Data-driven gasification combustion fault diagnosis method and gasification combustion machine

Abstract

This invention belongs to the field of intelligent manufacturing equipment industry and industrial automation control technology, and discloses a data-driven gasification combustion fault diagnosis method and a gasification burner. This invention compensates for sensor drift using an adaptive extended Kalman filter, aligns dynamic transmission delays based on the fluid dynamics continuity equation, extracts spatiotemporal features using a graph attention network with injected thermodynamic priors and a gated recurrent unit, outputs the fault probability through a focus loss function, and finally introduces a Fisher information matrix to perform incremental fine-tuning of the neural network. This invention solves the false alarm problem caused by time-varying delays and sensor degradation under variable load conditions, improves diagnostic robustness and adaptability, and can be widely applied to gasification combustion systems in the intelligent manufacturing equipment industry.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing equipment industry and industrial automation control technology, specifically involving a data-driven gasification combustion fault diagnosis method and a gasification burner. Background Technology

[0002] In actual industrial operation, gasification combustors involve complex internal physicochemical reactions: solid fuels undergo multiple multiphase flow (solid-gas-heat) coupling stages, including heating and drying, volatile matter release and pyrolysis, reductive gasification of semi-coke (a strongly endothermic reaction), and secondary complete combustion of the combustible mixture (a strongly exothermic reaction). These systems exhibit highly nonlinear, nonstationary, and strongly spatiotemporally coupled characteristics among multiple variables.

[0003] With the development of Industrial Internet of Things (IIoT) and Artificial Intelligence (AI) technologies, fault diagnosis of gasification combustors has evolved from manual experience to purely data-driven algorithms. However, existing fault diagnosis technologies still have the following shortcomings.

[0004] Gasification combustors are typical spatially distributed parameter systems. Changes in the physical state at the front end of the gasification chamber are transmitted to the rear end of the combustion chamber via the flow of the syngas medium. Existing diagnostic models based on time-series analysis (including Long Short-Term Memory networks, bidirectional gated loop units, or time-delay mutual information alignment methods using fixed time windows) implicitly assume that the physical causal transmission time between various physical field variables is a fixed constant. However, in actual industrial variable load conditions, the total mass flow rate and volumetric flow rate of the fluid medium dynamically change with primary air volume adjustments, feed rate variations, and fluctuations in pyrolysis gas yield. According to the fluid mechanics continuity equation, changes in flow velocity inevitably lead to changes in the residence time of the gas within a fixed geometric volume; that is, the physical transmission delay between variables is time-varying. Forcing the alignment of multidimensional sensor data using a fixed time window can cause phase misalignment of cross-spatial features under dynamic conditions, leading the neural network to extract contradictory false features during variable load phases, thus increasing the false alarm rate.

[0005] In the extreme high-temperature environment (800°C to 1200°C), alternating strong reducing or oxidizing conditions, and high dust concentrations inside a gasification burner, contact sensors (especially thermocouples) experience material degradation. The thermocouple wires undergo selective oxidation or grain growth in the high-temperature reducing atmosphere, leading to irreversible physical drift in their Seebeck coefficient. Simultaneously, ash accumulation and slagging of the protective sheath increase the thermal response time constant and amplify measurement noise. Most existing technologies treat sensors as ideal observers, employing only moving averages, static low-pass filtering, or fixed thresholds to remove outliers. Even when Kalman filtering is introduced for state estimation, the system process noise covariance matrix and measurement noise covariance matrix are typically set as empirical constants. This static assumption ignores the physical degradation process of the sensor over its service life. When a sensor enters the later stages of its lifespan, its actual measurement variance far exceeds the initial calibration value. Static filters still assign excessively high confidence weights, causing the final state estimation results to diverge, potentially misjudging the sensor's slow degradation as a gradual system failure (such as coking).

[0006] Traditional graph attention networks rely entirely on data-driven backpropagation algorithms for attention weight allocation. In gasification combustion scenarios, heat transfer (radiation, convection, conduction) and mass transfer (fluid flow) have strict physical directions and limitations. For example, the impact of pressure fluctuations at the bottom of the gasification chamber on the oxygen content at the tail of the combustion chamber is inevitably limited by the aerodynamic drag and volumetric buffering effect of the pipes. Existing graph attention networks lack physical laws (such as thermal resistance network topology and fluid network admittance matrices) as prior edge constraints in the early training stages. This causes the model to easily get trapped in local optima when facing small-sample faults or strong background noise, assigning high false attention weights to two physically uncoupled, distant, noisy nodes. This reduces the interpretability of the diagnostic model. Summary of the Invention

[0007] The purpose of this invention is to provide a data-driven gasification combustion fault diagnosis method and a gasification burner, which solves the problems of high false alarm rate, poor interpretability and inability to adaptively evolve in fault diagnosis caused by multiphase flow dynamic time delay, sensor physical degradation, lack of physical priors in pure data-driven operation and baseline drift in gasification burners under variable load conditions.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A data-driven gasification combustion fault diagnosis method includes the following steps: Step 1: Synchronously acquire multi-modal sensor data from the gasification burner, perform anti-aliasing low-pass filtering and downsampling processing on the high-frequency pressure signal, and align it with the low-frequency sensor data; Step 2: Construct an adaptive extended Kalman filter, setting the measurement noise covariance as a joint parameter of the sensor service time and the local variance of the signal, and using the theoretical temperature obtained by solving the thermodynamic balance equation as a virtual observation value to compensate for the drift bias of the temperature sensor. Step 3: Complete the missing fuel flow data based on the chemical stoichiometry equation, and generate virtual fault samples by solving the transient energy balance differential equation containing dynamic gasification efficiency and deriving the temperature decay trajectory under different air-fuel ratios. Step 4: Calculate the dynamic transmission delay between sensor nodes based on the fluid dynamics continuity equation, and perform phase shifting and alignment on the multi-source sensor data according to the dynamic transmission delay; Step 5: Construct a dual-branch feature extraction network, use a graph attention network with injected thermodynamic impedance prior matrix to extract spatial coupling features, and use an attention network with sparse addition mask and gated recurrent units to extract temporal evolution features. Step 6: Fuse the spatial coupling features and the temporal evolution features, and output the failure probability distribution through a focus loss function that includes a class balance factor; Step 7: When the model update is triggered, the Fisher information matrix is ​​introduced into the target loss function as a penalty term, and the biased samples are used to perform incremental fine-tuning of the model.

[0009] Furthermore, in step 1, the specific method for performing anti-aliasing low-pass filtering and downsampling processing on the high-frequency pressure signal is as follows: first, the parasitic high-frequency electromagnetic aliasing noise is eliminated by using a Butterworth low-pass filter with a cutoff frequency matched to the Nyquist sampling theorem, and then the downsampled pressure data sequence is obtained by sampling with a set time slice step size.

[0010] Furthermore, in step 2, the formula for calculating the measurement noise covariance of the adaptive extended Kalman filter is:

[0011] In the formula, For the first The dynamic measurement noise covariance at any given time is dimensionless. The noise variance is measured as a baseline and is dimensionless. The material aging coefficient is expressed in units of 1000 ppm. ; The cumulative service time of the sensor, in units of ; The local fluctuation sensitivity coefficient is dimensionless. For the past The local variance of the measured signal within a time window, dimensionless; For discrete-time indexing; The time window length is expressed in units of the number of sampling points. This is the measured signal.

[0012] Furthermore, in step 3, the calculation formula for completing the missing fuel flow data based on the stoichiometric ratio equation is as follows:

[0013] In the formula, The complete fuel mass flow rate, in units of ; For actual measured air mass flow rate, the unit is... ; This is a stoichiometric ratio, dimensionless. The theoretical air-fuel ratio is dimensionless.

[0014] Furthermore, in step 3, the transient energy balance differential equation is:

[0015] In the formula, This is the density of syngas, in units of... ; The effective volume of the gasification chamber is expressed in units of... ; The specific heat capacity of syngas at constant pressure, in units of... ; Temperature of the vaporization chamber, in units of or ; For time variables, the unit is ; Thermal power input to fuel, in units of ; The dynamic gasification efficiency is dimensionless. Syngas mass flow rate, in units of ; This refers to the feed temperature, in units of... or ; For heat loss, the unit is heat dissipation. When the chemical equivalence ratio At that time, the dynamic gasification efficiency Take a constant value of 0.75; when hour, .

[0016] Furthermore, in step 4, the formula for calculating the dynamic transmission delay between sensor nodes is:

[0017] In the formula, For the first The flow field changes from node to node Passed to node Dynamic latency, in units of ; For the node With the node The effective physical volume between them, in units of ; For real-time temperature The density of syngas at the specified value is given in units of 1. ; For the first The total mass flow rate of the gas passing through the system at all times, in units of ; and For sensor node indexing; For time variables, the unit is ; For the first Temperature at discrete time points, in units of or .

[0018] Furthermore, in step 5, the formula for calculating the bias value of the thermodynamic impedance prior matrix in the graph attention network is:

[0019] In the formula, For nodes With nodes The physical prior bias between them is dimensionless; The hyperparameter is a physical prior adjustment parameter, which is dimensionless and is set to 5.0 in this method; For the node With the node The normalized theoretical combined thermal resistance is dimensionless; the theoretical combined thermal resistance is obtained by adding the spatial thermal conduction resistance and the convective heat transfer resistance, and the normalization adopts minimum-maximum normalization to the interval. .

[0020] Furthermore, in step 5, the calculation formula for the attention network with a sparse addition mask is:

[0021] In the formula, This is the self-attention output matrix; For query matrix; The key matrix; It is a value matrix; Let be the transpose of the key matrix; The dimension of the key vector; It is a normalized exponential function; It is a sparse mask matrix; the sparse mask matrix The rule for setting elements in the middle is: when the position is... With position satisfy (in It is a non-negative integer, that is When ), mask element Values When this condition is not met, the mask element... Values .

[0022] Furthermore, in step 6, the formula for calculating the focus loss function is:

[0023] In the formula, The focal loss value; For the model to class The predicted probability value; For real category labels; As a category balance factor; For focusing parameters, this method uses 2.0; the category balance factor value for the normal category is set to be less than the category balance factor value for the fault category.

[0024] Furthermore, in step 6, before outputting the fault probability distribution, the feature vector is processed using a conditional batch normalization layer, as shown in the formula:

[0025] In the formula, The output feature tensor after conditional batch normalization; The input feature tensor; The average value of the characteristic batches; The standard deviation of the characteristic batch; To prevent positive numbers with a denominator of zero, take... ; This is a heat load command signal, a dimensionless percentage. The dynamic scaling factor is modulated by the heat load command signal; It is the dynamic translation factor modulated by the heat load command signal.

[0026] Furthermore, in step 6, a domain discriminator is also introduced, and a gradient inversion layer is inserted between the domain discriminator and the feature extraction network. During forward propagation, the gradient returned by the domain discriminator is multiplied by a negative constant during backpropagation to filter out the load information contained in the feature tensor.

[0027] Furthermore, after outputting the fault probability distribution in step 6, the integral gradient method is used to calculate the integral gradient contribution of the sensor channel to the output probability, as shown in the formula:

[0028] In the formula, For the first The integral gradient contribution of each sensor channel; Input tensors for current practical multidimensional sensors; The first tensor input to the actual multidimensional sensor One element; The baseline input tensor for fault-free operation is constructed by taking the average of sensor data from 1024 consecutive time steps within the last 24 hours when the model judges the system to be in normal operating condition and the output confidence level is greater than 0.95. The first input tensor of the baseline One element; To map to a specific fault category The prediction probability function; Index for fault categories; For integration variables; For network output to the first Partial derivatives of each sensor channel; The sensor channel index is used; the calculated integral gradient contribution is normalized and mapped to a two-dimensional spatiotemporal heat map.

[0029] Furthermore, in step 7, the formula for calculating the loss function of the Fisher information matrix is ​​as follows:

[0030] In the formula, This is the total loss function for the incremental fine-tuning phase; The focus loss term is calculated using the newly acquired bias samples; The current model is being optimized for the [number]th [period]. The weight parameters of each neuron; The old network weight parameters before triggering fine-tuning; These are the diagonal elements of the Fisher information matrix calculated based on data from the old model. The calculation only considers the weight parameters of the last two fully connected layers of the network. The regularization strength hyperparameter is set to 3000 in this method; This is the index for neuron weight parameters.

[0031] In addition, the present invention also discloses a gasification burner, including a fault diagnosis system, the fault diagnosis system comprising: Memory, used to store computer programs; When the processor executes the computer program stored in the memory, it implements the data-driven gasification combustion fault diagnosis method as described above.

[0032] Compared with the prior art, the present invention has the following beneficial effects: This invention, based on the dynamic mapping of tensor time delay in the fluid continuity equation, eliminates phase misalignment characteristics caused by flow velocity variations, thus reducing the false alarm rate under varying load conditions. An adaptive Kalman noise covariance model coupled with aging time and local variance is introduced, enabling the algorithm to tolerate sensor degradation and extending the effective diagnostic service life of high-temperature measurement points. A thermal resistance network is injected as a priori bias into the initial topology of the graph neural network, and energy balance violation derivation is performed during virtual sample generation, allowing the model to lock in physically reasonable evolutionary solutions even with fewer fault samples. Information geometry theory is applied in the parameter space, relying on the Fisher information matrix to construct an anti-forgetting stiffness field, enabling the diagnostic system to have adaptive evolutionary capabilities across fuel and wear cycles. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0034] Figure 1 This is an overall flowchart of the method described in this invention.

[0035] Figure 2 This is a flowchart of the physical self-consistent data enhancement and virtual fault sample generation process of the present invention.

[0036] Figure 3 This is one of the system operation interface diagrams when using this invention.

[0037] Figure 4 This is the second diagram of the system operation interface when using this invention. Detailed Implementation

[0038] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0039] The following is in conjunction with the appendix Figures 1-4 The embodiments of the present invention will be described in detail below.

[0040] This invention discloses a data-driven gasification combustion fault diagnosis method, comprising the following steps: Step 1: Hardware-level clock synchronization acquisition and anti-aliasing preprocessing for multi-source heterogeneous parameters.

[0041] A multimodal sensor network is deployed at key physical nodes of the gasification burner, including the fuel supply system, primary and secondary air ducts, gasification chamber, combustion chamber, and tail flue, covering parameters such as pressure, temperature, flow rate, and gas composition. For signals with significant frequency characteristics, a microsecond-level timestamp marking mechanism based on a high-precision hardware clock of a programmable logic controller is employed.

[0042] For pressure signals containing dynamic flow field characteristics, a digital anti-aliasing low-pass filter matching the Nyquist sampling theorem is applied before downsampling to a globally uniform time resolution (set to 10Hz in this method). Specifically, a fourth-order Butterworth low-pass filter is used with a cutoff frequency of 45Hz to eliminate parasitic high-frequency electromagnetic aliasing noise while preserving the fluid pressure pulsation envelope characteristics. Subsequent decimation is performed with a time slice step of 0.1 seconds to obtain the downsampled pressure data sequence.

[0043] Step 2: Adaptive Extended Kalman Filter State Estimation with Dual Perception of Lifetime and Operating Condition.

[0044] To address the nonlinear thermoelectric potential drift problem in temperature sensors operating in thermal environments, an adaptive extended Kalman filter (EPF) under non-stationary noise conditions is constructed. The theoretical temperature calculated from the dynamic thermal balance equation of the gasification combustion system is used as a virtual observation benchmark, and the measurement noise covariance matrix of the filter is dynamically adjusted in real time. This matrix is ​​not only a function of the sensor's service time (characterizing the material aging rate) but also coupled with the short-term local variance of the signal (characterizing sudden disturbances in the flow field). Through this design, the system can adaptively and smoothly transfer the confidence weight from the measured data to the physical equation prediction data during sensor physical degradation, outputting a debiased and denoised true temperature state sequence.

[0045] Step 3: Physical self-consistent data enhancement based on transient energy balance and dynamic gasification efficiency.

[0046] Under the constraint of small-sample learning due to the scarcity of actual fault samples, reverse physical completion is performed on data gaps caused by failures such as feed system failures, based on the stoichiometric ratio conservation equation. Simultaneously, to expand the high-value boundary fault samples, the transient energy balance differential equation, which includes a nonlinear decay term of dynamic gasification efficiency, is solved by artificially intervening in the input parameters (such as simulating air-fuel ratio misalignment caused by primary air blower inverter stall). Numerical methods such as the forward Euler method are used to derive the temperature evolution trajectories of the gasification chamber and combustion chamber caused by energy balance disruption, ensuring that the generated virtual fault samples satisfy the first laws of thermodynamics.

[0047] Step 4: Dynamic time delay mapping and tensor alignment based on fluid continuity equation.

[0048] Before inputting multimodal time-series signals into the deep learning network, spatial coherence misalignment caused by changes in fluid medium velocity is eliminated. A physical time-varying delay model based on mass flow rate is established. According to the continuity equation in fluid mechanics, fluid flows between two physical nodes. With nodes The physical transmission time between them is determined by the effective geometric volume of the pipeline and the real-time fluid volumetric flow rate:

[0049] During online diagnosis, the system calculates the local dynamic delay using the diagnosis window, and uses this as the translation step size to dynamically misalign and splice the multi-channel time series tensors to ensure that the tensors finally input to the subsequent feature extraction network are physically aligned.

[0050] Step 5: Extraction of two-branch spatiotemporal features from physical prior injection.

[0051] 5.1 Physical Prior Graph Attention Network Extracts Spatial Features: Each aligned sensor channel is defined as a graph structure. One of the nodes When calculating the initial attention coefficients between nodes, a physical prior bias matrix based on the system thermal resistance network model and mass flow topology is injected. Attention coefficients including prior constraints The calculation is as follows:

[0052] After normalizing and updating the node features, a fixed-length physical space coupling feature vector is obtained through global average pooling.

[0053] 5.2 Parallel Extraction of Temporal Features Using Sparse Transformer and Bidirectional Gated Recurrent Units: The second branch processes the tensor-aligned temporal stream. Two layers of bidirectional gated recurrent units are used to extract local slope evolution and short-to-medium-term evolution features; simultaneously, a sparse Transformer is used in parallel to extract long-range dependencies. In the computation of multi-head self-attention, an additive masking mechanism is employed to reduce computational complexity.

[0054] Step 6: High-order tensor dynamic fusion and focus loss decision.

[0055] The extracted spatial-physical coupling features are concatenated with medium- to long-range temporal evolution features and input into a multi-layer fully connected network for nonlinear mapping. To address the class imbalance problem in real-world industrial environments, where there are massive amounts of normal samples and scarce samples of specific faults, a focal loss function is used for gradient backpropagation calculation.

[0056] Step 7: Adaptive migration of operating conditions based on conditional batch normalization and adversarial gradient.

[0057] To suppress false fault characteristics caused by drastic load fluctuations, boiler load command signals are introduced as conditional parameters into the batch normalization layer of the deep network to implement conditional batch normalization:

[0058] In the formula, The output feature tensor after conditional batch normalization; The input feature tensor; The average value of the characteristic batches; The standard deviation of the characteristic batch; To prevent positive numbers with a denominator of zero, take... ; This is a heat load command signal, a dimensionless percentage. The dynamic scaling factor is modulated by the heat load command signal; It is the dynamic translation factor modulated by the heat load command signal.

[0059] Simultaneously, a domain discriminator is integrated at the network's bottom layer. By inserting a gradient inversion layer between the domain discriminator and the feature extraction network, the feature extraction network is forced to engage in a minimax game with the domain discriminator, ultimately filtering out the specific load information contained in the feature tensor and extracting domain-invariant fault features that characterize the device's health status.

[0060] Step 8: Integral gradient generation of interpretable diagnostic evidence chain.

[0061] After the fault decision is output, the integral gradient method is used to accurately calculate the contribution of each physical sensor channel in the input space to the final output probability within the diagnostic time window:

[0062] The calculated contribution is mapped to a two-dimensional spatiotemporal heat map, pointing to the starting position and evolution source of the fault in physical space.

[0063] Step 9: Incremental learning to resist forgetting based on elastic weights.

[0064] When equipment undergoes fuel changes or experiences long-term liner wear, accumulating a certain number of false alarms or biased samples, an incremental fine-tuning mechanism is triggered. To protect existing diagnostic knowledge from being erased, the target loss function during the fine-tuning phase is reconstructed into a penalty term that includes the diagonal elements of the Fisher information matrix:

[0065] This mechanism is equivalent to applying a non-uniform stiffness field in the parameter space: the weight parameters that are critical to the diagnosis of old faults are locked, and only redundant parameters that are not sensitive to historical tasks are allowed to be updated freely to adapt to the new operating baseline.

[0066] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further illustrated below with reference to specific implementation examples.

[0067] Example 1: This example demonstrates the complete deployment of the hierarchical fault diagnosis system of the present invention on an industrial-grade biomass gasification combustion unit with a rated thermal power of 4MW. The gasification chamber of this unit adopts a double-layer water-cooled wall structure, with a 150mm thick refractory ceramic fiber lining on the inner side (thermal conductivity is non-linear and increases with temperature). The system's design baseline operating temperature range is 800°C to 1100°C. The combustion chamber is equipped with a strong swirl burner. The bottom of the unit is equipped with a continuous slag removal system consisting of an automated reciprocating grate and a closed spiral slag discharger.

[0068] To support the physical deduction of subsequent algorithms in the dimensions of fluid dynamics and thermodynamics, this embodiment first calibrates the global boundary of the system and establishes a high-density three-dimensional spatial sensor perception matrix.

[0069] To ensure the convergence of all physical bias terms and mathematical solutions of differential equations in the diagnostic model, the system-level physical constants and hyperparameters shown in Table 1 must be fixed during the system initialization phase.

[0070] Table 1 Global boundary constraints and hyperparameter calibration matrix of gasification burner fault diagnosis system;

[0071] Table 2 lists the 18 sensor nodes deployed in this embodiment. The three-dimensional spatial relative coordinates of each sensor are centered at the bottom center of the gasification chamber. This provides the basis for constructing the thermodynamic impedance prior matrix in subsequent graph attention networks. It provides a spatial geometric benchmark.

[0072] Table 2. Spatial Topology and Accuracy Attribute Benchmark Table for Multi-Dimensional Sensors in Gasification and Combustion

[0073] The system is equipped with a central diagnostic edge computing server (configured with an Intel Xeon 3.0GHz processor, 64GB ECC memory, and an NVIDIA A100 Tensor Core GPU computing card). All front-end analog signals are connected to a Siemens S7-1500 programmable logic controller via shielded twisted-pair cables. To overcome Ethernet communication jitter, a hardware clock generator at the bottom layer of the programmable logic controller is used to add a microsecond-level absolute timestamp to each raw sampled data packet.

[0074] For the high-frequency pressure signals (raw sampling rate) of nodes 7, 8, 13, 14, and 15 in Table 2 This not only includes macroscopic hydrostatic pressure but also superimposed high-frequency acoustic pulsations induced by combustion chamber thermoacoustic oscillations. Under the constraint of a globally uniform 10Hz time resolution required by the feature extraction branch, if arithmetic mean downsampling is directly performed, according to the Nyquist-Shannon sampling theorem, parasitic high-frequency electromagnetic noise above 5Hz will be aliased and refracted into the low-frequency useful frequency band, contaminating the flow field characteristics.

[0075] This step implements anti-aliasing preprocessing logic: First, analyze the pressure time-series data stream. Implement a cutoff frequency of A fourth-order Butterworth low-pass digital filter. The squared amplitude response function of the Butterworth filter is:

[0076] in Let the filter order be . The angular cutoff frequency is used. After filtering, high-frequency components that may cause aliasing are removed. Then, precise decimation is performed with a time slice step of 0.1 seconds to obtain a 10Hz low-frequency pressure situation tensor that is anti-aliasing and preserves physical characteristics. This tensor is synchronized with the temperature and flow signals with zero phase deviation on the absolute timestamp.

[0077] Because the temperature sensors at nodes 1 to 3 of the vaporization chamber are in a harsh reducing environment, Seebeck coefficient drift may occur after hundreds of hours of operation. This embodiment constructs an adaptive state observer to decouple the nonlinearity throughout the measurement point's lifetime.

[0078] Taking the local state variable evolution of node 1 as an example, the first... State vector at discrete time:

[0079] in The true thermodynamic temperature that is being obscured. This is the sensor hardware attenuation drift bias.

[0080] Establish a partial differential model for state transition (based on the minimum time step discretization of the heat conduction continuity equation):

[0081] Calculate the Jacobian matrix of the transition process :

[0082] Dynamic measurement noise covariance scalar (For scalar observations) The evolution equation for ) is defined as:

[0083] in: It is the baseline variance in Table 1; The degradation factor, the time integral term, represents the parabolic decrease in the system's confidence in its measurement accuracy as the square of the number of hours the sensor is exposed to high temperatures increases; the third term involves a sliding window. The time is set to 60 seconds. When local turbulent disturbances occur in the flow field of the vaporization chamber, the variance within the window increases, which temporarily reduces the measurement confidence.

[0084] Based on this, the prior state prediction and covariance prediction steps of the adaptive extended Kalman filter are as follows:

[0085] Combined with the theoretically predicted temperature derived from the transient energy balance differential equation in step 3 below. As a benchmark, through the Kalman gain matrix Perform state fusion correction:

[0086] because Increases in the later stages of the life cycle. It will approach 0, resulting in the final corrected true temperature state. It mainly follows the forward derivation trajectory based on the first thermal equilibrium equation.

[0087] In actual operation of the gasification burner, the feed rate sensor (node ​​10) on the drive shaft of the feed screw conveyor often loses data due to mechanical jamming caused by dust or communication interruption of the frequency converter. This system abandons mean interpolation or linear interpolation and directly calls the inverse analysis based on the chemical stoichiometry conservation equation.

[0088] When the system detects the loss of signal at node 10, it triggers closed-loop repair calculations. The measured oxygen content in the tailpipe exhaust is then read. (Node 9) and total air mass flow rate (The sum of nodes 11 and 12). Based on the assumption of complete combustion and gas component equilibrium, calculate the real-time stoichiometric ratio for the current operating condition. :

[0089] For example, if a feeding signal interruption is detected at the 500th hour, the data will be read synchronously at this time. , Substituting into the formula, we get:

[0090] because The physical definition of air volume is the ratio of actual air volume to theoretical air volume, i.e. The fuel mass flow rate can be directly calculated from the above.

[0091] The system will reconstruct the physical value Fill the corresponding channel of the input tensor.

[0092] To address the convergence problem of deep learning models under imbalanced sample conditions, this invention generates boundary fault samples such as air-fuel ratio misalignment and sudden feed rate drops. A numerical calculation method is then used to force a solution of the transient energy balance differential equation.

[0093] Taking the simulated oxygen deficiency (air-fuel ratio imbalance) fault caused by partial stall of the primary air fan as an example, a disturbance command is artificially set: the chemical stoichiometric ratio is adjusted. The value was reduced from the normal value of 1.0 to 0.80 (i.e., 20% hypoxia).

[0094] Calling the transient energy balance equation that includes dynamic gasification efficiency:

[0095] Given normal initial steady-state temperature Take the time step. (Consistent with the system sampling rate), the differential term is discretized using the forward Euler method:

[0096] When a disturbance occurs ( When the oxygen level drops to 0.80, the system is in an oxygen-deficient state. The dynamic gasification efficiency is calculated according to the nonlinear decay law defined in this invention:

[0097] Substitute the system rated parameters: fuel thermal power Unchanged; as the primary air volume decreases, the syngas flow rate... The heat loss decreased from the rated 1.5 kg / s to 1.25 kg / s; System heat capacity constant The calculation is as follows:

[0098] exist (At the moment the disturbance occurs), temperature : System heat-generating end (first term on the right side of the equation): .

[0099] The system removes heat (the second and third terms on the right side of the equation): .

[0100] Calculate the rate of temperature change:

[0101] This value is theoretically too large, but in actual engineering calculations, due to the thermal inertia of the medium, the heat storage of the furnace wall, and the kinetic delay of the chemical reaction, the actual rate of temperature change will be significantly lower than this instantaneous value. This method introduces a thermal inertia correction coefficient during discrete iteration. (dimensions are) The effective rate of change is corrected to:

[0102] The corrected initial rate of change is approximately This conforms to the thermal inertia characteristics of the gasification chamber. The computer automatically iteratively solves for 1024 consecutive time steps to obtain a path from... Exponential decay and eventually stabilization Temperature approximation curves on the left and right.

[0103] This temperature vector, derived from physics, is compared with the formula... The generated theoretical oxygen content vector is time-aligned and superimposed with Gaussian background white noise under normal operating conditions (mean 0, variance ). That is, to generate virtual fault samples that are physically and logically self-consistent.

[0104] This method abandons pure mathematical warping (such as dynamic time warping) and adopts dynamic time delay mapping constrained by continuity equations.

[0105] Establish a physical causal chain: a single airflow fluctuation (cause, node 11). Changes in the reduction reaction rate in the gasification chamber The syngas flow passes through the gasification chamber and the combustion chamber space The oxygen content at the tail responds (result, node 9).

[0106] Call the flow rate-delay integral model described in step 4:

[0107] Assume the effective air duct and furnace volume between nodes 11 and 9. The average density of syngas at the current temperature. .

[0108] Operating condition 1: 100% full load operation. Total gas mass flow rate of the system. .

[0109]

[0110] The preprocessing module rounds the delay to 1 second (or uses linear interpolation to maintain decimal precision). When constructing the window tensor for the deep network to read, the time index of the primary airflow channel is shifted forward by 1 step.

[0111] Operating Condition 2: 50% low load operation. The total gas mass flow rate of the system decreases to... .

[0112]

[0113] The preprocessing module dynamically updates the translation step size to 3 seconds.

[0114] The specific implementation of dynamic phase misalignment splicing: in the new tensor provided to the network middle:

[0115] This operation ensures that, regardless of fluctuations in unit load, the cause of the result (air volume) and the result (oxygen content) remain aligned on the matrix row vectors (slices of the same time).

[0116] After completing the dynamic alignment of the physical phase, this system performs nonlinear coupling feature extraction in the spatial dimension.

[0117] (a) Initial tensor mapping of nodes (1D-CNN feature compression).

[0118] For the input tensor after dynamic delay alignment The time step Total number of sensor nodes For each node channel Perform one-dimensional convolution and global average pooling separately.

[0119] Define convolution kernel operations:

[0120] in: For the first Time series after node alignment; It is a one-dimensional convolution kernel with a size of The number of output channels is ; It is the bias vector; Represents the discrete convolution operator; The convolutional feature map uses "same padding" to preserve sequence length.

[0121] Then global average pooling is performed:

[0122] Get each node Initial physical space feature vector The feature vectors of the 18 nodes together constitute the input matrix of the graph neural network. .

[0123] (II) Construction logic and implementation of physical prior bias matrix.

[0124] This embodiment is based on the first principles of heat transfer, utilizing the three-dimensional physical coordinates in Table 2. Construct a thermal resistance network and transform it into a prior bias matrix for injecting a graph attention network. .

[0125] Define nodes With nodes Theoretical combined thermal resistance between :

[0126] in: For space thermal conduction resistance; For convective heat transfer resistance; The Euclidean physical distance between nodes (unit: m); The equivalent thermal conductivity of the refractory lining and the gas is taken as 0.2 W / (m·K). The convective heat transfer coefficient within the system is taken as 15 W / (m²·K).

[0127] To make the thermal resistance dimensionless and incorporate it into the network, the thermal resistance values ​​of all node pairs are first calculated, and then min-max normalization is applied:

[0128] Physical prior bias terms:

[0129] in When node With nodes When the physical distance is extremely large and there is no strong convection direct coupling, , When two nodes are physically adjacent and their thermal resistance approaches zero, , .

[0130] (iii) Attention flow computation with physical constraints.

[0131] Will Inject into a single-layer graph attention network.

[0132] Attention score calculation:

[0133] Normalized calculation:

[0134] Node feature map update:

[0135] Finally, for all updated node features Perform global average pooling at the node level, with an output dimension of... Spatial coupling physical feature vector .

[0136] Parallel to spatial feature branches, input tensor Simultaneously, it enters the dual branch of temporal feature extraction.

[0137] (a) Two-way gated cyclic unit captures local evolution rate.

[0138] Two layers of bidirectional gated recurrent units are configured, with 128 neurons in the hidden layer. At each time step... The forward and backward gated loop units calculate the hidden state respectively. and .

[0139] Take the last time step The states are spliced ​​together:

[0140] Then reduce the dimensionality to another layer through a fully connected layer. Dimension, to obtain the short-to-medium-term evolutionary feature vector .

[0141] (ii) Sparse addition mask Transformer computation logic.

[0142] To capture the slow coking process, global dependencies spanning up to 1024 steps need to be analyzed. This embodiment performs LogSparse sparse masking through linear algebra matrix operations.

[0143] Will Projected through a linear mapping layer (in A learnable category token vector is appended to the beginning of the sequence. This makes the length of the input sequence become After injecting learnable positional encodings, the system enters a multi-head self-attention layer.

[0144] For one of the attention heads, calculate the query, key, and value matrix: , , All dimensions (This embodiment) ).

[0145] Constructing a logarithmic sparse addition mask matrix :initialization It is a matrix with all negative infinity. According to the rule, if the position... and location satisfy ( )or (Self-loop), then update its value to .

[0146] For example, for the target location It is only allowed to be with These historical step sizes, which conform to an exponentially decreasing pattern, are interacted with, corresponding to... The value assigned is The remaining positions are as follows: Keep .

[0147] Perform self-attention computation of the addition mask:

[0148] because The generated attention values ​​are usually in Between, when the mask at the corresponding position is At that time, and as ;go through After activation, the weights of these irrelevant positions are reset to zero.

[0149] Finally, the output vector at position 0 (i.e., the class token position) in the sequence is extracted as the long-range physical dependency feature vector. .

[0150] 10. High-dimensional manifold fusion and classification decision output based on focus loss (corresponding to the specific implementation of step 6).

[0151] After spatiotemporal three-dimensional decoupling and extraction, the diagnostic network converges the feature flow here.

[0152] Perform tensor splicing and fusion: The input is a decision network consisting of three fully connected neurons (128, 64, 5), and the input is processed through... Layer mapping to probability distribution tensor These correspond to: normal operating conditions (0), feeding system blockage (1), coking of the gasification chamber lining (2), partial blockage of the nozzle (3), and air-fuel ratio imbalance (4).

[0153] Backpropagation of focus loss under severe class imbalance.

[0154] Since normal operating data accounts for more than 99% of the gasification burner's operating time, if the standard cross-entropy loss function is used, the network will be overwhelmed by a massive amount of normal samples, causing the gradient direction of a few fault classes to be deflected.

[0155] This embodiment implements focus loss:

[0156] in: The training batch size (set to 64); For the first Each sample in its true label Corresponding predicted probability value ; As the prior class balancing factor, take the normal class as... Take fault type ; .

[0157] In the actual operation of the gasification burner, the heat load command (node ​​18, The load fluctuates frequently between 30% and 100%. Changes in load can cause a global drift in the baseline of the system's temperature and flow fields.

[0158] In this embodiment, a conditional batch normalization layer is used instead of a traditional batch normalization layer in the hidden layer of a fully connected decision network.

[0159] Get real-time load instructions This is used as a priori condition parameter input. The formula for calculating the forward propagation with conditional batch normalization is:

[0160] in: This is the feature tensor output by the previous layer of the neural network; and These are the mean and standard deviation of the features within the current Mini-Batch, respectively. ; The normalized and rescaled feature tensor; and These are the dynamic scaling and translation factors modulated by the load command. This embodiment uses a miniature multilayer perceptron to... Mapped to these two parameters:

[0161] To further extract the load information from the features, in the feature fusion layer An additional domain discriminator subnetwork is then added. The goal of the domain discriminator is to predict the current system load range (range 0: 30-50%, range 1: 50-80%, range 2: 80-100%) based on the features.

[0162] A gradient inversion layer is inserted between it and the main feature extraction network. During forward propagation, the gradient inversion layer acts as an identity mapping; during backpropagation, when calculating the gradient, the gradient inversion layer multiplies the gradient returned by the domain discriminator by a negative constant. (Take 0.1):

[0163] Through this minimax game, the feature extraction network is forced to find a latent space representation that can accurately distinguish faults while preventing the discriminator from accurately predicting the load.

[0164] An integral gradient axiomatic attribution method is employed. A baseline input is set for fault-free operation. The construction method is as follows: Take the average of sensor data from 1024 consecutive time steps within the last 24 hours, where the model classifies the system as operating under normal conditions and the output confidence level is greater than 0.95. The current real-time sensor multidimensional spatiotemporal tensor is... .

[0165] Model for the first The prediction probability function for this type of fault is: .

[0166] No. The integral gradient contribution of each sensor channel is calculated as follows:

[0167] In the implementation of digital discretization in engineering, the integration path is divided into... Step size:

[0168] The calculated absolute contribution of the 18 sensors Normalize the data and map it to a red-blue two-color heatmap.

[0169] Example 2: This example verifies the model's lifetime evolution capability in the face of system fuel changes. After 3000 hours of stable operation in Example 1, the plant switched the fuel from pine pellets to high-moisture rice husk briquettes. This resulted in an overall decrease of approximately 40°C in the reference temperature of the gasification chamber under normal operating conditions, and the change in syngas composition caused a shift in the pressure pulsation spectrum.

[0170] Before incremental learning, the model misinterpreted this long-term baseline drift as a slow air-fuel ratio imbalance, causing the false alarm rate to rise to 22%.

[0171] Elastic weight consolidation closed-loop activation: The system accumulated 500 false alarm deviation samples of rice husk working conditions that had been corrected by manual intervention in the background. The system then triggered an incremental learning program.

[0172] Extract the parameter set of the historical network (a model based on pine particle convergence). Using a small subset (500 samples) of the historical pine sample dataset, the diagonal approximation of the Hessian matrix for the network's predicted log-likelihood with respect to the weights is calculated; this is the Fisher information matrix. Its diagonal elements... The calculation is as follows:

[0173] This element The larger the value, the higher the value. Neuron weights Maintaining the accuracy of historical fault diagnosis for pine wood particles is crucial. To reduce computational overhead, this embodiment only calculates the Fisher information matrix for the weight parameters of the last two fully connected layers of the network.

[0174] The new objective loss function for the rice husk bias samples is reconstructed as follows:

[0175] in .

[0176] When performing stochastic gradient descent, for the weights of the low-level convolutional and graph attention networks responsible for identifying absolute physical faults (such as pressure spikes caused by complete nozzle blockage), The values ​​are too large, so the system locks these parameters; however, for the high-level weights in the classifier responsible for fine-tuning the temperature baseline threshold, The size is small, and the system allows it to deflect to accommodate the new fuel.

[0177] After 5 training cycles (learning rate 0.0001), the updated model retained a 98.6% recognition rate for various faults in early pine wood particles while reducing the false alarm rate of the rice husk baseline from 22% to below 0.8%.

[0178] Comparative Example 1: Traditional Deep Learning Joint Architecture (Long Short-Term Memory Network + Convolutional Neural Network, without physical constraints). The model structure uses a two-layer Long Short-Term Memory Network (256 hidden units) combined with a one-dimensional convolutional neural network. The input does not undergo dynamic time-delay hydrodynamic alignment (fixed 2-second alignment); it does not include extended Kalman filtering (using the original temperature with drift noise); it lacks graph attention physical topology; and it has no elastic weight consolidation protection, only simple full-scale fine-tuning.

[0179] Comparative Example 2: Static Principal Component Analysis and Support Vector Machine. Feature compression was performed on the sensor time window (average over 60 seconds). Principal component analysis was used to extract the principal feature vectors with a cumulative contribution rate of 95%, which were then fed into a support vector machine with a radial basis function kernel for static point classification.

[0180] Table 3. Comprehensive Comparison Test Table of Core Performance of Fault Diagnosis Methods for Gasification Combustion Engines;

[0181] The above test data is based on on-site data collected and offline playback tests during the continuous operation of the same 4MW biomass gasification combustion unit for 6 months.

[0182] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0183] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that 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 data-driven gasification combustion fault diagnosis method, characterized in that, Includes the following steps: Step 1: Synchronously acquire multi-modal sensor data from the gasification burner, perform anti-aliasing low-pass filtering and downsampling processing on the high-frequency pressure signal, and align it with the low-frequency sensor data; Step 2: Construct an adaptive extended Kalman filter, setting the measurement noise covariance as a joint parameter of the sensor service time and the local variance of the signal, and using the theoretical temperature obtained by solving the thermodynamic balance equation as a virtual observation value to compensate for the drift bias of the temperature sensor. Step 3: Complete the missing fuel flow data based on the chemical stoichiometry equation, and generate virtual fault samples by solving the transient energy balance differential equation containing dynamic gasification efficiency and deriving the temperature decay trajectory under different air-fuel ratios. Step 4: Calculate the dynamic transmission delay between sensor nodes based on the fluid dynamics continuity equation, and perform phase shifting and alignment on the multi-source sensor data according to the dynamic transmission delay; Step 5: Construct a dual-branch feature extraction network, use a graph attention network with injected thermodynamic impedance prior matrix to extract spatial coupling features, and use an attention network with sparse addition mask and gated recurrent units to extract temporal evolution features. Step 6: Fuse the spatial coupling features and the temporal evolution features, and output the failure probability distribution through a focus loss function that includes a class balance factor; Step 7: When the model update is triggered, the Fisher information matrix is ​​introduced into the target loss function as a penalty term, and the biased samples are used to perform incremental fine-tuning of the model.

2. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 1, the specific method for performing anti-aliasing low-pass filtering and downsampling processing on the high-frequency pressure signal is as follows: First, the parasitic high-frequency electromagnetic aliasing noise is eliminated by Butterworth low-pass filter with cutoff frequency matched to the Nyquist sampling theorem, and then the downsampled pressure data sequence is obtained by sampling with a set time slice step size.

3. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 2, the measurement noise covariance of the adaptive extended Kalman filter is calculated as follows: based on the baseline measurement noise variance, the product of the material aging coefficient and the square of the cumulative service time of the sensor is added, plus the product of the local fluctuation sensitivity coefficient and the local variance of the measured signal within the past fixed time window is added, and the summation is used to obtain the dynamic measurement noise covariance at the current moment.

4. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 3, the specific method for supplementing the missing fuel flow data based on the stoichiometric ratio equation is as follows: divide the measured air mass flow by the product of the stoichiometric ratio and the theoretical air-fuel ratio, and use the resulting quotient as the supplemented fuel mass flow.

5. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 3, the transient energy balance differential equation is: In the formula, This is the density of syngas, in units of... ; The effective volume of the gasification chamber is expressed in units of... ; The specific heat capacity of syngas at constant pressure, in units of... ; Temperature of the vaporization chamber, in units of or ; For time variables, the unit is ; Thermal power input to fuel, in units of ; The dynamic gasification efficiency is dimensionless. Syngas mass flow rate, in units of ; This refers to the feed temperature, in units of... or ; For heat loss, the unit is heat dissipation. ; When the chemical equivalence ratio At that time, the dynamic gasification efficiency Take a constant value of 0.75; when hour, .

6. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 4, the specific method for calculating the dynamic transmission delay between sensor nodes is as follows: multiply the effective physical volume between the nodes by the syngas density at the real-time temperature, and then divide by the total mass flow rate of the gas passing through the system at the current moment. The resulting quotient is used as the dynamic delay of the flow field change being transmitted from one node to another.

7. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 5, the formula for calculating the bias value of the thermodynamic impedance prior matrix in the graph attention network is: In the formula, For nodes With nodes The physical prior bias between them is dimensionless; The hyperparameter is a physical prior adjustment parameter, which is dimensionless and is set to 5.0 in this method; For the node With the node The normalized theoretical combined thermal resistance is dimensionless; the theoretical combined thermal resistance is obtained by adding the spatial thermal conduction resistance and the convective heat transfer resistance, and the normalization adopts minimum-maximum normalization to the interval. .

8. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 5, the attention network with sparse additive mask is calculated as follows: the self-attention output matrix is ​​equal to the product of the query matrix and the transpose of the key matrix, divided by the square root of the key vector dimension, plus the sparse mask matrix, the result is processed by a normalized exponential function, and then multiplied by the value matrix; the setting rule for the elements in the sparse mask matrix is: when position i and position j satisfy j equal to i minus 2 to the power of p, where p is a non-negative integer, the mask element takes the value of 0; when this condition is not met, the mask element takes the value of negative infinity.

9. The data-driven gasification combustion fault diagnosis method according to claim 1, characterized in that, In step 6, the focus loss function is calculated as follows: the focus loss value is equal to the category balance factor multiplied by 1 minus the focus parameter of the model's predicted probability value of the true category raised to the power of 1, and then multiplied by the negative of the natural logarithm of the predicted probability value; the focus parameter is set to 2.0; the category balance factor value of the normal category is set to be less than the category balance factor value of the fault category.

10. A gasification burner, characterized in that: Includes a fault diagnosis system, the fault diagnosis system comprising: Memory, used to store computer programs; The processor, when executing the computer program stored in the memory, implements the data-driven gasification combustion fault diagnosis method according to any one of claims 1 to 9.

Images