Machine learning based steam turbine unit fault early warning method and system

Abstract

The application discloses a steam turbine unit fault early warning method and system based on machine learning, and particularly relates to the field of fault monitoring, and comprises the following steps: I. In the operation process of the steam turbine unit, multi-source multi-modal operation data of the steam turbine unit are collected, and the multi-source multi-modal operation data are time-space aligned through a unified time stamp and space calibration mechanism; II. The multi-modal data are preprocessed, and meanwhile, based on the preprocessed multi-modal data, the topological and coupling relationship between the equipment components of the steam turbine unit is modeled; the application can make the early warning no longer depend on obvious overrun signals, effectively realize early and forward-looking risk identification, reduce false alarms caused by single sensor abnormalities or noises, improve the stability and reliability of the early warning results, take into account the physical mechanism and data driving, enhance the engineering credibility, avoid the fixed threshold failure problem, avoid the data island problem, and make the overall early warning capability of the system constantly evolve with the operation time.

Description

Technical Field

[0001] This invention relates to the field of fault monitoring, specifically to a method and system for early warning of turbine unit faults based on machine learning. Background Technology

[0002] As critical rotating equipment in power, petrochemical, and energy systems, the operational safety and reliability of steam turbine units directly impact production continuity and system stability. With increasing unit capacity and more complex operating conditions, traditional fault monitoring and early warning methods based on empirical rules or single signal thresholds are gradually revealing their limitations in addressing early-stage, hidden faults, complex faults, and individual equipment variations. On the one hand, in the nascent stage of a fault, steam turbine units often exhibit only weak anomalies in multiple physical quantities in space and time, making accurate identification difficult using a single sensor or simple statistical indicator. On the other hand, differences in manufacturing deviations, operating environments, and aging processes among different units make it difficult to apply uniform thresholds or fixed models in the long term. In recent years, machine learning technology has received widespread attention in the field of industrial equipment condition monitoring, demonstrating significant advantages in multi-source data fusion, complex pattern recognition, and nonlinear modeling. However, existing machine learning-based fault early warning methods still generally suffer from high dependence on high-quality fault samples, insufficient model generalization ability, lack of physical interpretability, and difficulty in adapting to long-term equipment evolution, hindering their engineering application in large-scale critical equipment. Therefore, it is necessary to propose a machine learning-based method and system for early warning of turbine unit faults.

[0003] Existing turbine generator set fault early warning methods and systems rely solely on obvious over-limit signals, failing to achieve early and proactive risk identification. False alarms due to single sensor anomalies or noise are frequent, reducing the stability and reliability of early warning results. They also fail to balance physical mechanisms with data-driven approaches, lowering engineering credibility and making them prone to fixed threshold failures. Furthermore, they cannot allow the overall early warning capability of the system to continuously evolve over time. Therefore, we propose a machine learning-based turbine generator set fault early warning method and system. Summary of the Invention

[0004] The purpose of this invention is to solve the problem, and to propose a machine learning-based method and system for early warning of turbine unit faults.

[0005] In a first aspect of this invention, a method and system for early warning of turbine unit faults based on machine learning are first proposed, the method comprising: Ⅰ. During the operation of the steam turbine unit, multi-source and multi-modal operation data of the steam turbine unit are collected, and the multi-source and multi-modal operation data are spatiotemporally aligned through a unified timestamp and spatial calibration mechanism; II. Preprocess multimodal data, and model the topology and coupling relationships between turbine unit equipment components based on the preprocessed multimodal data; III. Based on the topology and coupling relationships between equipment components, a simulation operation model of the steam turbine unit is constructed, and the health status of each steam turbine unit is dynamically evolved and predicted according to the operating conditions of the steam turbine unit; IV. Real-time comparison between the actual unit status and the predicted health status of the turbine unit, identify and determine potential fault risks based on the comparison results, and generate graded early warning results. V. After triggering the early warning, analyze the causal relationship between multimodal operation data, identify abnormal variables and their transmission paths, and generate interpretable fault causal reasoning results; VI. After completing the causal analysis, combine historical maintenance cases and expert knowledge to generate an interactive decision-making solution for operators, and provide real-time feedback on on-site inspection results during the execution process.

[0006] As a further aspect of the present invention, the specific steps for preprocessing the multimodal data in step II are as follows: S1.1: Based on the nameplate parameters, operating procedures and historical stable operating condition statistics of each equipment in the steam turbine unit, physical allowable ranges are set for various sensors. Then, the operating data of each equipment in the steam turbine unit are collected through various sensors, and data outside the preset physical allowable range are marked as invalid values. Then, a weighted smoothing method within a local time window is used to filter and denoise each operating data. S1.2: Monitor the changing trends of each running data at adjacent time points in real time, and simultaneously count the missing values ​​in each running data. Then, interpolate and compensate for invalid and missing values ​​in each running data. After that, establish a sliding time window with a preset length and width and set the step size of the sliding time window. Move the sliding time window step by step according to the set step size. S1.3: After each round of movement of the sliding time window, calculate the standardized deviation of each running data within the corresponding sliding time window. If the standardized deviation is higher than the preset threshold, the corresponding running data is judged to be abnormal and is removed or re-marked as missing. At the same time, interpolation is performed to fill in the abnormal data marked as missing. S1.4: Select a set of time axes as a unified time scale, interpolate or downsample the low-frequency or high-frequency running data, align the running data in time, and then map the running data of each mode to the same high-dimensional tensor with time index and spatial location as coordinate axes to form a unified spatiotemporal data representation.

[0007] As a further aspect of the present invention, the specific calculation formula for the weighted smoothing described in S1.1 is as follows: ; In the formula, This represents the denoised output value at time j. This represents the original valid running data within the k-th neighborhood centered at time j; M represents the total number of sampling points within the time window. This represents the weight coefficient corresponding to the k-th neighborhood sampling point; The specific calculation formula for the interpolation compensation described in S1.2 is as follows: ; In the formula, The estimated value after compensation at time t; This represents the missing value at time t; After representing missing points The observed value at time; Before the missing point The observed value at time; This represents the length of the time interval.

[0008] As a further aspect of the present invention, the specific steps for modeling the topology and coupling relationships between turbine unit equipment components in step II are as follows: S2.1: Each physical field generated during the operation of the steam turbine unit is treated as an independent channel and stacked in a uniform time order. Then, the multimodal operation data is input into the corresponding independent channel according to the corresponding physical field type. The operation data of each mode is processed into a continuous frame sequence with a fixed step size to establish an input tensor that can be directly input into a three-dimensional convolutional network. S2.2: Input each input tensor into the three-dimensional convolutional network. The three-dimensional convolutional network slides the three-dimensional convolutional kernel on each input tensor according to the preset time step. At the same time, it performs joint weighting on the local spatial region and adjacent time frames of each input tensor and outputs the response value of the corresponding feature map at the corresponding spatial position and time. S2.3: Perform element-wise nonlinear activation on each response value to obtain the corresponding spatiotemporal features after nonlinear activation. Then, aggregate the spatiotemporal features after activation within a preset time and space range to generate aggregated features for each spatiotemporal region. At the same time, retain abnormal responses through statistical summarization. Then, use each component of the turbine unit as nodes in the graph, and the mechanical connections, thermal coupling, or fluid coupling relationships between components as edges. Construct the corresponding equipment topology graph based on the generated nodes and edges. S2.4: Perform feature propagation calculation on the equipment topology map, and simulate the diffusion of each physical field anomaly based on the equipment topology map to obtain the feature representation of the corresponding component of the turbine unit. Then, splice or weightedly combine the spatiotemporal features of the physical field and the topology propagation features of the component according to the correspondence to establish a joint feature containing information on "spatial distribution-temporal evolution-structural coupling".

[0009] As a further aspect of the present invention, the specific form of the input tensor in S2.1 is as follows: ; In the formula, Represents spatial coordinates The values ​​at the time index w and the physical channel d; The data scale normalization and physical quantity mapping function representing the d-th physical channel; Represents the d-th physical channel in the corresponding spatial coordinates Running data on time index w; The specific calculation formula for the joint weighting described in S2.2 is as follows: ; In the formula, Represents the spatial location of the output feature map of the 3D convolution kernel. The response value at time index k; Represents the spatial offset of the 3D convolution kernel The weights at time offset r and channel c; P and Q represent the spatial dimensions of the 3D convolution kernel; R represents the temporal depth of the 3D convolution kernel; C represents the number of input physics channels. The specific calculation formula for the aggregation feature mentioned in S2.3 is as follows: ; In the formula, The aggregated feature vector representing the s-th spatiotemporal region; Represents the set of indices within the s-th predefined spatiotemporal region; This represents the number of feature points contained in the s-th predefined spatiotemporal region; Represents spatial location Spatiotemporal features at time index z.

[0010] As a further aspect of the present invention, the physical field mentioned in S2.1 specifically includes: temperature field, stress field, flow velocity field, etc.

[0011] As a further aspect of the present invention, the specific steps for constructing the simulation operation model of the steam turbine unit in step III are as follows: S3.1: Based on thermodynamics, rotor dynamics and energy conversion mechanism, select each continuous variable that can completely describe the operating state of the turbine unit, and establish the corresponding mechanism state according to the selected continuous variables. Construct the state space of the corresponding turbine unit according to the established mechanism state level to build the simulation operation model of the corresponding turbine unit. Then set the corresponding measurement equation to convert each mechanism state into actual observation data. S3.2: The joint features are mapped sequentially to the input or correction interface of the simulation operation model according to their physical meaning. Based on the received joint features, the simulation operation model performs turbine unit operation simulation according to the preset time step and outputs the real-time simulation data of the corresponding turbine unit within the simulation time. The real-time simulation data is converted into predicted observation data through measurement equations. Then, the actual observation data that is consistent with the simulation time of the simulation operation model is arranged into a vector form according to sensor type and measurement point number. S3.3: Perform element-wise difference between the predicted observation data and the actual observation data corresponding to each sensor to form a deviation vector containing direction and amplitude information. Then, smooth and suppress each generated deviation vector. Then, sort the joint features from oldest to newest according to time and select the joint features corresponding to the smoothed deviation vector as the input dataset. S3.4: Input the input dataset into the simulation model. The simulation model adopts a multi-layer mapping structure, which transforms each joint feature in the input dataset into residual prediction values ​​layer by layer. Then, the actual observed residual values ​​are compared with the corresponding residual prediction values, and the residual prediction loss value at the corresponding time is calculated. Then, the residual prediction loss value is returned to the simulation model based on the backpropagation algorithm, and the simulation model parameters are updated. S3.5: Iteratively update the simulation model parameters until the residual prediction loss value converges to the preset range, then stop updating. At the same time, synchronize the real-time multimodal operation data of the turbine unit to the simulation model. Then, the simulation model outputs the twin state of the corresponding turbine unit, corrected model parameters and health latent variables, and plots the historical evolution trajectory of the corresponding turbine unit. Based on the preset abnormal threshold, mark the time points of the operation state that exceed the abnormal threshold.

[0012] As a further aspect of the present invention, the specific calculation formula for the residual prediction loss value in S3.4 is as follows: ; In the formula, R represents the residual prediction loss value at time k; R represents the dimension of the residual vector. This represents the j-th component of the true residual vector at time k. This represents the j-th component in the prediction residual vector at time k.

[0013] In a second aspect of the present invention, a machine learning-based turbine unit fault early warning system is proposed, comprising: a multi-source sensing module, a processing and control module, a feature learning module, a state representation module, a mapping and prediction module, an update learning module, a risk determination module, an analysis and interpretation module, a decision support module, and a feedback and update module. The multi-source sensing module is used to collect multi-source, multi-scale operating data of various equipment in the steam turbine unit; The processing control module is used to perform timestamp calibration, sampling alignment and spatial calibration on multi-source operating data, and to perform quality assessment and preprocessing on various types of operating data after spatiotemporal alignment; The feature learning module is used to convert the preprocessed multi-source running data into a unified spatiotemporal representation, organize the multi-source data into corresponding structured data, and extract the features of each structured data. The state characterization module quantifies the current health status of the turbine unit based on the extracted structured data features, and obtains the performance level of the turbine unit at different operating stages. The mapping prediction module constructs and runs a simulation operation model of the steam turbine unit, maps real-time sensing data into the model, and predicts the evolution trend of each piece of equipment in the steam turbine unit under the current operating conditions. The update learning module optimizes the parameters of the simulation operation model by leveraging the experience of multiple units, based on the aging of various equipment in the steam turbine unit, changes in operating conditions, and new types of faults. The risk assessment module is used to compare and analyze the actual monitoring status with the digital twin prediction results, assess the deviation between the two and its development trend, and generate graded early warning information. The analysis and interpretation module is used to perform correlation analysis on various types of operational data after an early warning is issued, identify the influence relationship between abnormal variables, and locate the root cause. The decision support module generates decision support solutions for on-site personnel based on anomaly analysis results and historical operation and maintenance experience. After the turbine unit is repaired or dealt with, the feedback update module collects on-site feedback information, stores the actual fault results and handling experience in the system knowledge base, and updates the simulation and anomaly identification parameters.

[0014] As a further aspect of the present invention, the specific steps of the update learning module in collaboratively optimizing the simulation operation model parameters using multi-unit experience are as follows: S4.1: Treat the data of each historical steam turbine unit under each operating stage or different fault type as an independent learning task, and determine the feature-state mapping relationship of each historical steam turbine unit under different conditions. Each task includes input samples and corresponding health or fault characterization targets. Then, extract a support sample subset of a preset size from each historical learning task. S4.2: Input the supporting sample subset into the simulation operation model of each turbine unit, and calculate the loss gradient of each simulation operation model in the supporting sample subset starting from the current model parameters. Then, perform a constrained update based on the loss gradient to obtain the temporary adaptation parameters in the corresponding historical learning task. After that, use the sample data in the same task that did not participate in the fast update to evaluate the predictive performance of the temporary adaptation parameters and generate the fast adaptation evaluation loss for the corresponding historical learning task. S4.3: Weight the rapid adaptation evaluation losses of all historical learning tasks to construct a cross-task overall objective function. Then, calculate the gradient of the overall objective function with respect to the parameters of each simulation model. Based on the calculated gradient, update the parameters of each simulation model. Then, upload the updated parameters of each simulation model to the central node for a global update. Finally, distribute the globally updated parameters to the current turbine unit simulation models and replace the original parameters. S4.4: When a new steam turbine unit is added, a set of time windows is randomly selected from the continuous operation data of the new steam turbine unit, and the samples within the time window are filtered and organized. Based on the operating status or expert rules, corresponding supervision information is generated to construct a small sample dataset for rapid learning. S4.5: Calculate the deviation between the new unit small sample dataset and the historical learning task samples in the feature space. If the deviation exceeds the preset tolerance threshold, adjust it to the tolerance threshold. Then, input the small sample dataset into the globally updated simulation model and update the simulation model parameters locally. Calculate the difference between the locally updated parameters and the globally updated parameters and use it as the local parameter update amount. S4.6: Encrypt or perturb the update amount of each local parameter before uploading it to the central node. Then, according to the data scale or reliability of each newly added turbine unit, the uploaded update information is weighted and averaged to generate new global parameters. The updated parameters are then distributed to the simulation operation model of each turbine unit, replacing the original parameters, and waiting for rapid adaptation to subsequent new units or new faults.

[0015] As a further aspect of the present invention, the learning task described in S4.1 is specifically manifested in the following form: ; In the formula, This represents the learning task corresponding to the a-th historical unit or historical operating condition. This represents the feature representation of the b-th input sample in the a-th learning task; This represents the supervised objective for the b-th input sample in the a-th learning task; Represents learning tasks The number of samples in the sample.

[0016] The beneficial effects of this invention are: This invention proposes a machine learning-based fault early warning method and system for steam turbine units. Based on equipment nameplate parameters, operating procedures, and historical stable operating conditions, physical allowable ranges are set for various sensors. The collected multi-source operating data is preprocessed, and then different physical fields are mapped as independent channels into a unified high-dimensional tensor using time and spatial location as coordinates. This tensor is then input into a three-dimensional convolutional network in a continuous frame sequence to extract key features of the physical fields in spatial distribution and temporal evolution. Graph structure propagation is used to simulate the coupling and diffusion of anomalies between components. Spatiotemporal features and topological features are concatenated or weighted to form a joint feature that simultaneously characterizes spatial distribution, temporal evolution, and structural coupling relationships. The joint features are then used as input and mapped to a simulation model built based on various physical mechanisms. This model is used for residual learning, adaptive parameter updates, and digital twin state evolution. It continuously participates in model optimization and global knowledge updates during multi-task learning and rapid adaptation to small samples of new units. This enables early warning to no longer rely on obvious over-limit signals, effectively achieves early and forward-looking risk identification, reduces false alarms caused by single sensor anomalies or noise, improves the stability and reliability of early warning results, takes into account both physical mechanisms and data-driven approaches, enhances engineering credibility, avoids the problem of fixed threshold failure, avoids data silos, and allows the overall early warning capability of the system to continuously evolve over time. Attached Figure Description

[0017] The present invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 A flowchart illustrating a machine learning-based turbine unit fault early warning method provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a machine learning-based turbine unit fault early warning system provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] 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.

[0021] This invention provides a machine learning-based method and system for early warning of turbine unit faults. See also... Figure 1 , Figure 1A flowchart illustrating a machine learning-based fault early warning method for steam turbine units provided in this embodiment of the invention. The method includes the following steps: During the operation of the steam turbine unit, multi-source and multi-modal operating data of the steam turbine unit are collected, and the multi-source and multi-modal operating data are spatiotemporally aligned through a unified timestamp and spatial calibration mechanism.

[0022] Multimodal data is preprocessed, and based on the preprocessed multimodal data, the topology and coupling relationships between turbine unit equipment components are modeled.

[0023] Specifically, based on the nameplate parameters, operating procedures, and historical stable operating condition statistics of each piece of equipment in the turbine unit, physical allowable ranges are set for various sensors. Then, operating data of each piece of equipment in the turbine unit is collected through these sensors, and data outside the preset physical allowable ranges are marked as invalid values. Next, a weighted smoothing method within a local time window is used to filter and denoise the operating data, and the changing trends of the operating data at adjacent time points are monitored in real time. Simultaneously, missing values ​​in each operating data point are statistically analyzed, and interpolation compensation is performed for invalid and missing values. Finally, a sliding time window with preset length and width is established, and the step size of this sliding time window is set. According to the set step size, the sliding time window is moved step by step. After each round of movement of the sliding time window, the standardization deviation of each running data in the corresponding time window is calculated. If the standardization deviation is higher than the preset threshold, the corresponding running data is judged to be abnormal and is removed or re-marked as missing. At the same time, the abnormal data marked as missing is filled by interpolation. A set of time axes is selected as a unified time scale, and low-frequency or high-frequency running data is interpolated or downsampled. The running data is time-aligned. Then, using time index and spatial position as coordinate axes, the running data of each modality is mapped to the same high-dimensional tensor to form a unified spatiotemporal data representation.

[0024] Specifically, each physical field generated during the operation of the steam turbine unit is treated as an independent channel and stacked according to a uniform time sequence. Then, multimodal operation data is input into the corresponding independent channel according to the type of physical field. Each modal operation data is processed into a continuous frame sequence with a fixed step size to establish input tensors that can be directly input into a 3D convolutional network. Each input tensor is then input into the 3D convolutional network, where the 3D convolutional network gradually slides the 3D convolutional kernel across each input tensor according to a preset time step. Simultaneously, it performs joint weighting on the local spatial regions and adjacent time frames of each input tensor and outputs the response values ​​of the corresponding feature maps at the corresponding spatial locations and times. Element-wise nonlinear activation is then applied to each response value to obtain nonlinear activation. The corresponding spatiotemporal features are then aggregated within a preset time and space range to generate aggregated features for each spatiotemporal region. Simultaneously, abnormal responses are retained through statistical summarization. Then, each component of the turbine unit is used as a node in the graph, and the mechanical connections, thermal coupling, or fluid coupling relationships between components are used as edges. Based on the generated nodes and edges, a corresponding equipment topology graph is constructed. Feature propagation calculations are performed on the equipment topology graph, and the diffusion simulation of each physical field anomaly is performed based on the equipment topology graph to obtain the feature representation of the corresponding component of the turbine unit. The spatiotemporal features of the physical field and the topology propagation features of the components are spliced ​​or weighted according to the correspondence to establish a joint feature containing information on "spatial distribution-temporal evolution-structural coupling".

[0025] In this embodiment, the physical fields specifically include: temperature field, stress field, flow velocity field, etc. The specific calculation formula for weighted smoothing is as follows: ; In the formula, This represents the denoised output value at time j. This represents the original valid running data within the k-th neighborhood centered at time j; M represents the total number of sampling points within the time window. This represents the weight coefficient corresponding to the k-th neighborhood sampling point; The specific calculation formula for interpolation compensation is as follows: ; In the formula, The estimated value after compensation at time t; This represents the missing value at time t; After representing missing points The observed value at time; Before the missing point The observed value at time; Represents the length of the time interval;

[0026] The specific representation of the input tensor is as follows: ; In the formula, Represents spatial coordinates The values ​​at the time index w and the physical channel d; The data scale normalization and physical quantity mapping function representing the d-th physical channel; Represents the d-th physical channel in the corresponding spatial coordinates Running data on time index w; The specific calculation formula for joint weighting is as follows: ; In the formula, Represents the spatial location of the output feature map of the 3D convolution kernel. The response value at time index k; Represents the spatial offset of the 3D convolution kernel The weights at time offset r and channel c; P and Q represent the spatial dimensions of the 3D convolution kernel; R represents the temporal depth of the 3D convolution kernel; C represents the number of input physics channels. The specific calculation formula for aggregation features is as follows: ; In the formula, The aggregated feature vector representing the s-th spatiotemporal region; Represents the set of indices within the s-th predefined spatiotemporal region; This represents the number of feature points contained in the s-th predefined spatiotemporal region; Represents spatial location Spatiotemporal features at time index z.

[0027] Based on the topology and coupling relationships between equipment components, a simulation operation model of the steam turbine unit is constructed, and the health status of each steam turbine unit is dynamically evolved and predicted according to the operating conditions of the steam turbine unit.

[0028] Specifically, based on thermodynamics, rotor dynamics, and energy conversion mechanisms, continuous variables that can fully describe the operating state of the turbine unit are selected. Corresponding mechanistic states are established based on these selected continuous variables. The state space of the corresponding turbine unit is constructed based on these established mechanistic state levels to build a simulation operation model for the turbine unit. Then, corresponding measurement equations are set to convert each mechanistic state into actual observation data. The joint features are mapped sequentially to the input or correction interface of the simulation operation model according to their physical meaning. Based on the received joint features, the simulation operation model performs turbine unit operation simulation according to a preset time step and outputs real-time simulation data of the corresponding turbine unit within the simulation duration. The real-time simulation data is converted into predicted observation data through measurement equations. Then, the actual observation data consistent with the simulation time of the simulation operation model are arranged into a vector form according to sensor type and measurement point number. Element-by-element difference is performed between the predicted observation data and the actual observation data corresponding to each sensor to form a deviation vector containing direction and amplitude information. The generated deviation vectors are then smoothed and suppressed. The joint features are then sorted from oldest to newest according to time, and the joint features corresponding to the smoothed deviation vectors are selected as the input dataset. The input dataset is then input into the simulation model. The simulation model adopts a multi-layer mapping structure, transforming each joint feature in the input dataset into a residual prediction value layer by layer. The actual observed residual values ​​are then compared with the corresponding residual prediction values, and the residual prediction loss value at the corresponding time is calculated. The residual prediction loss value is then returned to the simulation model based on the backpropagation algorithm, and the simulation model parameters are updated. The simulation model parameters are iteratively updated repeatedly until the residual prediction loss value converges to a preset range, at which point the update stops. At the same time, the real-time multimodal operation data of the turbine unit is synchronized to the simulation model. The simulation model then outputs the twin state of the corresponding turbine unit, the corrected model parameters, and the health latent variables. It also plots the historical evolution trajectory of the corresponding turbine unit and marks the time points of operation states that exceed the preset anomaly threshold based on the preset anomaly threshold.

[0029] It should be noted that the specific formula for calculating the residual prediction loss value is as follows: ; In the formula, R represents the residual prediction loss value at time k; R represents the dimension of the residual vector. This represents the j-th component of the true residual vector at time k. This represents the j-th component in the prediction residual vector at time k.

[0030] The system compares the actual unit status with the predicted health status of the turbine unit in real time, identifies and determines potential fault risks based on the comparison results, and generates graded early warning results.

[0031] After an early warning is triggered, the causal relationships between multimodal operational data are analyzed, abnormal variables and their transmission paths are identified, and interpretable causal reasoning results for the fault are generated.

[0032] After completing the causal analysis, and combining historical maintenance cases and expert knowledge, an interactive decision-making solution is generated for operators, and the results of on-site inspections are fed back in real time during the execution process.

[0033] Based on the same inventive concept, embodiments of the present invention also provide a machine learning-based turbine unit fault early warning system. See also Figure 2 , Figure 2 The schematic diagram of the machine learning-based turbine unit fault early warning system provided in the embodiment of the present invention includes: a multi-source sensing module, a processing and control module, a feature learning module, a state representation module, a mapping and prediction module, an update learning module, a risk determination module, an analysis and interpretation module, a decision support module, and a feedback and update module.

[0034] The multi-source sensing module is used to collect multi-source, multi-scale operating data from various equipment of the turbine unit; the processing and control module is used to perform timestamp calibration, sampling alignment, and spatial calibration on the multi-source operating data, and to perform quality assessment and preprocessing on various types of operating data after spatiotemporal alignment; the feature learning module is used to convert the preprocessed multi-source operating data into a unified spatiotemporal expression, organize the multi-source data into corresponding structured data, and extract the features of each structured data; the state characterization module quantifies the current health status of the turbine unit based on the extracted structured data features, and obtains the performance level of the turbine unit at different operating stages.

[0035] The mapping prediction module constructs and runs a simulation operation model of the turbine unit, maps real-time sensing data into the model, and predicts the evolution trend of each piece of equipment in the turbine unit under the current operating conditions. The update learning module optimizes the parameters of the simulation operation model in a collaborative manner based on the aging of each piece of equipment in the turbine unit, changes in operating conditions, and new faults, using experience from multiple units.

[0036] Specifically, data from each historical turbine unit under different operating stages or fault types is treated as an independent learning task. The feature-state mapping relationship for each historical turbine unit under different conditions is determined. Each task includes input samples and corresponding health or fault characterization targets. A pre-defined subset of support samples is extracted from each historical learning task and input into the simulation model of each turbine unit. Starting with the current model parameters, the loss gradient of each simulation model on the support sample subset is calculated, and a constrained update is performed based on the loss gradient to obtain temporary adaptation parameters within the corresponding historical learning task. Then, sample data from the same task that did not participate in the rapid update are used to evaluate the predictive performance of the temporary adaptation parameters, generating a rapid adaptation evaluation loss for the corresponding historical learning task. The rapid adaptation evaluation losses from all historical learning tasks are weighted and combined to construct a cross-task overall objective function. The gradient of the overall objective function with respect to the parameters of each simulation model is then calculated. Based on the calculated gradient, the parameters of each simulation model are updated, and the updated simulation model parameters are uploaded to the central node. A global update is performed, and the updated parameters are distributed to the current simulation operation models of each turbine unit, replacing the original parameters. When a new turbine unit is added, a time window is randomly selected from the continuous operation data of the new turbine unit, and the samples within the time window are filtered and organized. Corresponding supervision information is generated based on the operating status or expert rules to construct a small sample dataset for rapid learning. The deviation between the new unit's small sample dataset and the historical learning task samples in the feature space is calculated. If the deviation exceeds a preset tolerance threshold, it is adjusted to the tolerance threshold. Then, the small sample dataset is... The sample dataset is input into the globally updated simulation model, and the simulation model parameters are locally updated. The difference between the locally updated parameters and the globally updated parameters is calculated and used as the local parameter update amount. Each local parameter update amount is encrypted or perturbed and then uploaded to the central node. Then, according to the data scale or reliability of each newly added turbine unit, the uploaded update information is weighted and averaged to generate new global parameters. The updated parameters are then distributed to the simulation models of each turbine unit again, replacing the original parameters, and waiting for rapid adaptation to subsequent new units or new faults.

[0037] The risk assessment module is used to compare and analyze the actual monitoring status with the digital twin prediction results, assess the deviation between the two and its development trend, and generate graded early warning information; the analysis and interpretation module is used to perform correlation analysis on various types of operational data after the early warning is generated, identify the influence relationship between abnormal variables, and locate the root cause.

[0038] The decision support module generates decision support solutions for on-site personnel based on anomaly analysis results and historical operation and maintenance experience; the feedback update module collects on-site feedback information after the turbine unit is repaired or dealt with, stores the actual fault results and handling experience in the system knowledge base, and updates the simulation and anomaly identification parameters.

[0039] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A machine learning-based method for early warning of turbine unit faults, characterized in that, Includes the following steps: Ⅰ. During the operation of the steam turbine unit, multi-source and multi-modal operation data of the steam turbine unit are collected, and the multi-source and multi-modal operation data are spatiotemporally aligned through a unified timestamp and spatial calibration mechanism; II. Preprocess multimodal data, and model the topology and coupling relationships between turbine unit equipment components based on the preprocessed multimodal data; III. Based on the topology and coupling relationships between equipment components, a simulation operation model of the steam turbine unit is constructed, and the health status of each steam turbine unit is dynamically evolved and predicted according to the operating conditions of the steam turbine unit; IV. Real-time comparison between the actual unit status and the predicted health status of the turbine unit, identify and determine potential fault risks based on the comparison results, and generate graded early warning results. V. After triggering the early warning, analyze the causal relationship between multimodal operation data, identify abnormal variables and their transmission paths, and generate interpretable fault causal reasoning results; VI. After completing the causal analysis, combine historical maintenance cases and expert knowledge to generate interactive decision-making solutions for operators, and provide real-time feedback on on-site inspection results during the execution process; The specific steps for modeling the topology and coupling relationships between the components of the steam turbine unit, as described in step II, are as follows: S2.1: Each physical field generated during the operation of the steam turbine unit is treated as an independent channel and stacked in a uniform time order. Then, the multimodal operation data is input into the corresponding independent channel according to the corresponding physical field type. The operation data of each mode is processed into a continuous frame sequence with a fixed step size to establish an input tensor that can be directly input into a three-dimensional convolutional network. S2.2: Input each input tensor into the three-dimensional convolutional network. The three-dimensional convolutional network slides the three-dimensional convolutional kernel on each input tensor according to the preset time step. At the same time, it performs joint weighting on the local spatial region and adjacent time frames of each input tensor and outputs the response value of the corresponding feature map at the corresponding spatial position and time. S2.3: Perform element-wise nonlinear activation on each response value to obtain the corresponding spatiotemporal features after nonlinear activation. Then, aggregate the spatiotemporal features after activation within a preset time and space range to generate aggregated features for each spatiotemporal region. At the same time, retain abnormal responses through statistical summarization. Then, use each component of the turbine unit as nodes in the graph, and the mechanical connections, thermal coupling, or fluid coupling relationships between components as edges. Construct the corresponding equipment topology graph based on the generated nodes and edges. S2.4: Perform feature propagation calculation on the equipment topology map, and simulate the diffusion of each physical field anomaly based on the equipment topology map to obtain the feature representation of the corresponding component of the turbine unit. Then, splice or weightedly combine the spatiotemporal features of the physical field and the topology propagation features of the component according to the correspondence to establish a joint feature containing information on "spatial distribution-temporal evolution-structural coupling".

2. The machine learning-based turbine unit fault early warning method according to claim 1, characterized in that, The specific steps for preprocessing multimodal data in step II are as follows: S1.1: Based on the nameplate parameters, operating procedures and historical stable operating condition statistics of each equipment in the steam turbine unit, physical allowable ranges are set for various sensors. Then, the operating data of each equipment in the steam turbine unit are collected through various sensors, and data outside the preset physical allowable range are marked as invalid values. Then, a weighted smoothing method within a local time window is used to filter and denoise each operating data. S1.2: Monitor the changing trends of each running data at adjacent time points in real time, and simultaneously count the missing values ​​in each running data. Then, interpolate and compensate for invalid and missing values ​​in each running data. After that, establish a sliding time window with a preset length and width and set the step size of the sliding time window. Move the sliding time window step by step according to the set step size. S1.3: After each round of movement of the sliding time window, calculate the standardized deviation of each running data within the corresponding sliding time window. If the standardized deviation is higher than the preset threshold, the corresponding running data is judged to be abnormal and is removed or re-marked as missing. At the same time, interpolation is performed to fill in the abnormal data marked as missing. S1.4: Select a set of time axes as a unified time scale, interpolate or downsample the low-frequency or high-frequency running data, align the running data in time, and then map the running data of each mode to the same high-dimensional tensor with time index and spatial location as coordinate axes to form a unified spatiotemporal data representation.

3. The machine learning-based early warning method for turbine unit faults according to claim 1, characterized in that, The specific representation of the input tensor described in S2.1 is as follows: ； In the formula, Represents spatial coordinates The values ​​at the time index w and the physical channel d; The data scale normalization and physical quantity mapping function representing the d-th physical channel; Represents the d-th physical channel in the corresponding spatial coordinates Running data on time index w; The specific calculation formula for the joint weighting described in S2.2 is as follows: ； In the formula, Represents the spatial location of the output feature map of the 3D convolution kernel. The response value at time index k; Represents the spatial offset of the 3D convolution kernel The weights at time offset r and channel c; P and Q represent the spatial dimensions of the 3D convolution kernel; R represents the temporal depth of the 3D convolution kernel; C represents the number of input physics channels. The specific calculation formula for the aggregation feature mentioned in S2.3 is as follows: ； In the formula, The aggregated feature vector representing the s-th spatiotemporal region; Represents the set of indices within the s-th predefined spatiotemporal region; This represents the number of feature points contained in the s-th predefined spatiotemporal region; Represents spatial location Spatiotemporal features at time index z.

4. The machine learning-based fault early warning method for steam turbine units according to claim 1, characterized in that, The specific steps for constructing the simulation operation model of the steam turbine unit as described in step III are as follows: S3.1: Based on thermodynamics, rotor dynamics and energy conversion mechanism, select each continuous variable that can completely describe the operating state of the turbine unit, and establish the corresponding mechanism state according to the selected continuous variables. Construct the state space of the corresponding turbine unit according to the established mechanism state level to build the simulation operation model of the corresponding turbine unit. Then set the corresponding measurement equation to convert each mechanism state into actual observation data. S3.2: The joint features are mapped sequentially to the input or correction interface of the simulation operation model according to their physical meaning. Based on the received joint features, the simulation operation model performs turbine unit operation simulation according to the preset time step and outputs the real-time simulation data of the corresponding turbine unit within the simulation time. The real-time simulation data is converted into predicted observation data through measurement equations. Then, the actual observation data that is consistent with the simulation time of the simulation operation model is arranged into a vector form according to sensor type and measurement point number. S3.3: Perform element-wise difference between the predicted observation data and the actual observation data corresponding to each sensor to form a deviation vector containing direction and amplitude information. Then, smooth and suppress each generated deviation vector. Then, sort the joint features from oldest to newest according to time and select the joint features corresponding to the smoothed deviation vector as the input dataset. S3.4: Input the input dataset into the simulation model. The simulation model adopts a multi-layer mapping structure, which transforms each joint feature in the input dataset into residual prediction values ​​layer by layer. Then, the actual observed residual values ​​are compared with the corresponding residual prediction values, and the residual prediction loss value at the corresponding time is calculated. Then, the residual prediction loss value is returned to the simulation model based on the backpropagation algorithm, and the simulation model parameters are updated. S3.5: Iteratively update the simulation model parameters until the residual prediction loss value converges to the preset range, then stop updating. At the same time, synchronize the real-time multimodal operation data of the turbine unit to the simulation model. Then, the simulation model outputs the twin state of the corresponding turbine unit, corrected model parameters and health latent variables, and plots the historical evolution trajectory of the corresponding turbine unit. Based on the preset abnormal threshold, mark the time points of the operation state that exceed the abnormal threshold.

5. A machine learning-based turbine generator set fault early warning system, used to implement the machine learning-based turbine generator set fault early warning method according to any one of claims 1-4, characterized in that, include: The system includes a multi-source perception module, a processing and control module, a feature learning module, a state representation module, a mapping and prediction module, an update and learning module, a risk assessment module, an analysis and interpretation module, a decision support module, and a feedback and update module. The multi-source sensing module is used to collect multi-source, multi-scale operating data of various equipment in the steam turbine unit; The processing control module is used to perform timestamp calibration, sampling alignment and spatial calibration on multi-source operating data, and to perform quality assessment and preprocessing on various types of operating data after spatiotemporal alignment; The feature learning module is used to convert the preprocessed multi-source running data into a unified spatiotemporal representation, organize the multi-source data into corresponding structured data, and extract the features of each structured data. The state characterization module quantifies the current health status of the turbine unit based on the extracted structured data features, and obtains the performance level of the turbine unit at different operating stages. The mapping prediction module constructs and runs a simulation operation model of the steam turbine unit, maps real-time sensing data into the model, and predicts the evolution trend of each piece of equipment in the steam turbine unit under the current operating conditions. The update learning module optimizes the parameters of the simulation operation model by leveraging the experience of multiple units, based on the aging of various equipment in the steam turbine unit, changes in operating conditions, and new types of faults. The risk assessment module is used to compare and analyze the actual monitoring status with the digital twin prediction results, assess the deviation between the two and its development trend, and generate graded early warning information. The analysis and interpretation module is used to perform correlation analysis on various types of operational data after an early warning is issued, identify the influence relationship between abnormal variables, and locate the root cause. The decision support module generates decision support solutions for on-site personnel based on anomaly analysis results and historical operation and maintenance experience. After the turbine unit is repaired or dealt with, the feedback update module collects on-site feedback information, stores the actual fault results and handling experience in the system knowledge base, and updates the simulation and anomaly identification parameters.

6. The machine learning-based turbine unit fault early warning system according to claim 5, characterized in that, The specific steps of the update learning module in collaboratively optimizing the simulation operation model parameters using multi-unit experience are as follows: S4.1: Treat the data of each historical steam turbine unit under each operating stage or different fault type as an independent learning task, and determine the feature-state mapping relationship of each historical steam turbine unit under different conditions. Each task includes input samples and corresponding health or fault characterization targets. Then, extract a support sample subset of a preset size from each historical learning task. S4.2: Input the supporting sample subset into the simulation operation model of each turbine unit, and calculate the loss gradient of each simulation operation model in the supporting sample subset starting from the current model parameters. Then, perform a constrained update based on the loss gradient to obtain the temporary adaptation parameters in the corresponding historical learning task. After that, use the sample data in the same task that did not participate in the fast update to evaluate the predictive performance of the temporary adaptation parameters and generate the fast adaptation evaluation loss for the corresponding historical learning task. S4.3: Weight the rapid adaptation evaluation losses of all historical learning tasks to construct a cross-task overall objective function. Then, calculate the gradient of the overall objective function with respect to the parameters of each simulation model. Based on the calculated gradient, update the parameters of each simulation model. Then, upload the updated parameters of each simulation model to the central node for a global update. Finally, distribute the globally updated parameters to the current turbine unit simulation models and replace the original parameters. S4.4: When a new steam turbine unit is added, a set of time windows is randomly selected from the continuous operation data of the new steam turbine unit, and the samples within the time window are filtered and organized. Based on the operating status or expert rules, corresponding supervision information is generated to construct a small sample dataset for rapid learning. S4.5: Calculate the deviation between the new unit small sample dataset and the historical learning task samples in the feature space. If the deviation exceeds the preset tolerance threshold, adjust it to the tolerance threshold. Then, input the small sample dataset into the globally updated simulation model and update the simulation model parameters locally. Calculate the difference between the locally updated parameters and the globally updated parameters and use it as the local parameter update amount. S4.6: Encrypt or perturb the update amount of each local parameter before uploading it to the central node. Then, according to the data scale or reliability of each newly added turbine unit, the uploaded update information is weighted and averaged to generate new global parameters. The updated parameters are then distributed to the simulation operation model of each turbine unit, replacing the original parameters, and waiting for rapid adaptation to subsequent new units or new faults.

Images