A method and system for intelligent assessment of the operating status of natural gas generator sets

Abstract

This invention belongs to the field of intelligent status assessment technology for generator sets, and discloses a method and system for intelligent status assessment of natural gas generator sets. The method includes: acquiring multi-physics field operation data; reconstructing the phase space of the multi-physics field operation data based on different time scales to obtain a set of physically perceived state point clouds; performing complex filtering on the point clouds at each time scale, and conducting multi-scale topological evolution analysis of the filtered complex sequences; inputting the topological feature vectors into a topological random forest model for training, and performing joint inference to output the posterior distribution of fault modes; calculating a generator set health index, and combining the generator set health index with a dynamic early warning threshold to achieve the assessment of the operating status of the natural gas generator set. This invention can deeply explore the dynamic evolution laws of different physical processes such as combustion, mechanical, thermal, and degradation, and effectively capture early weak faults and multi-field coupling anomalies that are difficult to identify using traditional methods.

Description

Technical Field

[0001] This invention relates to the field of intelligent status assessment technology for generator sets, and particularly to a method and system for intelligent status assessment of natural gas generator sets. Background Technology

[0002] Natural gas generator sets are highly efficient and clean power generation devices that use natural gas as fuel, drive generators through gas turbines, and often combine with waste heat boilers to achieve a combined cycle. They have advantages such as low emissions, high efficiency, fast start-up and shutdown, and good peak-shaving performance, and have become an important supporting power source in modern power systems. However, gas turbines operate under extreme conditions of high temperature, high pressure, and high speed for extended periods. Therefore, assessing the operating status and health management of natural gas generator sets has become a key technical means to ensure their safe and reliable operation.

[0003] However, most existing methods for assessing the operating status of natural gas generator sets are based on single sensor thresholds or traditional statistical features, which make it difficult to deeply explore the dynamic evolution of different physical processes such as combustion, mechanical, thermal, and degradation. Furthermore, they are not robust to noise and varying operating conditions, which can easily lead to false alarms and missed alarms. Traditional methods have significant limitations in identifying early weak faults and multi-field coupling anomalies, and cannot effectively capture the implicit topological structures in multi-dimensional data. At the same time, single-scale feature extraction methods result in incomplete information, lack an effective fusion mechanism for multi-scale information, poor interpretability of diagnostic results, and inability to quantify uncertainty, making it difficult to meet the actual needs of natural gas generator sets for high-reliability and high-precision intelligent status assessment.

[0004] Therefore, how to provide a method and system for intelligent assessment of the operating status of natural gas generator sets is an urgent problem to be solved. Summary of the Invention

[0005] This invention provides a method and system for intelligent assessment of the operating status of a natural gas generator set, in order to solve the aforementioned technical problems in the prior art.

[0006] According to a first aspect of the present invention, a method for intelligent assessment of the operating status of a natural gas generator set is provided.

[0007] In one embodiment, a smart condition assessment method for natural gas generator set operation includes:

[0008] Multiphysics operation data is acquired, and the data is divided into several time scales according to the fault physical characteristics of the natural gas generator set. Based on different time scales, the multiphysics operation data is reconstructed in phase space to obtain a set of physical perception state point clouds.

[0009] Complex filtering is performed on each time-scale point cloud in the state point cloud set to obtain a filtered complex sequence. Multi-scale unit operation state topology evolution analysis is then performed on the filtered complex sequence to obtain the topology feature vector.

[0010] The topological feature vectors at different time scales are input into the topological random forest model for training, and the preliminary operational failure probability at each time scale is output. The preliminary operational failure probability at each time scale is then jointly inferred to output the posterior distribution of the failure mode.

[0011] The unit health index is based on the fault mode posterior distribution and is combined with the dynamic early warning threshold to realize the operation status assessment of natural gas generator units.

[0012] In one embodiment, multiphysics operation data is acquired, divided into several time scales based on the fault physical characteristics of the natural gas generator set, and phase space reconstruction is performed on the multiphysics operation data based on different time scales to obtain a set of physical sensing state point clouds, including:

[0013] Acquire multi-physics field operation data of natural gas generator sets, perform physical validity verification, multi-rate synchronization and operating condition normalization on the multi-physics field operation data, and obtain standardized multi-physics field time sequence signals;

[0014] Several timescales include: combustion dynamics, rotating machinery, thermodynamic cycles, and long-term performance degradation timescales;

[0015] Based on the physical characteristics of generator set faults, four time scales are divided: combustion dynamics, rotating machinery, thermodynamic cycle, and long-term performance degradation. At each time scale, a multivariate combination of dominant physical processes is selected by combining physical conservation laws and fault mechanisms. Furthermore, the time delay of the standardized multiphysics field time series signal is embedded with physical perception by combining kinematic and thermodynamic constraints to generate the initial state point cloud set for each time scale.

[0016] Topological denoising, physical manifold projection, and uniform resampling optimization are performed on the point clouds at each time scale in the initial state point cloud set to obtain an optimized state point cloud set. Cross-scale time indexing and physical association edge construction are then performed on the optimized state point cloud set to obtain a multi-scale physically-aware state point cloud set.

[0017] In one embodiment, complex filtering is performed on each time-scale point cloud in the state point cloud set to obtain a filtered complex sequence. Multi-scale unit operation state topology evolution analysis is then performed on the filtered complex sequence to obtain a topology feature vector including:

[0018] The complex type and distance metric are determined based on the physical characteristics of each time scale, and a filtered complex sequence of unit operating status is generated by combining the inherent dimensions and physical constraints of the state point cloud set.

[0019] The filtered complex sequence is subjected to continuous homology parallel computation to generate a multi-scale state persistence graph, and the multi-scale state persistence graph is transformed into an initial topological feature set for the topological evolution of the unit's operating state.

[0020] The topological features in the initial topological feature set are concatenated to obtain the concatenated topological features. The concatenated topological features are then standardized according to the physical range of historical normal operating conditions at each scale to obtain a fixed-dimensional topological feature vector.

[0021] In one embodiment, generating a filtered complex sequence of unit operating states by combining the intrinsic dimensions and physical constraints of the state point cloud set includes:

[0022] The physical perceived distance between all point pairs in the state point cloud set is calculated based on the distance metric, forming a pairwise distance matrix. The lower and upper limits of the filtering parameters are set from the pairwise distance matrix in combination with the inherent dimension and physical constraints of the state point cloud set, thus determining the range of the filtering parameters.

[0023] The step size strategy is selected based on the scale dynamics characteristics, and the range of filtering parameters is discretized and sampled based on the step size strategy to generate a discrete sequence of filtering parameters.

[0024] Based on the complex type and physical sensing distance, a complex is constructed for each filter parameter in the discrete filter parameter sequence to obtain the filtered complex sequence.

[0025] In one embodiment, performing persistent cohomological parallel computation on the filtered complex sequence to generate a multi-scale state persistence graph includes:

[0026] Configure the continuous cohomology calculation parameters according to the time scale type of the filtered complex sequence, and calculate the birth and death filtering values ​​of all topological generators in parallel for each time scale of the filtered complex sequence based on the continuous cohomology calculation parameters.

[0027] The birth and death filter values ​​of all topological generators extracted at each time scale are classified and organized according to the topological dimension. Then, the noise features of the persistence of the organized birth and death filter values ​​are removed to obtain the effective topological generators.

[0028] Each valid topology generator is encapsulated into a topology generator lifecycle tuple according to a preset format. All topology generator lifecycle tuples are then summarized according to each time scale to obtain a standardized state persistence graph. The standardized state persistence graphs at all scales are then integrated to obtain a multi-scale state persistence graph.

[0029] In one embodiment, the initial topological feature set for transforming the multi-scale state persistence graph into the unit operating state topology evolution includes:

[0030] The initial topological feature set includes: the first set of topological features, the second set of topological features, the third set of topological features, and the fourth set of topological features;

[0031] The multi-scale state persistence graph is separated according to the homology dimension to obtain zero-order, first-order and second-order state persistence subgraphs, and statistical features are extracted from each state persistence subgraph to obtain the first set of topological features.

[0032] Gaussian kernel density estimation is used to transform the state persistent subgraph into a state persistent image with a fixed resolution, and the state persistent image is flattened into a second set of topological features.

[0033] The Betty curve is calculated on the filtering parameter axis of the multi-scale state persistence graph, and a third set of topological features is extracted based on the Betty curve.

[0034] Multidimensional topological entropy is calculated based on the lifetime distribution of the state-persistent subgraph, and a fourth set of topological features is generated based on the multidimensional topological entropy calculation results.

[0035] In one embodiment, topological feature vectors at different time scales are input into a topological random forest model for training, outputting preliminary operational failure probabilities at each time scale. Joint inference is then performed on these preliminary operational failure probabilities at each time scale to output the posterior distribution of the failure mode, including:

[0036] Construct a topological random forest by expanding and grouping the topological feature vectors to obtain training samples and feature subspaces. Train the topological decision tree in the topological random forest based on the training samples and feature subspaces to obtain the trained topological decision tree.

[0037] The trained topology decision tree outputs the failure probability of a single tree, and combines the out-of-bag sample error and topology stability of the topology decision tree to evaluate the prediction error and weight of each topology decision tree. The prediction error and weight of each topology decision tree are weighted and averaged to obtain the preliminary operational failure probability at each time scale.

[0038] The association weights are set based on the physical correlation strength between each time scale and the unit's failure mode. The initial operational failure probability at each time scale is multiplied by the corresponding association weight, and then summed and normalized to obtain the posterior distribution of the failure mode of the unit's operating status.

[0039] In one embodiment, a topological random forest is constructed, and the topological feature vectors are expanded and grouped to obtain training samples and feature subspaces. The topological decision trees in the topological random forest are then trained based on the training samples and feature subspaces to obtain trained topological decision trees, including:

[0040] A hierarchical topological random forest is constructed based on the time scale and a preset number of topological decision trees. The topological feature vector is expanded using the topology-aware bootstrap sampling method to generate training samples.

[0041] Based on the homology dimension, the topological feature vectors of each time scale are grouped to obtain several semantic feature groups. Features are then randomly selected from these semantic feature groups to form a feature subspace.

[0042] Based on the Wasserstein distance splitting criterion, the topology decision tree is recursively split and node-grown using the dimension in the feature subspace on the training samples to obtain the trained topology decision tree.

[0043] In one embodiment, assessing the operating status of a natural gas generator set by calculating the unit health index based on the posterior distribution of failure modes and combining the unit health index with a dynamic early warning threshold includes:

[0044] Based on the pre-set health scores for each failure mode, the probabilities of each type of failure mode posterior distribution are used as weights of the corresponding health scores for weighted averaging to obtain the unit health index.

[0045] Identify the current operating condition based on the current operating parameters of the natural gas generator set, and determine the dynamic early warning threshold based on the current operating condition;

[0046] The unit health index is compared with the dynamic early warning threshold, and the health status of the natural gas generator unit is determined based on the comparison results.

[0047] According to a second aspect of the present invention, a smart operating status assessment system for natural gas generator sets is provided.

[0048] In one embodiment, the intelligent operating status assessment system for the natural gas generator set includes:

[0049] The state point cloud construction module is used to acquire multi-physics field operation data. Based on the fault physical characteristics of the natural gas generator set, it is divided into several time scales. Based on different time scales, the phase space of the multi-physics field operation data is reconstructed to obtain a set of state point clouds for physical perception.

[0050] The state topology evolution analysis module is used to perform complex filtering on each time scale point cloud in the state point cloud set to obtain a filtered complex sequence, and to perform multi-scale unit operation state topology evolution analysis on the filtered complex sequence to obtain topology feature vectors.

[0051] The failure probability prediction module is used to input the topological feature vectors at different time scales into the topological random forest model for training, output the preliminary failure probability at each time scale, and perform joint inference on the preliminary failure probability at each time scale to output the posterior distribution of the failure mode.

[0052] The operation status assessment module is used to assess the operation status of natural gas generator sets by combining the unit health index with dynamic early warning thresholds based on the fault mode posterior distribution computer health index.

[0053] According to a third aspect of the present invention, a computer device is provided.

[0054] In some embodiments, the computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described intelligent operating status assessment method for natural gas generator sets.

[0055] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0056] In one embodiment, a computer program is stored on a computer-readable storage medium, and when the computer program is executed by a processor, it implements the steps of the above-described intelligent operating status assessment method for natural gas generator sets.

[0057] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0058] 1. This invention, through multi-timescale partitioning and phase space reconstruction based on physical perception, can deeply explore the dynamic evolution laws of different physical processes such as combustion, mechanics, thermodynamics, and degradation. Utilizing the inherent robustness of topological data analysis to noise and varying operating conditions, it effectively captures early, subtle faults and multi-field coupling anomalies that are difficult to identify using traditional methods. Simultaneously, by employing a multi-scale topological random forest and joint inference mechanism, it not only avoids the incompleteness of single-scale information but also achieves interpretable output of fault modes and quantification of uncertainty. Combined with adaptive adjustment of dynamic thresholds, it significantly improves the accuracy, reliability, and early warning capability of unit condition assessment under complex operating conditions, providing revolutionary technical support for predictive maintenance and intelligent operation of natural gas generator units.

[0059] 2. This invention achieves the identification of the operating status of natural gas generator sets through multi-physics field data fusion and multi-timescale phase space reconstruction. By constructing a physically-sensory state point cloud, it effectively decouples the coupled processes of combustion, mechanics, thermodynamics, and degradation, significantly improving the targeting of feature representation. It utilizes topological random forest to perform noise-resistant modeling of topological features at various scales and combines it with physical correlation weights for joint inference, which not only improves the detection rate of early minor faults but also enhances the interpretability of the model. Furthermore, by linking the health index with the dynamic early warning threshold that is adaptive to operating conditions, it avoids false alarms and missed alarms under changing operating conditions caused by traditional fixed thresholds, greatly improving the safety of unit operation and the level of intelligent operation and maintenance.

[0060] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0061] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0062] Figure 1 This is a flowchart illustrating an intelligent operating status assessment method for a natural gas generator set according to an exemplary embodiment;

[0063] Figure 2 This is a schematic diagram illustrating the structure of an intelligent operating status assessment system for a natural gas generator set according to an exemplary embodiment;

[0064] Figure 3 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment.

[0065] Figure label:

[0066] 201. State point cloud construction module; 202. State topology evolution analysis module; 203. Operational failure probability prediction module; 204. Operational status assessment module. Detailed Implementation

[0067] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0068] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0069] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0070] Figure 1 An embodiment of the intelligent operating status assessment method for a natural gas generator set according to the present invention is shown.

[0071] In this optional embodiment, the intelligent operating status assessment method for the natural gas generator set includes:

[0072] Step S101: Obtain multi-physics operation data, divide it into several time scales according to the fault physical characteristics of the natural gas generator set, and reconstruct the phase space of the multi-physics operation data based on different time scales to obtain a set of physical perception state point clouds.

[0073] Step S102: Perform complex filtering on each time scale point cloud in the state point cloud set to obtain a filtered complex sequence, and perform multi-scale unit operation state topology evolution analysis on the filtered complex sequence to obtain a topology feature vector;

[0074] Step S103: Input the topological feature vectors of different time scales into the topological random forest model for training, output the preliminary operational failure probability of each time scale, and perform joint inference on the preliminary operational failure probability of each time scale to output the posterior distribution of the failure mode.

[0075] Step S104: Based on the fault mode posterior distribution computer group health index, and combining the group health index with the dynamic early warning threshold, the operating status assessment of the natural gas generator set is realized.

[0076] In this optional embodiment, multiphysics operation data is acquired, and the data is divided into several time scales based on the fault physical characteristics of the natural gas generator set. Phase space reconstruction is performed on the multiphysics operation data based on different time scales to obtain a set of physically sensed state point clouds, including:

[0077] Acquire multi-physics field operation data of natural gas generator sets, perform physical validity verification, multi-rate synchronization and operating condition normalization on the multi-physics field operation data, and obtain standardized multi-physics field time sequence signals;

[0078] Several timescales include: combustion dynamics, rotating machinery, thermodynamic cycles, and long-term performance degradation timescales;

[0079] Based on the physical characteristics of generator set faults, four time scales are divided: combustion dynamics, rotating machinery, thermodynamic cycle, and long-term performance degradation. At each time scale, a multivariate combination of dominant physical processes is selected by combining physical conservation laws and fault mechanisms. Furthermore, the time delay of the standardized multiphysics field time series signal is embedded with physical perception by combining kinematic and thermodynamic constraints to generate the initial state point cloud set for each time scale.

[0080] Topological denoising, physical manifold projection, and uniform resampling optimization are performed on the point clouds at each time scale in the initial state point cloud set to obtain an optimized state point cloud set. Cross-scale time indexing and physical association edge construction are then performed on the optimized state point cloud set to obtain a multi-scale physically-aware state point cloud set.

[0081] In this optional embodiment, complex filtering is performed on each time-scale point cloud in the state point cloud set to obtain a filtered complex sequence. Multi-scale unit operation state topology evolution analysis is then performed on the filtered complex sequence to obtain a topology feature vector including:

[0082] The complex type and distance metric are determined based on the physical characteristics of each time scale, and a filtered complex sequence of unit operating status is generated by combining the inherent dimensions and physical constraints of the state point cloud set.

[0083] The filtered complex sequence is subjected to continuous homology parallel computation to generate a multi-scale state persistence graph, and the multi-scale state persistence graph is transformed into an initial topological feature set for the topological evolution of the unit's operating state.

[0084] The topological features in the initial topological feature set are concatenated to obtain the concatenated topological features. The concatenated topological features are then standardized according to the physical range of historical normal operating conditions at each scale to obtain a fixed-dimensional topological feature vector.

[0085] In this optional embodiment, generating a filtered complex sequence of unit operating states by combining the intrinsic dimensions and physical constraints of the state point cloud set includes:

[0086] The physical perceived distance between all point pairs in the state point cloud set is calculated based on the distance metric, forming a pairwise distance matrix. The lower and upper limits of the filtering parameters are set from the pairwise distance matrix in combination with the inherent dimension and physical constraints of the state point cloud set, thus determining the range of the filtering parameters.

[0087] The step size strategy is selected based on the scale dynamics characteristics, and the range of filtering parameters is discretized and sampled based on the step size strategy to generate a discrete sequence of filtering parameters.

[0088] Based on the complex type and physical sensing distance, a complex is constructed for each filter parameter in the discrete filter parameter sequence to obtain the filtered complex sequence.

[0089] In this optional embodiment, performing continuous cohomological parallel computation on the filtered complex sequence to generate a multi-scale state persistence graph includes:

[0090] Configure the continuous cohomology calculation parameters according to the time scale type of the filtered complex sequence, and calculate the birth and death filtering values ​​of all topological generators in parallel for each time scale of the filtered complex sequence based on the continuous cohomology calculation parameters.

[0091] The birth and death filter values ​​of all topological generators extracted at each time scale are classified and organized according to the topological dimension. Then, the noise features of the persistence of the organized birth and death filter values ​​are removed to obtain the effective topological generators.

[0092] Each valid topology generator is encapsulated into a topology generator lifecycle tuple according to a preset format. All topology generator lifecycle tuples are then summarized according to each time scale to obtain a standardized state persistence graph. The standardized state persistence graphs at all scales are then integrated to obtain a multi-scale state persistence graph.

[0093] In this optional embodiment, the initial topological feature set for transforming the multi-scale state persistence graph into the unit operating state topology evolution includes:

[0094] The initial topological feature set includes: the first set of topological features, the second set of topological features, the third set of topological features, and the fourth set of topological features;

[0095] The multi-scale state persistence graph is separated according to the homology dimension to obtain zero-order, first-order and second-order state persistence subgraphs, and statistical features are extracted from each state persistence subgraph to obtain the first set of topological features.

[0096] Gaussian kernel density estimation is used to transform the state persistent subgraph into a state persistent image with a fixed resolution, and the state persistent image is flattened into a second set of topological features.

[0097] The Betty curve is calculated on the filtering parameter axis of the multi-scale state persistence graph, and a third set of topological features is extracted based on the Betty curve.

[0098] Multidimensional topological entropy is calculated based on the lifetime distribution of the state-persistent subgraph, and a fourth set of topological features is generated based on the multidimensional topological entropy calculation results.

[0099] In this optional embodiment, topological feature vectors at different time scales are input into a topological random forest model for training, outputting preliminary operational failure probabilities at each time scale. Joint inference is then performed on these preliminary operational failure probabilities at each time scale to output the posterior distribution of the failure mode, including:

[0100] Construct a topological random forest by expanding and grouping the topological feature vectors to obtain training samples and feature subspaces. Train the topological decision tree in the topological random forest based on the training samples and feature subspaces to obtain the trained topological decision tree.

[0101] The trained topology decision tree outputs the failure probability of a single tree, and combines the out-of-bag sample error and topology stability of the topology decision tree to evaluate the prediction error and weight of each topology decision tree. The prediction error and weight of each topology decision tree are weighted and averaged to obtain the preliminary operational failure probability at each time scale.

[0102] The association weights are set based on the physical correlation strength between each time scale and the unit's failure mode. The initial operational failure probability at each time scale is multiplied by the corresponding association weight, and then summed and normalized to obtain the posterior distribution of the failure mode of the unit's operating status.

[0103] In this optional embodiment, a topological random forest is constructed, and the topological feature vectors are expanded and grouped to obtain training samples and feature subspaces. The topological decision trees in the topological random forest are then trained based on the training samples and feature subspaces to obtain trained topological decision trees, including:

[0104] A hierarchical topological random forest is constructed based on the time scale and a preset number of topological decision trees. The topological feature vector is expanded using the topology-aware bootstrap sampling method to generate training samples.

[0105] Based on the homology dimension, the topological feature vectors of each time scale are grouped to obtain several semantic feature groups. Features are then randomly selected from these semantic feature groups to form a feature subspace.

[0106] Based on the Wasserstein distance splitting criterion, the topology decision tree is recursively split and node-grown using the dimension in the feature subspace on the training samples to obtain the trained topology decision tree.

[0107] In this optional embodiment, the assessment of the operating status of a natural gas generator set based on the fault mode posterior distribution computer health index, and combining the unit health index with a dynamic early warning threshold, includes:

[0108] Based on the pre-set health scores for each failure mode, the probabilities of each type of failure mode posterior distribution are used as weights of the corresponding health scores for weighted averaging to obtain the unit health index.

[0109] Identify the current operating condition based on the current operating parameters of the natural gas generator set, and determine the dynamic early warning threshold based on the current operating condition;

[0110] The unit health index is compared with the dynamic early warning threshold, and the health status of the natural gas generator unit is determined based on the comparison results.

[0111] Figure 2 An embodiment of the intelligent operating status assessment system for a natural gas generator set according to the present invention is shown.

[0112] In this optional embodiment, the natural gas generator set operates an intelligent condition assessment system, including:

[0113] The state point cloud construction module 201 is used to acquire multi-physics field operation data. It divides the natural gas generator set into several time scales according to the fault physical characteristics of the natural gas generator set. Based on different time scales, it performs phase space reconstruction on the multi-physics field operation data to obtain a set of physically perceived state point clouds.

[0114] The state topology evolution analysis module 202 is used to perform complex filtering on each time scale point cloud in the state point cloud set to obtain a filtered complex sequence, and to perform multi-scale unit operation state topology evolution analysis on the filtered complex sequence to obtain a topology feature vector.

[0115] The failure probability prediction module 203 is used to input the topological feature vectors of different time scales into the topological random forest model for training, output the preliminary failure probability of each time scale, and perform joint inference on the preliminary failure probability of each time scale to output the posterior distribution of the failure mode.

[0116] The operation status assessment module 204 is used to assess the operation status of natural gas generator sets by combining the unit health index with the dynamic early warning threshold based on the fault mode posterior distribution computer health index.

[0117] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0118] It should be further explained that corresponding sensors are deployed for the five major physical fields of vibration, temperature, pressure, speed, and chemistry: for the vibration field, acceleration or velocity sensors are deployed at locations such as bearing housings to sample at 10-20kHz; for the temperature field, thermocouples are deployed at locations such as the combustion chamber outlet to sample at 1-10Hz; for the pressure field, dynamic pressure sensors are deployed at the compressor inlet and outlet to sample at 5-20kHz; for the speed field, encoders are deployed on the turbine shaft to sample at 1-5kHz; and for the chemical field, gas analyzers are deployed in the exhaust pipe to sample at 0.1-1Hz. The raw signals of key physical quantities such as vibration acceleration, temperature, static pressure, speed, and NOx concentration are collected, which is the multi-physics field operating data. The original signal is physically validated to remove outliers exceeding reasonable limits and to mark sensor failure periods. Kalman filtering is then used to synchronize multi-rate signals. Using the time axis of the high-frequency reference signal as an anchor point, low-frequency asynchronous signals are mapped to a unified time axis through state estimation, while simultaneously filtering out sampling noise to ensure the continuity and consistency of physical quantities. High-frequency signals are downsampled using anti-aliasing filtering, and low-frequency signals are upsampled using cubic spline interpolation. Measurements under actual operating conditions are corrected to standard reference conditions using thermodynamic similarity criteria to achieve operating condition normalization, eliminating the influence of environmental and load changes, and ensuring the comparability of performance parameters under different operating conditions, resulting in a standardized multiphysics time-series signal.

[0119] It should be further explained that, based on the physical characteristics of natural gas generator set failures, the unit operation process is divided into four time scales: a combustion dynamic scale with a window of 50-200ms, focusing on rapid physical processes such as combustion instability and thermoacoustic oscillations; a rotating machinery scale with a window of 0.5-2s, focusing on shaft-related processes such as rotor dynamics and gear meshing; a thermodynamic cycle scale with a window of 30-120s, covering thermal inertial processes such as start-up or shutdown and load changes; and a performance degradation scale with a window of 1-7 days, capturing slow deterioration processes such as fouling and corrosion.

[0120] It should be further explained that the phase space reconstruction with physical constraints for the combustion dynamic scale is based on mass or energy conservation constraints. Combustion chamber pressure pulsation, pressure change rate, exhaust temperature, fuel flow rate, and pressure gradient are selected as embedded variables. The delay time is determined to be twice the reciprocal of the thermoacoustic mode frequency according to the sampling theorem. Combined with the thermoacoustic attractor dimension, the embedding dimension is determined to be 15, constructing a high-dimensional phase space point cloud. The phase space reconstruction for the rotating machinery scale follows Newton's second law of rotor dynamics. Horizontal or vertical vibration displacement, velocity, instantaneous rotational speed, and bond phase pulse are selected as embedded variables. The delay time is set to one-quarter of the rotational frequency, and the embedding dimension is 8, matching the rotor dynamics degrees of freedom. At the same time, the vibration displacement is converted to polar coordinates, and the number of rotations of the shaft center trajectory is calculated as a topologically invariant feature. The phase space reconstruction for the thermodynamic cycle scale is based on the Brayton cycle isentropic constraint. Pressure ratio, temperature ratio, flow rate ratio, thermal efficiency, and load rate are selected as embedded variables. The delay time is one-fifth of the thermal inertia time constant, and the embedding dimension is 5. Noise points that violate the isentropic constraint are eliminated. The phase space reconstruction of the performance degradation scale is based on the monotonic and irreversible characteristics of the degradation process. Efficiency loss, compressor pressure loss, exhaust temperature margin, and fuel consumption accumulation are selected as embedded variables. The delay time is set to 24 hours to eliminate intraday load fluctuations. The embedding dimension is 3, and only the zero-order topological features reflecting connectivity are retained, while higher-order cohomology structures are ignored.

[0121] It should be further explained that, for the point clouds at each time scale in the initial state point cloud set, short-lifetime cycles with lifetimes below the adaptive threshold are identified and removed, while long-lifetime topological features reflecting the real physical structure are retained. A low-dimensional embedded manifold is constructed using the physical constraints corresponding to this scale, such as combustion phase conservation, rotor trajectory periodicity, or thermodynamic cycle closure, and the point cloud is orthogonally projected onto this manifold to eliminate measurement redundancy and enhance physical consistency. A farthest point sampling (FPS) strategy based on curvature or density perception is adopted on the projected manifold to ensure that the point cloud is evenly distributed and its quantity is controllable in key geometric regions. Based on this, the point clouds at each scale are aligned and interpolated using the timestamps of the original time series signals as a reference to form a unified time grid. According to the multi-physics coupling mechanism, such as combustion pressure disturbance driving mechanical vibration and thermal fatigue accumulation affecting performance degradation, physical correlation edges are constructed between corresponding time points of adjacent or related scales, thereby generating a multi-scale physically-aware state point cloud set that combines time alignment, topological fidelity, and cross-scale physical semantics.

[0122] It should be further explained that differentiated filtering schemes were formulated for point clouds at four scales: high-dimensional dense combustion dynamics, mid-dimensional periodic rotating machinery, low-dimensional sparse thermodynamic cycles, and low-dimensional monotonic performance degradation. Specifically, the combustion dynamics scale adopted a Vietoris-Rips (VR) complex with filtering parameters of [0.001, 0.5] and logarithmically increasing step size; the rotating machinery scale combined VR complex with periodic constraints, with filtering parameters of [0.01, 2.0] and linearly increasing step size; the thermodynamic cycle scale adopted an Alpha complex with filtering parameters of [0.1, 10.0] and adaptively adjusted step size based on the Delaunay edge length distribution; and the performance degradation scale adopted a linear interpolation complex with filtering parameters of [1.0, 50.0] and coarse-grained increasing step size, with a filter parameter variation of 2.0. Simultaneously, it incorporates physical perception optimization, namely, for point clouds at the rotating machinery scale, it integrates Euclidean distance and periodic distance, i.e., calculates the distance matrix based on the circumferential topology of the axis trajectory; for point clouds at the thermodynamic scale, it calculates weighted distances based on fault sensitivity weighted physical quantities, i.e., pressure ratio, temperature ratio, etc.; and it adaptively determines the number of filtering steps through the intrinsic dimension of the point cloud, making the number of high-dimensional steps more, and finally generating a range of filtering parameters specific to each scale.

[0123] It should be further explained that in practical applications, a multi-scale parallel computing architecture can be built based on Ripser or GUDHI tools, allocating independent processes to the four scales. Specifically, the combustion dynamics scale uses a sparse approximation to accelerate VR complex calculations for high-dimensional point clouds, retaining only topological features with a lifetime > 0.01 to filter numerical noise; the rotating machinery scale calculates complete persistent cohomology and post-processes to identify the H1 loop features corresponding to the axis trajectory; the thermodynamic cycle scale uses an Alpha complex to improve geometric accuracy; and the performance degradation scale only calculates H0, reflecting connectivity, ignoring higher-order cohomology. After calculation, the persistent graph format is unified as "(dimension, (birth, death))", outputting persistent graphs for each scale (PD0 / H0, PD1 / H1, PD2 / H2), providing a foundation for subsequent feature extraction. k A k-dimensional homology group is used to measure k-dimensional "holes" or connected structures in space, where H0 represents connected components, H1 represents independent loops or holes, and H2 represents cavities or voids. PD stands for Persistence Diagram, which records the birth and death scales of homology features and is the core visualization and analysis object of persistent homology.

[0124] It should be further explained that the first set of topological features are the statistics of the persistent subgraphs of each homology dimension, such as the mean birth time, mean lifespan, maximum lifespan, lifespan standard deviation, total persistence, and number of long-life features; the second set is a fixed-resolution persistent image generated by estimating the Gaussian kernel density of each persistent subgraph, such as a pixel density vector obtained by flattening a 32×32 image; the third set is dynamic indicators extracted based on the Betti curve on the filtering parameter axis, including the curve peak, area under the curve, rising or falling slope, number of inflection points, and correlation of Betti numbers across multiple scales; the fourth set is derived from the multi-dimensional topological entropy of the lifespan distribution, including Shannon entropy, rate of change of entropy across time windows, and joint entropy or mutual information between different homology dimensions, which together comprehensively characterize the topological characteristics of the unit's operating state from four dimensions: structural saliency, spatial distribution, evolutionary dynamics, and system complexity.

[0125] It should be further explained that the birth and death filter values ​​of all topological generators extracted at each time scale are classified and organized according to the topological dimension, namely H0 connected components, H1 annular holes, and H2 cavities. For combustion or mechanical scales, the birth or death values ​​of H0, H1, and H2 are retained. For thermodynamic scales, only the valid values ​​of H0 and H1 are selected. For degradation scales, only the birth or death values ​​of H0 are retained, and all noise features with a persistence (death value - birth value) < 0.01 are removed. Each valid topological generator is encapsulated into a tuple of (topological dimension, (birth filter value, death filter value)) in a unified format. Features with an infinite death value are marked as persistent core topological structures. The corresponding tuple sets are summarized according to the four scales of combustion dynamics, rotating machinery, thermodynamic cycle, and performance degradation to form a standardized persistence graph containing the mapping relationship between the topological dimension and the birth or death filter value at each scale. The persistence graphs of all scales are integrated to obtain a multi-scale persistence graph. This persistence graph completely depicts the entire life cycle evolution process of topological features from formation (birth) to disappearance (death) at different time scales. Each valid topology generator is encapsulated into a topology generator lifecycle tuple according to a preset format. The uniform format of this tuple is (topology dimension identifier, (birth filter value, death filter value)). The first element is the topology dimension identifier, where 0 represents H0 connected components, 1 represents H1 loops and holes, and 2 represents H2 cavities. The second element is a binary sub-tuple that stores the birth filter value (the filter parameter when the feature is formed) and the death filter value (the filter parameter when the feature disappears) of the topology generator in sequence. For core topology features with an infinite death value, the death filter value is marked as inf. Only valid topology generators with a persistence of (death value - birth value) ≥ 0.01 are encapsulated. After removing noisy features, the standardized encapsulation is completed.

[0126] It should be further explained that the topological random forest model, also known as the random forest model, adopts a hierarchical ensemble architecture, containing four independent sub-forests corresponding to four physical time scales: combustion dynamics, rotating machinery, thermodynamic cycles, and performance degradation. Each sub-forest consists of 50 topological decision trees with a maximum depth of 15. Node splitting is achieved through recursive binary search. There are no direct connections between the sub-forests; they only interact in the subsequent fusion stage. The training process uses a CART-style greedy splitting algorithm. Supervised scenarios use Gini impurity as the criterion, while unsupervised scenarios use topological density purity. Key hyperparameters, such as the number of trees (50), depth (15), feature subsampling rate (0.7), persistence threshold (≥0.01), and perturbation intensity (σ=0.01), are all determined through 5-fold time series cross-validation. The model input is the standardized topological feature vectors at each scale, and the output is the initial operational failure probability and confidence level.

[0127] It should be further explained that during the training of the topological decision tree, the out-of-bag (OOB) samples of each tree are used for dual evaluation. On the one hand, their classification error is calculated, i.e., the inconsistency rate between the true label of the out-of-bag sample and the predicted result. On the other hand, their topological stability is evaluated. Specifically, a small perturbation is injected into the continuous graph of the out-of-bag samples, such as adding Gaussian noise with a standard deviation of 0.01 to the birth or death coordinates. The perturbed features are recalculated and forward propagated to the tree. If the change in the predicted probability, measured by the L1 norm, is less than a preset tolerance threshold, such as 0.05, then the tree is determined to be robust to the topological perturbation. The ensemble weight of the tree is set as (1-OOB error) × topological stability score and normalized to the interval [0.1,1] to avoid zero weight. Meanwhile, the node splitting criterion is based on the Wasserstein distance. That is, after traversing possible thresholds along the candidate feature dimension and dividing the samples into left and right subsets, two corresponding persistent graph sets are constructed. The 1-Wasserstein distance between the two sets is calculated as a measure of distribution difference. The feature and threshold pair that maximizes this distance is selected for splitting. Since the larger the Wasserstein distance, the more separable the left and right child nodes are in the topology, the splitting is effective and enhances the fault detection capability. To improve efficiency, an approximate Wasserstein distance is actually used, such as the Sinkhorn algorithm or an L2 distance proxy based on persistent images to meet real-time requirements.

[0128] It should be further explained that the physical correlation strength between each time scale and the unit failure mode is determined by combining domain expert knowledge and historical failure statistics to construct a 4×5 correlation weight matrix, that is, the four scales correspond to the five types of failure modes. For example, the combustion dynamic scale assigns a high weight of 0.9 to the combustion instability mode, a medium weight of 0.4 to the mechanical failure mode, and only 0.1 to the severe degradation mode. This weight is used to adjust the contribution of each scale when multi-scale fusion.

[0129] It should be further explained that the pre-set health scores for each failure mode are quantified based on their impact on unit safety, availability, and maintenance urgency. Specifically, they are set as follows: "Normal" = 100 points, "Combustion Instability" = 60 points, "Mechanical Failure" = 50 points, "Coupled Failure" = 30 points, and "Severe Degradation" = 10 points. This score vector is weighted and averaged with the posterior probability of the failure mode to generate a unit health index of 0-100.

[0130] It should be further explained that the determination of the dynamic early warning threshold depends on the real-time identification of the current operating parameters of the natural gas generator set, including load rate, speed fluctuation, exhaust temperature gradient, start-up and shutdown status, etc. That is, the system determines the current operating stage through a rule engine or lightweight classifier, such as steady-state full load, variable load transition, start-up and shutdown process, etc. If it is in a highly dynamic operating condition, such as a load change rate > 5% per minute, the early warning threshold will be appropriately lowered by 5-10 points to tolerate normal fluctuations; if it is in stable operation and the topology entropy change rate increases for three consecutive windows, the threshold will be raised by 3-8 points to enhance early sensitivity, thereby realizing a hierarchical early warning mechanism that adaptively matches the operating status.

[0131] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above-described embodiment of the intelligent state assessment method for natural gas generator set operation.

[0132] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described embodiments of the intelligent state assessment method for natural gas generator set operation.

[0134] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described embodiments of the intelligent operating status assessment method for natural gas generator sets.

[0135] Those skilled in the art will understand that implementing all or part of the processes in the above-described intelligent state assessment method for natural gas generator set operation can be accomplished by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above-described intelligent state assessment method for natural gas generator set operation. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0136] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for intelligently assessing the operating status of a natural gas generator set, characterized in that, include: Multiphysics operation data is acquired and divided into several time scales based on the fault physical characteristics of the natural gas generator set. Phase space reconstruction is performed on the multiphysics operation data based on different time scales to obtain a set of physically sensed state point clouds, including: Acquire multi-physics field operation data of natural gas generator sets, perform physical validity verification, multi-rate synchronization and operating condition normalization on the multi-physics field operation data, and obtain standardized multi-physics field time sequence signals; The aforementioned timescales include: combustion dynamics, rotating machinery, thermodynamic cycles, and long-term performance degradation timescales; Based on the physical characteristics of generator set faults, four time scales are divided: combustion dynamics, rotating machinery, thermodynamic cycle, and long-term performance degradation. At each time scale, a multivariate combination of dominant physical processes is selected by combining physical conservation laws and fault mechanisms. Furthermore, the time delay of the standardized multiphysics field time series signal is embedded with physical perception by combining kinematic and thermodynamic constraints to generate the initial state point cloud set for each time scale. Topological denoising, physical manifold projection and uniform resampling optimization are performed on the point clouds at each time scale in the initial state point cloud set to obtain an optimized state point cloud set. Cross-scale time indexing and physical association edge construction are then performed on the optimized state point cloud set to obtain a multi-scale physically-aware state point cloud set. Complex filtering is performed on the point cloud at each time scale in the state point cloud set to obtain a filtered complex sequence. Multi-scale unit operation state topology evolution analysis is then performed on the filtered complex sequence to obtain a topology feature vector, including: The complex type and distance metric are determined based on the physical characteristics of each time scale, and a filtered complex sequence of unit operating status is generated by combining the inherent dimensions and physical constraints of the state point cloud set. The filtered complex sequence is subjected to continuous homology parallel computation to generate a multi-scale state persistence graph, and the multi-scale state persistence graph is transformed into an initial topological feature set for the topological evolution of the unit's operating state. The topological features in the initial topological feature set are concatenated to obtain the concatenated topological features. The concatenated topological features are then standardized according to the physical range of the historical normal working conditions at each scale to obtain a fixed-dimensional topological feature vector. The topological feature vectors at different time scales are input into the topological random forest model for training, and the preliminary operational failure probability at each time scale is output. The preliminary operational failure probability at each time scale is then jointly inferred to output the posterior distribution of the failure mode. The unit health index is based on the fault mode posterior distribution and is combined with the dynamic early warning threshold to realize the operation status assessment of natural gas generator units.

2. The intelligent operating status assessment method for natural gas generator sets according to claim 1, characterized in that, The filtered complex sequence for generating unit operating states by combining the intrinsic dimensions and physical constraints of the state point cloud set includes: The physical perceived distance between all point pairs in the state point cloud set is calculated based on the distance metric, forming a pairwise distance matrix. The lower and upper limits of the filtering parameters are set from the pairwise distance matrix in combination with the inherent dimension and physical constraints of the state point cloud set, thus determining the range of the filtering parameters. The step size strategy is selected based on the scale dynamics characteristics, and the range of filtering parameters is discretized and sampled based on the step size strategy to generate a discrete sequence of filtering parameters. Based on the complex type and physical sensing distance, a complex is constructed for each filter parameter in the discrete filter parameter sequence to obtain the filtered complex sequence.

3. The intelligent operating status assessment method for natural gas generator sets according to claim 1, characterized in that, The step of performing continuous homology parallel computation on the filtered complex sequence to generate a multi-scale state persistence graph includes: Configure the continuous cohomology calculation parameters according to the time scale type of the filtered complex sequence, and calculate the birth and death filtering values ​​of all topological generators in parallel for each time scale of the filtered complex sequence based on the continuous cohomology calculation parameters. The birth and death filter values ​​of all topological generators extracted at each time scale are classified and organized according to the topological dimension. Then, the noise features of the persistence of the organized birth and death filter values ​​are removed to obtain the effective topological generators. Each valid topology generator is encapsulated into a topology generator lifecycle tuple according to a preset format. All topology generator lifecycle tuples are then summarized according to each time scale to obtain a standardized state persistence graph. The standardized state persistence graphs at all scales are then integrated to obtain a multi-scale state persistence graph.

4. The intelligent operating status assessment method for natural gas generator sets according to claim 1, characterized in that, The initial topological feature set for transforming the multi-scale state persistence graph into the unit operating state topological evolution includes: The initial topological feature set includes: a first set of topological features, a second set of topological features, a third set of topological features, and a fourth set of topological features; The multi-scale state persistence graph is separated according to the homology dimension to obtain zero-order, first-order and second-order state persistence subgraphs, and statistical features are extracted from each state persistence subgraph to obtain the first set of topological features. Gaussian kernel density estimation is used to transform the state persistent subgraph into a state persistent image with a fixed resolution, and the state persistent image is flattened into a second set of topological features. The Betty curve is calculated on the filtering parameter axis of the multi-scale state persistence graph, and a third set of topological features is extracted based on the Betty curve. Multidimensional topological entropy is calculated based on the lifetime distribution of the state-persistent subgraph, and a fourth set of topological features is generated based on the multidimensional topological entropy calculation results.

5. The intelligent operating status assessment method for natural gas generator sets according to claim 1, characterized in that, The process involves inputting topological feature vectors at different time scales into a topological random forest model for training, outputting preliminary operational failure probabilities at each time scale, and jointly inferring the preliminary operational failure probabilities at each time scale to output the posterior distribution of the failure mode, including: Construct a topological random forest by expanding and grouping the topological feature vectors to obtain training samples and feature subspaces. Train the topological decision tree in the topological random forest based on the training samples and feature subspaces to obtain the trained topological decision tree. The trained topology decision tree outputs the failure probability of a single tree, and combines the out-of-bag sample error and topology stability of the topology decision tree to evaluate the prediction error and weight of each topology decision tree. The prediction error and weight of each topology decision tree are weighted and averaged to obtain the preliminary operational failure probability at each time scale. The association weights are set based on the physical correlation strength between each time scale and the unit's failure mode. The initial operational failure probability at each time scale is multiplied by the corresponding association weight, and then summed and normalized to obtain the posterior distribution of the failure mode of the unit's operating status.

6. The intelligent operating status assessment method for natural gas generator sets according to claim 5, characterized in that, The construction of the topological random forest involves expanding and grouping the topological feature vectors to obtain training samples and feature subspaces. The topological decision trees in the topological random forest are then trained based on the training samples and feature subspaces, resulting in trained topological decision trees, including: A hierarchical topological random forest is constructed based on the time scale and a preset number of topological decision trees. The topological feature vector is expanded using the topology-aware bootstrap sampling method to generate training samples. Based on the homology dimension, the topological feature vectors of each time scale are grouped to obtain several semantic feature groups. Features are then randomly selected from these semantic feature groups to form a feature subspace. Based on the Wasserstein distance splitting criterion, the topology decision tree is recursively split and node-grown using the dimension in the feature subspace on the training samples to obtain the trained topology decision tree.

7. The intelligent operating status assessment method for natural gas generator sets according to claim 1, characterized in that, The method of using a fault mode posterior distribution-based computer group health index and combining the group health index with dynamic early warning thresholds to assess the operating status of natural gas generator sets includes: Based on the pre-set health scores for each failure mode, the probabilities of each type of failure mode posterior distribution are used as weights of the corresponding health scores for weighted averaging to obtain the unit health index. Identify the current operating condition based on the current operating parameters of the natural gas generator set, and determine the dynamic early warning threshold based on the current operating condition; The unit health index is compared with the dynamic early warning threshold, and the health status of the natural gas generator unit is determined based on the comparison results.

8. A smart operating status assessment system for natural gas generator sets, characterized in that, The intelligent operating status assessment system for the natural gas generator set includes: The state point cloud construction module is used to acquire multiphysics operation data. Based on the fault physical characteristics of the natural gas generator set, it divides the data into several time scales. Phase space reconstruction is performed on the multiphysics operation data based on different time scales to obtain a set of physically perceived state point clouds, including: Acquire multi-physics field operation data of natural gas generator sets, perform physical validity verification, multi-rate synchronization and operating condition normalization on the multi-physics field operation data, and obtain standardized multi-physics field time sequence signals; The aforementioned timescales include: combustion dynamics, rotating machinery, thermodynamic cycles, and long-term performance degradation timescales; Based on the physical characteristics of generator set faults, four time scales are divided: combustion dynamics, rotating machinery, thermodynamic cycle, and long-term performance degradation. At each time scale, a multivariate combination of dominant physical processes is selected by combining physical conservation laws and fault mechanisms. Furthermore, the time delay of the standardized multiphysics field time series signal is embedded with physical perception by combining kinematic and thermodynamic constraints to generate the initial state point cloud set for each time scale. Topological denoising, physical manifold projection and uniform resampling optimization are performed on the point clouds at each time scale in the initial state point cloud set to obtain an optimized state point cloud set. Cross-scale time indexing and physical association edge construction are then performed on the optimized state point cloud set to obtain a multi-scale physically-aware state point cloud set. The state topology evolution analysis module is used to perform complex filtering on the point cloud at each time scale in the state point cloud set to obtain a filtered complex sequence, and then perform multi-scale unit operation state topology evolution analysis on the filtered complex sequence to obtain a topology feature vector, including: The complex type and distance metric are determined based on the physical characteristics of each time scale, and a filtered complex sequence of unit operating status is generated by combining the inherent dimensions and physical constraints of the state point cloud set. The filtered complex sequence is subjected to continuous homology parallel computation to generate a multi-scale state persistence graph, and the multi-scale state persistence graph is transformed into an initial topological feature set for the topological evolution of the unit's operating state. The topological features in the initial topological feature set are concatenated to obtain the concatenated topological features. The concatenated topological features are then standardized according to the physical range of the historical normal working conditions at each scale to obtain a fixed-dimensional topological feature vector. The failure probability prediction module is used to input the topological feature vectors at different time scales into the topological random forest model for training, output the preliminary failure probability at each time scale, and perform joint inference on the preliminary failure probability at each time scale to output the posterior distribution of the failure mode. The operation status assessment module is used to assess the operation status of natural gas generator sets by combining the unit health index with dynamic early warning thresholds based on the fault mode posterior distribution computer health index.

Images