level warning will be triggered;

[0067] (3) Model consistency verification is to establish the neural network mapping relationship between sensor parameters, detect faults by the deviation between the model prediction value and the actual value, and implement the third layer of verification;

[0068] Training a neural network agent model: Input x is the relevant sensor data, and output y is the target sensor value;

[0069] Its loss function is:

[0070] ;

[0071] Then calculate the residuals:

[0072] ;

[0073] when Then the fault is confirmed. The statistical threshold is based on historical normal data. The principles are established;

[0074] in, Let be the reading of sensor i at time t; The distance between sensors i and j is represented by δ; δ is the consistency threshold. Score the sensor malfunction; For neural network proxy models; For the relevant sensor input vector; To predict residuals.

[0075] Furthermore: When a sensor is confirmed to be faulty, recovery data is generated using a surrogate model.

[0076] ;

[0077] The surrogate model minimizes on the training data To ensure that the recovered data conforms to physical laws;

[0078] in, For neural network proxy models; For the relevant sensor input vector; For the recovery of sensor data.

[0079] Furthermore: the intelligent feature extraction module calculates physical characteristics such as combustion efficiency and thermal stress distribution based on the first principles of combustion science;

[0080] The formula for calculating combustion efficiency η is as follows:

[0081] ;

[0082] in, LHV is the fuel mass flow rate, and LHV is the lower heating value. For heat loss;

[0083] thermal stress The calculation formula is:

[0084] ;

[0085] Where E is Young's modulus; The coefficient of thermal expansion; This refers to the temperature difference.

[0086] The intelligent feature extraction module dynamically selects the subset of features most relevant to the current state through an adaptive attention mechanism.

[0087] Furthermore, the health status assessment module constructs a hierarchical Bayesian network that can assess the health status of key subsystems separately; by integrating the assessment results of each subsystem through fuzzy comprehensive evaluation, an overall health score is obtained, thereby outputting the overall health status assessment result.

[0088] Furthermore, the specific implementation process of the health status assessment module is as follows:

[0089] Let f be the eigenvector of subsystem k (k=1,…,K). k ;

[0090] (1) In the constructed hierarchical Bayesian network, the latent variable of health status is defined. Where 0 = healthy, 1 = faulty;

[0091] Conditional probability tables are learned from historical data:

[0092] ;

[0093] The marginal likelihood function is:

[0094] ;

[0095] The posterior probability formula is:

[0096] ;

[0097] Gaussian mixture model (GMM) was used for fitting, and the parameters were estimated by maximum likelihood.

[0098] (2) Fuzzy comprehensive evaluation:

[0099] The subsystem health score is then: ;

[0100] Constructing the evaluation matrix: V represents the set of comments;

[0101] The membership function is calculated as follows:

[0102] ;

[0103] Weight vector w = [w1, ..., wk] T The overall score is:

[0104] ;

[0105] in, For the fuzzy synthesis operator, v = [v1, ..., vV] T Numericalize the comments into vectors;

[0106] Summarizing the above formulas, the Bayesian posterior probability is:

[0107] ;

[0108] The fuzzy comprehensive score is:

[0109] ;

[0110] in, For the hidden variable representing the health status of subsystem k; Let be the eigenvector of subsystem k; The conditional probability density of features; This is the prior probability; For health reasons; Fault prior; , Let represent the mean and covariance of the GMM under healthy conditions, respectively; The subsystem health score is [0,1]. Rate Membership degree of comment v; , , is the parameter of the triangular membership function; w is the subsystem weight vector; v is the numerical vector of the evaluation; HI is the overall health index.

[0111] Furthermore, the specific implementation process of the remaining useful life prediction module is as follows:

[0112] (1) A deep spatiotemporal sequence network with an encoder-decoder structure is used to capture the spatiotemporal dependencies of sensor data;

[0113] Let the degradation feature sequence be It's truly RUL for RUL * ;

[0114] The encoder calculation formula is as follows:

[0115] ;

[0116] in, This represents the encoder's final hidden state. For encoder LSTM network;

[0117] The decoder calculation formula is:

[0118] ;

[0119] ;

[0120] The loss function is:

[0121] ;

[0122] in, Output for damage accumulation model; For the decoder LSTM network; This is the physical regularization coefficient;

[0123] (2) Establish a life consumption model for key components based on damage accumulation theory and degradation physics analysis;

[0124] Using the Palmgren-Miner rule, the formula for physical model enhancement is:

[0125] ;

[0126] in, This represents the number of load cycles. D represents the corresponding fatigue life; D represents the damage ratio.

[0127] When D=1, the RUL mapping fails, therefore ; From feature x t Predict using regression models;

[0128] (3) Generate a probability distribution prediction of remaining lifetime using Monte Carlo dropout technique;

[0129] If dropout is enabled during testing, run the test M times; then

[0130] ;

[0131] in, These are the network parameters under the dropout mask; The m-th Monte Carlo test used RUL;

[0132] The RUL distribution is as follows:

[0133] ;

[0134] In summary, the formula for damage accumulation is:

[0135] ;

[0136] The formula for estimating RUL is:

[0137] ; Let the probability be the degradation. ;

[0138] The probability distribution is as follows:

[0139] ;

[0140] in, , ;

[0141] in, is the mean of the RUL distribution; Let be the standard deviation of the RUL distribution;

[0142] Furthermore: the risk perception decision-making module learns the reward distribution of decisions based on a quantitative regression deep Q-network; constructs a multi-objective optimization problem, simultaneously considering safety, economy, environmental protection, and reliability; simulates the long-term effects of decision-making schemes through digital twin simulation; and can provide decision explanations and clearly demonstrate the basis and expected effects of recommended decisions.

[0143] Furthermore: the quality report from the sensor verification and recovery module guides the uncertainty quantification and management module to adjust the confidence level assessment; the uncertainty quantification and management module provides risk measurement support for all analysis and decision-making modules.

[0144] The present invention has the following advantages: the present invention improves the comprehensiveness of state perception through the adaptive weighting algorithm of the multimodal sensor fusion module; the dual-path network of the intelligent feature extraction module enables collaborative analysis of multi-dimensional features; the probabilistic framework of the remaining useful life prediction module provides uncertainty quantification, significantly improving the reliability of early warning; and the three-layer architecture of the sensor verification and recovery module ensures the fault tolerance of the system.

[0145] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description

[0146] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0147] Figure 1 This is an architecture diagram of a burner fault prediction and health management system provided in an embodiment of this application. Detailed Implementation

[0148] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments to the present invention based on the above-described content.

[0149] Please see Figure 1 A burner fault prediction and health management system includes a multimodal sensor fusion module, a sensor verification and recovery module, an intelligent feature extraction module, a digital twin simulation module, a health status assessment module, a remaining useful life prediction module, a risk perception decision-making module, an uncertainty quantification and management module, and an adaptive control module; the modules are connected through standardized data interfaces and communication protocols.

[0150] The multimodal sensor fusion module collects raw data from various sensors, including temperature, pressure, vibration, and emission components. It performs unified processing on the collected multi-source heterogeneous raw data and outputs the processed raw data to the intelligent feature extraction module and the sensor verification and recovery module.

[0151] Among them, for sensors such as temperature, pressure, vibration, and emission components deployed in key parts of the burner, the multi-modal sensor fusion module collects raw data through an industrial bus; an adaptive weighted fusion algorithm is used to dynamically weight and integrate the readings of multiple sensors for the same physical quantity (such as temperature), with the weights adaptively adjusted based on the real-time reliability of the sensors; a timestamp synchronization mechanism is used to solve the problem of inconsistent timing of data at different sampling rates, and linear interpolation is used to achieve time alignment to ensure the time consistency of data at different sampling rates; Kalman filtering technology is used to eliminate random noise.

[0152] In this embodiment, the multimodal sensor fusion process is as follows:

[0153] Suppose there are n sensors for the same physical quantity, and the measurement value of the i-th sensor at time k is... The true value is x(k); its calculation process is divided into timestamp synchronization, adaptive weighted fusion and Kalman filter noise reduction.

[0154] (1) Timestamp synchronization:

[0155] Different sensor sampling rates cause time point misalignment. Let the target time point be t, and the two most recent sampling points of sensor i be... and .

[0156] The synchronization value at time t is obtained through linear interpolation:

[0157] ;

[0158] This process ensures that all sensor data is aligned on a unified timeline;

[0159] (2) Adaptive weighted fusion:

[0160] Weight Dynamic updates are based on historical sensor errors. The variance estimate of the i-th sensor within window W is defined as:

[0161] ;

[0162] in, It is the fusion estimate from the previous moment;

[0163] Weights are inversely proportional to variance (the smaller the variance, the higher the reliability):

[0164] ;

[0165] Calculation of fusion estimate:

[0166] ;

[0167] Among them, weight normalization ensures Online variance updates enable the system to adapt to sensor degradation.

[0168] (3) Kalman filtering noise reduction:

[0169] After fusion As the observation input, a discrete Kalman filter is applied;

[0170] Then its state equation is:

[0171] ;

[0172] The observation equation is:

[0173] ;

[0174] Perform filtering recursion; then the prediction step is:

[0175] ;

[0176] ;

[0177] The update steps are as follows:

[0178] ;

[0179] ;

[0180] ;

[0181] Q and R are calibrated using historical data, and the output is filtered. This is the fused data after denoising.

[0182] The calculation formulas are summarized as follows:

[0183] ;

[0184] Weighted fusion:

[0185] ;

[0186] The Kalman filter calculation formula is:

[0187] ;

[0188] ;

[0189] ;

[0190] in, These are the sensor values ​​after time synchronization; , Indicates the adjacent time points of sensor i; This represents the variance estimate of sensor i within window W; For sensor i, the adaptive weights; These are the estimated values ​​of the physical quantities after fusion; This is the true state vector; Here, A is the observation vector; H is the state transition matrix; Q is the observation matrix; and R is the process noise variance. The state estimation error covariance; is the Kalman gain; I is the identity matrix.

[0191] The sensor verification and recovery module adopts a three-layer architecture of physical rationality check, signal consistency analysis and model consistency verification. It analyzes the processed raw data and outputs a quality report. For sensors confirmed to be faulty, it starts a neural network proxy model to generate recovery data to ensure the continuous and reliable operation of the system when sensors are abnormal.

[0192] In this embodiment, the sensor reading sequence is assumed to be... ;

[0193] The physical rationality check involves setting the readings of each sensor within a physically reasonable range to perform the first layer of verification.

[0194] Define the physical boundaries of the sensor [L] i U i For example, the upper limit U of the temperature sensor. i =1500℃; L i U i These represent the lower and upper physical limits of sensor i, respectively;

[0195] The fault determination rule is: if or If it is, then it is marked as a potential fault.

[0196] Signal consistency analysis, on the other hand, uses cluster analysis to detect the consistency between redundant sensors, thus implementing a second layer of verification.

[0197] For a redundant sensor group (e.g., M sensors of the same type), the pairwise Euclidean distance is calculated using the following formula:

[0198] ;

[0199] Then k-means clustering is applied to... Grouping: if a sensor is more than δ away from most other sensors, it is considered inconsistent; the threshold is set based on historical data.

[0200] Therefore, the sensor fault score is:

[0201] ;

[0202] in, For indicator functions; when If this occurs, a second-level warning will be triggered.

[0203] Model consistency verification involves establishing a neural network mapping relationship between sensor parameters, detecting faults by the deviation between model predictions and actual values, and implementing a third layer of verification.

[0204] Training a neural network agent model: Input x is the relevant sensor data, and output y is the target sensor value;

[0205] Its loss function is:

[0206] ;

[0207] Then calculate the residuals:

[0208] ;

[0209] when Then the fault is confirmed. The statistical threshold is based on historical normal data. The principles have been established.

[0210] In the event of a failure, use a proxy model to generate recovery data:

[0211] ;

[0212] The surrogate model minimizes on the training data To ensure that the recovered data conforms to physical laws;

[0213] in, Let be the reading of sensor i at time t; The distance between sensors i and j is represented by δ; δ is the consistency threshold. Score the sensor malfunction; For neural network proxy models; For the relevant sensor input vector; To predict residuals; For the recovery of sensor data.

[0214] The intelligent feature extraction module processes time-domain and frequency-domain features separately through a dual-path deep network. Combined with physical model-guided features and an adaptive attention mechanism, it extracts feature vectors that comprehensively reflect the health status of the burner from the raw data processed by the multimodal sensor fusion module.

[0215] The intelligent feature extraction module constructs a dual-path deep network, in which the time-domain path extracts time series features through a long short-term memory network, and the frequency-domain path extracts spectral features through wavelet transform and fast Fourier transform.

[0216] In addition, the intelligent feature extraction module calculates physical characteristics such as combustion efficiency and thermal stress distribution based on the first principles of combustion science;

[0217] The formula for calculating combustion efficiency η is as follows:

[0218] ;

[0219] in, LHV is the fuel mass flow rate, and LHV is the lower heating value. For heat loss;

[0220] thermal stress The calculation formula is:

[0221] ;

[0222] Where E is Young's modulus; The coefficient of thermal expansion; This refers to the temperature difference.

[0223] The intelligent feature extraction module dynamically selects the subset of features most relevant to the current state through an adaptive attention mechanism.

[0224] The digital twin simulation module constructs a multi-physics coupled virtual model of the burner. Through real-time data assimilation and model order reduction techniques, it achieves accurate mapping between the physical entity and the virtual model, providing a simulation environment for condition assessment, life prediction, and decision verification.

[0225] The health status assessment module receives the extracted feature vectors, and then quantifies the health status of the burner and its key components based on hierarchical Bayesian networks and fuzzy comprehensive evaluation methods, and outputs the health status assessment results.

[0226] By employing transfer learning and online adaptive mechanisms, the accuracy and adaptability of the evaluation model can be improved.

[0227] The constructed hierarchical Bayesian network can assess the health status of key subsystems such as the combustion chamber, ignition system, and fuel supply system. By integrating the assessment results of each subsystem through fuzzy comprehensive evaluation, an overall health score is obtained, thereby outputting the overall health status assessment result.

[0228] Let f be the eigenvector of subsystem k (k=1,…,K). k ;

[0229] (1) In the constructed hierarchical Bayesian network, the latent variable of health status is defined. Where 0 = healthy, 1 = faulty;

[0230] Conditional probability tables (CPTs) learn from historical data:

[0231] ;

[0232] The marginal likelihood function is:

[0233] ;

[0234] The formula for the posterior probability (health confidence) is:

[0235] ;

[0236] A Gaussian mixture model (GMM) was used for fitting, and the parameters were estimated by maximum likelihood.

[0237] (2) Fuzzy comprehensive evaluation (system level):

[0238] The subsystem health score is then: ;

[0239] Constructing the evaluation matrix: V represents a set of evaluation criteria, such as excellent, good, average, and poor.

[0240] The membership function (triangle) is calculated using the following formula:

[0241] ;

[0242] Weight vector w = [w1, ..., wk] T The overall score is:

[0243] ;

[0244] in, For fuzzy composition operators (such as weighted averages), v = [v1, ..., vV] T Numericalize the comments into vectors;

[0245] Transfer learning fine-tunes GMM parameters using source domain data; online adaptive updates. To adapt to new data.

[0246] Summarizing the above formulas, the Bayesian posterior probability is:

[0247] ;

[0248] The fuzzy comprehensive score is:

[0249] ;

[0250] in, For the hidden variable representing the health status of subsystem k; Let be the eigenvector of subsystem k; The conditional probability density of features; This is the prior probability; For health reasons; Fault prior; , Let represent the mean and covariance of the GMM under healthy conditions, respectively; The subsystem health score is [0,1]. Rate Membership degree of comment v; , , is the parameter of the triangular membership function; w is the subsystem weight vector; v is the numerical vector of the evaluation; HI is the overall health index.

[0251] The remaining useful life prediction module receives the extracted feature vectors, uses a deep spatiotemporal sequence network combined with a physical model enhancement method, and outputs the remaining useful life prediction results with the generated remaining useful life probability distribution. It adapts to the degradation characteristics of different components through multi-timescale modeling.

[0252] Specifically, the remaining useful life prediction module uses a deep spatiotemporal sequence network with an encoder-decoder structure to capture the spatiotemporal dependencies of sensor data;

[0253] Let the degradation feature sequence be It's truly RUL for RUL * ;

[0254] The formula for calculating the encoder (LSTM) is:

[0255] ;

[0256] in, This represents the encoder's final hidden state. For encoder LSTM network;

[0257] The calculation formula for the decoder (LSTM + physical constraints) is as follows:

[0258] ;

[0259] ;

[0260] The loss function is:

[0261] ;

[0262] in, Output for damage accumulation model; For the decoder LSTM network; This is the physical regularization coefficient.

[0263] The remaining useful life prediction module establishes a life consumption model for key components based on damage accumulation theory and degradation physics analysis.

[0264] Using the Palmgren-Miner rule, the formula for physical model enhancement (damage accumulation) is:

[0265] ;

[0266] in, This represents the number of load cycles. D represents the corresponding fatigue life; D represents the damage ratio.

[0267] When D=1, the RUL mapping fails, therefore ; From feature x t Predict using regression models.

[0268] Probability distribution predictions of remaining lifetime are generated using Monte Carlo dropout techniques;

[0269] If dropout is enabled during testing, run the test M times; then

[0270] ;

[0271] in, These are the network parameters under the dropout mask; The mth Monte Carlo test used RUL;

[0272] The RUL distribution is as follows:

[0273] ;

[0274] Here, dropout simulates the Bayesian approximation, which converges to the true posterior when M is sufficiently large.

[0275] In summary, after combining the calculation formulas, the damage accumulation formula is as follows:

[0276] ;

[0277] The formula for estimating RUL is:

[0278] ; Let the probability be the degradation. ;

[0279] The probability distribution is as follows:

[0280] ;

[0281] in, , ;

[0282] in, is the mean of the RUL distribution; Let be the standard deviation of the RUL distribution.

[0283] The risk perception and decision-making module receives health status assessment results and remaining life prediction results. Based on distributed reinforcement learning and multi-objective optimization algorithms, it generates the optimal operation and maintenance strategy while considering the uncertainty of prediction. Furthermore, it improves the reliability and acceptability of decisions through a decision interpretation mechanism and a human-machine collaboration framework.

[0284] Specifically: it learns the reward distribution of decisions based on quantized regression deep Q-networks; it constructs a multi-objective optimization problem that considers safety, economy, environmental protection, and reliability; it simulates the long-term effects of decision-making schemes through digital twin simulation; in addition, it provides decision explanations and clearly demonstrates the basis and expected effects of recommended decisions.

[0285] The uncertainty quantification and management module uses Bayesian deep learning and Monte Carlo sampling techniques to quantify uncertainty in data, models, and predictions throughout the entire process, providing a comprehensive uncertainty measurement for risk perception and decision-making.

[0286] The uncertainty quantification and management module can perform uncertainty source analysis, identify major sources of uncertainty, and adaptively adjust the conservatism of system behavior based on the level of uncertainty.

[0287] In addition, the quality report from the sensor verification and recovery module guides the uncertainty quantification and management module in adjusting its confidence level assessment; the uncertainty quantification and management module provides risk measurement support for all analysis and decision-making modules.

[0288] The adaptive control module, based on model predictive control and online learning algorithms, transforms intelligent decisions into precise control actions. Through multi-objective coordinated control and fault-tolerant strategies, it achieves real-time optimization of the combustion process and emergency handling of faults.

[0289] The adaptive control module uses recursive least squares to update the burner dynamic model online; it implements multi-objective coordinated control to balance multiple objectives such as efficiency, emissions, and lifespan; it designs a fault-tolerant control strategy to maintain basic operation when sensors or actuators fail; and it feeds back the execution efficiency to the system through newly acquired data from various sensors.

[0290] This invention enhances the comprehensiveness of state perception through the adaptive weighting algorithm of the multimodal sensor fusion module; sets up a dual-path network for the intelligent feature extraction module, enabling collaborative analysis of multi-dimensional features; provides uncertainty quantification through the probabilistic framework of the remaining useful life prediction module, significantly improving the reliability of early warning; and ensures system fault tolerance through the three-layer architecture of the sensor verification and recovery module.

[0291] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A burner fault prediction and health management system, characterized in that, include: The multimodal sensor fusion module collects raw data from multiple types of sensors, processes it uniformly, and outputs the processed raw data to the intelligent feature extraction module and the sensor verification and recovery module. The sensor verification and recovery module analyzes the processed raw data and outputs a quality report; The intelligent feature extraction module extracts feature vectors from the processed raw data that can comprehensively reflect the health status of the burner. The digital twin simulation module constructs a multi-physics coupled virtual model of the burner, providing a simulation environment for condition assessment, life prediction, and decision verification. The health status assessment module receives the extracted feature vectors, and then quantifies the health status of the burner and its key components based on hierarchical Bayesian networks and fuzzy comprehensive evaluation methods, and outputs the health status assessment results. The remaining useful life prediction module receives the extracted feature vectors and outputs the prediction results of the generated remaining useful life probability distribution; and adapts to the degradation characteristics of different components through multi-timescale modeling. The risk perception and decision-making module receives health status assessment results and remaining life prediction results, and generates the optimal maintenance decision based on distributed reinforcement learning and multi-objective optimization algorithms, taking into account the uncertainty of prediction. The uncertainty quantification and management module quantifies uncertainty in data, models, and forecasts throughout the entire process, providing a comprehensive uncertainty measurement for risk perception and decision-making. The adaptive control module translates maintenance decisions into precise control actions; the execution results are fed back to the system through the multimodal sensor fusion module. The intelligent feature extraction module calculates physical characteristics such as combustion efficiency and thermal stress distribution based on the first principles of combustion science. The formula for calculating combustion efficiency η is as follows: ； in, LHV is the fuel mass flow rate, and LHV is the lower heating value. For heat loss; thermal stress The calculation formula is: ； Where E is Young's modulus; The coefficient of thermal expansion; For temperature difference; The intelligent feature extraction module dynamically selects the subset of features most relevant to the current state through an adaptive attention mechanism.

2. The burner fault prediction and health management system according to claim 1, characterized in that, The fusion process of the multimodal sensor fusion module is as follows: Suppose there are n sensors for the same physical quantity, and the measurement value of the i-th sensor at time k is... The true value is x(k); its calculation process is divided into timestamp synchronization, adaptive weighted fusion and Kalman filter noise reduction; (1) Timestamp synchronization: Different sensor sampling rates cause time point misalignment. Let the target time point be t, and the two most recent sampling points of sensor i be... and ; The synchronization value at time t is obtained through linear interpolation: ； This process ensures that all sensor data is aligned on a unified timeline; (2) Adaptive weighted fusion: Weight Based on dynamic updates of sensor historical errors; the variance estimate of the i-th sensor within window W is defined as: ； in, It is the fusion estimate from the previous moment; Weights are inversely proportional to variance: ； Calculation of fusion estimate: ； Among them, weight normalization ensures Online variance updates enable the system to adapt to sensor degradation. (3) Kalman filtering noise reduction: After fusion As the observation input, a discrete Kalman filter is applied; Then its state equation is: ； The observation equation is: ； Perform filtering recursion; then the prediction step is: ； ； The update steps are as follows: ； ； ； Q and R are calibrated using historical data, and the output is filtered. The fused data after denoising; The calculation formulas are summarized as follows: ； Weighted fusion: ； The Kalman filter calculation formula is: ； ； ； in, These are the sensor values ​​after time synchronization; , Indicates the adjacent time points of sensor i; This represents the variance estimate of sensor i within window W; For sensor i, the adaptive weights; These are the estimated values ​​of the physical quantities after fusion; This is the true state vector; Here, A is the observation vector; H is the state transition matrix; Q is the observation matrix; and R is the process noise variance. The state estimation error covariance; is the Kalman gain; I is the identity matrix.

3. The burner fault prediction and health management system according to claim 1, characterized in that, The specific verification process for the sensor verification and recovery module is as follows: Let the sensor reading sequence be ; (1) Physical rationality check is to set the readings of each sensor within the physical rationality range and implement the first layer of verification; Define the physical boundaries of the sensor [L] i U i ];L i U i These represent the lower and upper physical limits of sensor i, respectively; The fault determination rule is: if or If so, it is marked as a potential fault; (2) Signal consistency analysis is performed by cluster analysis to detect the consistency between redundant sensors and to implement the second layer of verification; For redundant sensor groups, the pairwise Euclidean distance is calculated using the following formula: ； Then k-means clustering is applied to... Grouping: If a sensor is more than δ away from most other sensors, it is considered inconsistent; the threshold is set based on historical data. Therefore, the sensor fault score is: ； in, For indicator functions; when If this occurs, a second-level warning will be triggered; (3) Model consistency verification is to establish the neural network mapping relationship between sensor parameters, detect faults by the deviation between the model prediction value and the actual value, and implement the third layer of verification; Training a neural network agent model: Input x is the relevant sensor data, and output y is the target sensor value; Its loss function is: ； Then calculate the residuals: ； when Then the fault is confirmed. The statistical threshold is based on historical normal data. The principles are established; in, Let be the reading of sensor i at time t; The distance between sensors i and j is represented by δ; δ is the consistency threshold. Score the sensor malfunction; For neural network proxy models; For the relevant sensor input vector; To predict residuals.

4. The burner fault prediction and health management system according to claim 3, characterized in that, When a sensor is identified as faulty, recovery data is generated using a surrogate model: ； The surrogate model minimizes on the training data To ensure that the recovered data conforms to physical laws; in, For neural network proxy models; For the relevant sensor input vector; For the recovery of sensor data.

5. The burner fault prediction and health management system according to claim 1, characterized in that, The health status assessment module constructs a hierarchical Bayesian network that can assess the health status of key subsystems separately; by integrating the assessment results of each subsystem through fuzzy comprehensive evaluation, an overall health score is obtained, thereby outputting the overall health status assessment result.

6. The burner fault prediction and health management system according to claim 5, characterized in that, The specific implementation process of the health status assessment module is as follows: Let f be the eigenvector of subsystem k (k=1,…,K). k ; (1) In the constructed hierarchical Bayesian network, the latent variable of health status is defined. Where 0 = healthy, 1 = faulty; Conditional probability tables are learned from historical data: ； The marginal likelihood function is: ； The posterior probability formula is: ； Gaussian mixture model (GMM) was used for fitting, and the parameters were estimated by maximum likelihood. (2) Fuzzy comprehensive evaluation: The subsystem health score is then: ; Constructing the evaluation matrix: V represents the set of comments; The membership function is calculated as follows: ； Weight vector w = [w1, ..., wk] T The overall score is: ； in, For the fuzzy synthesis operator, v = [v1, ..., vV] T Numericalize the comments into vectors; Summarizing the above formulas, the Bayesian posterior probability is: ； The fuzzy comprehensive score is: ; in, For the hidden variable representing the health status of subsystem k; Let be the eigenvector of subsystem k; The conditional probability density of features; This is the prior probability; Health comes first; Fault prior; , Let represent the mean and covariance of the GMM under healthy conditions, respectively; The subsystem health score is [0,1]. Rate Membership degree of comment v; , , is the parameter of the triangular membership function; w is the subsystem weight vector; v is the numerical vector of the evaluation; HI is the overall health index.

7. The burner fault prediction and health management system according to claim 1, characterized in that, The specific implementation process of the remaining useful life prediction module is as follows: (1) A deep spatiotemporal sequence network with an encoder-decoder structure is used to capture the spatiotemporal dependencies of sensor data; Let the degradation feature sequence be It's truly RUL for RUL * ; The encoder calculation formula is as follows: ； in, This is the final hidden state of the encoder; For encoder LSTM network; The decoder calculation formula is: ； ； The loss function is: ； in, Output for damage accumulation model; For the decoder LSTM network; This is the physical regularization coefficient; (2) Establish a life consumption model for key components based on damage accumulation theory and degradation physics analysis; Using the Palmgren-Miner rule, the formula for physical model enhancement is: ； in, This represents the number of load cycles. D represents the corresponding fatigue life; D represents the damage ratio. When D=1, the RUL mapping fails, therefore ; From feature x t Predict using regression models; (3) Generate a probability distribution prediction of remaining lifetime using Monte Carlo dropout technique; If dropout is enabled during testing, run the test M times; then ； in, These are the network parameters under the dropout mask; The m-th Monte Carlo test used RUL; The RUL distribution is as follows: ； In summary, the formula for damage accumulation is: ； The formula for estimating RUL is: ; The probability of degradation. ; The probability distribution is as follows: ； in, , ; in, is the mean of the RUL distribution; Let be the standard deviation of the RUL distribution.

8. The burner fault prediction and health management system according to claim 1, characterized in that, The risk perception and decision-making module learns the reward distribution of decisions based on a quantitative regression deep Q-network; it constructs a multi-objective optimization problem, simultaneously considering safety, economy, environmental protection, and reliability; it simulates the long-term effects of decision-making schemes through digital twin simulation; it can provide decision explanations and clearly demonstrate the basis and expected effects of recommended decisions.

9. The burner fault prediction and health management system according to claim 1, characterized in that, The quality report from the sensor verification and recovery module guides the uncertainty quantification and management module in adjusting the confidence level assessment; the uncertainty quantification and management module provides risk measurement support for all analysis and decision-making modules.

Images