A generator fault early warning method, device and storage medium
By combining multi-resolution wavelet transform and directed graph neural network with dynamic calibration and feature fusion, the accuracy and robustness of generator fault diagnosis under varying operating conditions are solved, achieving efficient early warning and global collaborative perception of early and minor faults, and improving the intelligent operation and maintenance level of generator sets.
Patent Information
- Application Number
- CN202611319432.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing generator fault diagnosis technologies have poor generalization and adaptability under varying operating conditions, cannot effectively capture early and weak fault characteristics, and fail to fully utilize the spatial topological response correlation between multiple measurement points, resulting in high false alarm rates and low recognition rates.
Multi-resolution continuous wavelet transform is used to extract time-frequency feature matrix, and directed graph neural network is used to extract multi-measurement point collaborative spatial features. A dynamic calibration function is constructed through operating condition parameters to remove operating condition interference. A dynamic weighting mechanism is introduced to fuse features and generate multi-dimensional fused features to output early warning results.
It significantly improves the accuracy and robustness of generator fault early warning under complex and variable operating conditions, realizes high-confidence early warning of early and weak faults, enhances the ability to identify complex coupled faults, reduces reliance on expert experience, and improves the efficiency of intelligent operation and maintenance.
Smart Images

Figure CN123068932A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a generator fault early warning method, device, and storage medium, belonging to the field of generator equipment condition monitoring and intelligent operation and maintenance technology. Background Technology
[0002] The main technical solutions in the field of generator condition monitoring and fault diagnosis are as follows:
[0003] Option 1: Technology based on traditional time-domain / frequency-domain analysis and fixed threshold alarm
[0004] Signals are acquired through individual vibration sensors, and harmonic components or time-domain peak values are extracted using Fourier transform. Fixed vibration amplitude alarm thresholds are then set based on industry standards or historical experience. The technical process is as follows: "Independent signal acquisition → FFT analysis to obtain amplitude → Fixed threshold comparison → Over-limit alarm".
[0005] Option 2: Diagnostic techniques based on artificial feature engineering and traditional machine learning
[0006] Statistical features such as skewness, kurtosis, and margin indices, as well as envelope spectrum and power spectrum features, are extracted from vibration signals at various measuring points. After feature dimensionality reduction, these features are input into models such as Support Vector Machine (SVM) or Random Forest (RF) for fault classification. The technical process is as follows: "Signal acquisition → Single-channel manual feature extraction → Feature dimensionality reduction → Machine learning classification".
[0007] Option 3: Diagnostic technology based on deep learning of one-dimensional signals at a single measurement point
[0008] The one-dimensional vibration time series signal or its spectrum obtained from a single point is directly input into a deep learning model such as a one-dimensional convolutional neural network (1D-CNN) to perform end-to-end fault feature extraction and state identification. The technical process is as follows: "data acquisition → sample segmentation → deep network adaptive learning → output classification results".
[0009] The existing technology has the following drawbacks:
[0010] Option 1 drawbacks: This method has low computational cost and is the most widely used in industry, but it is essentially a static, one-dimensional threshold judgment. It does not take into account the normal impact of generator active power fluctuations on the fundamental frequency amplitude, nor can it capture weak transient impacts outside the fundamental frequency, resulting in an extremely high false alarm rate of the system under varying operating conditions.
[0011] Option 2's shortcomings: While this option introduces pattern recognition capabilities compared to Option 1, its feature extraction process is highly dependent on expert experience, and the extracted static features are difficult to withstand drastic changes in operating conditions. Furthermore, the features of each channel are typically simply pieced together, failing to establish a physical mechanism-based correlation model between measurement points.
[0012] Option 3 drawbacks: While this method boasts a high degree of automation, its network structure typically lacks an explicit load condition decoupling module. Once the model is trained on a dataset under fixed loads, its performance deteriorates sharply (poor generalization ability) when faced with variable load conditions unseen in actual production. Furthermore, such models are often trained only on single-channel data, neglecting the interconnectivity of multiple sensors in the spatial topology, thus limiting their ability to perceive the global state of complex systems.
[0013] Based on the above-mentioned existing technical solutions, their main inherent drawbacks and limitations in achieving accurate early warning under complex generator operating conditions are as follows:
[0014] (1) Lack of resistance to operating condition interference and poor generalization and adaptability. Changes in generator speed and active power during operation will cause natural drift of the vibration signal baseline. The static threshold of Scheme 1, the shallow model of Scheme 2, and the conventional end-to-end deep network of Scheme 3 all lack explicit modeling of operating condition variables and "operating condition normalization" processing mechanism. This results in the boundary obtained by the model under specific operating conditions being unable to adapt to the data distribution shift caused by load fluctuations, and is very likely to misjudge normal operating condition fluctuations as faults, and cannot meet the early warning requirements under variable operating condition scenarios.
[0015] (2) The feature extraction dimension is single and the resolution is insufficient, making it insensitive to early weak changes. Early faults often only show weak impact characteristics within a specific very short time window or high frequency band. Existing solutions mostly use global time-frequency analysis or conventional convolution kernels, lacking multi-resolution analysis methods. They are difficult to accurately characterize transient impacts in signals and easily submerge early fault characteristics in strong rotating fundamental frequencies and background noise, causing the system to only issue an alarm when the fault has developed to the middle and late stages, thus losing the meaning of "early warning".
[0016] (3) The spatial topological response correlation between multiple measurement points is ignored, resulting in incomplete information perception. Existing technologies, whether setting thresholds or extracting features, are mostly conducted independently at the single measurement point level, which is akin to "the blind men and the elephant." The vibration state of the generator stator and rotor system has strong spatial correlation (such as the vibration phase difference and mode shape correlation coefficient between different measurement points). Existing solutions fail to effectively utilize graph topology and other methods to model the spatial correlation of multi-source sensors, and cannot fully explore the potential of multi-dimensional data collaboration to improve early warning performance, resulting in limited recognition rate for complex composite faults. Summary of the Invention
[0017] The purpose of this invention is to overcome the shortcomings of the prior art and provide a generator fault early warning method, device and storage medium, which effectively overcomes the problems of variable operating condition interference, difficulty in capturing early weak features and lack of spatial correlation of multiple measurement points, and significantly improves the accuracy and robustness of fault early warning.
[0018] To achieve the above objectives, the present invention is implemented using the following technical solution:
[0019] In a first aspect, the present invention provides a generator fault early warning method, comprising:
[0020] Acquire the multi-source vibration signal sequence of the generator;
[0021] The multi-source vibration signal sequence is used as input, and time-frequency decomposition is performed using multi-resolution continuous wavelet transform to extract the time-frequency feature matrix that characterizes weak transient impacts.
[0022] The time-frequency feature matrix is input into a pre-constructed directed graph neural network, and spatial features of multi-measurement point collaboration are extracted through message passing and feature aggregation between nodes.
[0023] Synchronously acquire the real-time operating condition parameters of the generator, construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features.
[0024] The desensitized spatiotemporal features are dynamically and adaptively fused with pre-extracted traditional hand-crafted features to generate multi-dimensional fused features;
[0025] The generator status assessment and early warning results are output based on the multi-dimensional fusion characteristics.
[0026] Furthermore, using the multi-source vibration signal sequence as input, time-frequency decomposition is performed using multi-resolution continuous wavelet transform to extract the time-frequency feature matrix characterizing the weak transient impact, including:
[0027] The multi-source vibration signal sequence is decomposed into multiple scales using multi-resolution continuous wavelet transform. The mathematical expression for the continuous wavelet transform is as follows:
[0028] ;
[0029] In the formula, a is the scaling factor, which controls the scaling to correspond to the frequency, and b is the translation factor, which controls the time position. (t) is the complex conjugate of the mother wavelet function;
[0030] By optimizing the two-dimensional scaling and translation parameters of the multi-source vibration signal sequence in the frequency and time domains, the weak energy mutation matrix caused by early local faults in a specific high-frequency band is extracted, and the time-frequency feature matrix is output.
[0031] Furthermore, the time-frequency feature matrix is input into a pre-constructed directed graph neural network. Through message passing and feature aggregation between nodes, multi-point collaborative spatial features are extracted, including:
[0032] The time-frequency feature matrix is used as the initial features of the nodes in the input layer of the directed graph neural network. The directed graph neural network is pre-constructed by mapping the sensing points of each part of the generator to nodes in a graph data structure. The graph topology model is represented as G=(V,E,A), where the set of nodes is... There are N vibration measurement points, the edge set E represents the physical and energy conduction paths between the measurement points, and the adjacency matrix A characterizes the correlation strength between the measurement points.
[0033] The directed graph neural network uses a learnable adjacency matrix to characterize the physical connections and energy conduction path strength between measurement points; its first... The feature state update rule for a layer is expressed as follows:
[0034] ;
[0035] In the formula, For the first The node feature matrix of the layer For the first The node feature matrix of the layer For adjacency matrices with added self-connections, Let it be its corresponding degree matrix. The weight parameter matrix is a learnable matrix. It is a non-linear activation function;
[0036] After multi-layer graph convolution operation, spatial features including vibration phase difference and mode shape correlation coefficient under synchronous response of multiple measurement points are extracted, and the spatial features that deeply integrate the spatial correlation of measurement points are output.
[0037] Furthermore, the real-time operating condition parameters include the generator's active power and speed; the operating condition segment calibration and normalization process refers to generating the expected mean and variance under the current operating condition based on the active power and speed mapping, and using this as a benchmark to perform numerical scaling and offset desensitization on the spatiotemporal features, removing the nonlinear drift interference of the vibration response baseline caused by dynamic load changes, and generating desensitized spatiotemporal features.
[0038] Furthermore, the desensitized spatiotemporal features are dynamically and adaptively fused with pre-extracted traditional handcrafted features to generate multidimensional fused features, including:
[0039] The time-domain statistical features extracted based on time-domain statistical analysis are obtained as the traditional manual features. The time-domain statistical features include skewness, kurtosis, and margin indices.
[0040] By introducing an attention-driven dynamic weighting mechanism, the saliency scores of the desensitized spatiotemporal features and the traditional handmade features are calculated at different fault evolution stages. The fusion weight coefficients of each feature are dynamically calculated and output using an attention network. Based on the fusion weight coefficients, the heterogeneous features are adaptively and optimally matched to generate the multidimensional fusion features.
[0041] Furthermore, the step of outputting generator status assessment and early warning results based on the multi-dimensional fusion features includes:
[0042] The multidimensional fusion features are input into a pre-trained classifier, which outputs a device-level health assessment index and a corresponding warning level. The warning level includes four levels: normal, attention, abnormal, and severe. The health assessment index is a continuous value between 0 and 1, used to characterize the degree of generator degradation.
[0043] In a second aspect, the present invention provides a generator fault early warning system for implementing the generator fault early warning method described in any one of the foregoing claims, comprising:
[0044] The data acquisition module is used to acquire the multi-source vibration signal sequence of the generator and the real-time operating parameters.
[0045] The time-frequency feature extraction module is used to take the multi-source vibration signal sequence as input, perform time-frequency decomposition using multi-resolution continuous wavelet transform, and extract the time-frequency feature matrix that characterizes the weak transient impact.
[0046] The spatial feature extraction module is used to input the time-frequency feature matrix into a pre-constructed directed graph neural network, and extract the spatial features of multiple measurement points through message passing and feature aggregation between nodes.
[0047] The operating condition normalization module is used to construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization processing on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features.
[0048] The adaptive fusion module is used to dynamically and adaptively fuse the desensitized spatiotemporal features with the pre-extracted traditional handmade features to generate multi-dimensional fused features.
[0049] The early warning output module is used to output generator status assessment and early warning results based on the multi-dimensional fusion features.
[0050] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.
[0051] Fourthly, the present invention provides an electronic device, comprising:
[0052] Memory, used to store computer programs / instructions;
[0053] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.
[0054] Fifthly, the present invention provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of any of the methods described above.
[0055] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0056] 1. This invention provides a generator fault early warning method, device and storage medium. By constructing a full-link closed-loop early warning architecture of "time-frequency feature extraction, spatial feature extraction, operating condition normalization, dynamic weight fusion and early warning output", it performs systematic processing on multi-source vibration signals of generators from local transient impact capture to global spatial collaborative perception, and from operating condition interference stripping to heterogeneous feature adaptive fusion. This achieves high-confidence early warning of early weak faults of generators in complex variable operating conditions, and significantly improves cross-scenario generalization ability and operation and maintenance intelligence level.
[0057] 2. This invention employs multi-resolution continuous wavelet transform to perform multi-scale time-frequency decomposition of multi-source vibration signals of generators, and optimizes two-dimensional scaling and translation parameters in the frequency and time domains to extract the energy weak mutation matrix within a specific high-frequency band. This enables high-resolution capture of early weak transient impacts hidden in strong fundamental frequencies and background noise, breaking through the bottleneck of insensitivity to early weak mutations and gaining a critical time window for condition-based maintenance.
[0058] 3. This invention maps sensing points at various locations to nodes in a graph data structure and constructs a directed graph neural network. The time-frequency feature matrix is used as the initial feature input for the nodes. Through multi-layer graph convolution operations, message passing and feature aggregation between nodes are performed, and spatial features, including vibration phase difference and mode shape correlation coefficient under synchronous response of multiple measuring points, are extracted. This realizes high-dimensional feature reconstruction from isolated single-point monitoring to global spatial collaboration, overcomes the "blind men and the elephant" information island phenomenon in traditional schemes, and greatly enhances the system's comprehensive identification capability for complex coupled faults.
[0059] 4. This invention constructs a dynamic calibration function by synchronously accessing real-time active power and speed parameters of the generator. Based on the expected mean and variance under the current operating conditions, the spatiotemporal features are numerically scaled and desensitized by offset. This removes the nonlinear drift interference of the vibration response baseline caused by dynamic load changes from the mathematical level, solving the problem of frequent false alarms and missed alarms caused by the baseline drift of the early warning system in variable operating conditions. It significantly improves the robustness and cross-scenario generalization ability of the model in complex operating scenarios such as deep peak shaving.
[0060] 5. This invention introduces an attention-driven dynamic weighting mechanism to calculate the significance scores of the desensitized spatiotemporal features and traditional manual features at different fault evolution stages. It uses an attention network to dynamically output the fusion weight coefficients of each feature and performs adaptive optimal matching, realizing the scientific and intelligent fusion of heterogeneous features. The entire process does not require manual intervention in feature selection or frequent threshold adjustments, forming a highly automated "perception-calibration-fusion-decision" closed-loop system. This significantly reduces the reliance on expert operation and maintenance experience and significantly improves the intelligent operation and maintenance efficiency of generator sets. Attached Figure Description
[0061] Figure 1 This is a flowchart of a generator fault early warning method provided in an embodiment of the present invention. Detailed Implementation
[0062] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.
[0063] Example 1
[0064] This embodiment introduces a generator fault early warning method, including:
[0065] Acquire the multi-source vibration signal sequence of the generator;
[0066] The multi-source vibration signal sequence is used as input, and time-frequency decomposition is performed using multi-resolution continuous wavelet transform to extract the time-frequency feature matrix that characterizes weak transient impacts.
[0067] The time-frequency feature matrix is input into a pre-constructed directed graph neural network, and spatial features of multi-measurement point collaboration are extracted through message passing and feature aggregation between nodes.
[0068] Synchronously acquire the real-time operating condition parameters of the generator, construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features.
[0069] The desensitized spatiotemporal features are dynamically and adaptively fused with pre-extracted traditional hand-crafted features to generate multi-dimensional fused features;
[0070] The generator status assessment and early warning results are output based on the multi-dimensional fusion characteristics.
[0071] The generator fault early warning method provided in this embodiment involves the following steps in its application process:
[0072] Step 1: Time-frequency feature extraction and weak mutation capture based on multi-resolution wavelet basis;
[0073] When a generator operates under varying operating conditions, conventional single-time-domain or frequency-domain features are insufficient to extract transient changes in early faults. This step involves acquiring a multi-source vibration signal sequence of the generator. Multi-scale spatial decomposition is performed using the Continuous Wavelet Transform (CWT) with multi-resolution resolution. The mathematical expression for the wavelet transform is:
[0074]
[0075] In the formula, a is the scaling factor (controlling the scaling to correspond to the frequency), and b is the translation factor (controlling the time position). (t) is the complex conjugate of the mother wavelet function. This is achieved by performing operations in the frequency and time domains... Two-dimensional parameter optimization is performed, which can accurately characterize the transient impacts hidden in the signal and the weak energy changes caused by early hidden faults, and output a time-frequency feature matrix with high fidelity resolution. This enables the extracted signal features to have preliminary noise resistance at the underlying layer.
[0076] Step 2: Multi-point collaboration and spatial topology modeling based on directed graph neural networks;
[0077] For complex multi-point measurement scenarios involving generators, this scheme constructs a directed graph space topology model G=(V,E,A). The node set... Representing N vibration measurement points, the edge set E represents the physical and energy conduction paths between the measurement points, and the adjacency matrix... This step characterizes the correlation strength between measurement points. To extract higher-order spatial features such as vibration phase difference and mode shape correlation coefficient under multi-point synchronous response, a graph convolutional network (GCN) is used for message passing and aggregation of node information. The feature state update rule for a layer is expressed as follows:
[0078]
[0079] In the formula, For the first The node feature matrix of the layer, For adjacency matrices with added self-connections, Let it be its corresponding degree matrix. The weight parameter matrix is a learnable matrix. It is a non-linear activation function. After multi-layer graph convolution, the output is a feature representation that deeply integrates the spatial correlation of measurement points. .
[0080] Step 3: Operating condition segment calibration and normalization desensitization for unsteady operating parameters;
[0081] To address the vibration response baseline drift issue caused by unsteady-state operation, this solution employs a condition normalization algorithm to eliminate variable operating condition interference. The system synchronously accesses the generator control terminal command stream and, combined with real-time key state parameters such as active power and speed, constructs a dynamic optimization mapping matrix to perform fine-grained condition segmentation calibration on the previously extracted spatiotemporal vibration features. This normalization mechanism effectively isolates and eliminates the nonlinear effects of dynamic changes in grid load and speed fluctuations on the absolute amplitude of generator vibration.
[0082] Step 4: Introduce multi-dimensional parameter adaptive fusion and intelligent early warning with a dynamic weighting mechanism;
[0083] Building upon high-fidelity spatiotemporal features, this method further integrates traditional handcrafted features such as time-domain statistics across different fault modes. To address the varying representational contributions of heterogeneous parameters under different fault modes, a dynamic weighting mechanism is introduced within the diagnostic model. This mechanism adaptively allocates and adjusts the weights of feature vectors across each dimension based on real-time operating conditions and feature saliency. Finally, based on the aforementioned multi-dimensional feature fusion and operating condition adaptive mechanism, the system outputs equipment-level health assessments and early warning levels. This architecture significantly enhances the robustness of the early warning algorithm, enabling the model to maintain excellent generalization ability and high-confidence early warning accuracy even under complex and unknown production conditions.
[0084] Compared with the prior art, this embodiment has the following significant advantages:
[0085] Superior Adaptability and Anti-interference Capability: By introducing an operating condition normalization mechanism and a dynamic calibration function, this invention can automatically isolate the nonlinear interference of active power fluctuations and speed changes on the vibration baseline from a physical perspective. This solves the problem of frequent false alarms and missed alarms caused by reference drift in the early warning system under complex operating scenarios such as deep peak shaving of generators, and significantly improves the robustness of the model under varying operating conditions and its cross-scenario generalization ability.
[0086] Extremely high-resolution early-stage weak fault detection capability: By leveraging multi-resolution wavelet bases to deeply characterize non-stationary signals, this invention can accurately extract hidden transient impulses and weak energy mutations caused by early-stage concealed faults. This high-resolution sensing overcomes the limitations of traditional fixed threshold or conventional spectral analysis in its insensitivity to abrupt signals, ensuring reliable early warning at the fault's nascent stage and gaining a critical time window for condition-based maintenance.
[0087] Achieving holographic perception and coordination at the spatial topology level: By utilizing directed graph neural networks to model the physical and energy transmission relationships between measurement points, this invention transforms isolated sensor data into high-order features with spatial topological consistency, uncovering the correlation between vibration phase differences and mode shapes between measurement points. This leap from "single-point monitoring" to "global coordination" overcomes the phenomenon of information silos and greatly enhances the system's comprehensive ability to identify complex coupled faults.
[0088] The scientific and highly intelligent nature of decision fusion: By introducing an attention-driven dynamic weighting mechanism, an adaptive optimal ratio of time-frequency features, spatial features, and traditional manual features is achieved. The entire process requires no manual intervention in feature selection or frequent manual threshold adjustments, forming a highly automated "perception-calibration-fusion-decision" closed-loop system. This significantly reduces reliance on expert operation and maintenance experience and substantially improves the intelligent operation and maintenance efficiency of generator sets.
[0089] Example 2
[0090] This embodiment provides a generator fault early warning system for implementing the generator fault early warning method described in any one of Embodiment 1, including:
[0091] The data acquisition module is used to acquire the multi-source vibration signal sequence of the generator and the real-time operating parameters.
[0092] The time-frequency feature extraction module is used to take the multi-source vibration signal sequence as input, perform time-frequency decomposition using multi-resolution continuous wavelet transform, and extract the time-frequency feature matrix that characterizes the weak transient impact.
[0093] The spatial feature extraction module is used to input the time-frequency feature matrix into a pre-constructed directed graph neural network, and extract the spatial features of multiple measurement points through message passing and feature aggregation between nodes.
[0094] The operating condition normalization module is used to construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization processing on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features.
[0095] The adaptive fusion module is used to dynamically and adaptively fuse the desensitized spatiotemporal features with the pre-extracted traditional handmade features to generate multi-dimensional fused features.
[0096] The early warning output module is used to output generator status assessment and early warning results based on the multi-dimensional fusion features.
[0097] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.
[0098] Example 3
[0099] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.
[0100] Example 4
[0101] This embodiment provides an electronic device, including:
[0102] Memory, used to store computer programs / instructions;
[0103] A processor for executing the computer program / instructions to implement the steps of the method described in Embodiment 1.
[0104] Example 5
[0105] This embodiment provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the method described in Embodiment 1.
[0106] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
[0107] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0108] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.
Claims
1. A generator fault early warning method, characterized in that, include: Acquire the multi-source vibration signal sequence of the generator; The multi-source vibration signal sequence is used as input, and time-frequency decomposition is performed using multi-resolution continuous wavelet transform to extract the time-frequency feature matrix that characterizes weak transient impacts. The time-frequency feature matrix is input into a pre-constructed directed graph neural network, and spatial features of multi-measurement point collaboration are extracted through message passing and feature aggregation between nodes. Synchronously acquire the real-time operating condition parameters of the generator, construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features. The desensitized spatiotemporal features are dynamically and adaptively fused with pre-extracted traditional hand-crafted features to generate multi-dimensional fused features; The generator status assessment and early warning results are output based on the multi-dimensional fusion characteristics.
2. The generator fault early warning method according to claim 1, characterized in that, Using the multi-source vibration signal sequence as input, time-frequency decomposition is performed using multi-resolution continuous wavelet transform to extract the time-frequency feature matrix characterizing weak transient impacts, including: The multi-source vibration signal sequence is decomposed into multiple scales using multi-resolution continuous wavelet transform. The mathematical expression for the continuous wavelet transform is as follows: ; In the formula, a is the scaling factor, which controls the scaling to correspond to the frequency, and b is the translation factor, which controls the time position. (t) is the complex conjugate of the mother wavelet function; By optimizing the two-dimensional scaling and translation parameters of the multi-source vibration signal sequence in the frequency and time domains, the weak energy mutation matrix caused by early local faults in a specific high-frequency band is extracted, and the time-frequency feature matrix is output.
3. The generator fault early warning method according to claim 1, characterized in that, The time-frequency feature matrix is input into a pre-constructed directed graph neural network. Through message passing and feature aggregation between nodes, multi-point collaborative spatial features are extracted, including: The time-frequency feature matrix is used as the initial features of the nodes in the input layer of the directed graph neural network. The directed graph neural network is pre-constructed by mapping the sensing points of each part of the generator to nodes in a graph data structure. The graph topology model is represented as G=(V,E,A), where the set of nodes is... There are N vibration measurement points, the edge set E represents the physical and energy conduction paths between the measurement points, and the adjacency matrix A characterizes the correlation strength between the measurement points. The directed graph neural network uses a learnable adjacency matrix to characterize the physical connections and energy conduction path strength between measurement points; its first... The feature state update rule for a layer is expressed as follows: ; In the formula, For the first The node feature matrix of the layer For the first The node feature matrix of the layer For adjacency matrices with added self-connections, Let it be its corresponding degree matrix. The weight parameter matrix is a learnable matrix. It is a non-linear activation function; After multi-layer graph convolution operation, spatial features including vibration phase difference and mode shape correlation coefficient under synchronous response of multiple measurement points are extracted, and the spatial features that deeply integrate the spatial correlation of measurement points are output.
4. The generator fault early warning method according to claim 1, characterized in that, The real-time operating parameters include the generator's active power and speed; the operating condition segment calibration and normalization process refers to generating the expected mean and variance under the current operating condition based on the active power and speed mapping, and using this as a benchmark to perform numerical scaling and offset desensitization on the spatiotemporal features, removing the nonlinear drift interference of the vibration response baseline caused by dynamic load changes, and generating desensitized spatiotemporal features.
5. The generator fault early warning method according to claim 1, characterized in that, The desensitized spatiotemporal features are dynamically and adaptively fused with pre-extracted traditional handcrafted features to generate multidimensional fused features, including: The time-domain statistical features extracted based on time-domain statistical analysis are obtained as the traditional manual features. The time-domain statistical features include skewness, kurtosis, and margin indices. By introducing an attention-driven dynamic weighting mechanism, the saliency scores of the desensitized spatiotemporal features and the traditional handmade features are calculated at different fault evolution stages. The fusion weight coefficients of each feature are dynamically calculated and output using an attention network. Based on the fusion weight coefficients, the heterogeneous features are adaptively and optimally matched to generate the multidimensional fusion features.
6. The generator fault early warning method according to claim 1, characterized in that, The step of outputting generator status assessment and early warning results based on the multi-dimensional fusion features includes: The multidimensional fusion features are input into a pre-trained classifier, which outputs a device-level health assessment index and a corresponding warning level. The warning level includes four levels: normal, attention, abnormal, and severe. The health assessment index is a continuous value between 0 and 1, used to characterize the degree of generator degradation.
7. A generator fault early warning system, characterized in that, include: The data acquisition module is used to acquire the multi-source vibration signal sequence of the generator and the real-time operating parameters. The time-frequency feature extraction module is used to take the multi-source vibration signal sequence as input, perform time-frequency decomposition using multi-resolution continuous wavelet transform, and extract the time-frequency feature matrix that characterizes the weak transient impact. The spatial feature extraction module is used to input the time-frequency feature matrix into a pre-constructed directed graph neural network, and extract the spatial features of multiple measurement points through message passing and feature aggregation between nodes. The operating condition normalization module is used to construct a dynamic calibration function based on the operating condition parameters, perform operating condition segmentation calibration and normalization processing on the spatiotemporal features obtained by fusing the time-frequency feature matrix and the spatial features, remove the interference caused by operating condition fluctuations, and generate desensitized spatiotemporal features. The adaptive fusion module is used to dynamically and adaptively fuse the desensitized spatiotemporal features with the pre-extracted traditional handmade features to generate multi-dimensional fusion features. The early warning output module is used to output generator status assessment and early warning results based on the multi-dimensional fusion features.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-6.
9. An electronic device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-6.
10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.
