Boiler atmosphere field state monitoring method

Abstract

The invention relates to the field of electrical digital data processing, in particular to a boiler atmosphere field state monitoring method, which comprises the following steps: collecting concentration data of combustible gases in a boiler water wall area through measuring equipment, carrying out space-time calibration on the concentration data of multiple measuring points, constructing a data processing model based on a graph convolutional network, and training the data processing model. Obtaining a trained data processing model; the method comprises the following steps: collecting concentration data of each combustible gas in a boiler water-cooled wall area through measuring equipment to obtain new multi-measuring-point concentration data; and after the steps are executed, inputting the data into a trained data processing model to obtain a combustible gas predicted concentration matrix of each measuring point at a future moment, continuously monitoring the combustible gas predicted concentration matrix of each measuring point at the future moment, and when the predicted concentration of any measuring point exceeds a safety threshold value, triggering an early warning signal and positioning a high-risk area. An existing boiler atmosphere field state monitoring method has the problem of low prediction credibility, and the method provided by the invention is high in prediction credibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electric digital data processing, in particular to a boiler atmosphere field state monitoring method. BACKGROUND

[0002] Thermal power generation is a power generation method that converts heat energy generated by combustible materials during combustion into electric energy through a power generation device, and is also the most common power generation method.

[0003] Thermal power plants mostly use power station boilers to obtain the heat energy of combustible materials. When the power station boiler is running, the water wall of the power station boiler and its surrounding area always have incomplete combustion or local high-temperature reduction conditions, which causes H2S, CO, SO2 and other toxic combustible gases to accumulate and diffuse in small gaps and wall boundary layers. When the concentration of combustible gas exceeds the safety threshold, it will induce furnace explosion, water wall high-temperature corrosion or unplanned shutdown, causing double risks of equipment damage and personnel injury. In order to reduce the occurrence of such situations, people will monitor the atmosphere field state of the boiler water wall and its surrounding area.

[0004] The existing boiler atmosphere field state monitoring method mainly collects boiler atmosphere field data through dispersedly arranged sensors. The sampling period is long, the measurement points are sparse, and different sensors have different installation positions, pipeline lengths and response times, which causes the data to be spatially misaligned and temporally out of phase, resulting in that the concentration field obtained by the central DCS cannot truly reflect the three-dimensional distribution and is difficult to capture transient leaks in time. At the same time, the existing graph neural network based on fixed graph structure or pure time series deep data processing model regards the sensor relationship as static and completely relies on data fitting for physical laws when modeling gas diffusion. As a result, non-physical solutions that violate the diffusion equation may occur in the sparse sensor coverage area, and the response to sudden concentration peaks is lagging.

[0005] Therefore, the existing boiler atmosphere field state monitoring method has the problem of low prediction credibility. SUMMARY

[0006] The technical problem to be solved by the present application is to overcome the shortcomings of the above-mentioned prior art and provide a boiler atmosphere field state monitoring method with high prediction credibility.

[0007] To solve the above technical problems, the present application provides a boiler atmosphere field state monitoring method, which comprises:

[0008] S1. Collecting concentration data of each combustible gas in the boiler water wall area by a measuring device to obtain multi-measurement-point concentration data;

[0009] S2. Spatio-temporal calibration of the multi-measurement-point concentration data to obtain a calibrated concentration vector;

[0010] S3. Construct a data processing model based on graph convolutional networks;

[0011] S4. Train the data processing model to obtain the trained data processing model;

[0012] S5. Collect the concentration data of each combustible gas in the boiler water-cooled wall area through the measuring equipment to obtain new multi-point concentration data; after executing S2, input it into the trained data processing model to obtain the H2S / CO predicted concentration matrix of each measuring point at future time; continuously monitor the H2S / CO predicted concentration matrix of each measuring point at future time, and when the predicted concentration of any measuring point exceeds the safety threshold, trigger the early warning signal and locate the high-risk area.

[0013] As a further improvement of the present invention: the measuring device is a distributed gas sensor array or a physical probe array.

[0014] As a further improvement of the present invention: step S2 includes:

[0015] S201. Based on the spatial distance between two measuring points and the Pearson correlation coefficient of the historical concentration sequences of the two measuring points, calculate the elements of the spatial calibration matrix and generate the spatial calibration matrix;

[0016] S202. Calculate the elements of the time delay matrix using the cross-correlation function, compensate the original concentration by combining the concentration change rate, and use the spatial calibration matrix elements and time delay matrix elements of the spatial neighborhood measurement point set to perform weighted aggregation on the original concentration vector to generate a calibrated concentration vector.

[0017] As a further improvement of the present invention: step S3 includes:

[0018] S301. Generate a dynamic adjacency matrix based on the scalar gradient value of each gas, and use the dynamic adjacency matrix for data processing;

[0019] S302. Based on the adjacency matrix at each scale, the feature matrix output by the l-th layer of the graph convolutional network at each time step is obtained, and the feature matrix output by the l-th layer of the graph convolutional network at each time step is used for data processing;

[0020] S303. Output the spatiotemporal fusion feature matrix based on the attention weights of each measurement point at historical moments;

[0021] S304. Calculate the predicted concentration matrix for each time point;

[0022] S305. Obtain the prediction loss function based on the prediction concentration matrix and the prediction concentration spatial gradient matrix;

[0023] S306. Generate a physical constraint loss term based on the predicted concentration change rate and the discrete Laplace operator value;

[0024] Preferably, step S301 includes:

[0025] S3011. Based on the calibration concentration matrix and the spatial location of the measuring points, calculate the gradient vector of each measuring point at different times, and use the scalar concentration values ​​and spatial distances of the spatial neighborhood measuring point set to calculate the scalar gradient value of each gas.

[0026] S3012. After concatenating the calibration concentration matrix and the concentration space gradient matrix, the matrix is ​​mapped to the latent space through a linear projection matrix. The output matrix is ​​then processed by the hyperbolic tangent activation function, the dynamic adjacency weight matrix, and the dimension transformation matrix. The output matrix is ​​then row-normalized using the Softmax function to generate the dynamic adjacency matrix.

[0027] Preferably, step S302 includes:

[0028] S3021. Based on the dynamic adjacency matrix and the k-hop mask matrix, the elements of the multi-scale adjacency matrix are calculated through the Hadamard product to obtain the adjacency matrix at each scale;

[0029] S3022. Perform normalized graph convolution on the adjacency matrices at each scale, sum them, and then process them through the Sigmoid activation function. Generate a gated signal vector based on the global average pooling of the input feature matrix. After copying it into a matrix, obtain the output feature matrix through Hadamard product modulation.

[0030] Preferably, step S303 includes:

[0031] S3031. Based on the feature matrix at the current time and the feature matrix at the historical time, generate the query matrix and the key matrix by querying the projection matrix and the key projection matrix, and calculate the attention weight of each measurement point at the historical time τ using the row vector inner product;

[0032] S3032. The historical feature matrix is ​​aggregated by weighting attention weights and added to the current feature matrix. The spatiotemporal fusion feature matrix is ​​then output through layer normalization.

[0033] Preferably, step S305 includes:

[0034] S3051. Based on the spatiotemporal fusion feature matrix, the predicted concentration matrix and the predicted concentration spatial gradient matrix are output synchronously through feature encoding weight matrix and modified linear unit activation function, and then linear calculation of the prediction weight matrix.

[0035] S3052. The error between the predicted concentration matrix and the true concentration matrix is ​​constrained by the L2 norm, and the error between the predicted concentration space gradient matrix and the true concentration space gradient matrix is ​​constrained by the L1 norm. The gradient loss weight coefficient balances the two losses to form the prediction loss function.

[0036] Preferably, step S306 includes:

[0037] S3061. Based on the predicted concentration and spatial distance between measurement points, the discrete Laplace operator value of each measurement point is calculated using the predicted concentration value and spatial distance of the spatial neighborhood measurement point set;

[0038] S3062. Based on the predicted concentration change rate and the discrete Laplace operator value, the residual of the diffusion equation is calculated using the diffusion coefficient, and the physical constraint loss term is generated by weighted summation using the inverse of the spatial distance.

[0039] As a further improvement of the present invention: step S4 includes:

[0040] S401. Integrated total loss function based on physical constraint weight coefficients;

[0041] S402. Data processing model parameter update and iterative training;

[0042] S403. Determining the stopping iteration condition;

[0043] S404. Save and output the optimal data processing model to obtain the trained data processing model.

[0044] The beneficial effects of the present invention are as follows: The boiler atmosphere field condition monitoring method provided by the present invention has high predictive reliability.

[0045] First, this method combines the spatial topology matrix and the sensor response delay matrix to apply to the raw concentration data, thereby eliminating spatial misalignment and temporal offset in one go, making the data from multiple measurement points truly synchronized and isomorphic, and improving the reliability of the basic data.

[0046] Secondly, this method utilizes the real-time concentration gradient to drive the adaptive evolution of the adjacency matrix, and the graph edge weights change in real time with the gas diffusion path, breaking through the bottleneck that conventional graph convolution's "fixed connection" cannot characterize transient diffusion.

[0047] Furthermore, this method directly embeds the discrete residuals of the diffusion equation into the loss function, forcing the predicted concentration field to satisfy the mass conservation and diffusion law, avoiding "non-physical solutions" in the sparse regions of the sensor, and improving the rationality of the prediction.

[0048] Finally, this method simultaneously outputs the future concentration value and its spatial gradient, and designs an L1-L2 mixed loss, which significantly enhances the model's response speed to leakage mutations, achieves dual early warning of "trend + magnitude", and improves the risk response speed.

[0049] In summary, after optimization in various aspects, this method can yield prediction results with high reliability. This allows boiler plant staff to make more reliable responses based on the prediction results. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the principle of the present invention;

[0051] Figure 2 This is a diagram illustrating the impact of the spatiotemporal calibration of this invention on gas concentration measurement.

[0052] Figure 3 This is a comparative analysis diagram of the static adjacency matrix and the dynamic adjacency matrix of the present invention;

[0053] Figure 4 This is a diagram illustrating the impact of physical constraints on sparse region prediction in this invention.

[0054] Figure 5 This is a diagram illustrating the impact of physical constraints on the prediction results in this invention.

[0055] Figure 6 This is an analysis diagram of the multi-scale graph convolution feature extraction effect of the present invention. Detailed Implementation

[0056] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0057] like Figure 1 As shown, the boiler atmosphere field condition monitoring method provided by the present invention includes:

[0058] S1. Collect the concentration data of each combustible gas in the water-cooled wall area of ​​the boiler using measuring equipment to obtain multi-point concentration data;

[0059] The measuring equipment is deployed in the boiler water-cooled wall area, and the measuring equipment is a distributed gas sensor array or a physical probe array;

[0060] Combustible gases include H2S, CO, and SO2. H2S and CO are preferred. The following explanation uses these two gases as examples, but CO and SO2 can also be collected in practical applications.

[0061] When using a distributed gas sensor array to periodically collect concentration data of H2S and CO, the distributed gas sensor array is fixedly installed in the key monitoring area of ​​the water-cooled wall according to a preset spatial topology. The dual-channel gas concentration data of each measuring point is collected synchronously every 1 second. The unit of the concentration data is ppm, forming a raw concentration time series dataset. The raw concentration time series dataset is transmitted to the central processing unit through the industrial bus and stored as a structured data matrix (i.e., multi-measuring-point concentration data) according to the measuring point number and timestamp. The matrix rows correspond to the measuring point index, the columns correspond to the time sampling points, and the channel dimensions record the H2S and CO concentration values ​​respectively.

[0062] When using a physical probe array to collect concentration data of H2S and CO, the physical probe array is deployed in the boiler water-cooled wall area according to a preset spatial topology. Each probe transmits the extracted gas to the central measurement module in real time. The module then performs dual-channel gas concentration analysis (unit ppm) at a cycle of 1 second, generating a structured concentration matrix (i.e., multi-point concentration data) with timestamps and measurement point numbers.

[0063] S2. Perform spatiotemporal calibration on the concentration data from multiple measurement points to obtain the calibrated concentration vector;

[0064] S201. Calibrate spatial topology;

[0065] Based on the spatial distance between two measuring points and the Pearson correlation coefficient of the historical concentration sequences of the two measuring points, the elements of the spatial calibration matrix are calculated to generate a spatial calibration matrix that preserves the spatial topological relationship between the measuring points and eliminates spatial feature confusion.

[0066] The element in the i-th row and j-th column of the spatial calibration matrix Γ is represented as Γ ij , Γ ij Used to characterize the spatial calibration weight between the i-th measurement point and the j-th measurement point;

[0067] Γ ij The calculation formula is:

[0068]

[0069] In the formula, i is the measurement point index; j is the measurement point index that is different from i; d ij d represents the spatial distance between the i-th and j-th measuring points, in meters. ij Euclidean distance is used to quantify the degree of physical spatial separation between measurement points, ensuring that neighboring measurement points have higher weights and preserving the topological structure; σ d This is a scaling factor for spatial distance, measured in meters, used to control the rate attenuation of the distance's influence. It can be set to σ. d =5; corr(c i ,c j ) represents the Pearson correlation coefficient between the historical concentration sequences of the i-th and j-th measurement points, characterizing the correlation between the concentration changes at the two measurement points; c i This represents the historical concentration sequence of the i-th measuring point, in vector form, obtained through periodic sampling by the sensor. It refers to the time-series data of H2S and CO concentrations collected at the i-th measuring point over a past period, such as data within the past hour; c j The historical concentration sequence of the j-th measurement point;

[0070] S202. Calculate the calibrated concentration vector;

[0071] The elements of the time delay matrix are calculated by cross-correlation function, and the original concentration is compensated by the concentration change rate. The original concentration vector is weighted and aggregated by the spatial calibration matrix elements and time delay matrix elements of the spatial neighborhood measurement point set to generate a calibrated concentration vector (the calibrated concentration vector corresponds to the calibrated concentrations of the two gases [H2S, CO]) to achieve time synchronization compensation and eliminate the influence of sensor response delay.

[0072] The element ΔT in the i-th row and j-th column of the time delay matrix ΔT ij The calculation formula is:

[0073]

[0074] In the formula, arg max τ This represents the value of τ that maximizes the function value; τ is the index of the time delay variable. It is the cross-correlation function between the concentration at the i-th measurement point at time t and the concentration at the j-th measurement point at time t-τ. The cross-correlation function is a mathematical tool in signal processing that measures the similarity between two time series under different time delays. and Similarity between two time series; Let be the concentration vector of the i-th measuring point with time t as the center of the time window sequence, corresponding to the concentrations of the two gases [H2S, CO]. The window sequence can have 0 time steps. Let be the concentration vector of the j-th measuring point with time window t-τ as the center, corresponding to the concentrations of the two gases [H2S, CO]. The window sequence can have 0 time steps; t is the index of the current time; τ is the index of the historical time.

[0075] The concentration vector calibrated at time t for the i-th measuring point The calculation formula is:

[0076]

[0077] In the formula, N(i) represents the set of spatial neighboring measurement points of the i-th measurement point, and the range of the neighborhood is taken as the radius threshold; For the j-th measuring point at t-ΔT ij The original concentration vector at time ΔT corresponds to the original concentrations of the two gases [H2S, CO]; ij Let be the element in the i-th row and j-th column of the time delay matrix ΔT; The sign for partial derivatives; Let be the original concentration vector of the j-th measuring point at time t; t is the index of the current time. Let be the concentration change rate vector of the j-th measuring point at time t, corresponding to the concentration change rates of the two gases [H2S, CO]. It is approximately calculated using the central difference method, and the sampling time interval Δt calculated by the central difference method is consistent with the sampling interval of the sensor.

[0078] The H2S and CO concentration data in the boiler water-cooled wall region suffer from uneven spatial distribution and asynchronous temporal acquisition. Conventional methods directly stitch together data from multiple measuring points, which not only ignores the positional relationship of sensors within the boiler space, leading to spatial feature confusion, but also fails to account for the time delay caused by differences in the response speeds of different sensors, affecting time synchronization. S2 performs joint spatiotemporal calibration of the original concentration data by constructing a spatial calibration matrix and a time delay matrix, eliminating the effects of spatial topology confusion and sensor response delay. Figure 2 As shown, the effectiveness of S2 is verified by analyzing 3D spatiotemporal concentration field experimental diagrams. Figure 2 This includes the horizontal axis (position, in meters), the vertical axis (time, in seconds), and the depth axis (hydrogen sulfide concentration, in parts per million, ppm). Figure 2 The 3D surface plot can intuitively show the spatial location and temporal distribution of hydrogen sulfide concentration inside the boiler. Figure 2 The paper presents both uncalibrated data and calibrated data using S2 calibration. The true concentration field exhibits a smooth diffusion waveform, consistent with the physical law of gas diffusion from the center outwards. The uncalibrated data represents the lower surface, which shows obvious waveform misalignment and spatial distortion. The calibrated data represents the upper surface, with a stable waveform and normal timing phase. Experimental results demonstrate that the S2 spatial calibration matrix and time delay compensation mechanism can effectively solve the problems of sensor position deviation and response delay.

[0079] S3. Construct a data processing model based on graph convolutional networks;

[0080] S301. Generate a dynamic adjacency matrix based on the scalar gradient value of each gas, and use the dynamic adjacency matrix for data processing;

[0081] S3011. Calculate the scalar gradient value for each gas;

[0082] Based on the calibration concentration matrix and the spatial location of the measuring points, the gradient vector of each measuring point at different times is calculated. Using the scalar concentration values ​​and spatial distances of the spatial neighborhood measuring point set, the scalar gradient value of each gas is calculated.

[0083] Gradient matrix integration is performed, concatenating the gradient values ​​of the two gases at each measuring point into a vector to characterize the local concentration change trend; thus, the gradient vector of the i-th measuring point at time t is obtained. Vertically stacking the gradients of all measurement points yields the spatial gradient matrix of concentration at all measurement points at time t.

[0084] The gradient vector of the i-th measurement point at time t The calculation formula is:

[0085]

[0086] In the formula, Indicates matrix transpose; Let be the scalar gradient value of the m-th gas at the i-th measuring point at time t;

[0087] The scalar gradient value of the m-th gas at the i-th measuring point at time t. The calculation formula is:

[0088]

[0089] In the formula, m represents the index of H2S and CO respectively, m=1 refers to the concentration of H2S, and m=2 refers to the concentration of CO. Let m be the scalar concentration value of the m-th gas at the j-th measuring point at time t; Let be the scalar concentration value of the m-th gas at the i-th measuring point at time t. The element in the i-th row and m-th column;

[0090] S3012. Generate a dynamic adjacency matrix;

[0091] The calibration concentration matrix and the concentration space gradient matrix are concatenated and mapped to the latent space through a linear projection matrix. After processing by the hyperbolic tangent activation function, dynamic adjacency weight matrix and dimension transformation matrix, the output matrix is ​​obtained. The output matrix is ​​then normalized row-wise by the Softmax function to generate a dynamic adjacency matrix. The elements of the dynamic adjacency matrix represent the diffusion correlation strength between measurement points. The dimension is N×N, where N is the total number of measurement points.

[0092] The dynamic adjacency matrix A at time t t The calculation formula is:

[0093]

[0094] In the formula, Soft(·) is the Softmax function, calculated using row normalization; W ce Let be the dimension transformation matrix, and be trainable parameters, which transform an N×D [structure / structure]. The matrix transformation is to a dimension of N×N, where D is the output hidden layer dimension of the linear projection matrix Θ, and N is the total number of measurement points; W a`T` is the dynamic adjacency weight matrix, a trainable parameter used to capture higher-order interactions between features. Its dimension is D×D, and D can be set to 64. `tanh(·)` is the hyperbolic tangent activation function, limiting the output range to [-1, 1]. `Θ` is the linear projection matrix, a trainable parameter with a dimension of 4×D, which concatenates features of dimension N×4. Mapped to a latent space of dimension N×D; || represents the channel concatenation operation; Let be the spatial gradient matrix of the concentration at all measuring points at time t, derived from all measuring points. It is constructed by vertical stacking, that is, It is a matrix form representing the vertically stacked gradients of all measurement points. It is the vector form of the gradient at a single measurement point;

[0095] Conventional graph convolutional networks use static adjacency matrices to model sensor relationships, which cannot adapt to the dynamic diffusion characteristics of gas concentrations within a boiler over time and space, leading to inaccurate spatial dependency modeling. S301, however, dynamically generates an adaptive adjacency matrix by fusing calibrated concentration data with its spatial gradient information, thus more accurately capturing spatial dependencies during gas diffusion. Experiments comparing the fundamental differences between conventional static adjacency matrices and S301's dynamic adjacency matrix yield a graph structure thermogram comparison. Figure 3 As shown, the horizontal and vertical axes of each heatmap are the measurement point indices (i.e., sensor index numbers), and the color intensity represents the adjacency weight (i.e., diffusion correlation strength). The darker the color, the higher the diffusion correlation strength, meaning a larger weight and stronger gas diffusion coupling between sensors. The leftmost heatmap is a static adjacency matrix based on a fixed spatial distance, which presents a uniform and symmetrical pattern and cannot reflect the dynamic characteristics of gas diffusion. The three heatmaps on the right show the dynamic adjacency matrices at 10 seconds, 20 seconds, and 30 seconds, respectively. They all show a dispersed block shape, reflecting the changes in the gas diffusion path at 10, 20, and 30 seconds. This indicates that the dynamic adjacency matrix generation module of S301 can adaptively capture the time-varying characteristics of gas diffusion in the boiler, overcoming the shortcomings of conventional fixed diagram structures that cannot adapt to dynamic diffusion processes.

[0096] S302. Based on the adjacency matrix at each scale, the feature matrix output by the l-th layer of the graph convolutional network at each time step is obtained, and the feature matrix output by the l-th layer of the graph convolutional network at each time step is used for data processing;

[0097] S3021. Generate adjacency matrices at various scales;

[0098] Based on the dynamic adjacency matrix and k-hop mask matrix, the elements of the multi-scale adjacency matrix are calculated through the Hadamard product to obtain the adjacency matrix at each scale. The adjacency matrix at each scale is used to encode the diffusion correlation strength at different scales and expand the receptive field of graph convolution.

[0099] The adjacency matrix at time t and scale k is represented as: This is used to encode the diffusion association strength between measurement points at time t within the range of k-hop neighbors. k-hop is the shortest path length between nodes / measurement points; for example, k=1 represents direct neighbors, k=2 represents neighbors of neighbors, and k-hop neighbors refer to pairs of measurement points whose shortest path distance is equal to k. The adjacency matrix at time t is at scale k. The calculation formula is:

[0100]

[0101] In the formula, k is the scale index, k = 1, 2, ..., K; K is the maximum scale number, which can be set to K = 3; ⊙ represents the Hadamard product, i.e., element-wise multiplication; M k Let be a k-hop mask matrix. If the j-th measurement point is a k-hop neighbor of the i-th measurement point, the element in the i-th row and j-th column of the k-hop mask matrix is ​​1; otherwise, it is 0.

[0102] S3022. Obtain the feature matrix output by the l-th layer of the graph convolutional network at each time step;

[0103] Normalized graph convolution is performed on the adjacency matrices at each scale. After summation, the matrix is ​​processed by the Sigmoid activation function. A gated signal vector is generated based on the global average pooling of the input feature matrix. After being copied into a matrix, the output feature matrix is ​​obtained by Hadamard product modulation. The output feature matrix is ​​used to capture multi-scale diffusion features.

[0104] The feature matrix output by the l-th layer of the graph convolutional network at time t. Represented as:

[0105]

[0106] In the formula, Sig(·) is the Sigmoid activation function; The -1 / 2 power of the degree matrix at scale k is a diagonal matrix used for normalization; D k It is a degree matrix with a scale of k; Let be the feature matrix output by the (l-1)th layer of the graph convolutional network at time t, which is also the feature matrix input to the lth layer; is the graph convolution weight matrix at scale k, and are trainable parameters; ⊙ represents the Hadamard product, which applies the gated signal to the feature at each measurement point; M N [·] indicates that the vector is copied N times to form a matrix; GAP(·) is a global average pooling operation, which outputs a gated signal vector with a dimension of 1×F; l is the layer index of the graph convolutional network; This indicates that the dimension is 1×F. The vector is copied N times and broadcast as an N×F matrix; F is the characteristic matrix. Feature dimensions;

[0107] The degree matrix D at scale k k Represented as:

[0108]

[0109] In the formula, diag(·) is the diagonal matrix formed by extracting the diagonal elements of the vector; For the element in the i-th row and j-th column of the adjacency matrix at k-scale at time t, quantify the diffusion correlation strength between the i-th and j-th measurement points at k-scale. The larger the value, the stronger the hash correlation. This represents the degree of the i-th measurement point at time t on scale k, which is the sum of the connection weights of all k-hop neighbors;

[0110] Conventional graph convolution operations typically use neighborhood information at a single scale, making it difficult to effectively capture the characteristics of different ranges involved in the diffusion of H2S and CO gases in the boiler atmosphere. This results in the data processing model being unable to collaboratively model local details and global patterns. S302, on the other hand, employs a multi-scale adjacency matrix and gating mechanism to construct a graph convolutional layer with an adaptive receptive field, simultaneously extracting local detail features and global diffusion features, enabling the data processing model to collaboratively model local details and global patterns.

[0111] S303. Output the spatiotemporal fusion feature matrix based on the attention weights of each measurement point at historical moments;

[0112] S3031. Calculate the attention weight for each measurement point at historical moments;

[0113] Based on the feature matrix at the current time and the feature matrix at historical time, a query matrix and a key matrix are generated by querying the projection matrix and the key projection matrix. The attention weight α at historical time τ for each measurement point is calculated using the row vector inner product. τ α τ Characterizing the importance of features at time τ to features at time t;

[0114] The query matrix q at time t t The calculation formula is:

[0115]

[0116] In the formula, W q The weight matrix for querying the projection is a trainable parameter; Let be the feature matrix output by the Lth layer of the graph convolutional network at time t, with dimensions N×F; L is the number of layers in the graph convolutional network.

[0117] The bond matrix k at time τ τ The calculation formula is:

[0118]

[0119] In the formula, W k The weight matrix for the key projection is a trainable parameter. Let F be the feature matrix output by the Lth layer of the graph convolutional network at time τ, with dimensions N×F.

[0120] Attention weight α at historical time τ τ The calculation formula is:

[0121]

[0122] In the formula, q t Let k be the query matrix at time t; τ Let F be the key matrix at time τ; F is the characteristic matrix. The feature dimension; T is the size of the time window, which limits the range of historical moments involved in the calculation to [tT, t-1], and can be set to T = 50; <·,·> represent vector dot product operations. When calculating attention weights, row vector dot products are used to perform vector dot product operations independently on each measurement point; For t a The key matrix at time step;

[0123] Attention weight α τ During calculation, the feature vector of each measurement point needs to be processed independently. For the i-th measurement point, the query vector at time t of the i-th measurement point... Let the dimension be F, and let F be the key vector at time τ of the i-th measurement point. Let F be the dimension, and let F be the attention weight of the i-th measurement point at time τ in history. The calculation formula is:

[0124]

[0125] In the formula, Let be the query vector for the i-th measurement point at time t; Let be the key vector at time τ of the i-th measurement point; For the i-th measurement point t a The key vector at time step;

[0126] S3032. Output the spatiotemporal fusion feature matrix;

[0127] The historical feature matrix is ​​aggregated by weighting attention weights and added to the current feature matrix. The spatiotemporal fusion feature matrix is ​​then output through layer normalization to enhance spatiotemporal continuity.

[0128] Feature matrix of spatiotemporal fusion at time t Represented as:

[0129]

[0130] In the formula, Lay(·) is the layer normalization operation; T is the time window size; W c The projection matrix is ​​a trainable parameter that maps historical features to a space compatible with the current features.

[0131] The residual connection represents the term, which retains the original information and alleviates gradient vanishing by adding aggregated historical features to the current features. Furthermore, it stabilizes the feature distribution through layer normalization, which improves training stability. At the same time, it strengthens the spatiotemporal continuity through cross attention, so that the historical diffusion pattern can dynamically guide the current state modeling and avoid feature incoherence caused by time breakpoints.

[0132] If characterizing historical features With current features For large scale differences, a learnable scaling gate vector can be added to ensure a stable feature distribution in the fused features. That is, the layer normalization operation uses the input combined with the scaling gate vector β, expressed as:

[0133]

[0134] In the formula, β is the scaling gate vector, which is a trainable parameter; The term represents a weighted aggregation of historical features, using attention weight α. τ Dynamically fuse historical moment features, value projection matrix W c Project historical features into a space that is compatible with current features.

[0135] Conventional time-series modeling methods struggle to effectively correlate spatial diffusion patterns at different historical moments within a boiler atmosphere field with the current state. This results in historical diffusion information being unable to effectively guide current predictions and also ignores the continuous temporal and spatial variations in gas concentration. S303 addresses this by fusing features from the current moment with those from historical moments through a cross-attention mechanism, dynamically capturing spatiotemporal dependencies and enhancing the continuity of feature representation.

[0136] S304. Calculate the predicted concentration matrix for each time point;

[0137] Feature matrix that integrates spatiotemporal fusion By mapping to the concentration space through a fully connected layer, the spatiotemporally fused feature matrix is ​​obtained. Each measurement point's feature vector is processed independently, outputting the H2S and CO concentration values ​​for the corresponding measurement point, forming the predicted concentration matrix at time t. In this fully connected layer, through the fully connected layer weights W f Bias b with fully connected layer f Feature matrix of spatiotemporal fusion conduct Linear calculations of this form yield the predicted concentration matrix at time t. Each row corresponds to the predicted concentration of [H2S, CO] for one measurement point. Where W... f and b f To ensure that the output dimension matches the number of measurement points and the types of gases, the parameters are designed to be trainable.

[0138] S305. Based on the predicted concentration matrix and the predicted concentration spatial gradient matrix, a prediction loss function is obtained to enhance the prediction of concentration gradient;

[0139] S3051. Calculate the predicted concentration matrix and the predicted concentration spatial gradient matrix;

[0140] Based on the spatiotemporal fusion feature matrix, the system processes the feature encoding weight matrix and the modified linear unit activation function, followed by linear calculation of the prediction weight matrix. This simultaneously outputs the predicted concentration matrix and the predicted concentration spatial gradient matrix, enhancing the capture of gas diffusion trends. This is represented as follows:

[0141]

[0142] In the formula, This is the vertical concatenation operator, which stacks two matrices into one matrix. The concentration matrix is ​​predicted for time t+1, containing the [H2S,CO] concentrations at N measurement points; The predicted spatial gradient matrix of concentration at time t+1 characterizes the concentration change trend at each measuring point. Its physical meaning is the spatial change trend of concentration output by the data processing model, compared with the actual spatial gradient matrix of concentration. Consistency constraints are applied to the loss function; W p The prediction weight matrix is ​​denoted by , which are trainable parameters; ReLU(·) is the modified linear unit activation function; W e The weight matrix for feature encoding is a trainable parameter;

[0143] S3052. Calculate the prediction loss function;

[0144] The error between the predicted concentration matrix and the true concentration matrix is ​​constrained by the L2 norm, and the error between the predicted concentration space gradient matrix and the true concentration space gradient matrix is ​​constrained by the L1 norm. The gradient loss weight coefficient balances the two losses to form the prediction loss function.

[0145] Prediction loss function L pred The calculation formula is:

[0146]

[0147] In the formula, ||·||2 represents the L2 norm, which constrains the accuracy of the absolute concentration value; Predict the concentration matrix at time t+1; Let be the true concentration matrix at time t+1; ||·||1 represents the L1 norm, which constrains the consistency of the gradient direction; λ is the gradient loss weight coefficient, which can be set to 0.5 to balance the contributions of the two loss terms. The spatial gradient matrix of the true concentration at time t+1 is obtained by the central difference method.

[0148] Conventional concentration prediction methods focus only on the absolute value of gas concentration, ignoring spatial trends and resulting in a lag in the response of data processing models to dynamic changes in gas diffusion, making it difficult to accurately predict abrupt concentration changes. S305 constructs a gradient-enhanced prediction mechanism by jointly predicting future concentration values ​​and their spatial gradients, thereby strengthening the data processing model's ability to capture gas diffusion trends.

[0149] S306. Based on the predicted concentration change rate and the discrete Laplace operator value, generate a physical constraint loss term to regularize the diffusion physical constraint;

[0150] S3061. Calculate the discrete Laplace operator value for each measurement point.

[0151] Based on the predicted concentration and spatial distance between measurement points, the discrete Laplace operator value of each measurement point is calculated using the predicted concentration value and spatial distance of the spatial neighborhood measurement point set, which characterizes the spatial curvature of the concentration.

[0152] Discrete Laplace operator value at the i-th measurement point The discrete Laplacian operator value at the i-th measurement point characterizes the curvature of the concentration space. The calculation formula is:

[0153]

[0154] In the formula, Let be the predicted concentration at the j-th measuring point, specifically the predicted concentration at time t+1; Let be the predicted concentration at the i-th measuring point, specifically the predicted concentration at time t+1;

[0155] S3062. Construct physical constraint loss;

[0156] Based on the predicted concentration change rate and the discrete Laplace operator value, the residual of the diffusion equation is calculated through the diffusion coefficient. The physical constraint loss term is generated by weighted summation using the inverse of the spatial distance, which forces the predicted concentration to conform to the diffusion law.

[0157] Physical constraint loss term L phy The calculation formula is:

[0158]

[0159] In the formula, D represents the predicted concentration change rate at the i-th measuring point; a Let D be the diffusion coefficient. a =0.2, determined by the physical properties of the gas. Summation operation ∑ i Iterate through all measuring points inside the boiler and perform the summation operation ∑ j∈N(i) Iterate through all spatial neighbors of the i-th measurement point.

[0160] Predicted concentration change rate at the i-th measuring point The calculation formula is:

[0161]

[0162] In the formula, The predicted concentration at the i-th measuring point at time t+1; Let be the predicted concentration at the i-th measuring point at time t;

[0163] Purely data-driven concentration prediction models may produce results that violate the fundamental physical laws of gas diffusion, leading to physically unreliable predictions and potentially unreasonable predictions in areas with sparse sensor coverage. S306 employs physical constraints based on the diffusion equation to force the data processing model's predictions to conform to the fundamental physical laws of gas diffusion, thus enhancing the physical reliability of the predictions. The impact of the physical constraint mechanism on the prediction results is verified by obtaining contour maps. Figure 4 As shown in the figure, seven red dots mark the sensor locations, and the background color displays the actual hydrogen sulfide concentration field inside the boiler (unit: parts per million, ppm). The red dashed line represents the prediction result of the unconstrained data processing model. It can be seen that an abnormally high concentration circle appears in the sparse sensor area, violating the diffusion law. The blue solid line represents the prediction result of S306, whose contour lines are basically consistent with the actual field height, especially maintaining a reasonable distribution in the sensorless area in the upper right corner. Its prediction results are all relatively accurate. It can be seen that the physical constraints of the diffusion equation of S306 can effectively prevent the data-driven data processing model from producing non-physical solutions, reduce the prediction error in areas with insufficient sensor coverage, and improve the reliability of the monitoring system.

[0164] To verify the effect of physical constraint regularization on improving the physical rationality of the prediction results, the influence of physical constraints on the prediction results is further analyzed, and the following results are obtained. Figure 5 . Figure 5The horizontal axis represents time (in seconds), and the vertical axis represents carbon monoxide concentration (in parts per million, ppm). The black solid line represents the actual concentration, the magenta dashed line represents the prediction without physical constraints, and the cyan solid line represents the prediction with physical constraints (i.e., the S306 technique). The light red shaded area highlights areas of physical anomalies. It can be seen that in areas of physical anomalies, the prediction without physical constraints exhibits significant fluctuations that do not conform to physical laws, with the magenta curve deviating from the true value. The cyan solid line, however, closely matches the actual concentration. Particularly after approximately 35 seconds, the prediction without physical constraints produces non-physical oscillations, while the prediction with physical constraints maintains stable tracking. This demonstrates that the physical constraint loss term based on the diffusion equation effectively forces the prediction results to conform to the basic laws of gas diffusion, eliminating non-physical explanations in sparse areas of the sensor.

[0165] S4. Train the data processing model to obtain the trained data processing model;

[0166] S401. Integrated total loss function based on physical constraint weight coefficients;

[0167] S4011. Calculate the physical constraint weight coefficients;

[0168] Based on the L2 norm of the gradient vector of the training iteration step and the gradient vector of the prediction loss function and the physical constraint loss term, the physical constraint weight coefficient is calculated through the exponential decay term and the gradient sensitivity adjustment term, and the dynamic change of the weight is controlled by the initial physical constraint weight and the decay period.

[0169] The physical constraint weights γ in the t-th training iteration t The calculation formula is:

[0170]

[0171] In the formula, γ0 is the initial physical constraint weight, which can be set to γ0 = 0.8; t represents the current training iteration step; T decay The decay period can be set as T. decay =100, used to control the rate of decay of the weight exponent; Gradient operator; For the prediction loss function L pred The gradient is in vector form and represents the prediction loss function L. pred The gradient vector relative to all trainable parameters of the data processing model; For the physical constraint loss term L phy The gradient is in vector form and represents the physical constraint loss term L. phy The gradient vector relative to all trainable parameters of the data processing model;

[0172] The term represents the exponentially decaying term, which reduces the strength of physical constraints as the number of iterations increases, thus avoiding overfitting to physical laws and ignoring data features in the later stages.

[0173] The term represents a gradient-sensitive adjustment term, which increases its weight when the gradient of the prediction loss is large, prioritizing the optimization of the main task and ensuring training stability.

[0174] S4012. Integrated total loss function;

[0175] By predicting the loss function L pred With physical constraint loss term L phy The weighted physical constraint loss term, calculated by adding the adaptive weight coefficients, is represented as γ. t L phy The sum of these values ​​represents the total loss function, Loss, which serves as the final optimization objective for training the data processing model.

[0176] S402. Data processing model parameter update and iterative training;

[0177] Based on the total loss function Loss, the backpropagation algorithm is used to calculate the gradient values ​​of all trainable parameters in the data processing model. The Adam optimizer is used to update the network parameters according to the gradient direction. The initial learning rate is set to 0.005, and the batch size is 32 time steps for training samples.

[0178] Each iteration randomly selects a batch of samples from the training set and sequentially performs four operations: forward propagation to calculate the predicted concentration, loss function calculation, backpropagation, and trainable parameter update.

[0179] The training process employs a sliding time window strategy, using calibration concentration data from T+1 consecutive time points as a single input-output sample. The data from the first T time points is input into the data processing model to predict the concentration at time T+1. After every 100 iterations, the performance of the current data processing model is evaluated on the validation set and the loss value is recorded.

[0180] Fixed-weight multi-task loss functions cannot adapt to the differences in the variation patterns of the two losses during training, which may lead to difficulties in training the data processing model, slow convergence, or excessive bias towards satisfying physical constraints, thus reducing prediction accuracy. S402 employs a time-varying adaptive weight mechanism to dynamically balance the contributions of prediction loss and physical constraint loss to the total loss, optimizing the stability of the training process. To verify the feature enhancement capability of multi-scale graph convolution, multi-scale feature extraction effect analysis was performed, yielding... Figure 6 . Figure 6The left image shows the result of the original feature distribution extraction, where sensor feature values ​​(color intensities) exhibit a discrete distribution. The right image shows the features after multi-scale graph convolution processing. Solid lines connect first-order neighbor relationships (distance < 5 meters), and dashed lines connect second-order neighbor relationships. Original isolated feature points (such as the two blue dots in the lower left corner) form feature associations after neighbor aggregation. The feature value range remains the same, but the distribution is smoother and conforms to the physical diffusion law. Experiments demonstrate that multi-scale gated graph convolution can effectively fuse local details and global patterns, providing more representative features for concentration prediction.

[0181] S403. Determining the stopping iteration condition;

[0182] During training, the changes in the total loss function value on the validation set are continuously monitored. Iteration stops when the following dual conditions are met for five consecutive validation set evaluations:

[0183] 1) If the decrease in the total loss value is less than the threshold of 0.01, it indicates that the data processing model has converged;

[0184] 2) Physical constraint loss term L phy The values ​​are stable within a tolerance range of ±0.05, ensuring that the prediction results conform to the physical laws of gas diffusion.

[0185] In addition, if the above conditions are not triggered even after the number of iterations reaches the preset limit of 1000, the training will be forcibly terminated and the current optimal data processing model will be output. The optimal data processing model is selected by choosing the data processing model with the smallest total loss value in the historical iteration process. After each verification evaluation, if the total loss value of the verification set hits a new historical low, a temporary copy of the data processing model will be established as a candidate optimal data processing model.

[0186] S404. Save and output the optimal data processing model to obtain the trained data processing model;

[0187] After training terminates, the data processing model with the minimum total loss of the validation set (i.e., the optimal data processing model) is selected from all candidate data processing model replicas as the final data processing model.

[0188] S5. Collect the concentration data of each combustible gas in the boiler water-cooled wall area through the measuring equipment to obtain new multi-point concentration data; after executing S2, input it into the trained data processing model to obtain the combustible gas prediction concentration matrix (i.e. H2S / CO prediction concentration matrix) of each measuring point at future time; continuously monitor the combustible gas prediction concentration matrix of each measuring point at future time, and when the predicted concentration of any measuring point exceeds the safety threshold, trigger the early warning signal and locate the high-risk area to realize the monitoring of the boiler atmosphere field status and early warning of dangerous situations.

Claims

1. A method for monitoring the state of a boiler atmosphere field, characterized in that, include: S1. Collect the concentration data of each combustible gas in the water-cooled wall area of ​​the boiler using measuring equipment to obtain multi-point concentration data; S2. Perform spatiotemporal calibration on the concentration data from multiple measurement points to obtain the calibrated concentration vector; S3. Construct a data processing model based on graph convolutional networks; S4. Train the data processing model to obtain the trained data processing model; S5. Collect the concentration data of each combustible gas in the boiler water-cooled wall area through the measuring equipment to obtain new multi-point concentration data; after executing S2, input it into the trained data processing model to obtain the combustible gas prediction concentration matrix of each measuring point at future time; continuously monitor the combustible gas prediction concentration matrix of each measuring point at future time, and when the predicted concentration of any measuring point exceeds the safety threshold, trigger the early warning signal and locate the high-risk area.

2. The boiler atmosphere field condition monitoring method according to claim 1, characterized in that, The measuring device is a distributed gas sensor array or a physical probe array.

3. The boiler atmosphere field condition monitoring method according to claim 1, characterized in that, The steps in S2 include: S201. Based on the spatial distance between two measuring points and the Pearson correlation coefficient of the historical concentration sequences of the two measuring points, calculate the elements of the spatial calibration matrix and generate the spatial calibration matrix; S202. Calculate the elements of the time delay matrix using the cross-correlation function, compensate the original concentration by combining the concentration change rate, and use the spatial calibration matrix elements and time delay matrix elements of the spatial neighborhood measurement point set to perform weighted aggregation on the original concentration vector to generate a calibrated concentration vector.

4. The boiler atmosphere field condition monitoring method according to claim 1, characterized in that, The steps in S3 include: S301. Generate a dynamic adjacency matrix based on the scalar gradient value of each gas, and use the dynamic adjacency matrix for data processing; S302. Based on the adjacency matrix at each scale, the feature matrix output by the l-th layer of the graph convolutional network at each time step is obtained, and the feature matrix output by the l-th layer of the graph convolutional network at each time step is used for data processing; S303. Output the spatiotemporal fusion feature matrix based on the attention weights of each measurement point at historical moments; S304. Calculate the predicted concentration matrix for each time point; S305. Obtain the prediction loss function based on the prediction concentration matrix and the prediction concentration spatial gradient matrix; S306. Generate a physical constraint loss term based on the predicted concentration change rate and the discrete Laplace operator value.

5. The boiler atmosphere field condition monitoring method according to claim 4, characterized in that, The steps in S301 include: S3011. Based on the calibration concentration matrix and the spatial location of the measuring points, calculate the gradient vector of each measuring point at different times, and use the scalar concentration values ​​and spatial distances of the spatial neighborhood measuring point set to calculate the scalar gradient value of each gas. S3012. After concatenating the calibration concentration matrix and the concentration space gradient matrix, the matrix is ​​mapped to the latent space through a linear projection matrix. The output matrix is ​​then processed by the hyperbolic tangent activation function, the dynamic adjacency weight matrix, and the dimension transformation matrix. The output matrix is ​​then row-normalized using the Softmax function to generate the dynamic adjacency matrix.

6. The boiler atmosphere field condition monitoring method according to claim 4, characterized in that, The steps in S302 include: S3021. Based on the dynamic adjacency matrix and the k-hop mask matrix, the elements of the multi-scale adjacency matrix are calculated through the Hadamard product to obtain the adjacency matrix at each scale; S3022. Perform normalized graph convolution on the adjacency matrices at each scale, sum them, and then process them through the Sigmoid activation function. Generate a gated signal vector based on the global average pooling of the input feature matrix. After copying it into a matrix, obtain the output feature matrix through Hadamard product modulation.

7. The boiler atmosphere field condition monitoring method according to claim 4, characterized in that, The steps in S303 include: S3031. Based on the feature matrix at the current time and the feature matrix at the historical time, generate the query matrix and the key matrix by querying the projection matrix and the key projection matrix, and calculate the attention weight of each measurement point at the historical time τ using the row vector inner product; S3032. The historical feature matrix is ​​aggregated by weighting attention weights and added to the current feature matrix. The spatiotemporal fusion feature matrix is ​​then output through layer normalization.

8. The boiler atmosphere field condition monitoring method according to claim 4, characterized in that, The steps in S305 include: S3051. Based on the spatiotemporal fusion feature matrix, the predicted concentration matrix and the predicted concentration spatial gradient matrix are output synchronously through feature encoding weight matrix and modified linear unit activation function, and then linear calculation of the prediction weight matrix. S3052. The error between the predicted concentration matrix and the true concentration matrix is ​​constrained by the L2 norm, and the error between the predicted concentration space gradient matrix and the true concentration space gradient matrix is ​​constrained by the L1 norm. The gradient loss weight coefficient balances the two losses to form the prediction loss function.

9. A method for monitoring the state of a boiler atmosphere field according to claim 4, characterized in that, The steps in S306 include: S3061. Based on the predicted concentration and spatial distance between measurement points, the discrete Laplace operator value of each measurement point is calculated using the predicted concentration value and spatial distance of the spatial neighborhood measurement point set; S3062. Based on the predicted concentration change rate and the discrete Laplace operator value, the residual of the diffusion equation is calculated using the diffusion coefficient, and the physical constraint loss term is generated by weighted summation using the inverse of the spatial distance.

10. A method for monitoring the state of a boiler atmosphere field according to claim 1, characterized in that, The steps in S4 include: S401. Integrated total loss function based on physical constraint weight coefficients; S402. Data processing model parameter update and iterative training; S403. Determining the stopping iteration condition; S404. Save and output the optimal data processing model to obtain the trained data processing model.

Images