Boiler burner fault early warning and diagnosis system based on machine learning

Abstract

The invention relates to the technical field of industrial intelligent monitoring, and discloses a boiler burner fault early warning and diagnosis system based on machine learning, and the system comprises a multi-modal data collection module which aligns a real-time flame video with sensor time sequence data; the space-time node construction module extracts features and constructs an initial graph structure containing visual and sensing nodes; the differentiable time-delay topology learning network module executes time-delay graph convolution and performs aggregation to generate a graph-level context vector; the self-supervised predictive coding training module constructs a predictive task and updates parameters by maximizing mutual information; and the manifold geometry fault diagnosis module calculates a reference manifold projection residual error early warning and analyzes the structure change of the adjacent matrix to generate a diagnosis result. According to the method, machine learning is utilized to explicitly model the physical time lag and multi-mode coupling relationship, the problem of false alarm under a dynamic working condition is effectively solved, and accurate early warning and interpretable fault tracing are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent monitoring technology, specifically a boiler burner fault early warning and diagnosis system based on machine learning. Background Technology

[0002] As the core equipment for energy conversion in coal-fired power plants, the combustion stability of boiler burners directly affects the operational safety and economic benefits of the unit. Machine learning, a technology that uses algorithms to automatically summarize patterns and build predictive models from massive amounts of historical data, has been increasingly introduced into the condition monitoring of complex industrial systems in recent years. By mining the massive amounts of data generated during burner operation, operators can be assisted in promptly identifying potential anomalies.

[0003] Existing burner monitoring technologies primarily rely on threshold alarms based on thermal parameters collected by distributed control systems, or on monitoring flame video using industrial television. With technological advancements, some solutions have begun to apply machine learning algorithms, such as using convolutional neural networks to classify flame images to determine combustion status, or using time-series models to predict trends in single-sensor data such as exhaust gas temperature. These methods have improved the automation level of monitoring to some extent.

[0004] However, existing technologies still have significant shortcomings in practical applications. First, the combustion process involves complex coupling of multiple physical fields such as light, heat, and fluid. Single visual or sensor data is insufficient to comprehensively characterize the system state, and simple splicing and fusion often ignores the deep correlations between heterogeneous data. Second, there is an objective physical transmission time lag from pulverized coal ignition to thermal parameter response. Most existing graph neural networks or correlation analysis models assume instantaneous coupling between nodes, making it difficult to effectively handle such dynamic time-delay causal relationships, leading to a high likelihood of false alarms under dynamic operating conditions such as varying loads. Furthermore, it is difficult to obtain sufficient fault label samples in industrial settings, and methods relying on supervised learning are ill-suited for unknown faults. Moreover, traditional models lack interpretable analysis of the root causes of faults.

[0005] Therefore, this invention proposes a boiler burner fault early warning and diagnosis system based on machine learning to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a boiler burner fault early warning and diagnosis system based on machine learning. This system solves the problems of existing technologies, such as difficulty in effectively integrating multimodal heterogeneous data with physical transmission time delays, distortion of causal relationship modeling under dynamic and changing operating conditions, and difficulty in achieving interpretable fault tracing when there are no fault samples.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a boiler burner fault early warning and diagnosis system based on machine learning, comprising: A multimodal data acquisition module is used to connect to a high-temperature industrial camera and a distributed control system at the boiler site to acquire real-time flame video streams and sensor time-series data, and to align the flame video streams and sensor time-series data with timestamps. The spatiotemporal node construction module is used to divide the flame video stream into multiple spatial grids and extract features as visual nodes, and to map the sensor time series data as sensing nodes, thereby constructing an initial graph structure node set containing visual nodes and sensing nodes. Differentiable time-delay topology learning network module is used to perform time-delay graph convolution operation on the initial graph structure node set using graph neural network and perform graph-level feature aggregation to generate a graph-level context vector representing the overall state of the current system. The self-supervised predictive coding training module is used to receive the graph-level context vector and construct a prediction task to predict the features of the sensing nodes at future times. It updates the graph neural network parameters in the differentiable time-delay topology learning network module by backpropagating with the goal of maximizing mutual information. The manifold geometry fault diagnosis module is used to construct a reference manifold under normal operating conditions using the graph neural network parameters updated by backpropagation, calculate the projection residual of the generated graph-level context vector on the reference manifold to generate a fault warning signal, and analyze the structural changes of the adjacency matrix inside the graph neural network to generate fault diagnosis results.

[0008] Preferably, the multimodal data acquisition module performs a timestamp alignment operation based on master-slave clocks, using the acquisition timestamp of each frame of the flame video stream as the main reference, and retrieves the record with the smallest absolute time difference in the sensor timing data; For video frames where no updated data is matched, the multimodal data acquisition module adopts a zero-order hold strategy, reuses the most recent valid sensor value, and outputs data synchronized in the time dimension.

[0009] Preferably, the spatiotemporal node construction module uses a deep convolutional neural network to extract the feature vector of each spatial grid, and defines the feature vector as the visual node; The spatiotemporal node construction module standardizes each physical quantity channel of the sensor time series data and uses a multilayer perceptron to project the scalar data onto the same feature dimension as the visual node, defining the projected feature representation as the sensing node.

[0010] Preferably, the graph neural network in the differentiable time-delay topology learning network module is defined by a set of learnable network parameters, which specifically include adjacency matrix parameters and time-delay matrix parameters; The differentiable time-delay topology learning network module uses the adjacency matrix parameters to generate an adjacency matrix representing the weights of causal connections between nodes, and applies a non-negative activation function to the time-delay matrix parameters to generate a time-delay matrix containing only non-negative values. The elements in the time-delay matrix represent the physical time delay required for node state transmission.

[0011] Preferably, the differentiable time-delay topology learning network module is configured with a feature queue to store historical node feature sequences and to perform differentiable linear interpolation sampling operations; The differentiable linear interpolation sampling operation includes calculating the integer part of the time delay parameter as a discrete index and the fractional part as interpolation coefficients, retrieving the feature vectors of two adjacent time moments from the feature queue using the discrete index, and using the interpolation coefficients to perform a weighted summation of the feature vectors of the two adjacent time moments to obtain the time delay feature vector, and using the time delay feature vector as input to perform the time delay graph convolution operation.

[0012] Preferably, the differentiable time-delay topology learning network module generates the graph-level context vector specifically by: Obtain the node feature representations output after performing the time-delay graph convolution operation through stacked graph convolutional layers; Perform graph-level feature aggregation operation, and use the global average pooling layer to calculate the average of the feature vectors of all nodes to generate a global average feature vector; The global average feature vector is nonlinearly mapped and its dimensions compressed using a multilayer perceptron network consisting of fully connected layers and nonlinear activation functions to obtain the graph-level context vector.

[0013] Preferably, the self-supervised predictive coding training module updates the graph neural network parameters in the differentiable time-delay topology learning network module, specifically including: The feature data within the next K time steps is obtained from the data stream. Aggregation calculation is performed on the feature vectors of all sensor nodes at each future time to generate the real sensor feature vector of the future time as the prediction target. Configure bilinear prediction matrices with different prediction step sizes, receive the graph-level context vector at the current time, and use the bilinear prediction matrix to calculate the similarity score between the graph-level context vector and the real sensing feature vector at the future time. A mini-batch contrastive learning strategy is adopted, which constructs a positive sample pair by combining the graph-level context vector of the same training sample at the current time with the real sensing feature vector at the future time, and constructs a negative sample pair by combining the real sensing feature vectors at the future time of other samples in the batch. The information noise contrastive estimation loss function is calculated based on the similarity score to achieve the goal of maximizing mutual information. The total loss function is obtained by summing the loss functions of all prediction steps. The gradient backpropagation algorithm is then executed based on the total loss function to update the adjacency matrix parameters and time delay matrix parameters that define the graph neural network structure.

[0014] Preferably, the manifold geometry fault diagnosis module constructs a reference manifold for normal operating conditions, specifically including: A sample set is constructed by retrieving graph-level context vectors under multiple normal operating conditions from the historical database, and the covariance matrix of the sample set is calculated. Eigenvalue decomposition is performed on the covariance matrix to extract the eigenvectors corresponding to the first d largest eigenvalues. The eigenvectors are used to construct an orthogonal basis matrix that defines a low-dimensional linear subspace, and the orthogonal basis matrix is ​​used as the reference manifold.

[0015] Preferably, the manifold geometry fault diagnosis module generates fault early warning signals specifically including: The projection reconstruction vector of the real-time generated graph-level context vector onto the reference manifold is calculated using the orthogonal basis matrix; Calculate the difference vector between the real-time generated graph-level context vector and the projection reconstruction vector, and calculate the squared Euclidean norm of the difference vector as the projection residual; The projection residual is compared with a threshold determined based on the statistical distribution of historical normal data. If the projection residual is greater than the threshold, the fault warning signal is generated.

[0016] Preferably, the manifold geometry fault diagnosis module generates fault diagnosis results specifically including: The gradient of the projection residual with respect to the adjacency matrix is ​​calculated using an automatic differentiation engine, and the gradient is then subjected to a Hadamard product operation with the adjacency matrix to obtain a structural contribution matrix that quantifies the degree of contribution of each connecting edge to the anomaly. Extract the index of the associated node pair corresponding to the element with the largest value in the structure contribution matrix; The fault type is determined based on the type of the associated node index: if all the indexes are visual nodes, it is determined to be an abnormal local flow field morphology; if all the indexes are sensor nodes, it is determined to be an abnormal thermal parameter association; if the indexes span both visual and sensor nodes, it is determined to be an abnormal multimodal coupling.

[0017] This invention provides a boiler burner fault early warning and diagnosis system based on machine learning. It has the following beneficial effects: 1. This invention achieves deep fusion of multimodal data from boiler burners by constructing a graph structure containing visual and sensor nodes. The system timestamps the flame video stream acquired by a high-temperature industrial camera with the time-series sensor data from the distributed control system, and uses machine learning techniques to extract spatial grid features and physical quantity features. This method effectively solves the problem of incomplete monitoring information from traditional single data sources, and can capture the complex correlation between flame morphology and thermal parameters, thereby providing a more comprehensive state characterization in the fault early warning stage and significantly improving the accuracy and robustness of the monitoring system.

[0018] 2. This invention introduces a differentiable time-delay topology learning network module to automatically identify the physical time lag relationships between nodes during combustion. Addressing the transmission delay characteristics from fuel combustion to flue gas temperature change within the boiler, the system utilizes a learnable time-delay matrix and linear interpolation sampling operations to explicitly model causal time-series misalignments within a machine learning model. This allows the fault diagnosis model to move beyond instantaneous correlations and accurately capture dynamic topologies that conform to physical laws, effectively avoiding false alarms caused by physical transmission lags and ensuring the timing accuracy of burner condition analysis.

[0019] 3. This invention achieves unsupervised anomaly detection and interpretable root cause localization through self-supervised predictive coding training and manifold geometry analysis. The system constructs a baseline manifold using normal operating data and generates sensitive fault warning signals by calculating projection residuals, without relying on a large number of hard-to-obtain fault samples. Furthermore, by combining an automatic differential engine to analyze the structural changes in the adjacency matrix, it can accurately locate visual regions or sensor parameters causing anomalies, providing intuitive and physically meaningful diagnostic results for boiler burner operation and maintenance, thus solving the problem of the lack of interpretability in traditional black-box models. Attached Figure Description

[0020] Figure 1 This is an architecture diagram of a machine learning-based boiler burner fault early warning and diagnosis system according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the process of multimodal data spatiotemporal graph structured representation according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a fault diagnosis and location process based on manifold geometric projection according to an embodiment of the present invention.

[0021] Among them, 101 is the multimodal data acquisition module; 102 is the spatiotemporal node construction module; 103 is the differentiable time-delay topology learning network module; 104 is the self-supervised predictive coding training module; and 105 is the manifold geometry fault diagnosis module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 This invention provides a machine learning-based boiler burner fault early warning and diagnosis system, including: a multimodal data acquisition module 101, a spatiotemporal node construction module 102, a differentiable time-delay topology learning network module 103, a self-supervised predictive coding training module 104, and a manifold geometry fault diagnosis module 105.

[0024] The multimodal data acquisition module 101 connects to the hardware equipment at the boiler site. It connects to a high-temperature industrial camera installed at the boiler burner's inspection port to acquire real-time flame video streams of the combustion zone. Simultaneously, it connects to the distributed control system (DCS) to acquire sensor time-series data related to the combustion process. The multimodal data acquisition module 101 timestamps the flame video stream and sensor time-series data to ensure synchronization of different modal data over time.

[0025] The spatiotemporal node construction module 102 is connected to the multimodal data acquisition module 101 and receives synchronized data. The spatiotemporal node construction module 102 processes the flame video stream, dividing each frame into multiple non-overlapping spatial grids. It extracts deep features from each spatial grid, defining each grid as an independent visual node. Simultaneously, it processes sensor time-series data, mapping the data from each physical quantity channel to a high-dimensional feature vector, defining each channel as an independent sensing node. The spatiotemporal node construction module 102 combines the visual nodes and sensing nodes to construct an initial graph structure node set containing information from all nodes.

[0026] The differentiable time-delay topology learning network module 103 is connected to the spatiotemporal node construction module 102. The differentiable time-delay topology learning network module 103 receives an initial graph structure node set. The differentiable time-delay topology learning network module 103 internally constructs a graph neural network, which is used to perform time-delay graph convolution operations on the initial graph structure node set and perform graph-level feature aggregation to generate a graph-level context vector representing the overall state of the current system.

[0027] The self-supervised predictive coding training module 104 is connected to the differentiable time-delay topology learning network module 103. The self-supervised predictive coding training module 104 receives graph-level context vectors and constructs a prediction task to predict the features of sensor nodes at future times. The self-supervised predictive coding training module 104 calculates the mutual information between the prediction results and the actual future data, and updates the graph neural network parameters in the differentiable time-delay topology learning network module 103 by backpropagation with the goal of maximizing the mutual information.

[0028] The manifold geometry fault diagnosis module 105 is connected to the differentiable time-delay topology learning network module 103 and the self-supervised predictive coding training module 104. The manifold geometry fault diagnosis module 105 uses backpropagation-updated graph neural network parameters to map historical data under normal operating conditions to a reference manifold in the feature space. The manifold geometry fault diagnosis module 105 calculates the projection residual of the real-time generated graph-level context vector onto the reference manifold. The manifold geometry fault diagnosis module 105 compares the projection residual with a preset comparison threshold; when the projection residual exceeds the comparison threshold, a fault warning signal is generated. Furthermore, the manifold geometry fault diagnosis module 105 analyzes the structural changes of the adjacency matrix within the graph neural network, identifies the associated node pairs that cause abnormal projection residuals, and generates fault diagnosis results indicating the fault type.

[0029] Through the collaborative work of the multimodal data acquisition module 101, the spatiotemporal node construction module 102, the differentiable time-delay topology learning network module 103, the self-supervised predictive coding training module 104, and the manifold geometry fault diagnosis module 105, this system can learn the spatiotemporal coupling law and physical hysteresis characteristics in the normal combustion process, realize timely early warning of early burner faults, and diagnose the cause of faults based on abnormal changes in the physical topology.

[0030] See attached document Figure 1 In order to achieve accurate perception of the operating status of the boiler burner, the multimodal data acquisition module 101 and the spatiotemporal node construction module 102 work together to convert the acquired unstructured flame video stream and structured sensor time-series data into computer-processable graph structure data.

[0031] The multimodal data acquisition module 101 is connected to the hardware equipment at the boiler site. The multimodal data acquisition module 101 is connected to a high-temperature industrial camera installed at the boiler burner viewing port to acquire real-time flame video streams of the combustion zone. Specifically, the high-temperature industrial camera continuously acquires image data of the burner throat region at a fixed frame rate (e.g., 25 frames / second) and a resolution of H×W (e.g., 640×480 pixels) using a sapphire high-temperature resistant lens. The multimodal data acquisition module 101 is also connected to a distributed control system (DCS) to acquire sensor timing data related to the combustion process. The multimodal data acquisition module 101 is configured to read sensor timing data directly associated with the specific burner at a fixed sampling frequency (e.g., 1Hz). The sensor timing data includes at least real-time values ​​of furnace negative pressure, primary air pressure, fuel flow rate, flue gas temperature, and combustion damper opening feedback.

[0032] Due to the inherent rate difference between the high-frequency acquisition characteristics of the flame video stream and the low-frequency acquisition characteristics of the sensor time-series data, the multimodal data acquisition module 101 internally performs a timestamp alignment operation based on the master-slave clock. This alignment operation uses the acquisition timestamp of each frame of the flame video stream. Using the primary reference, retrieve the timestamp from the received sensor data buffer queue. Seeking satisfaction And the sensor data record with the smallest absolute time difference, among which The preset time tolerance threshold is used (e.g., 500 milliseconds). For video frame moments where no updated data is matched within the tolerance range, the multimodal data acquisition module 101 adopts a zero-order hold strategy, that is, reuses the most recent valid sensor value, thereby outputting a multimodal dataset that is strictly synchronized in the time dimension.

[0033] The spatiotemporal node construction module 102 is connected to the multimodal data acquisition module 101 and receives synchronized data. The spatiotemporal node construction module 102 processes the flame video stream, dividing each frame into multiple non-overlapping spatial grids. Specifically, for any input video frame at time t... (Where G is the number of color channels), the spatiotemporal node construction module 102 uniformly divides it into P×P rectangular regions on the spatial plane, each region representing a local physical space of the combustion field (such as the flame root or edge backflow zone). Subsequently, the spatiotemporal node construction module 102 extracts the deep features of each spatial grid, defining each spatial grid as an independent visual node. This extraction process utilizes a deep convolutional neural network to process the image patches corresponding to each spatial grid. The calculation is performed using the following formula: ; Where i represents the grid index ( ), here (This represents the total number of spatial grids generated after dividing the image). This represents the convolution operation; The kernel weight matrix; For bias terms; This indicates a batch normalization operation; This represents a non-linear activation function (such as ReLU); the output This is a D-dimensional feature vector. The spatiotemporal node construction module 102 uses this feature vector... It is marked as the i-th visual node.

[0034] Simultaneously, the spatiotemporal node construction module 102 processes the sensor time-series data, mapping the data of each physical quantity channel to a high-dimensional feature vector, and defining each physical quantity channel as an independent sensing node. Specifically, suppose the sensor data at time t contains Q physical quantity channels, and the original numerical vector is... The simultaneous spatiotemporal node construction module 102 first performs Z-Score normalization: ;in, , To calculate the mean and standard deviation, a multilayer perceptron (MLP) is then used to process the standardized scalar values. Project onto the same feature dimension as the visual node: ; Where j represents the sensor channel index ( ); These are learnable mapping weights; This is the bias vector. (Generated) This is the feature representation of the j-th sensing node.

[0035] Finally, the spatiotemporal node construction module 102 combines visual nodes and sensor nodes to construct an initial graph structure node set containing information about all nodes. The initial graph structure node set refers to the set of graph vertices composed of all P² visual nodes and Q sensor nodes. Physically, it unifies the local spatial distribution information of the visual modality with the macroscopic physical quantity information of the sensor modality. In terms of data structure, it is represented by a single-dimensional... The characteristic matrix of .

[0036] See attached document Figure 1 and Figure 2 The differentiable time-delay topology learning network module 103 is connected to the spatiotemporal node construction module 102, and receives the initial graph structure node set at time t. This node set is represented as a feature matrix. Where M is the total number of nodes and D is the feature dimension. To support the tracing of historical states, the differentiable time-delay topology learning network module 103 is internally configured with a fixed-length first-in-first-out (FIFO) feature queue. This queue stores features from time step [time value missing]. Historical node feature sequence up to time t ,in This represents the maximum physical time lag allowed by the system. Whenever new data arrives, the earliest data at the front of the queue is removed, and the current data is pushed to the end of the queue.

[0037] The differentiable time-delay topology learning network module 103 internally constructs a graph neural network. The topology of this graph neural network is not fixed, but defined by a set of learnable network parameters, specifically including the adjacency matrix parameters. and time delay matrix parameters To transform unconstrained parameters into physically meaningful computational matrices, the differentiable time-delay topology learning network module 103 performs the following generation operation: Using adjacency matrix parameters Generate adjacency matrix A. Specifically, the adjacency matrix... elements in As a scalar, it numerically represents the causal connection weight of node j to node i; Using time delay matrix parameters Generate time delay matrix Specifically, to conform to the unidirectional flow characteristic of physical time, 103 differentiable time-delay topology learning network modules are used. Apply a non-negative activation function (such as ReLU) to generate the time delay matrix. Time delay matrix Each element in the equation is defined as a time delay parameter in scalar form. The non-negative activation function ensures The time delay parameter Numerically represent the physical time lag required for the state of node j to be transmitted to node i.

[0038] Due to time delay parameter Since the variables are continuous, while the data in the feature queue is stored in discrete time steps, direct indexing would lead to non-differentiability issues. Therefore, the differentiable time-delay topology learning network module 103 performs differentiable linear interpolation sampling operations. For any pair of related nodes (i,j), in order to obtain the value of node j at the delay time... The characteristic state, the differentiable time-delay topology learning network module 103 first calculates the discrete index. and interpolation coefficients : ; in, This indicates the floor function; The integer part; This is the decimal part. Then, the differentiable time-delay topology learning network module 103 retrieves the time from the feature queue. and time The feature vectors, using interpolation coefficients Calculate the time-delay eigenvector of node j : ; In the above formula, This indicates that node j stored in the feature queue represents the historical moment. The characteristic record. The time delay parameter is obtained through this formula. Included in interpolation coefficients The process involves computation, allowing the gradient of the loss function to be passed through... Backpropagation to This enables automatic learning of physical delay parameters.

[0039] After sampling the time-delay features of all node pairs, the differentiable time-delay topology learning network module 103 performs time-delay graph convolution operations. The graph neural network consists of L pre-configured graph convolutional layers stacked together, where L is an integer greater than or equal to 1 (e.g., L=3), representing the total number of layers in the network. For the l-th layer of the network... The feature update formula for node i is as follows: ; in, This represents the node output by the l-th layer; It is the time delay characteristic of node j obtained from the output of the previous layer and calculated using the aforementioned interpolation formula; It is the learnable linear transformation weight matrix of the l-th layer, used to project the feature dimension; It is the bias vector; It is a non-linear activation function. This operation enables information exchange between nodes based on physical time-delay topology.

[0040] After L layers of stacked time-delay graph convolution operations, the differentiable time-delay topology learning network module 103 obtains the final node feature representation. To transform discrete node-level features into a single vector representing the current overall state of the system, the differentiable time-delay topology learning network module 103 performs a graph-level feature aggregation operation. This operation is specifically calculated using a global average pooling layer cascaded multilayer perceptron, with the following formula: ; in, This indicates that the element-wise summation operation is performed on the feature vectors of all M nodes; This indicates that the summation result is averaged to generate a global average feature vector; This represents a multilayer perceptron network consisting of fully connected layers and nonlinear activation functions, used for nonlinear mapping and dimensionality compression of the globally averaged feature vector. The final calculated output... Defined as a graph-level context vector, this vector aggregates the visual morphological features and sensor parameter features of the entire burner system and encodes the time-delay causal dependencies between them.

[0041] See attached document Figure 1 The self-supervised predictive coding training module 104 establishes a data connection with the differentiable time-delay topology learning network module 103. The self-supervised predictive coding training module 104 receives the graph-level context vector at time t generated by the differentiable time-delay topology learning network module 103. The graph-level context vector The spatiotemporal features of all visual and sensor nodes at time t are aggregated. Based on this input, a self-supervised predictive coding training module 104 is used to construct a prediction task to predict the features of sensor nodes at future times. The purpose of this task is to use the flame morphology (visual information) and operating parameters (sensor information) observed at the current time to infer the changing trends of boiler thermal parameters (sensor information) at future times.

[0042] To construct the supervision signal (i.e., the true value) for this prediction task, the self-supervised predictive coding training module 104 obtains feature data for the next K time steps from the data stream (i.e., the time series set containing feature vectors of consecutive time points, which is output in real time and cached in chronological order by the aforementioned spatiotemporal node construction module 102). For the k-th future time step ( The self-supervised predictive encoding training module 104 first obtains the deep feature vectors of all sensor nodes at that time step. Let the set of sensor node indices be... It contains Q sensing nodes. The self-supervised predictive encoding training module 104 aggregates and calculates the feature vectors of the sensing nodes at each future time (i.e., time t+k) to generate the true sensing feature vectors for the future time. : ; in, Let be the feature vector extracted by the graph neural network for the j-th sensor node at time t+k. This step explicitly limits the prediction target to the aggregate state of the sensor nodes, ensuring that the model focuses on learning physical features that can explain the evolution of thermal parameters.

[0043] The self-supervised predictive encoding training module 104 utilizes the graph-level context vector at time t. The true sensing feature vector at future time t+k Make predictions. The self-supervised prediction encoding training module 104 is configured with bilinear prediction matrices corresponding to different prediction step sizes. (That is, it is internally configured with K independent bilinear prediction matrices). For a prediction step size k, the self-supervised predictive coding training module 104 receives the graph-level context vector at the current time step and uses the bilinear prediction matrix to calculate the similarity score between the graph-level context vector and the real sensing feature vector at the future time step. (i.e., the non-normalized logarithmic density ratio): ; in, Indicates vector transpose; This represents matrix multiplication. The formula quantifies the current context. With future real-time sensing feature vectors Consistency in the feature space.

[0044] The self-supervised predictive coding training module 104 measures prediction quality by calculating the Information Noise Contrast Estimation (InfoNCE) loss function. This self-supervised predictive coding training module 104 employs a mini-batch contrastive learning strategy to maximize mutual information. Let the training batch size be N, for the i-th sample in the batch ( Its context vector at time t is The true sensing feature vector at time t+k is At this point, the self-supervised predictive encoding training module 104 constructs a positive sample pair (i.e., the graph-level context vector of the same training sample at the current time and the real sensing feature vector at the future time) using the graph-level context vector of the same training sample at the current time. ), and construct negative sample pairs from the future real sensing feature vectors of other samples in the batch (i.e. , With the current context The self-supervised predictive encoding training module 104 calculates the information noise contrastive estimation loss function based on similarity scores. : ; The self-supervised predictive coding training module 104 sums the loss functions for all K prediction steps to obtain the total loss function. .

[0045] The self-supervised predictive encoding training module 104 is based on the total loss function. The gradient backpropagation algorithm is executed (here, the backpropagation algorithm can be an existing standard algorithm based on the chain rule to calculate the gradient, its function is to pass the error signal of the output layer back to the input layer layer by layer). The self-supervised predictive coding training module 104 calculates the gradient of the loss function with respect to the parameters of the preceding graph neural network and updates the graph neural network parameters in the differentiable time-delay topology learning network module 103 using an optimizer (such as Adam). The updated graph neural network parameters specifically include the adjacency matrix parameters used to define the graph structure. and the time delay matrix parameters used to define physical delay characteristics Through this training mechanism, the adjacency matrix A converges to a state that reflects the true physical coupling strength between nodes, while the time delay matrix... The value converges to reflect the actual physical transmission delay of the combustion reaction process (such as the lag time from pulverized coal combustion to flue gas temperature change), thereby enabling automatic identification of the dynamic topology of the system.

[0046] See attached document Figure 1 and Figure 3 The manifold geometry fault diagnosis module 105 is connected to the self-supervised predictive coding training module 104. During the initialization phase before system deployment, the manifold geometry fault diagnosis module 105 uses the graph neural network parameters updated via backpropagation from the self-supervised predictive coding training module 104 to map historical data under normal operating conditions to a baseline manifold in the feature space. The baseline manifold refers to a low-dimensional linear subspace spanned by graph-level context vectors under normal combustion conditions in the multi-dimensional feature space; it defines the statistical distribution boundary of the system when it is operating stably and conforms to physical laws.

[0047] Specifically, the manifold geometry fault diagnosis module 105 retrieves multiple graph-level context vectors under normal operating conditions from a historical database (i.e., a storage medium that stores graph-level context vectors generated during long-term stable operation of the system and processed by graph neural networks). A set of samples is composed of graph-level context vectors containing only the normal combustion state. The manifold geometry fault diagnosis module 105 calculates the covariance matrix of the sample set. : ; in, This represents the total number of historical normal samples involved in the calculation; Represents the graph-level context vector of the i-th historical normal sample; This represents the mean vector of all historical normal samples; This represents the matrix transpose operation. Subsequently, the manifold geometry fault diagnosis module 105 performs a covariance matrix operation. Perform eigenvalue decomposition to extract the eigenvectors of the top d largest eigenvalues ​​(i.e., the preset number of principal components or the dimension of the target low-dimensional subspace). Use these eigenvectors to construct an orthogonal basis matrix defining the low-dimensional linear subspace. The orthogonal basis matrix is ​​used as the reference manifold (the orthogonal basis matrix is ​​the mathematical expression of the reference manifold, which defines the statistical distribution boundary of the system when it is running stably and conforms to physical laws).

[0048] During the real-time monitoring phase, the manifold geometry fault diagnosis module 105 receives graph-level context vectors generated in real time by the differentiable time-delay topology learning network module 103. To quantify the degree to which the current state deviates from the normal state, the manifold geometry fault diagnosis module 105 calculates the projection residual of the real-time generated graph-level context vector onto the reference manifold.

[0049] The specific calculation process is as follows: The manifold geometry fault diagnosis module 105 first utilizes the orthogonal basis matrix... Calculate the real-time generated graph-level context vector Projection reconstruction vector on the reference manifold : ; Next, the manifold geometry fault diagnosis module 105 calculates the difference vector between the original vector (i.e., the real-time generated graph-level context vector) and the projection reconstruction vector to obtain the projection residual vector. : ; Subsequently, the manifold geometry fault diagnosis module 105 calculates the projected residual vector. The squared Euclidean norm is defined as the value of the projected residual. (i.e., abnormal score): ; Manifold geometry fault diagnosis module 105 compares projected residuals Compare with a preset comparison threshold. (Preset comparison threshold) These are numerical limits determined based on the statistical characteristics of the projected residual distribution of historical normal data (e.g., taking the 99th quantile of the residual distribution of historical normal data or a confidence boundary obtained from kernel density estimation). When an anomaly occurs, the system is identified and a fault warning signal is generated.

[0050] After a fault warning signal is generated, the manifold geometry fault diagnosis module 105 executes manifold projection decomposition technology to identify the specific physical source of the fault. This manifold projection decomposition technology identifies the root cause of the anomaly by analyzing the structural changes in the adjacency matrix within the graph neural network. Specifically, the adjacency matrix within the graph neural network refers to the adjacency matrix A located within the differentiable time-delay topology learning network module 103, which defines the information transmission path of the graph neural network. Analyzing structural changes involves calculating gradients to analyze which connection weights in the adjacency matrix contribute significantly to the current projection residual.

[0051] The manifold geometry fault diagnosis module 105 uses an automatic differentiation engine to calculate the gradient of the projected residual with respect to the adjacency matrix, and performs a Hadamard product operation on the gradient and the adjacency matrix to obtain a structural contribution matrix that quantifies the contribution of each connecting edge to the anomaly. The calculation formula is as follows: ; in, It is the projection residual. The gradient relative to the context vector (using the previously calculated projection residual vector) ); This term represents the Jacobian matrix of the context vector relative to the adjacency matrix A, and is obtained by backpropagation in the frozen graph neural network computation graph through an automatic differentiation engine. This represents the Hadamard product (element-wise multiplication). The physical meaning of this formula is: using the chain rule to backfeed abnormal deviations at the output to the topology at the input, quantifying the connection weights of node j pointing to node i in the adjacency matrix A for the current fault score. Sensitivity.

[0052] elements in The larger the value, the more likely the association between node j and node i is to be the main reason for the current system's deviation from the baseline manifold, i.e., identifying the associated node pairs that cause abnormal projection residuals. Finally, the manifold geometry fault diagnosis module 105 contributes to the structural contribution matrix. Perform numerical sorting and extract the previous values. The index of the node corresponding to the maximum value And generate fault diagnosis results indicating the fault type based on the node type: If index All are visual nodes (corresponding to image grids), and the diagnostic results are generated as local flow field morphology anomalies, which are located to specific image grid coordinates; If index All of these are sensor nodes (corresponding to physical quantities), and the diagnostic results are generated as abnormal correlations of thermal parameters. If index Across vision and sensing nodes, the diagnostic results are generated as multimodal coupling anomalies.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A boiler burner fault early warning and diagnosis system based on machine learning, characterized in that, include: A multimodal data acquisition module is used to connect to a high-temperature industrial camera and a distributed control system at the boiler site to acquire real-time flame video streams and sensor time-series data, and to align the flame video streams and sensor time-series data with timestamps. The spatiotemporal node construction module is used to divide the flame video stream into multiple spatial grids and extract features as visual nodes, and to map the sensor time series data as sensing nodes, thereby constructing an initial graph structure node set containing visual nodes and sensing nodes. Differentiable time-delay topology learning network module is used to perform time-delay graph convolution operation on the initial graph structure node set using graph neural network and perform graph-level feature aggregation to generate a graph-level context vector representing the overall state of the current system. The self-supervised predictive coding training module is used to receive the graph-level context vector and construct a prediction task to predict the features of the sensing nodes at future times. It updates the graph neural network parameters in the differentiable time-delay topology learning network module by backpropagating with the goal of maximizing mutual information. The manifold geometry fault diagnosis module is used to construct a reference manifold under normal operating conditions using the graph neural network parameters updated by backpropagation, calculate the projection residual of the generated graph-level context vector on the reference manifold to generate a fault warning signal, and analyze the structural changes of the adjacency matrix inside the graph neural network to generate fault diagnosis results.

2. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The multimodal data acquisition module performs a timestamp alignment operation based on the master-slave clock, using the acquisition timestamp of each frame of the flame video stream as the main reference, and retrieves the record with the smallest absolute time difference in the sensor timing data. For video frames where no updated data is matched, the multimodal data acquisition module adopts a zero-order hold strategy, reuses the most recent valid sensor value, and outputs data synchronized in the time dimension.

3. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The spatiotemporal node construction module uses a deep convolutional neural network to extract the feature vector of each spatial grid, and defines the feature vector as the visual node; The spatiotemporal node construction module standardizes each physical quantity channel of the sensor time series data and uses a multilayer perceptron to project the scalar data onto the same feature dimension as the visual node, defining the projected feature representation as the sensing node.

4. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The graph neural network in the differentiable time-delay topology learning network module is defined by a set of learnable network parameters, which specifically include adjacency matrix parameters and time-delay matrix parameters. The differentiable time-delay topology learning network module uses the adjacency matrix parameters to generate an adjacency matrix representing the weights of causal connections between nodes, and applies a non-negative activation function to the time-delay matrix parameters to generate a time-delay matrix containing only non-negative values. The elements in the time-delay matrix represent the physical time delay required for node state transmission.

5. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 4, characterized in that, The differentiable time-delay topology learning network module is configured with a feature queue to store historical node feature sequences and performs differentiable linear interpolation sampling operations. The differentiable linear interpolation sampling operation includes calculating the integer part of the time delay parameter as a discrete index and the fractional part as interpolation coefficients, retrieving the feature vectors of two adjacent time moments from the feature queue using the discrete index, and using the interpolation coefficients to perform a weighted summation of the feature vectors of the two adjacent time moments to obtain the time delay feature vector, and using the time delay feature vector as input to perform the time delay graph convolution operation.

6. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The differentiable time-delay topology learning network module generates the graph-level context vector specifically by: Obtain the node feature representations output after performing the time-delay graph convolution operation through stacked graph convolutional layers; Perform graph-level feature aggregation operation, and use the global average pooling layer to calculate the average of the feature vectors of all nodes to generate a global average feature vector; The global average feature vector is nonlinearly mapped and its dimensions compressed using a multilayer perceptron network consisting of fully connected layers and nonlinear activation functions to obtain the graph-level context vector.

7. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The self-supervised predictive coding training module updates the graph neural network parameters in the differentiable time-delay topology learning network module, specifically including: The feature data within the next K time steps is obtained from the data stream. Aggregation calculation is performed on the feature vectors of all sensor nodes at each future time to generate the real sensor feature vector of the future time as the prediction target. Configure bilinear prediction matrices with different prediction step sizes, receive the graph-level context vector at the current time, and use the bilinear prediction matrix to calculate the similarity score between the graph-level context vector and the real sensing feature vector at the future time. A mini-batch contrastive learning strategy is adopted, which constructs a positive sample pair by combining the graph-level context vector of the same training sample at the current time with the real sensing feature vector at the future time, and constructs a negative sample pair by combining the real sensing feature vectors at the future time of other samples in the batch. The information noise contrastive estimation loss function is calculated based on the similarity score to achieve the goal of maximizing mutual information. The total loss function is obtained by summing the loss functions of all prediction steps. The gradient backpropagation algorithm is then executed based on the total loss function to update the adjacency matrix parameters and time delay matrix parameters that define the graph neural network structure.

8. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The manifold geometry fault diagnosis module constructs a baseline manifold for normal operating conditions, specifically including: A sample set is constructed by retrieving graph-level context vectors under multiple normal operating conditions from the historical database, and the covariance matrix of the sample set is calculated. Eigenvalue decomposition is performed on the covariance matrix to extract the eigenvectors corresponding to the first d largest eigenvalues. The eigenvectors are used to construct an orthogonal basis matrix that defines a low-dimensional linear subspace, and the orthogonal basis matrix is ​​used as the reference manifold.

9. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 8, characterized in that, The manifold geometry fault diagnosis module generates fault early warning signals specifically including: The projection reconstruction vector of the real-time generated graph-level context vector onto the reference manifold is calculated using the orthogonal basis matrix; Calculate the difference vector between the real-time generated graph-level context vector and the projection reconstruction vector, and calculate the squared Euclidean norm of the difference vector as the projection residual; The projection residual is compared with a threshold determined based on the statistical distribution of historical normal data. If the projection residual is greater than the threshold, the fault warning signal is generated.

10. The machine learning-based boiler burner fault early warning and diagnosis system according to claim 1, characterized in that, The fault diagnosis results generated by the manifold geometry fault diagnosis module specifically include: The gradient of the projection residual with respect to the adjacency matrix is ​​calculated using an automatic differentiation engine, and the gradient is then subjected to a Hadamard product operation with the adjacency matrix to obtain a structural contribution matrix that quantifies the degree of contribution of each connecting edge to the anomaly. Extract the index of the associated node pair corresponding to the element with the largest value in the structure contribution matrix; The fault type is determined based on the type of the associated node index: if all the indexes are visual nodes, it is determined to be an abnormal local flow field morphology; if all the indexes are sensor nodes, it is determined to be an abnormal thermal parameter association; if the indexes span both visual and sensor nodes, it is determined to be an abnormal multimodal coupling.

Images