Power plant abnormality early warning method and system fusing multi-source time sequence sensing and spatial feature coordination

Abstract

The application relates to the technical field of power plant intelligent early warning, and discloses a power plant abnormal early warning method and system fusing multi-source time sequence sensing and spatial feature cooperation, which comprises the following steps: carrying out adaptive normalization of multi-source sensor data based on dynamic statistics; extracting multi-resolution time sequence features by using discrete wavelet transform and multi-branch Transformer, and carrying out weighted fusion through a gating network; combining process topology and data driving to construct a dynamic graph structure, and utilizing two-stage graph convolution to aggregate high-order spatial correlations among devices; carrying out deep nonlinear interaction of space-time features in a tensor space through Tucker decomposition; quantifying classification uncertainty based on an evidence network, and generating early warning actions in combination with an inverse reinforcement learning strategy. The application effectively solves the problem of multi-scale non-stationary and spatial strong coupling of power plant data, and improves the sensitivity and decision reliability of abnormal early warning by fusing expert strategies and uncertainty measurement.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology for power plants, specifically to a method and system for early warning of power plant anomalies that integrates multi-source temporal sensing and spatial feature coordination. Background Technology

[0002] As power plants play an increasingly prominent role in the modern energy supply system, their operational safety is directly related to public safety and the stability of the ecological environment. To ensure the stable operation of power generation equipment, modern power plants deploy digital control systems (DCS) to continuously collect data from multiple sensor sources (such as pressure, temperature, and flow rate) for real-time monitoring of equipment status. Analyzing this time-series data allows for timely identification of whether equipment is operating normally or facing potential risks, which is crucial for preventing serious safety accidents caused by delayed condition assessment.

[0003] Although artificial intelligence technology has been gradually applied to the field of power plant anomaly monitoring and has achieved automated identification of operating conditions to a certain extent, existing technologies still have limitations in the face of the complex operating environment and multi-source heterogeneous data of power plants. First, power plant sensor data generally exhibits multi-scale and non-stationary characteristics, with both long-term trends and instantaneous pulses. Existing monitoring models focus on feature extraction at a single scale, making it difficult to simultaneously capture slow wear trends and sudden transient disturbances, resulting in inaccurate feature representation and easy omission of weak early fault signals. Second, power plant systems have numerous devices with complex physical connections, and there are implicit and dynamically changing high-order functional relationships between devices. However, existing methods treat each sensor as an independent node or rely solely on static modeling based on a fixed physical topology, lacking effective mining of the dynamic collaborative relationships between devices, resulting in insufficient robustness of the model when facing complex system-level faults. Furthermore, at the decision-making and fusion level, existing feature fusion methods are relatively simple and cannot fully explore the deep interaction information between spatiotemporal features. Moreover, classification models based on traditional probability output (such as Softmax) cannot quantify the cognitive uncertainty in the reasoning process. When encountering new abnormal working conditions not included in the training set, the model will give unreliable high-confidence predictions. In addition, the lack of integration of expert risk preferences and decision-making logic results in a high false alarm rate in the system in practical applications, making it difficult to meet the high reliability requirements of safety supervision.

[0004] In summary, existing technologies still have significant shortcomings in accurately representing multi-scale temporal characteristics, effectively modeling dynamic spatial relationships between equipment, and quantifying the uncertainty of novel abnormal operating conditions. Therefore, there is an urgent need for a power plant anomaly early warning scheme that can integrate multi-source temporal perception with spatial feature coordination to achieve comprehensive perception and accurate decision-making for complex operating conditions, and provide reliable technical support for the safe operation of power plants. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a power plant anomaly early warning method and system that integrates multi-source temporal sensing and spatial feature synergy. This solves the problems of existing technologies, which struggle to simultaneously extract multi-scale non-stationary temporal features and model dynamic spatial relationships between equipment when dealing with multi-source heterogeneous data from power plants. Furthermore, the lack of deep feature interaction and uncertainty quantification in early warning decisions leads to low accuracy in operating condition identification, high false alarm rate, and weak generalization ability.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a power plant anomaly early warning method that integrates multi-source temporal sensing and spatial feature coordination, comprising the following steps: The data preprocessing module collects raw time-series data from multiple sources, performs time alignment, and adaptively normalizes the raw time-series data based on dynamic statistics, outputting a normalized data vector. The multi-source time-series sensing module performs multi-resolution decomposition on the normalized data vector, extracts the time-series features of each resolution component, performs weighted fusion, and outputs a time-series feature vector. The spatial feature collaboration module generates a dynamic graph structure based on the normalized data vector and historical data window, and aggregates the correlation information between devices based on the dynamic graph structure to output spatial feature vectors. The feature fusion module maps temporal feature vectors and spatial feature vectors to tensor space for interactive fusion, generating the final fused feature matrix; The early warning decision module calculates the expected probability of categories and the overall uncertainty index based on the final fused feature matrix, and generates early warning actions or silent actions in combination with the decision strategy.

[0008] Preferably, considering the inconsistent sampling frequencies and non-stationary distribution of industrial field data, the sampling delay between multiple sensors is eliminated during the data preprocessing stage by truncating the timestamps of the original time-series data to a uniform time granularity. Furthermore, unlike static mean-variance calculations, this method uses a continuously weighted exponential moving average algorithm to iteratively update the mean and variance. This approach can track the drift trend of data distribution in real time, using dynamically updated statistics to standardize and project the original data, ensuring the consistency of the input data distribution over long time spans and improving the model's adaptability to changes in operating conditions.

[0009] Preferably, to address the issue that a single scale cannot simultaneously capture transient faults and long-term trends, the multi-source time-series sensing module introduces a time-frequency analysis mechanism. First, discrete wavelet transform is used to decompose the data into orthogonal frequency components, separating low-frequency trend terms from high-frequency noise or abrupt changes. Then, multiple parallel Transformer encoder branches are constructed, each independently processing its respective frequency component. The Transformer architecture effectively captures long-distance time dependencies through a self-attention mechanism, while the frequency division strategy avoids interference from high-frequency noise in the extraction of low-frequency trend features. Finally, a gating network is used to calculate dynamic weights, performing point-by-point weighted summation of the independent encoded features of each component to achieve adaptive reconstruction of multi-resolution features.

[0010] Preferably, to address the complex physical connections and implicit coupling relationships among power plant equipment, the spatial feature collaboration module constructs a dynamic graph structure that integrates physical priors and data-driven approaches. On one hand, physical connections are initialized based on the process topology logic of the power plant equipment; on the other hand, a recurrent neural network is used to process historical data sequences, generating a learned correlation matrix reflecting the real-time operating status through nonlinear transformations of hidden states. After fusing and sparsifying these two approaches, the resulting dynamic graph structure contains both inherent physical connection constraints and reflects the dynamic correlations during operation. On this graph structure, a two-stage aggregation mechanism—from nodes to hyperedges and from hyperedges to nodes—is employed to simulate the interaction processes within and between subsystems, thereby extracting spatial features containing high-order collaborative relationships.

[0011] Preferably, in the multimodal feature fusion stage, to avoid simple concatenation operations destroying the intrinsic structure of spatiotemporal features, the feature fusion module employs a tensor interaction technique based on Tucker decomposition. High-dimensional features are mapped to a low-dimensional core space using temporal and spatial projection matrices, and multilinear product operations are performed using the core tensor. This process compresses the number of parameters while achieving deep nonlinear interaction between temporal and spatial features. Furthermore, a residual correction path is introduced, adding the linearly transformed features obtained by concatenating the original temporal and spatial features to the nonlinear features obtained through tensor interaction. Layer normalization effectively preserves complementary information in the original feature stream, preventing feature degradation in deep networks.

[0012] Preferably, to improve the reliability and interpretability of early warning decisions, the early warning decision module abandons the traditional point estimation output and instead adopts an evidence-based learning framework. By mapping the feature matrix to Dirichlet distribution parameters, it directly calculates the expected probability of each category and the overall uncertainty index, thus providing a high uncertainty warning when the model's confidence in judging the current operating condition is insufficient. Simultaneously, combined with maximum entropy inverse reinforcement learning technology, the implicit reward function is recovered from the expert's historical operational data. This allows the policy network to not only rely on data-driven probabilistic judgments when generating early warning actions but also to reproduce the expert's risk preferences and decision-making logic, effectively reducing the false alarm rate.

[0013] A second aspect of the present invention provides a power plant anomaly early warning system that integrates multi-source temporal sensing and spatial feature coordination, comprising: The data preprocessing module is used to collect raw time-series data from multiple sources, perform time alignment operations, and perform adaptive normalization processing on the raw time-series data based on dynamic statistics, outputting a normalized data vector. The multi-source time-series sensing module is used to decompose the normalized data vector into multiple resolutions, extract the time-series features of each resolution component, perform weighted fusion, and output a time-series feature vector. The spatial feature collaboration module is used to generate a dynamic graph structure based on the normalized data vector and historical data window, and aggregate the correlation information between devices based on the dynamic graph structure to output a spatial feature vector. The feature fusion module is used to map temporal feature vectors and spatial feature vectors to tensor space for interactive fusion, generating the final fused feature matrix. The early warning decision module is used to calculate the category probability and uncertainty index based on the final fused feature matrix, and generate early warning actions or silent actions in combination with decision-making strategies.

[0014] This invention provides a method and system for early warning of power plant anomalies that integrates multi-source temporal sensing and spatial feature coordination. It has the following beneficial effects: 1. This invention addresses the challenge of multi-scale and non-stationary characteristics of power plant time series data, which involve both long-term trends and instantaneous pulses, by employing a three-layer discrete wavelet transform decomposition and gated reconstruction network. This mechanism can dynamically allocate frequency band weights based on signal characteristics to focus on key features, avoiding the omission of weak anomalous signals caused by traditional single-scale analysis, and providing accurate time series feature support for the early identification of abnormal operating conditions.

[0015] 2. This invention designs a dynamic graph construction and two-stage graph convolution aggregation mechanism. By combining the process topology logic of power plant equipment with dynamic data similarity to collaboratively define the graph structure, it updates in real time according to operating conditions. This approach effectively captures the implicit many-to-many high-order functional coupling relationships between equipment, reduces the risk of misjudgment caused by insufficient spatial association modeling or rigid physical connection definitions, and improves the model's ability to perceive system-level anomalies.

[0016] 3. This invention utilizes low-rank tensor decomposition technology to achieve deep nonlinear fusion of temporal and spatial features, avoiding feature redundancy caused by simple splicing; further, by combining evidence networks and maximum entropy inverse reinforcement learning technology, it can not only effectively detect unknown new anomalies by quantifying the uncertainty of reasoning, but also integrate expert risk preferences into decision-making logic, thereby providing interpretable and reliable decision support for the safe operation of power plants while ensuring high accuracy. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature coordination according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the dynamic graph construction and aggregation process of the spatial feature collaboration module in an embodiment of the present invention. Figure 3 This is a schematic diagram of the tensor interaction principle of the feature fusion module in an embodiment of the present invention; Figure 4 This is a block diagram of the power plant anomaly early warning system integrating multi-source temporal sensing and spatial feature coordination according to an embodiment of the present invention. Figure 5 This is a comparison chart of the fault diagnosis performance indicators of different models in embodiments of the present invention; Figure 6 This is a schematic diagram of the dynamic response curve of the uncertainty index of the evidence network in an embodiment of the present invention.

[0018] The module includes: 10. Data preprocessing module; 20. Multi-source temporal sensing module; 30. Spatial feature collaboration module; 40. Feature fusion module; and 50. Early warning decision module. Detailed Implementation

[0019] 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.

[0020] See attached document Figure 4The present invention provides a power plant anomaly early warning system that integrates multi-source temporal sensing and spatial feature coordination. The system includes: a data preprocessing module 10, a multi-source temporal sensing module 20, a spatial feature coordination module 30, a feature fusion module 40, and an early warning decision module 50.

[0021] The data preprocessing module 10 is configured to communicate with the power plant's distributed control system (DCS) to receive raw time-series data from multiple sources. The data preprocessing module 10 performs timestamp truncation and alignment to eliminate sampling delay, calculates dynamic statistics using a continuously weighted exponential moving average algorithm, adaptively normalizes the raw time-series data, and outputs a preprocessed normalized data vector. ,in, The time is indicated. For the DCS data access interface and basic cleaning logic, those skilled in the art can use existing data acquisition cards or OPC protocol interfaces to implement them; the hardware connections are well-known technologies in this field and will not be elaborated upon here.

[0022] The multi-source time-series sensing module 20 is connected to the data preprocessing module 10 and is used to receive standardized data vectors. The multi-source time-series sensing module 20 is configured to decompose the input data into components of different frequencies using discrete wavelet transform, and extract the time-series features of each frequency band through multiple independent encoders. The multi-source time-series sensing module 20 further includes a gating network for calculating the dynamic weights of each frequency band and reconstructing the output time-series feature vector accordingly. .

[0023] Spatial feature coordination module 30 is connected to data preprocessing module 10 and is used to receive standardized data vectors. and historical data window. The spatial feature collaboration module 30 is configured to build and maintain a dynamic graph structure, the association matrix of which... Based on system functional logic and real-time data evolution, the spatial feature collaboration module 30 aggregates high-order correlation information between devices through two-stage graph convolution operations from nodes to edges and from edges to nodes, outputting a spatial feature vector. .

[0024] The feature fusion module 40 is connected to the multi-source temporal sensing module 20 and the spatial feature coordination module 30, respectively, and is used to receive temporal feature vectors. and spatial eigenvectors The feature fusion module 40 is configured to construct a spatiotemporal interaction tensor and use Tucker decomposition technology to project high-dimensional features onto a low-dimensional core space for nonlinear interaction, and then project them back into the output space to generate the final fused feature matrix. .

[0025] The early warning decision module 50 is connected to the feature fusion module 40 and is used to receive the final fused feature matrix. The early warning decision module 50 is configured to calculate the belief distribution and overall uncertainty index for each category based on evidence networks. Meanwhile, the early warning decision module 50 combines the expert reward function recovered by maximum entropy inverse reinforcement learning to dynamically generate early warning actions or silent actions based on belief distribution and uncertainty.

[0026] See attached document Figure 1 This invention provides a flowchart of a power plant anomaly early warning method that integrates multi-source temporal sensing and spatial feature collaboration. Based on the aforementioned system architecture, this method specifically includes the following steps: Step S1: Data Access and Adaptive Preprocessing. The data preprocessing module 10 acquires real-time observation data vectors from the power plant's DCS system. To address time discrepancies caused by inconsistent sampling frequencies among different sensors, the data preprocessing module 10 aligns the data by truncating timestamps to a uniform precision (e.g., seconds or milliseconds). Subsequently, the data preprocessing module 10 iteratively updates the mean using an exponential moving average method. and variance Based on this, the current data is standardized to generate a normalized data vector that adapts to the drift of the operating condition distribution. .

[0027] Step S2: Multi-resolution temporal feature extraction. The multi-source temporal sensing module 20 receives normalized data vectors. A multi-level discrete wavelet transform is performed on the component to generate multi-resolution components containing low-frequency approximation coefficients and high-frequency detail coefficients. For each component, the multi-source temporal sensing module 20 captures its unique temporal dependencies using an independent Transformer encoder. Subsequently, the importance weight of each component at the current time step is calculated through a gating mechanism, and the components are weighted and fused to obtain a temporal feature vector containing multi-scale information. .

[0028] Step S3: Dynamic spatial association feature extraction. The spatial feature collaboration module 30 initializes the device node graph based on the system functional logic and uses a recurrent neural network to update the association matrix between nodes according to recent data windows. Based on this dynamic correlation matrix, the spatial feature collaboration module 30 performs alternating graph convolution operations: first, it aggregates node features into edge features, and then it aggregates the updated edge features into node features. This process captures the complex collaborative relationships between devices that change with operating conditions, generating spatial feature vectors. .

[0029] Step S4: Spatiotemporal feature low-rank tensor fusion. The feature fusion module 40 fuses the temporal feature vectors... and spatial eigenvectors The mapping is in tensor form. The feature fusion module 40 does not directly concatenate features; instead, it uses Tucker decomposition to project the features onto the core tensor space. Within the core space, the feature fusion module 40 performs multi-linear rank operations to achieve deep interaction between temporal and spatial dimensions, ultimately outputting the final fused feature matrix. .

[0030] Step S5: Uncertainty Quantification and Early Warning Decision. The early warning decision module 50 will ultimately fuse the feature matrix. Input parameterization layer, output parameters of Dirichlet distribution Based on this parameter, the early warning decision module 50 calculates the probability that the current working condition belongs to each category and the overall uncertainty index of the model's judgment. The early warning decision module 50 further incorporates a pre-trained reward function. The system comprehensively assesses risks and uncertainties, selects the best option to issue an early warning signal or remain silent, and outputs the decision results to the power plant monitoring terminal.

[0031] In a typical implementation environment, the data preprocessing module 10 does not directly process the raw analog electrical signals, but instead establishes high-frequency data communication with the power plant's distributed control system (DCS) server via industrial Ethernet.

[0032] To address the inherent asynchronous nature of multi-source heterogeneous data during acquisition, the data preprocessing module 10 is configured to execute a discretization alignment strategy based on a global clock network. In a DCS system, different types of sensors (such as thermocouples, pressure transmitters, and flow orifice plates) have varying physical response times and sampling frequencies, and network transmission is subject to random jitter. Directly constructing the feature matrix based on the receiving order would result in the same row vector containing state values ​​from different physical moments, thereby disrupting the physical consistency of the instantaneous coupling relationships between variables. Therefore, this embodiment predefines a unified time alignment granularity. The time alignment granularity The value is usually set based on the minimum control cycle of the power plant control loop, ranging from 100ms to 1000ms, in order to balance computational efficiency and the ability to capture transient processes.

[0033] In specific data stream processing, for the first... Each sensor at the original data acquisition timestamp The uploaded measurement data packet is not only read by the data preprocessing module 10, but also its process values. Simultaneously, the system parses the OPC quality code attached to the data packet. The system only performs timestamp mapping if the quality code indicates valid data; if the quality code is bad, the data point is directly marked as missing. This quality code-based pre-screening mechanism prevents noise data from sensor malfunctions from contaminating subsequent models. For valid data that passes verification, it is mapped to discrete time grid points using the following formula. : ; in: This represents the discrete time grid points after mapping, i.e., the unified timestamp after data alignment; This indicates the original data acquisition timestamp carried in the data packet reported by the sensor; This indicates the preset time alignment granularity, used to define the minimum interval of the time grid; This represents the floor operation, used to map continuous time to the nearest past grid point. The physical meaning of this formula is to create an evenly spaced time grid that falls within the same grid window. All sensor measurements within the system are considered as synchronous data under the same operating condition snapshot.

[0034] To address the issue of sampling frequency being lower than time alignment granularity To address the issue of missing grid point data due to packet loss, this embodiment employs a forward-filling mechanism based on zero-order preservation. Considering the significant thermal inertia of power plant thermal processes, which are unlikely to undergo abrupt changes in a very short time, therefore, when... When data from a certain sensor is missing, the system automatically reuses that sensor's data. The effective value at any given time. This addresses the cold start issue that occurs during the initial system startup phase (i.e.,...). (i.e., no historical data), the data preprocessing module 10 is configured to read the setpoint or rated operating condition value corresponding to the sensor for initialization filling. After the above strict time alignment and completion processing, the system constructs a multi-source input tensor that is strictly synchronized in the time dimension and aligned in the spatial dimension. ,in The time window length, For the number of sensors, This provides a feature dimension, thus laying a precise data foundation for subsequently capturing the dynamic coupling relationships between multiple variables. As for the specific implementation of the underlying communication protocol, those skilled in the art can use the OPC UA or Modbus TCP standard library, which are well-known technologies in the field and will not be elaborated upon here.

[0035] After completing the temporal alignment of the multi-source data, the data preprocessing module 10 is configured to perform dynamic feature scaling for non-stationary streaming data.

[0036] Considering that power plant thermal processes are often accompanied by non-stationary operating conditions such as load fluctuations and peak shaving, the statistical distribution characteristics (mean and variance) of sensor data drift significantly over time. Traditional static Z-score normalization methods rely on pre-calculated fixed global statistics, which cannot adapt to such dynamic changes and are prone to generating erroneous feature mappings during operating condition transitions, thereby masking early subtle fault characteristics or causing false alarms. Therefore, this embodiment adopts a continuously weighted exponential moving average algorithm to construct a soft sliding window that advances along the time axis, tracking the central trend and dispersion of each sensor's data in real time.

[0037] In the specific computational logic, for each aligned moment... Observation data vector (This vector contains) Sensor measurements in each dimension, of which (Number of sensors), the system uses a first-order recursive filtering mechanism to update the adaptive mean vector. : ; in: Indicates time The updated mean reflects the system's baseline operating level under current conditions; Indicates the previous moment The mean; This represents the vector of observation data input at the current moment; This represents a smoothing factor, used to adjust the rate at which the model forgets historical information. In this embodiment, The value of is not set arbitrarily, but is determined based on the sampling frequency and the dominant time constant of the controlled object, and ranges from 0 to 1. to Smaller The value is suitable for filtering out high-frequency measurement noise during steady-state operation; a larger value is acceptable. The value is applicable to periods of rapid load change in order to reduce statistical lag.

[0038] Based on the updated mean The system further calculates the current variance. The calculation process directly utilizes the deviation between the current observed value and the latest mean, as shown in the following formula: ; in: Indicates time Updated variance; Indicates the previous moment The variance; This represents the instantaneous deviation energy of the current observation relative to the current dynamic baseline. The mean value is used here. Instead It participates in the calculation, aiming to more quickly capture the increase in volatility caused by sudden changes in operating conditions and improve the sensitivity to transient anomalies.

[0039] Based on the aforementioned real-time updated dynamic statistics, the data preprocessing module 10 performs standardized projection on the raw observation data to generate the final normalized data vector. : ; in: This represents a normalized data vector after normalization processing. Its physical meaning is to map physical quantities with different dimensions (such as temperature ℃, pressure MPa, and flow rate t / h) to a dimensionless relative deviation space of the same scale. This is the numerical stability constant, with a value of 1 x 10⁻⁶. -5 This is used to prevent variance from occurring when the sensor output is constant. Approaching zero triggers a division-by-zero computation anomaly, specifically for recursive algorithms. Regarding the initialization issue at time step, the system is configured to initialize using the sample mean and sample variance of the first sampling batch (e.g., the first 50 data points). and This ensures the numerical stability of the algorithm during its startup phase. Through this dynamic adaptive normalization path, the system can ensure that the feature data input to the subsequent neural network maintains a distribution characteristic of approximately zero mean and unit variance across the entire operating range, thereby improving the convergence efficiency and generalization ability of the deep learning model.

[0040] After the multi-source time-series sensing module 20 receives the normalized time-series data stream output by the data preprocessing module 10, the system enters the in-depth time-frequency domain feature analysis stage.

[0041] Given that power plant thermal process data is essentially a superposition of multiple dynamic processes at different time scales, for example, the slow intraday drift of main steam pressure (low-frequency trend) is often mixed with transient fluctuations (medium-frequency oscillations) caused by valve regulation and electromagnetic noise (high-frequency pulses) generated by electrical circuits.

[0042] Traditional single-timescale analysis struggles to accurately capture minute local mutations while preserving global trends, easily leading to fault characteristics being obscured by background signals. Therefore, this embodiment configures the analysis to perform Discrete Wavelet Transform (DWT) on the time series data of each sensor channel, constructing an orthogonal multi-resolution analysis architecture.

[0043] In a specific signal processing flow, the system extracts the time window length. Inner Normalized input sequence of each sensor The Mallat pyramid algorithm is used for three-level cascaded decomposition. To address the Gibbs effect or truncation error that easily occurs at the boundaries of convolution operations for finite-length time-series signals, this embodiment forces a symmetric extension strategy to pad the beginning and end of the sequence before performing the transformation, instead of zero-padding. This strategy ensures the continuity of the first derivative of the signal at the endpoints by mirroring the boundary data, thereby avoiding the introduction of false high-frequency artifacts during reconstruction.

[0044] In selecting the wavelet basis function, this embodiment uses the Daubechies4 (db4) wavelet. The physical basis for choosing this type of wavelet basis is that it possesses good compact support and asymmetry, and can capture the non-stationary transient waveforms unique to power plant mechanical equipment faults (such as blade breakage and water hammer impact) through convolution operations with the fewest coefficients. The number of decomposition layers is set to 3 ( The number of layers is determined based on the typical control response cycle of a power plant's DCS system (usually in the range of seconds to minutes), which is sufficient to separate high-frequency noise from effective control dynamics.

[0045] After three recursive low-pass filtering, high-pass filtering, and downsampling operations, the original sequence is deconstructed into a set of orthogonal frequency band coefficient vectors, mathematically expressed as: ; in: Indicates the first The normalized input sequence of each sensor within the current time window; This represents a three-layer discrete wavelet transform operator with symmetric continuation preprocessing; the output vector set contains four components with different frequency characteristics, and the physical meaning of each component is as follows: (Third-level approximation coefficients) correspond to the frequency range It mainly reflects the unit's steady-state operating baseline and slow heat accumulation trend, among which The sampling frequency of the original data; (Third layer detail factor) corresponding frequency range It mainly reflects the regulation oscillations of the control loop and the process fluctuations at a medium scale; (Second layer detail coefficients) Corresponding frequency range It mainly reflects the natural frequency vibration of mechanical parts; (First layer detail coefficients) Corresponding frequency range It primarily captures extremely short, transient pulses, acoustic emission signals, or early microcrack features.

[0046] Through the multi-resolution decomposition described above, the system deconstructs the original aliased signal into four physically meaningful orthogonal subspaces, providing clean feature inputs for subsequent independent frequency coding. For the specific convolution formulas and downsampling principles of the discrete wavelet transform, those skilled in the art can refer to relevant standard textbooks on digital signal processing, as these are well-known techniques in the field and will not be elaborated upon here.

[0047] In the multi-source time-series sensing module 20, the original time-series data is decomposed into four orthogonal frequency components through discrete wavelet transform. Afterwards, the system does not simply re-superimpose these components, but instead enters a process of learning independent features in frequency bands and dynamically weighted fusion.

[0048] Given the significant differences in the physical information carried by different frequency components, for example, the third-layer approximation coefficients... Typically, the first-layer detail coefficients reflect the energy accumulation trend of a thermal system, while they carry transient information such as mechanical shock or crack propagation. If a single network with shared weights is used for hybrid processing, it can easily lead to spectral crosstalk where high-frequency noise characteristics suppress low-frequency trend characteristics. Therefore, this embodiment is configured to construct four sets of parallel and parameter-independent Transformer encoder branches, corresponding to the four frequency components mentioned above.

[0049] Under this architecture, for any frequency component sequence (in The system maps wavelet coefficients in scalar form to a high-dimensional feature space through a linear projection layer and superimposes sinusoidal position codes. Each encoder branch contains... The computational units are stacked in layers, with each layer containing a multi-head self-attention mechanism and a feedforward neural network. This lightweight, independent, and parallel design allows the model to capture high-frequency transient impact features without interfering with each other, while preserving low-frequency long-term dependencies, ultimately outputting a four-dimensional temporal context feature tensor.

[0050] Given that the diagnostic value of each frequency component dynamically changes at different stages of equipment failure evolution (for example, high-frequency signals contain the highest information entropy at the moment of failure, while low-frequency trends are more critical during the stable period of failure), simple static averaging fusion cannot adapt to this time-varying characteristic. Therefore, this embodiment designs an adaptive gated reconstruction network to dynamically calculate the contribution weight of each frequency band at each time step based on the current global context.

[0051] In practical implementation, the system first uses a multilayer perceptron as a gating generator to calculate the normalized weights for each frequency band. For time... The weight calculation formula is as follows: ; in: Indicates at time The generated dynamic importance weight vector has a dimension of 4X1 and contains the weight values ​​corresponding to the four frequency components. This represents the normalization exponential function, used to map the raw output of the neural network to a probability distribution, ensuring that the sum of the weights is 1 and that all weights are non-negative, thereby introducing a competition mechanism between frequency bands. The learnable weight matrix of the gated network has dimensions of . This is used to map the concatenated global features to the weight space; This represents the concatenation operator along the feature channel dimension, used to merge four independent feature vectors into a single global context vector; These respectively represent the frequency components. (Low-frequency approximation) (Mid-low frequency details) (Mid-to-high frequency details) (High-frequency details) at any time Each of the independent encoded feature vectors has a dimension of 1. ; This represents the learnable bias vector of the gated network, which has a dimension of 4x1 and is used to adjust the activation threshold.

[0052] Based on the dynamic weights calculated above, the system then performs a point-by-point weighted summation of the features in each frequency band to reconstruct the final multi-source time-series feature vector. The calculation formula is as follows: ; in: Indicates at time The reconstructed final multi-source time-series feature vector, which integrates multi-resolution time-frequency information, serves as the input feature for subsequent fault diagnosis. These represent the dynamic importance weight vectors. The four scalar elements in the data correspond to time points. The following frequency components The adaptive attention coefficient has a value range of [0, 1]. These respectively represent the frequency components. (Low-frequency approximation) (Mid-low frequency details) (Mid-to-high frequency details) (High-frequency details) at any time Each vector is an independently encoded feature vector, and the dimension of each vector is 1. .

[0053] The physical significance of this reconstruction process lies in achieving dynamic focusing of frequency domain attention: when the system detects high-frequency vibration anomalies, the model automatically increases... The value of makes the multi-source temporal feature vector at that moment... The model is dominated by high-frequency information; conversely, during steady-state operation, the model automatically increases. The values ​​emphasize low-frequency trends. All internal parameters of the aforementioned encoder and gating network (including...) , Both the Transformer weights and other parameters are trainable parameters, and are jointly optimized end-to-end through the backpropagation algorithm during the model training phase. For details regarding the specific matrix operations of the Transformer, those skilled in the art can refer to the standard architecture proposed by Vaswani et al. or the standard implementation in deep learning frameworks, which are well-known technologies in the field and will not be elaborated upon here.

[0054] See attached document Figure 2 After the multi-source timing sensing module 20 outputs a timing vector that incorporates frequency domain features, the spatial feature coordination module 30 is further configured to construct and evolve the implicit topology between device nodes.

[0055] Given the complexity of the physical systems of large power plants, there are not only explicit static topologies based on pipe and valve connections, but also a large number of implicit dynamic relationships governed by thermodynamic mechanisms and control logic. For example, the feedwater pump and the economizer are physically far apart, but they have extremely strong state coupling in the thermodynamic cycle, and the strength of this coupling will fluctuate significantly with the rise and fall of the unit load.

[0056] Traditional graph neural networks based on fixed adjacency matrices struggle to characterize these time-varying characteristics that change with operating conditions. Therefore, this embodiment introduces a dynamic hypergraph architecture, which captures many-to-many high-order functional group relationships by real-time evolution of the association matrix.

[0057] In the initial stage of the evolution process, the system combines the power plant's piping and instrumentation diagram with an expert mechanism knowledge base to construct an initial correlation matrix. Specifically, if the devices monitored by the two sensors are directly connected in a physical pipeline or located in the same control loop, the corresponding matrix element is initialized to 1; otherwise, it is 0. This matrix serves as the logical starting point for evolution, providing prior knowledge of physical constraints for the dynamic graph evolution network and avoiding the blindness of purely data-driven methods in the cold start phase.

[0058] Based on the initial topology, this embodiment deploys a nonlinear evolutionary network that updates the topology in real time using historical state memory and current observation data streams. The evolutionary calculation formula for the dynamic correlation matrix is ​​as follows: ; in: Indicates time The generated association matrix has dimensions of .in For the number of sensors, The number of predefined hyperedges (i.e., functional subsystems). Elements in the matrix. The value takes the range (0, 1), which physically represents the first... The sensor, under the current operating conditions, belongs to the [number]th [sensor]. Confidence level or coupling strength of a functional subsystem; Indicates from arrive The input sequence at each time step contains characteristic information reflecting the current operating state of the system, and the window size is... The time steps are set to 10 to 50. The selection of this parameter is based on the typical time constant of the power plant's thermal process, ensuring that the dynamic graph evolution network can capture the complete causal chain from the occurrence of disturbances (such as changes in valve opening) to the reconstruction of the system's correlation. Indicates the previous moment The correlation matrix reflects the memory and continuity of the system's topological structure; This represents a nonlinear dynamic evolution function. In this embodiment, to fully disclose this general characteristic, the function is specifically configured as a composite neural network structure consisting of a gated recurrent unit (GRU), a linear projection layer, and a sigmoid activation layer. The specific implementation logic is as follows: The system first uses the correlation matrix from the previous time step. The input sequence is flattened and embedded into a low-dimensional hidden state vector through a fully connected layer; subsequently, the input sequence is processed using GRU units. The temporal context features are extracted and the hidden state is updated. Finally, the hidden state is reconstructed using linear projection of the output layer and Sigmoid nonlinear transformation. The matrix space. The reason for choosing GRU instead of LSTM is that GRU has fewer parameters while ensuring the ability to capture long-term dependencies, making it more suitable for the online computational needs of high-dimensional matrix evolution.

[0059] To prevent the generated correlation matrix from becoming too dense and introducing non-physical noise connections (i.e., avoiding false weak correlations between nodes that do not have actual thermal connections), the system... A sparsity threshold was further set at the output. (Range from 0.05 to 0.1). When the calculated correlation strength When a fault is detected, it is forcibly set to 0. Through this dynamic evolution mechanism, the system can adaptively reconstruct the connection topology between devices. For example, when a high-pressure heater disconnection fault is detected, the dynamic graph evolution network will automatically disconnect the association weights between relevant sensors and the main loop, thereby accurately reflecting the current process flow status and providing a precise structural basis for subsequent graph convolution operations. For the specific training process of the neural network, an end-to-end backpropagation algorithm is adopted, using the prediction error of the final task to guide the training. Updating parameters is a routine technical method for those skilled in the art, and will not be elaborated upon here.

[0060] The dynamic correlation matrix output by the dynamic graph evolution network reflects the implicit coupling relationships between devices under the current operating conditions. Subsequently, the spatial feature collaboration module 30 is configured to perform graph convolution operations based on bidirectional message passing from node to superedge and then to node.

[0061] Given that the interactions between power plant equipment are often not simple pairwise connections, but rather group collaborations at the functional subsystem level—for example, an anomaly in one coal mill can simultaneously affect multiple parameters such as primary air pressure, mill bowl differential pressure, and outlet temperature—traditional point-to-point graph convolution is insufficient to capture such high-order clustering features. Therefore, this embodiment employs a two-stage aggregation mechanism to simulate the information exchange process between the aggregation of component states within the thermal system to the subsystem and the feedback of subsystem states to the components.

[0062] In the specific computational process, the system will use the timing feature vector output by the multi-source timing sensing module 20 As initial node features Then, it begins multi-level iterative aggregation. To ensure the stability of numerical computation and prevent feature magnitude explosion in deep networks, the system performs a multi-level iterative aggregation on the correlation matrix before performing convolution operations. Normalization based on node degree and hyperedge degree (i.e., dividing by the square root of the degree matrix) is performed as a preprocessing step to ensure the conservation of energy during graph propagation.

[0063] The first phase is a bottom-up aggregation of nodes to hyperedges. During this process, the algorithm traverses each hyperedge representing a functional subsystem. The information from all the sensor nodes included in the subsystem is then weighted and aggregated to generate a macroscopic state representation of the subsystem. The calculation formula is as follows: ; in: Indicates the first The first in the layer The feature vector of a functional subsystem is a high-dimensional abstraction of the overall operating state of the subsystem (such as "the health of the condensate system"). Indicates belonging to the sensor set The first in One sensor; This represents the elements of the normalized correlation matrix, quantifying the sensor. For superedge Contribution weight; Indicates the first The first stage of the layer can learn a transformation matrix to map features from the node space to the hyperedge space; Indicates the first The first in the layer Input characteristics of each sensor; This represents an aggregation operator used to traverse the set of sensor nodes. All sensors And perform a weighted summation operation; The nonlinear activation function is represented by LeakyReLU (with the negative half-axis slope set to 0.2). The reason for this choice is that, compared to ReLUs, LeakyReLU allows for the weak propagation of negative gradients, which can effectively prevent the phenomenon of permanent neuron deactivation (i.e., gradient returning to zero, causing parameters to fail to update) when processing zero or negative value fluctuations (such as negative pressure signals) commonly found in power plant data.

[0064] The second stage is a top-down feedback from the hyperedge to the node. After acquiring the macroscopic features of the subsystem, the algorithm backscatters them back to each component node, enabling each sensor to perceive not only its own temporal changes but also the overall environmental context of its system. The formula for updating node features is as follows: ; in: Indicates the updated number Layer node feature vectors, which integrate global information from all functional subsystems to which they belong, realize feature dimensionality upgrade from a single point perspective to a system perspective; Indicates belonging to the hyperedge set The first in One functional subsystem; Indicates the first The second stage of the layer can learn a transformation matrix, which is used to map the features of the hyperedge space back to the node space; This represents a feedback aggregation operator used to traverse the set of hyperedges. All hyperedges And perform a weighted accumulation operation.

[0065] To achieve information diffusion from the local neighborhood to the global topology, the above two-stage aggregation process will be repeated. Next. In this embodiment, the number of network layers. The convolution depth is set to 2 to 3 layers. The physical basis for this parameter selection is that in the complex network of a power plant, fault propagation usually follows the path of "equipment, subsystem, associated subsystem, associated equipment" (e.g., feedwater pump failure, feedwater system anomaly, main steam system fluctuation, turbine vibration). A convolution depth of 2 to 3 layers is sufficient to cover this typical high-order propagation chain, while avoiding the oversmoothing phenomenon caused by an overly deep network, that is, preventing all node features from converging and losing the ability to distinguish specific fault points.

[0066] go through After layer iterations, the system outputs the final node feature matrix as the spatial feature vector. This feature deeply integrates the dynamic high-order collaborative relationships between devices, providing a decision-making basis that includes a plant-wide topology perspective for subsequent fault feature fusion. For the dimension matching and gradient backpropagation mechanism in matrix multiplication operations, those skilled in the art can refer to relevant standard literature on graph neural networks, which are well-known techniques in the field and will not be elaborated upon here.

[0067] See attached document Figure 3 The time-series feature vectors output by the multi-source time-series sensing module 20 were obtained respectively. Spatial feature vector output by spatial feature coordination module 30 Subsequently, the feature fusion module 40 is configured to perform a deep interactive operation based on low-rank tensor decomposition.

[0068] Given the complex spatiotemporal coupling characteristics of power plant system failure modes—that is, certain temporal trends (such as a sharp rise in temperature) only have fault indication significance under specific spatial topologies (such as upstream and downstream equipment in the same loop)—simple feature concatenation or linear addition can only preserve the independent attributes of features and cannot capture this multiplicative nonlinear dependency. Therefore, this embodiment constructs a factorized bilinear interaction network based on Tucker decomposition, aiming to simulate the high-order interaction logic between spatiotemporal features through a low-dimensional core tensor.

[0069] In the specific fusion process, the system considers the input temporal feature vector With spatial feature vectors These are two independent feature modes. Here, Indicates the number of sensors. This represents the unified feature channel dimension output by the multi-source temporal sensing module 20 and the spatial feature collaboration module 30. To achieve full-dimensional feature interaction with limited computational resources and to prevent parameter explosion and overfitting risks associated with high-dimensional dense tensors, this embodiment employs a parameter-efficient Tucker decomposition form to approximate the ideal full-temporal-spatial interaction tensor. The general mathematical principle of this interaction process can be expressed as the following approximate formula: ; in: This represents the theoretical all-space-time interaction tensor, with dimension . This represents the potential relationship space between all temporal and spatial channels, representing the pairwise combinations of all channels. This parameter represents the dimension of the target output feature. In this embodiment, the parameter is set according to the complexity of the fault type to be diagnosed. In order to ensure sufficient information to distinguish subtle early faults, it is set to 64 to 128. The low-rank approximation relationship means that this embodiment does not pursue a mathematically accurate reconstruction of the full-time-space interaction tensor, but actively filters out high-frequency noise and redundant information in the tensor through low-rank decomposition, retaining only the most essential time-space coupling structure, thereby improving the robustness of the factorized bilinear interaction network in complex power plant environments. (in () represents the n-modal product operator, a standard operator in tensor algebra, indicating that the tensor moves along the nth modulus. Performing a multiplication operation between each dimension and the matrix, for example... Specifically, this refers to the contraction operation between the core tensor and the matrix in the temporal dimension; The core tensor has dimensions of . This tensor is the core weight of the entire fusion network, and its internal elements Through training and learning, the specific interaction rules between temporal and spatial patterns were encoded. These represent the compression ranks of the three modes, respectively. To force the model to learn the most essential interaction features on the low-dimensional manifold and to achieve denoising, this embodiment sets... ,and The range of values ​​is to ; This represents the temporal projection matrix, used to map high-dimensional temporal features to a compact low-rank subspace; This represents the spatial projection matrix, used to map features of a high-dimensional space to an isomorphic interactive subspace; This represents the output projection matrix, used to map the interacted low-dimensional features back to the target output dimension. .

[0070] Based on the above decomposition architecture, the actual computation process is designed as three tightly coupled pipeline steps: projection compression, core interaction, and output reconstruction.

[0071] Based on this, the system uses a projection matrix to map the original input features to a low-dimensional core space, thereby removing redundant information and reducing computational complexity. The projection calculation is as follows: ; ; in: This represents the dimensionality-reduced temporal latent vector matrix, with dimension 1. ; This represents the spatial latent vector matrix after dimensionality reduction, with dimension . ; This represents a random deactivation operation. To improve the model's generalization ability and prevent co-adaptation of specific neurons, this embodiment introduces a Dropout mechanism (dropout rate of 0.2 to 0.3) after projection. This step is physically equivalent to extracting the most representative temporal and spatial principal components, laying the foundation for subsequent efficient interactions.

[0072] Then, the core interaction phase begins. The system performs multilinear product operations using Einstein's summation convention, through core tensors. The compressed temporal and spatial features are deeply fused. Intermediate-state fusion features. The calculation formula is as follows: ; in: Indicates the first The sensor at the first Intermediate fusion features on each output latent factor; This represents the sensor node index, with a value ranging from 1 to... ; This indicates the output latent factor index, with a value range of 1 to... ; This represents the latent factor index of the time series, with values ​​ranging from 1 to... ; This represents the spatial latent factor index, with values ​​ranging from 1 to... ; Indicates the core tensor in The learnable weight parameter for coordinate position acts as a physical gating coefficient, determining the first... The time series latent factor and the first After the interaction of the spatial latent factors, the first... The contribution of each output latent factor. This mechanism enables the factorized bilinear interaction network to adaptively amplify key spatiotemporal combinations (such as load surges and main steam temperature channels) while suppressing irrelevant noise combinations.

[0073] Finally, in the output reconstruction stage, the system projects the interacted low-rank features back to the target feature space. Although the standard Tucker decomposition only includes linear projection operations, to compensate for the inability of linear transformations to handle systematic biases, this embodiment configures this step as an affine transformation layer containing bias terms. The final fused feature matrix... The calculation formula is as follows: ; in: This represents the final fused feature matrix, with dimension 1. , as input to the fault classifier; Indicates the output projection matrix The transpose of the matrix, with dimension Used to reduce feature dimensions from compressed rank Restore to the target output feature dimension ; Represents a learnable bias vector with dimension . Its number of sensors Broadcast addition is performed to correct systematic biases in the model.

[0074] Through this Tucker decomposition and fusion mechanism, this embodiment successfully maintains linear computational complexity. At the same time, it achieved the equivalent of The quadratic feature interaction capability. Regarding the initialization of network parameters, this embodiment focuses on the core tensor. and projection matrix All initializations use Xavier uniform distribution to maintain variance consistency during forward propagation and avoid gradient vanishing or exploding problems caused by multi-level multiplication operations.

[0075] After constructing the core interaction tensor through the Tucker decomposition logic described above, in order to transform the mathematically low-rank approximation into a neural network layer with strong nonlinear fitting capabilities, this embodiment designs a projection calculation process that includes nonlinear activation and residual correction paths.

[0076] Although core tensor operations can effectively capture the multiplicative coupling between temporal and spatial relationships, in actual power plant operation scenarios, there are often significant nonlinear response characteristics between fault features and monitoring data. For example, some vibration signals only exhibit fault correlation when the load exceeds a specific critical threshold, i.e., there is a physical activation threshold. To address this, the feature fusion module 40 is configured to introduce a nonlinear transform layer after tensor reconstruction and retain the basic features of the original signal through skip connections, thereby forming a complete deep fusion unit.

[0077] Intermediate fusion features obtained from prior calculations (lie in (Latent factor space of dimension), the system executes to output feature dimension to the target. The mapping and activation operations are performed. This step aims to decode abstract interactive latent factors into feature representations with specific physical orientations (such as the probability of a specific fault mode). The calculation formula is as follows: ; in: This represents the feature vector generated after nonlinear activation; This represents the intermediate fusion feature after core tensor compression interaction, with dimension 1. ,in For the number of sensors, To output the compressed rank of the latent factors; Indicates the output projection matrix The transpose of the matrix, with dimension Used for feature dimensionality upscaling; Represents a learnable bias vector with dimension . Its initial value is set to 0; The nonlinear activation function is used. Given that the sensor signal contains negative values ​​(representing fluctuations below the mean) after preprocessing, and to avoid neuron death during deep network training, this embodiment preferably uses a Gaussian error linear unit (GELU). Compared to ReLU, GELU introduces a stochastic regularization effect on the negative half-axis, which can better preserve weak negative trend information, making the factorized bilinear interaction network more sensitive in processing early fault signals.

[0078] Considering that deep feature interaction networks inevitably lose some statistical properties of the original signal (such as absolute amplitude) when extracting high-order semantics, and that this fundamental information is crucial for identifying hard faults (such as open circuits and over-range faults), this embodiment constructs a residual correction path to balance high-order interactive features and low-order original features. This path directly injects the original input features into the final output, allowing the model to degenerate into a linear concatenation model even in the worst case, thus ensuring the lower bound of fault diagnosis performance.

[0079] Combining layer normalization mechanism, the final output features are... The calculation logic is as follows: ; ; in: This represents unnormalized aggregated features, with dimension 1. ; This represents the feature vector generated after nonlinear activation; This indicates a feature concatenation operation, which concatenates features of dimension 1. Temporal feature vectors With spatial feature vectors By splicing together the channel dimensions, a The joint vector; Represents the residual projection matrix, with dimension 1. This matrix is ​​used to align the original features to the target dimension. It is trained using an end-to-end backpropagation algorithm, and its function is to adaptively learn the weight proportions of the original features in the final decision. Representing aggregation features respectively Mean and standard deviation along the channel dimension; This represents the numerical stability constant, used to prevent issues arising during long-term stable operation of the equipment (i.e., variance). This example assumes a division-by-zero calculation error occurs when the value approaches 0. 1x10 -5 This ensures the integrity of the algorithm logic under any extreme operating conditions; Denotes the learnable affine transformation parameters, all of which are dimensionless. These two parameters allow factorized bilinear interaction networks to recover the original distribution of features when necessary, preventing normalization operations from destroying the learned, physically meaningful feature magnitude information.

[0080] Through the aforementioned composite processing flow including projection, activation, residual connection, and normalization, this embodiment not only achieves deep interaction of spatiotemporal features but also constructs a numerically stable gradient propagation channel. During the model training phase, all the aforementioned weight matrices ( Both the fault classification cross-entropy loss function and the bias term are jointly updated using the Adam optimizer by minimizing the fault classification cross-entropy loss function. This end-to-end learning strategy ensures that the feature fusion process always adaptively adjusts to maximize fault discriminative power.

[0081] The feature fusion module 40 outputs a final fused feature matrix with high-order spatiotemporal semantics. Subsequently, in order to overcome the inherent defect of the traditional Softmax classifier in outputting an unreasonablely high confidence probability for unknown samples when facing unknown fault conditions not included in the training set, the early warning decision module 50 was configured to introduce an evidence theory framework to construct an evidence estimation network that can simultaneously output the fault category probability and the prediction uncertainty.

[0082] Unlike traditional neural networks that only output point estimates of probability, the core concept of this embodiment is to redefine the output layer of the neural network as a collector of virtual evidence, and then parameterize a Dirichlet distribution through this evidence, thereby quantifying the degree of certainty of the model's cognition of the current device state at the probability density level.

[0083] In its specific network implementation, the early warning decision module 50 utilizes a fully connected evidence layer to connect the dimension... The final fusion feature matrix Mapping to the category evidence space. This process no longer uses the Softmax function, which forces the output to compress into the (0, 1) interval, but instead employs a non-negative activation mechanism to compute the evidence vector. For the ... Each operating condition category (including normal operating conditions and preset fault conditions) corresponds to a Dirichlet distribution parameter. The calculation formula is as follows: ; in: This represents the working condition category index, with values ​​ranging from 1 to... ,in The total number of predefined typical power plant fault modes (e.g., including 10-20 categories such as normal, overheating, and leakage). This indicates the corresponding first layer in the fully connected data layer. The weight vector transpose of the working condition, with dimension . ; Indicates the corresponding number The bias scalar for similar operating conditions; This indicates a modified positive linear unit, which in this context represents the evidence non-negativity constraint, meaning that the support of sensor data for a certain type of fault cannot be physically negative. If the linear transformation result is less than 0, it is considered as lacking evidence support, and the network automatically truncates it to 0. The prior bias of the Dirichlet distribution ensures that even without any observational evidence (i.e., (Output is 0), Dirichlet distribution parameters The minimum value is always 1. This not only avoids the risk of the denominator being 0 in subsequent calculations mathematically, but also corresponds physically to an initial state without any class bias (i.e., the Dirichlet distribution degenerates into a uniform distribution), ensuring the rigor of the reasoning logic.

[0084] Based on the parameter vector obtained above This embodiment calculates multi-dimensional decision indicators through statistical transformations to assess the credibility of the prediction results. Total strength of evidence. The calculation is as follows: ; in: Representing the overall strength of evidence, this metric physically quantifies the overall degree of matching between the current sensor data and the distribution of historical training data, taking into account... Then there must be This eliminates the possibility of a division-by-zero error; This indicates the total number of predefined operating condition categories (e.g., including normal operating mode and typical fault modes such as overheating, leakage, and vibration). This represents the summation operator, used to calculate the sum of evidence parameters for all possible operating conditions.

[0085] Based on this, the expected value of the category probability Compared with overall uncertainty indicators The calculation is as follows: ; ; in, This represents the expected value of the category probability, and its physical meaning is the evidence estimation network's determination that the current device state belongs to the [missing category]. The average predicted probability of a given working condition differs from traditional probabilities; this value is explicitly constrained by the normalization of the total amount of evidence. The overall uncertainty index is defined as (0, 1], and this index is the core basis for judging unknown faults in this embodiment.

[0086] Decision logic: The system is configured to monitor this cognitive uncertainty indicator. When Indicates a higher probability of a certain type of fault (e.g., >0.8) but At the same time, a high confidence level (e.g., >0.5) indicates that the evidence estimation network is in a state of "classification has been achieved but evidence is insufficient" (due to data noise or sensor failure). The system will trigger a low confidence warning instead of an immediate alarm. Only when... Exceeding the preset threshold and A fault diagnosis is only considered valid when the value is below a preset uncertainty threshold (ranging from 0.1 to 0.2).

[0087] To enable the evidence estimation network to output the aforementioned parameterized distribution, this embodiment employs an evidence learning strategy based on Bayesian risk or mean squared error for parameter updates during the training phase. Specifically, the training objective is to minimize the difference between the predicted Dirichlet distribution and the true label distribution. For the part belonging to the... The optimization process will drive the evidence values ​​for the corresponding categories based on real training samples from similar working conditions. The training mechanism maximizes evidence values ​​for non-labeled categories while suppressing them from approaching zero. This mechanism forces the evidence estimation network to accumulate evidence only when the feature pattern highly matches known faults, thus ensuring that it automatically outputs a high level of overall uncertainty index when faced with unknown or ambiguous conditions due to a lack of matching evidence. The specific mathematical form for constructing such evidence loss functions can be implemented by referring to existing theories of evidence-based deep learning, and will not be elaborated here.

[0088] In the previous stage, we obtained the expected probability of each category. Compared with overall uncertainty indicators After the distributed output, in order to transform these statistical indicators into operating instructions that conform to the actual operation and maintenance specifications of the power plant, the early warning decision module 50 is further configured to execute a data-driven dynamic strategy optimization.

[0089] Given that power plant fault handling decisions often involve complex risk trade-offs—for example, when some non-critical sensors experience significant fluctuations but the overall system uncertainty is high—the optimal strategy is often not to immediately shut down and trigger an alarm, but rather to remain silent to gather more evidence. This non-linear decision boundary is difficult to precisely encode using fixed logical rules. Therefore, this embodiment constructs a decision strategy network that aims to learn the reward function implicit in operational behaviors by observing expert operation records in historical distributed control system (DCS) logs, thereby achieving an adaptive balance between risk and reward.

[0090] In the specific implementation of constructing the decision-making environment, the system first defines the key elements of the Markov decision-making process: state space. Defined as the joint vector of the output of the preceding evidence estimation network, denoted as: ,in The expected probability of each category of working condition. This represents an indicator of current overall uncertainty. (Action Space) Defined as a discrete set of operation instructions, this embodiment includes {trigger warning, remain silent}. As an extensible approach, the action space can also include {request manual review} to handle high-risk ambiguous situations.

[0091] To extract the optimal strategy from historical data, this example employs a maximum entropy inverse reinforcement learning framework. The core logic of this framework is that it assumes historical action trajectories statistically follow a Boltzmann distribution, meaning that trajectories with higher cumulative rewards have a greater probability of occurrence. The system extracts a series of expert states and action trajectories from historical storage. And the reward function is solved by maximizing the policy entropy. In this process, the key reward function... It is parameterized as a linear combination of state and action feature vectors, and its calculation formula is as follows: ; in: In the state space Lower execution action space The obtained reward function physically quantifies the system's tendency to take specific actions under specific operating conditions (e.g., receiving a positive reward when the probability of failure is high, and receiving a negative reward when the uncertainty is high). This represents the transpose of the learnable weight vector, with the same dimension as the feature vector. This weight vector essentially encodes the cost matrix in power plant operation and maintenance. It is iteratively updated through the IRL algorithm until the preference weights consistent with expert behavior are recovered. Represents the feature vectors of state and action.

[0092] To ensure the algorithm can capture high-dimensional decision logic, this embodiment demonstrates the construction of physically oriented interaction features, including: Underreporting risk characteristics: calculated as This is used to measure the potential risks of choosing to remain silent when the probability of failure is high. False alarm penalty characteristics: calculated as This is used to measure the unreliability of forced alarms when there is insufficient information (high uncertainty); Effective intervention characteristics: calculated as This is used to incentivize decisive action when there is high certainty and a high probability of failure.

[0093] in: This is an indicator function that takes the value 1 if the condition within the parentheses is met, and 0 otherwise. Let be the expected probability of the fault category.

[0094] The reward function that fits the expert preferences is obtained through the above process. Subsequently, in order to obtain a decision model that can be directly used for real-time reasoning, this embodiment uses the Proximal Policy Optimization (PPO) algorithm to perform forward training on the decision policy network.

[0095] Specifically, the decision-making policy network adopts an Actor-Critic dual-stream architecture: Policy sub-network: receiving state space As input, after passing through several fully connected layers and Tanh activation processing, the output is a probability distribution in the action space. This output distribution constitutes the current policy. , used to generate specific operation instructions; Value subnetwork: also receives state space Output a scalar value The role of this network is to evaluate the expected cumulative reward of the current state, which guides the gradient update direction of the policy network.

[0096] During the training phase, the system utilizes the learned reward function. A simulated reward signal is generated, and the policy update magnitude is limited by maximizing the objective function and introducing a truncation mechanism, thereby ensuring the numerical stability of the training process. This PPO-based training method is a well-known technology in the field, but in this embodiment, its core innovation lies in using a special reward function derived from IRL that includes an uncertainty penalty term, rather than a simple, manually set rule.

[0097] The optimal strategy after training convergence It is deployed in the online monitoring process. Unlike simple probability threshold judgments, this strategy comprehensively considers both failure probability and cognitive uncertainty to achieve dynamic adaptive risk warnings. The specific judgment logic is as follows: Deterministic fault response: when Exceeding a preset high-level threshold (e.g., 0.85) and the overall uncertainty index When the level is low (e.g., <0.2), the optimal strategy is... The system will trigger an early warning action with a high probability, instructing the system to immediately trigger an audible and visual alarm. Fuzzy state suppression: when At a moderate level (e.g., 0.4 to 0.7) but with an overall uncertainty index When the model score is high (e.g., >0.6), it indicates that the current data features are insufficient to support a clear conclusion. In this case, the optimal strategy... The learned weights will automatically suppress the output probability of triggering warning actions, instead outputting instructions to remain silent or suggest manual review. This mechanism effectively filters out false alarms caused by transient sensor interference or unknown operating conditions, reducing the workload of operators in handling invalid alarms.

[0098] Specific application examples: Implementation environment and scenario construction: This embodiment selects a 600MW supercritical coal-fired power generating unit as the verification object. The verification scenario focuses on a leakage fault in the high-pressure heater piping system. This fault is highly concealed in its early stages and is easily confused with water level fluctuations caused by load changes.

[0099] The data collection scope covers the water supply system, steam extraction system, and condensate drainage system, and a total of [number] samples were selected. Key sensor nodes (including feedwater temperature, extraction steam pressure, heater terminal differential pressure, and steam trap opening, etc.). Data sampling frequency is 1Hz, and time alignment granularity is set to [value missing]. .

[0100] Detailed explanation of the implementation process: The data preprocessing module 10 receives data from the DCS in real time. Assuming that... At that moment, the unit was in the load ramp-up phase, and the feedwater flow sensor readings fluctuated.

[0101] First, perform timestamp alignment: ; Subsequently, the dynamic mean is calculated using a continuously weighted exponential moving average algorithm. A smoothing factor is then set. To adapt to increased load conditions: ; The normalized data vector obtained by the system calculation It eliminates the overall trend deviation caused by load ramping, retaining only the instantaneous residuals relative to the current operating condition baseline.

[0102] The multi-source time-series sensing module 20 performs three-level wavelet decomposition on the normalized data.

[0103] Low-frequency components ( ): It captured the smooth curve of the unit's load increase.

[0104] High-frequency details ( ):exist Near the (fault initiation point), a weak high-frequency glitch signal caused by acoustic emission was extracted. The dynamic weights calculated by the gating network shifted at this point, automatically increasing... The weights of the time series feature vectors make the time series feature vectors The focus is on the transient impact caused by the leak.

[0105] Spatial Feature Coordination Module 30 Updates Association Matrix .

[0106] ; In the early stages of a leak, although changes from a single sensor are not significant, the coupling relationship between the extraction pressure and the steam trap opening becomes abnormal (pressure drop causes unintended valve opening adjustment). The dynamic graph evolution network captures this change, enhancing the correlation between these two nodes in the correlation matrix. The connection weights in the graph are aggregated into spatial feature vectors through graph convolution. This amplifies the spatial impact of local anomalies.

[0107] Evidence Network Decision Making and Uncertainty Response: Early warning decision module 50 is based on fused features Output the final judgment. See attached document. Figure 6 The data shown is explained as follows: When the fault occurred ( The physical fault has just occurred, and its characteristics have not yet fully manifested. At this point, the model calculates the expected probability of the fault category. It is only around 0.1, which does not reach the warning threshold.

[0108] High uncertainty region ( to As leakage characteristics become mixed with background noise (increased load interference), the model enters a cognitive confusion zone. At this point, although the expected probability of the fault category is... It started to rise slowly, but the overall uncertainty index... A sharp rise, in It reaches a peak value of 0.8 nearby. According to the formula... And the characteristics of false alarm penalties, at this time the overall uncertainty index This will generate strong negative feedback for the "triggered warning" action. Therefore, the system will execute the optimal strategy. By choosing to remain silent, the false alarms that are prone to occur in traditional models at this stage are effectively avoided.

[0109] Trigger warning signal ( ): As time went by Nearby, a pentagram marker appears in the image (indicated by the Chinese text "Trigger Warning Signal"). This pentagram represents the critical moment when the system makes its final alarm decision.

[0110] At this moment, the data satisfies the dual-judgment logic: High certainty: The overall uncertainty index is shown by the gray dashed line. It has fallen back to a low level ( This indicates that the model has collected sufficient evidence of failure; Probability met: Expected probability of the fault category indicated by the black solid line. It exceeded the "warning trigger threshold" (0.85) shown by the dotted horizontal line in the figure.

[0111] The pentagram symbol physically corresponds to the optimal strategy. When the output action switches from "remain silent" to "trigger warning", the system immediately sends an audible and visual alarm command for high-pressure heater leakage to the DCS terminal.

[0112] To verify the effectiveness of the fusion method proposed in this invention, a test set including typical operating conditions such as "normal, sensor drift, high-pressure heater leakage, and reheater tube wall overheating" was constructed and compared with existing mainstream methods.

[0113] Comparison model settings: CNN-LSTM: Traditional deep learning models only use convolution to extract local features, while LSTM extracts temporal features, lacking spatial graph structure and uncertainty quantification.

[0114] GAT-Trans: Based on a combination of Graph Attention Network (GAT) and Transformer, it has spatiotemporal modeling capabilities, but it uses Softmax for direct classification and lacks theoretical support for evidence.

[0115] The method of this invention integrates multi-source wavelet temporal sensing, dynamic graph evolution spatial coordination, and evidence network decision-making.

[0116] Performance index analysis: The experimental results are shown in the appendix. Figure 5 As shown in the figure. The black bars represent accuracy, the gray bars represent precision, and the white bars represent F1 scores.

[0117] Accuracy: The CNN-LSTM model achieved an accuracy of 89.4%, but its performance was the worst due to its inability to handle dynamic spatial relationships under varying conditions. GAT-Trans improved this to 93.1%. In contrast, the method of this invention, benefiting from the deep interaction of spatiotemporal features through Tucker tensor decomposition, achieved an accuracy of 98.7%, significantly outperforming the comparative models.

[0118] Accuracy: This reflects the model's ability to suppress false alarms. CNN-LSTM only achieves 88.2%, indicating its tendency to misjudge load fluctuations as faults. The method of this invention achieves 98.5%. This is directly attributed to the evidence theory and overall uncertainty index introduced in the early warning decision module 50. When features are ambiguous (as shown in the attached image) Figure 6 In areas of high uncertainty, the system actively remains silent, thereby reducing false alarms.

[0119] F1 score: As a comprehensive evaluation indicator, the method of this invention achieved 98.3%, which is more than 11 percentage points higher than CNN-LSTM (86.9%), proving that the system maintains extremely high reliability while ensuring a high detection rate.

[0120] in conclusion: In summary, this invention solves the problem of difficult extraction of complex spatiotemporal coupled features in power plants by introducing multi-resolution temporal sensing and dynamic graph evolution mechanisms; it achieves low-rank high-order feature fusion through Tucker decomposition; and in particular, the uncertainty decision-making mechanism based on evidence theory successfully solves the problem of false alarms during the transition process of changing operating conditions, which has extremely high engineering value in practical power plant applications.

Claims

1. A power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy, characterized in that, Includes the following steps: The data preprocessing module (10) collects the raw time series data from the multi-source sensors, performs time alignment operation, and performs adaptive normalization processing on the raw time series data based on dynamic statistics, and outputs a normalized data vector. The multi-source time-series sensing module (20) performs multi-resolution decomposition on the normalized data vector, extracts the time-series features of each resolution component, performs weighted fusion, and outputs a time-series feature vector. The spatial feature collaboration module (30) generates a dynamic graph structure based on the normalized data vector and historical data window, and aggregates the association information between devices based on the dynamic graph structure to output a spatial feature vector; The feature fusion module (40) maps the temporal feature vector and the spatial feature vector to the tensor space for interactive fusion to generate the final fused feature matrix; The early warning decision module (50) calculates the expected probability of categories and the overall uncertainty index based on the final fusion feature matrix, and generates early warning actions or silent actions in combination with the decision strategy.

2. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy as described in claim 1, characterized in that, The data preprocessing module (10) performs time alignment and adaptive normalization operations, specifically including the following steps: The timestamps of the original time-series data are truncated to a uniform time granularity to eliminate sampling delay; The mean and variance are iteratively updated using a continuously weighted exponential moving average algorithm, which serves as the dynamic statistic. The original time-series data is standardized and projected using the mean and variance to generate the normalized data vector.

3. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature coordination according to claim 1, characterized in that, The steps of the multi-source temporal sensing module (20) to extract the temporal features of each resolution component specifically include: The normalized data vector is decomposed into orthogonal frequency components containing low-frequency approximation coefficients and high-frequency detail coefficients using discrete wavelet transform; Construct multiple sets of parallel Transformer encoder branches, each corresponding to a frequency component in the orthogonal frequency components; The time dependencies of each frequency component are captured by the Transformer encoder branch, and the independent encoded feature vectors corresponding to each frequency component are output.

4. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy according to claim 3, characterized in that, The weighted fusion steps performed by the multi-source time-series sensing module (20) specifically include: The global features obtained by concatenating all the independent coded feature vectors are received using a gated network, and the dynamic weight of each frequency component at the current time is calculated. The time-series feature vector is reconstructed by using the dynamic weights to perform point-by-point weighted summation on each of the independent encoded feature vectors.

5. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy according to claim 1, characterized in that, The steps of the spatial feature coordination module (30) in generating the dynamic graph structure based on the normalized data vector and historical data window specifically include: Construct an initial association matrix based on the equipment and process topology logic of a power plant; The input sequence of the historical data window is processed using a recurrent neural network to extract temporal context features and update the hidden state; The updated hidden states are used to generate a learning-type correlation matrix that reflects the dynamic dependencies of the data through nonlinear transformation; The initial correlation matrix and the learned correlation matrix are merged, and the merged matrix is ​​sparsified to obtain the dynamic graph structure.

6. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy according to claim 1, characterized in that, The spatial feature collaboration module (30) specifically includes the following steps in aggregating the association information between devices: First-stage aggregation from node to hyperedge: Traverse the hyperedges representing the functional subsystems, and aggregate the information of the sensor nodes contained in the hyperedges according to the correlation matrix to generate the macroscopic state representation of the functional subsystems. Perform the second-stage aggregation from the hyperedge to the node: The macroscopic state representation of the functional subsystem is fed back to each of the sensor nodes to update the features of the sensor nodes. The first-stage aggregation and the second-stage aggregation are executed alternately to output the spatial feature vector containing the dynamic high-order collaborative relationships between devices.

7. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy according to claim 1, characterized in that, The specific steps of the feature fusion module (40) in generating the final fused feature matrix include: Using the Tucker decomposition technique, the temporal feature vector and the spatial feature vector are mapped to a low-dimensional core space through temporal projection matrix and spatial projection matrix, respectively. The core tensor is used to perform multilinear product operations on the compressed temporal and spatial features to generate intermediate fusion features; The intermediate fused features are mapped back to the target output dimension using the output projection matrix, and a bias term is superimposed to obtain the features after nonlinear interaction.

8. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature coordination according to claim 7, characterized in that, The feature fusion module (40) is also used to perform a residual correction operation, the steps of which specifically include: The features resulting from the nonlinear interaction are processed through an activation function to obtain intermediate hidden states; The temporal feature vector and the spatial feature vector are concatenated, and a linear transformation is performed using the residual weight matrix to obtain linear features; The intermediate hidden state is added to the linear feature, and the result is subjected to layer normalization to generate the final fused feature matrix.

9. The power plant anomaly early warning method integrating multi-source temporal sensing and spatial feature synergy according to claim 1, characterized in that, The steps for the early warning decision module (50) to generate an early warning action specifically include: The final fused feature matrix is ​​mapped to Dirichlet distribution parameters using an evidence network; Calculate the expected probability of each category and the overall uncertainty index based on the Dirichlet distribution parameters; Recover the reward function from expert historical operation data using maximum entropy inverse reinforcement learning; The expected probability of the category and the overall uncertainty index are input into the policy network, and the policy network, in conjunction with the reward function, outputs the warning action or the silent action.

10. A power plant anomaly early warning system integrating multi-source temporal sensing and spatial feature coordination, characterized in that, The power plant anomaly early warning method based on the fusion of multi-source temporal sensing and spatial feature coordination as described in any one of claims 1-9 includes: The data preprocessing module (10) is used to collect raw time series data from multiple sources of sensors, perform time alignment operations, and perform adaptive normalization processing on the raw time series data based on dynamic statistics, and output a normalized data vector. The multi-source time-series sensing module (20) is used to perform multi-resolution decomposition on the normalized data vector, extract the time-series features of each resolution component and perform weighted fusion, and output the time-series feature vector. The spatial feature collaboration module (30) is used to generate a dynamic graph structure based on the normalized data vector and the historical data window, and to aggregate the correlation information between devices based on the dynamic graph structure and output the spatial feature vector. The feature fusion module (40) is used to map the temporal feature vector and the spatial feature vector to the tensor space for interactive fusion to generate the final fused feature matrix; The early warning decision module (50) is used to calculate the category probability and uncertainty index based on the final fused feature matrix, and generate early warning actions or silent actions in combination with the decision strategy.

Images