Pre-start detection and control system of diesel generating set

Abstract

The invention discloses a pre-start detection and control system for a diesel generating set, belongs to the technical field of diesel generating sets, and solves the problem that the diesel generating set is damaged under a series of influences due to the fact that composite faults such as fuel oxidation and sensor signal drifting are recognized and the diesel generating set is directly started. Comprising a data processing module, and a fuel oil system detection module, a combustion system detection module, an emission system detection module and a control system information extraction module which are in communication connection with the data processing module, the fuel oil system detection module, the combustion system detection module, the emission system detection module and the control system information extraction module collect parameters and transmit the parameters to the data processing module. According to the method, during long-term operation of the diesel generating set, due to the fact that fuel oil is oxidized to form colloid to block an oil way, the oil supply pressure fluctuates, the combustion efficiency is reduced, signal drifting of an oxygen sensor and the like covers real faults, and misjudgment is caused, the problem is solved through multi-dimensional cooperative detection.

Description

Technical Field

[0001] The invention relates to the technical field of diesel generator sets, and in particular to a pre-start detection and control system for diesel generator sets. Background Art

[0002] A diesel generator set is a power generation device that uses diesel as fuel and converts mechanical energy into electrical energy by driving a generator with a diesel engine. It is mainly composed of a diesel engine, a generator, a control box, a fuel tank, a storage battery for starting and control, a protective device and other components. As the core of the power, the performance of the diesel engine directly affects the efficiency and reliability of the entire generator set. The generator is responsible for converting mechanical energy into electrical energy and is the core part of power generation. The control box is used to monitor and control the entire power generation process to ensure the stable and safe operation of the generator set.

[0003] The working principle of a diesel generator set is relatively simple but highly efficient. When a diesel engine is running, mechanical energy is generated by burning diesel, which further drives the generator to rotate, thereby generating electrical energy. Specifically, in the diesel engine cylinder, the clean air filtered by the air filter is fully mixed with the high-pressure atomized diesel injected from the injector. Under the squeezing of the piston upward, the volume is reduced, the temperature rises rapidly, and the ignition point of the diesel is reached. The diesel is ignited, the mixed gas burns violently, the volume expands rapidly, and the piston is pushed downward, which is called "work". Each cylinder works in a certain order, and the thrust acting on the piston is converted into a force that drives the crankshaft to rotate through the connecting rod, thereby driving the crankshaft to rotate, and then driving the generator to generate electricity.

[0004] Diesel generator sets have the advantages of high reliability, quick start and easy maintenance. They can operate stably in harsh environments and are a reliable power guarantee. They can start and supply power quickly when needed to meet the power needs in emergency situations. In addition, their structure is relatively simple, which makes maintenance easier and reduces operating costs.

[0005] Due to these characteristics, diesel generator sets are widely used in many fields. In the power industry, it is used as a backup power source to ensure rapid power supply when the main power source fails, ensuring the stable operation of the power grid.

[0006] During storage and use, fuel undergoes oxidation reactions due to factors such as temperature, humidity, and air, producing colloid and other sediments that clog fuel system components, affect the normal flow and injection of fuel, and lead to incomplete combustion. At the same time, the oxygen sensor is subject to long-term effects of high temperature and exhaust gas chemicals, and its sensitivity and accuracy decrease, causing output signal drift. At this point, the diesel generator set has a compound fault of fuel oxidation and sensor signal drift.

[0007] When these two faults exist at the same time, the oxygen sensor signal drifts, causing the ECU to receive incorrect oxygen content information, mistakenly believing that the mixture is too lean and increasing the amount of fuel injection, making the mixture richer. This will lead to a series of consequences: First, the amount of fuel injection increases, the air in the combustion chamber is insufficient, the fuel cannot be completely burned, and a large amount of carbon particles are produced, which are discharged in the form of black smoke, resulting in black smoke; second, the incompletely burned fuel forms carbon deposits in the combustion chamber and other parts, aggravating the formation of carbon deposits and affecting the normal operation of the engine; third, some of the incompletely burned fuel continues to burn in the exhaust pipe, resulting in afterburning, and the high-temperature exhaust gas damages the DPF, causing its filtering effect to decline or even fail.

[0008] However, this complex fault is difficult to identify because: on the one hand, fuel oxidation and sensor signal drift occur gradually, and their initial impact on engine performance is not obvious, and they are somewhat hidden; on the other hand, the two are interrelated. Fuel oxidation causes incomplete combustion, which affects the oxygen sensor feedback signal, and the oxygen sensor signal drift causes the ECU to incorrectly adjust the injection amount, further aggravating the impact of fuel oxidation and making the fault manifestation complex.

[0009] Therefore, it is necessary to identify complex faults such as fuel oxidation and sensor signal drift to avoid the diesel generator set being started directly and causing damage to the diesel generator set due to a series of influences.

[0010] Therefore, a diesel generator set pre-start detection and control system is proposed to solve or alleviate the above problems. Summary of the invention

[0011] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a pre-start detection and control system for a diesel generator set.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] A pre-start inspection and control system for a diesel generator set includes a data processing module, and a fuel system inspection module, a combustion system inspection module, an emission system inspection module, and a control system information extraction module that are communicatively connected thereto;

[0014] The fuel system detection module, combustion system detection module, emission system detection module, and control system information extraction module collect parameters and transmit them to the data processing module;

[0015] The data processing module receives the parameters transmitted by each detection module, and performs consistency alignment on the spatiotemporal data of multiple sensors through the Lie group differential method to eliminate the spatiotemporal mismatch in the sensor network;

[0016] The nonlinear dynamic characteristics caused by fuel oxidation are extracted using high-order recursive quantitative analysis, and the fuel viscoelastic degradation process is quantified through multi-scale Cauchy-Green tensor analysis.

[0017] A fast approximation algorithm based on the curvature change of the Fisher information matrix and the Wasserstein distance is used to detect sensor signal drift.

[0018] Using high-order tensor decomposition and tensor singular value entropy, multi-physical domain features are integrated and a joint feature tensor is generated.

[0019] The decision boundary is optimized by diffeomorphic mapping and the composite fault classification is realized by combining nonlinear support tensor machine.

[0020] Based on the conservation law residual verification and parameter elastic network dynamic parameter adjustment mechanism, the physical rationality and robustness of the algorithm output are ensured.

[0021] Preferably, the fuel system detection module includes a fuel flow meter and a fuel pressure sensor, the combustion system detection module includes an oxygen sensor and an exhaust temperature sensor, the emission system detection module includes a cylinder pressure sensor and a NOx sensor, the control system information extraction module extracts the ECU air-fuel ratio correction value and injection pulse width instruction, and the data processing module includes a processor.

[0022] Preferably, the data processing module receives parameters transmitted by each detection module, performs consistency alignment on multi-sensor spatiotemporal data by Lie group differential method, and eliminates spatiotemporal mismatch in the sensor network, including the following steps:

[0023] The sensor network is defined as a Lie group structure, which consists of a three-dimensional rigid motion group and a sensor signal space, where the three-dimensional rigid motion group contains a rotation matrix and a translation vector;

[0024] The spatiotemporal registration objective function is constructed to achieve alignment by minimizing the weighted distance of all sensor signals on the Riemann manifold, where the weighting coefficient is determined according to the signal-to-noise ratio of each sensor and the Riemann distance is calculated based on the logarithmic difference integral of the signal power spectral density.

[0025] The Lie algebraic gradient descent method is used to iteratively optimize the group parameters. In each iteration, the group action parameters are updated according to the gradient direction of the loss function. The learning rate is fixed at 0.01, and the iteration termination condition is that the change in the loss function value for two consecutive times is less than the set threshold.

[0026] Preferably, the method of extracting the nonlinear dynamic characteristics caused by fuel oxidation by high-order recursive quantitative analysis and quantifying the fuel viscoelastic degradation process by multi-scale Cauchy-Green tensor analysis comprises the following steps:

[0027] The phase space of the fuel pressure signal is reconstructed, the time delay is determined by the mutual information method, and a multi-dimensional recursive matrix is ​​constructed. The recursive entropy spectrum is calculated by statistically analyzing the probability distribution of the diagonal length in the recursive matrix. When the embedding dimension is 5, if the entropy value drops by more than 20%, it is determined that the fuel is oxidized.

[0028] The velocity field gradient is extracted from the fuel flow signal, and the Cauchy-Green tensor of the flow mapping Jacobian matrix is ​​calculated. The damage index is constructed by comparing the relative deviation between the real-time eigenvalue and the benchmark value. If the index exceeds 30% for more than 5 sampling cycles, the fuel viscoelastic degradation is confirmed.

[0029] Preferably, the fast approximation algorithm based on the curvature change of the Fisher information matrix and the Wasserstein distance for detecting the sensor signal drift comprises the following steps:

[0030] Assuming that the oxygen sensor signal follows a Gaussian distribution, the curvature tensor of its Fisher information matrix is ​​calculated, and the normal operating condition baseline curvature is established through Monte Carlo simulation. The Frobenius norm of the curvature change is detected in real time, and a drift warning is triggered when it exceeds the preset threshold.

[0031] An empirical distribution is constructed for NOx sensor signals, and the Wasserstein distance between the NOx sensor signal and the theoretical distribution is quickly and approximately calculated using the Sinkhorn iterative algorithm. Combined with the dynamically adjusted statistical threshold, the sensor is judged to be abnormal.

[0032] Preferably, the method of utilizing high-order tensor decomposition and tensor singular value entropy to fuse multi-physical domain features and generate a joint feature tensor comprises the following steps:

[0033] Construct a three-dimensional joint feature tensor, where the dimensions correspond to the number of sensors, feature type, and number of sliding time windows, and each element is generated by a windowed feature extraction function;

[0034] Tucker decomposition is used to decompose high-order tensors into the product of core tensors and factor matrices. The system health status is evaluated by counting the proportion of non-zero elements in the core tensor and calculating the entropy of the tensor singular value distribution. If the entropy is lower than the preset level, a compound fault is confirmed.

[0035] Preferably, the method of optimizing the decision boundary by differential homeomorphism mapping and implementing compound fault classification in combination with a nonlinear support tensor machine comprises the following steps:

[0036] A nonlinear state space mapping is constructed, and the decision boundary is optimized by balancing the mapping smoothness and classification accuracy in the objective function, and the spectral method with a cutoff frequency of 10 is used to solve it.

[0037] An exponential kernel function based on tensor Frobenius distance is defined, supporting tensors are selected from the historical fault library, and the classifier weights are solved through the sequential minimum optimization algorithm. The final decision function is the symbolic function of weighted kernel similarity.

[0038] Preferably, the residual verification based on conservation law and the dynamic parameter adjustment mechanism of parameter elastic network ensure the physical rationality and robustness of the algorithm output, including the following steps:

[0039] The residual is calculated in real time based on the law of conservation of mass, and the judgment threshold is dynamically adjusted in combination with the load change rate. If the residual exceeds the threshold, the model parameters are reset;

[0040] A non-convex optimization rule with an exponential decay factor is used to update the feature weights, constraining the absolute value of the weights not to exceed a fixed value, and automatically expanding the constraint boundaries every 24 hours to maintain the flexibility of the algorithm.

[0041] The present invention has the following beneficial effects:

[0042] In the long-term operation of the diesel generator set, the present invention solves this problem through multi-dimensional collaborative detection because the oxidation of fuel will form colloid to block the oil circuit, resulting in fluctuations in the fuel supply pressure and decreased combustion efficiency, and the signal drift of the oxygen sensor and other sensors will cover up the real fault and cause misjudgment;

[0043] 1. Lie group differential alignment eliminates the temporal and spatial deviations of multiple sensors and ensures data consistency;

[0044] 2. High-order recursive analysis extracts the phase space entropy decay characteristics from the fuel pressure signal and quantifies the degree of oxidation;

[0045] 3. Information geometry detection identifies sensor drift through changes in signal distribution curvature and Wasserstein distance shift;

[0046] 4. Tensor fusion correlates and analyzes the multi-physical characteristics of combustion, emission, and control systems to locate fault coupling points;

[0047] 5. Differentiomorphic classification constructs clear decision boundaries in nonlinear space and separates complex fault modes;

[0048] 6. Conservation law verification ensures that the test results comply with mass-energy conservation and avoids physical contradictions. Through multi-scale feature decoupling and dynamic threshold adjustment, accurate distinction between oxidation and drift is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 It is a structural block diagram of the present invention.

[0051] 1. Fuel system detection module; 101. Fuel pressure sensor; 102. Fuel flow meter; 2. Combustion system detection module; 201. Oxygen sensor; 202. Exhaust temperature sensor; 3. Emission system detection module; 301. Cylinder pressure sensor; 302. NOx sensor; 4. Control system information extraction module; 5. Data processing module. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0053] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0055] In the description of the present invention, it should be understood that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or are the orientations or positional relationships in which the product of the invention is conventionally placed when in use, or are the orientations or positional relationships conventionally understood by those skilled in the art. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0056] Furthermore, the terms “first”, “second”, “third”, etc. are merely used for distinguishing descriptions and are not to be understood as indicating or implying relative importance.

[0057] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0058] A diesel generator set pre-start detection and control system, such as Figure 1 As shown, it includes a data processing module 5, and a fuel system detection module 1, a combustion system detection module 2, an emission system detection module 3, and a control system information extraction module 4 that are communicatively connected thereto, the fuel system detection module 1 includes a fuel flow meter 102 and a fuel pressure sensor 101, the combustion system detection module 2 includes an oxygen sensor 201 and an exhaust temperature sensor 202, the emission system detection module 3 includes a cylinder pressure sensor 301 and a NOx sensor 302, the control system information extraction module 4 extracts an ECU air-fuel ratio correction value and an injection pulse width instruction, and the data processing module 5 includes a processor;

[0059] The fuel system detection module 1, the combustion system detection module 2, the emission system detection module 3, and the control system information extraction module 4 collect parameters and transmit them to the data processing module 5;

[0060] The data processing module 5 receives the parameters transmitted by each detection module, and aligns the spatiotemporal data of multiple sensors through the Lie group differential method to eliminate the spatiotemporal mismatch in the sensor network;

[0061] The nonlinear dynamic characteristics caused by fuel oxidation are extracted using high-order recursive quantitative analysis, and the fuel viscoelastic degradation process is quantified through multi-scale Cauchy-Green tensor analysis.

[0062] A fast approximation algorithm based on the curvature change of the Fisher information matrix and the Wasserstein distance is used to detect sensor signal drift.

[0063] Using high-order tensor decomposition and tensor singular value entropy, multi-physical domain features are integrated and a joint feature tensor is generated.

[0064] The decision boundary is optimized by diffeomorphic mapping and the composite fault classification is realized by combining nonlinear support tensor machine.

[0065] Based on the conservation law residual verification and parameter elastic network dynamic parameter adjustment mechanism, the physical rationality and robustness of the algorithm output are ensured.

[0066] During the long-term operation of diesel generator sets, fuel oxidation will form colloidal deposits, leading to oil line blockage, fuel supply pressure fluctuations and reduced combustion efficiency. The signal drift of key sensors such as oxygen sensor 201 will mask the real fault characteristics and even mislead the control strategy, forming a vicious cycle of "fault masking-misadjustment-secondary damage". The above-mentioned module is used to collect parameters, and multi-dimensional detection and analysis are performed after parameter collection to solve the coupling interference between fuel oxidation and sensor drift, improve the fault separation accuracy, and provide reliable protection for the intelligent operation and maintenance of diesel units.

[0067] The sensor network is modeled as a joint space of a rigid body motion group and a signal space through Lie group differential theory. The Riemann manifold distance minimization is used to achieve sub-millisecond alignment of multi-source data, eliminate the time-space mismatch caused by mechanical vibration and electromagnetic interference, and complete the time-space alignment. At the same time, wavelet packet-variational mode decomposition is combined to remove high-frequency noise and baseline drift, retain the real fault characteristics, and perform noise suppression.

[0068] Based on the phase space reconstruction and recursive quantitative analysis of the fuel pressure signal, the chaotic characteristic attenuation caused by oxidation is quantified, the velocity field distortion of the fuel flow is analyzed by the Cauchy-Green tensor, the viscoelastic damage index is calculated, and the early oxidation degradation is accurately captured.

[0069] Then, the Fisher information matrix of the sensor signal distribution is constructed using information geometry theory, and the drift is detected by curvature change. The signal distribution offset is quantified using the Wasserstein distance, and the impact of load fluctuations is dynamically compensated. The residual verification of the mass-energy conservation law is further used to ensure that the fault judgment complies with the laws of thermodynamics, and physical verification is performed to avoid false alarms.

[0070] Finally, Tucker tensor decomposition is used to fuse the multi-dimensional characteristics of combustion, emission, and control systems, the fault coupling point is located by singular value entropy, and the classification boundary is optimized in curved space using differential homeomorphism mapping to solve the aliasing problem of traditional linear methods in composite faults, thereby ensuring that the diesel generator set can finally eliminate the composite fault.

[0071] Preferably, the data processing module 5 receives the parameters transmitted by each detection module, and performs consistency alignment on the spatiotemporal data of multiple sensors by using the Lie group differential method to eliminate the spatiotemporal mismatch in the sensor network, including the following steps:

[0072] The sensor network is defined as a Lie group structure, which consists of a three-dimensional rigid motion group and a sensor signal space, where the three-dimensional rigid motion group contains a rotation matrix and a translation vector. Specifically, the Lie group G = SE (3) × R n, where SE(3) is the rigid motion group, R n is the signal space;

[0073] The spatiotemporal registration objective function is constructed to achieve alignment by minimizing the weighted distance of all sensor signals on the Riemann manifold, where the weighting coefficient is determined according to the signal-to-noise ratio of each sensor, and the Riemann distance is calculated based on the logarithmic difference integral of the signal power spectral density. Specifically, the spatiotemporal registration objective function is: Among them, d M is the Riemann metric, φ g For group action, gradient descent is used to solve: g k+1 =g k exp G (-η·grad g L), η is the learning rate, grad g L is the gradient of the loss function with respect to the Lie group parameters;

[0074] The Lie algebraic gradient descent method is used to iteratively optimize the group parameters. In each iteration, the group action parameters are updated according to the gradient direction of the loss function. The learning rate is fixed at 0.01, and the iteration termination condition is that the change in the loss function value for two consecutive times is less than the set threshold.

[0075] In the above method steps, the sensor network is modeled as a joint space of a three-dimensional rigid body motion group and a signal space, accurately representing the spatial configuration changes of the sensor caused by mechanical vibration and installation posture differences. Taking the fuel pressure signal as a benchmark, the weighted distance of all sensor signals on the power spectrum Riemann manifold is minimized to eliminate time-space mismatch, ensure that the data of the oxygen sensor 201, NOx sensor 302 and other data are strictly synchronized in physical events, and use the Lie algebra parameterized group action to iteratively adjust the rotation matrix and translation vector along the gradient direction of the loss function. The learning rate is fixed to 0.01 to avoid oscillation, and the convergence condition is that the loss change is less than one millionth, achieving sub-millisecond alignment accuracy. This step solves the feature aliasing problem caused by sensor spatial offset and signal delay caused by oil circuit blockage vibration, provides high-quality data consistent in time and space for subsequent detection algorithms, separates the slight difference between fuel oxidation and drift, and reduces the false alarm rate.

[0076] Preferably, the nonlinear dynamic characteristics caused by fuel oxidation are extracted by high-order recursive quantitative analysis, and the fuel viscoelastic degradation process is quantified by multi-scale Cauchy-Green tensor analysis, including the following steps:

[0077] The phase space of the fuel pressure signal is reconstructed, the time delay is determined by the mutual information method, and a multidimensional recursive matrix is ​​constructed. The recursive entropy spectrum is calculated by statistically analyzing the probability distribution of the diagonal length in the recursive matrix. When the embedding dimension is 5, if the entropy value drops by more than 20%, it is determined that the fuel is oxidized. Specifically, the multidimensional recursive matrix is in, For the phase space trajectory embedded in dimension m, calculate the recursive entropy spectrum: When fuel oxidation produces H R (m,∈) decreases by >20% when m=5, is a phase space trajectory point with embedding dimension m, p k is the probability distribution of the diagonal length;

[0078] The velocity field gradient is extracted from the fuel flow signal, and the Cauchy-Green tensor of the flow mapping Jacobian matrix is ​​calculated. The damage index is constructed by comparing the relative deviation between the real-time eigenvalue and the reference value. If the index exceeds 30% for more than 5 sampling cycles, the fuel viscoelastic degradation is confirmed. Specifically, from the velocity field Extract the Cauchy-Green tensor Among them, F t The damage index is constructed as the Jacobian matrix of the flow map When DI>0.3, it indicates that the fuel viscoelasticity has deteriorated.

[0079] In the above method steps, the nonlinear dynamic feature extraction technology is used to accurately decouple the complex faults, and the differential effects of fuel oxidation and sensor drift are revealed through the nonlinear dynamic mechanism, providing physically explainable quantitative indicators for the separation of complex faults, with a detection sensitivity of 0.5% oxidation concentration and 0.3% drift bias.

[0080] The fuel pressure signal is mapped to a high-dimensional phase space, a recursive matrix reflecting the dynamic similarity of the system is constructed, the diagonal length distribution is statistically calculated, and the entropy value is calculated. Oxidation causes the fluidity of the fuel to decrease, and the pressure fluctuation shows a chaotic characteristic attenuation, which is manifested as a decrease in the recursive entropy value of more than 20%. The degree of oxidation is directly quantified, the velocity field gradient is extracted from the fuel flow signal, the Jacobian matrix of the flow mapping is constructed, and the eigenvalue of the Cauchy-Green tensor is calculated. Oxidized colloid increases the viscosity of the fuel, resulting in the accumulation of flow deformation energy. Viscoelastic degradation is identified by the deviation of the eigenvalue relative to the benchmark. The recursive entropy reflects the instability of macroscopic pressure fluctuations caused by oxidation, and the Cauchy-Green tensor captures microscopic flow distortion. The spatiotemporal correlation analysis of the two can distinguish between true oxidation and pseudo-fluctuations caused by sensor drift. For example, when the drift of the oxygen sensor 201 causes abnormal NOx residuals, if the Cauchy-Green damage index does not rise synchronously, it is determined to be drift interference rather than true oxidation.

[0081] Preferably, the sensor signal drift is detected based on a fast approximation algorithm based on the change of Fisher information matrix curvature and Wasserstein distance, comprising the following steps:

[0082] Assume that the signal of the oxygen sensor 201 follows a Gaussian distribution, calculate the curvature tensor of its Fisher information matrix, establish a benchmark curvature under normal conditions through Monte Carlo simulation, and detect the Frobenius norm of the curvature change in real time. When it exceeds the preset threshold, a drift warning is triggered. Specifically, assume that the sensor signal follows a parametric distribution p(s|θ), and calculate the FIM: Drift causes a change in the curvature of the parameter manifold: ΔK = ||Ric(L) - Ric(L 0 )|| F , determine that ΔK > K th Trigger a drift warning, where L ij (θ) is the Fisher information matrix, which measures the accuracy of parameter estimation, and Ric(L) is the Ricci curvature, which measures the geometric properties of the manifold;

[0083] Construct an empirical distribution for the signal of the NOx sensor 302, and use the Sinkhorn iteration algorithm to quickly approximate and calculate the Wasserstein distance between it and the theoretical distribution. Combine with a dynamically adjusted statistical threshold to determine whether the sensor is abnormal. Specifically, construct the Wasserstein distance of the signal distribution: Quickly approximate through the Sinkhorn algorithm: Threshold: γ is the regularization parameter of the Sinkhorn algorithm.

[0084] In a diesel generator set, the drift of the sensor signal will distort the true fault characteristics caused by fuel oxidation. The traditional threshold method has a high misjudgment rate because it ignores the structural changes in the signal distribution. The above method steps are used to achieve precise drift detection through information geometry theory.

[0085] Assume that the signal of the oxygen sensor 201 follows a Gaussian distribution, construct the geometric structure of the parameter space, calculate its curvature tensor. Under normal conditions, the curvature distribution is stable, while when the sensor drifts, the geometric deformation of the parameter space causes a significant change in the curvature. Establish a benchmark curvature through Monte Carlo simulation and detect the offset in real time to avoid misjudging the change in the distribution mean caused by fuel oxidation as drift.

[0086] For the NOx sensor 302, compare the real-time data distribution with the theoretical model, and use the Sinkhorn iteration algorithm to quickly calculate the Wasserstein distance between the two. This distance reflects the overall morphological difference of the distribution, and can distinguish the distribution broadening caused by drift from the distribution offset caused by oxidation. The dynamic threshold is adaptively adjusted in combination with the load change rate to suppress the interference of working condition fluctuations.

[0087] FIM curvature is sensitive to local deformations in parameter space, and Wasserstein distance captures global distribution differences. The combined judgment of the two can identify slow drifts, such as temperature drift, and step drifts, such as circuit aging. Specifically, for example, NOx mean shift caused by oxidation only triggers the Wasserstein alarm, while sensor gain drift triggers both curvature and distance alarms.

[0088] Preferably, high-order tensor decomposition and tensor singular value entropy are used to fuse multi-physical domain features and generate a joint feature tensor, including the following steps:

[0089] Construct a three-dimensional joint feature tensor, where the dimensions correspond to the number of sensors, feature type, and number of sliding time windows. Each element is generated by a windowed feature extraction function. The specific three-dimensional joint feature tensor T∈R N×M×K , where N is the number of sensors, M is the feature dimension, and K is the number of time windows;

[0090] Tucker decomposition is used to decompose the high-order tensor into the product of the core tensor and the factor matrix. Tucker decomposition is used: T≈g× 1 U (1) × 2 U (2) × 3 U (3) , the health status of the system is evaluated by counting the proportion of non-zero elements in the core tensor, where the sparsity of the core tensor g is S = ∥g∥ 0 / numel(g) reflects the health status of the system and calculates the entropy of the tensor singular value distribution. If the entropy is lower than the preset level, a composite fault is confirmed. Specifically, the tensor modal singular value entropy is calculated. Determine the composite fault causing H TSVE Decrease>15%, U (1) , U (2) , U (3) is the factor matrix of Tucker decomposition.

[0091] In diesel generator sets, the combined failure of fuel oxidation and sensor drift will cause multi-dimensional anomalies in the combustion, emission, and control systems. Traditional single-dimensional feature analysis is difficult to distinguish their coupling effects. In the above method steps, global fault characterization is achieved through multi-physical domain tensor fusion technology to improve detection accuracy.

[0092] The six-dimensional features of eight types of sensors, including the flow characteristics of fuel pressure reaction, the combustion state of oxygen sensor 201 reaction, the chemical reaction of NOx emission reaction, and the control response of ECU correction reaction, are integrated. The six-dimensional features include mean, variance, entropy, fractal dimension, spectral kurtosis, and Hurst exponent. They are stacked into three-dimensional tensors according to time windows to fully characterize the cross-scale dynamic behavior of the system. Fuel oxidation is manifested as an increase in the fractal dimension of pressure pulsation and a decrease in entropy value, while sensor drift causes an abnormal and sharp increase in the spectral kurtosis of the oxygen signal. The tensor structure can naturally distinguish the spatiotemporal patterns of the two types of faults, and then the high-order tensor is decomposed into the product of the core tensor and the factor matrix. The sparsity of the core tensor reflects the coupling strength of multiple physical fields. Under normal operating conditions, the combustion, flow, and control characteristics are weakly coupled, with a sparsity >0.4. Compound faults lead to strong correlation of features, with a sparsity <0.4, which directly indicates the fault coupling point. For example, fuel oxidation causes NOx model prediction deviation, which in turn leads to ECU correction abnormality.

[0093] Finally, the complexity of the system is quantified by the entropy of the singular value distribution. The normal combustion process presents a high entropy value due to turbulence, chemical reactions, etc. The coupling effect of oxidation and drift makes the system dynamics simpler, and the entropy value decreases by >15%. Combined with the coordinated judgment of sparsity and entropy value, oxidation-dominated faults and drift-dominated faults can be separated.

[0094] Preferably, the decision boundary is optimized by differential homeomorphism mapping, and the composite fault classification is realized by combining nonlinear support tensor machine, including the following steps:

[0095] Construct a nonlinear state space mapping, optimize the decision boundary by balancing the mapping smoothness and classification accuracy in the objective function, and solve it using the spectral method with a cutoff frequency of 10. Specifically, define the differential homeomorphism φ:R from the state space to the decision space d →R d ,satisfy: Among them, λ controls the curvature and the optimization objective is: Separate fault clusters in curved space to improve the clarity of classification boundaries;

[0096] Define an exponential kernel function based on the tensor Frobenius distance, select the supporting tensor from the historical fault library, solve the classifier weights through the sequential minimum optimization algorithm, and the final decision function is the symbolic function of the weighted kernel similarity. Specifically, the decision function is The kernel function K is based on the tensor Riemann metric K(T i ,T j )=exp(-γd FR (T i ,T j )), where d FRis the Frobenius-Riemann distance.

[0097] In diesel generator sets, the combined faults of fuel oxidation and sensor drift have feature aliasing in traditional Euclidean space, which makes it impossible for linear classification boundaries to effectively separate fault modes. The above method uses differential homeomorphism decision boundary optimization and nonlinear support tensor machine to achieve accurate classification in high-dimensional space. Through geometric space transformation and tensor similarity measurement, it solves the aliasing problem of combined faults in traditional feature space.

[0098] The original feature space is transformed into a curved space through nonlinear mapping. In this space, the pressure fluctuation pattern caused by fuel oxidation and the signal distribution offset caused by sensor drift show geometric separation characteristics. For example, the oxidation fault cluster appears spherical after mapping, while the drift fault unfolds along a specific manifold, which improves the boundary clarity.

[0099] By minimizing the energy functional in the mapping space, which includes the gradient smoothing term and the classification error term, the distribution morphology of the fault clusters in the curved space is constrained, and the spectral method with a cutoff frequency of 10 is used to solve it, suppressing high-frequency noise interference, ensuring smooth boundaries and adapting to dynamic working conditions. Then, a kernel function is constructed based on the tensor Frobenius distance to measure the global similarity between real-time data and samples in the historical fault library. The key support tensors are screened through the sequential minimum optimization algorithm and assigned high weights to enhance the sensitivity to complex faults. For example, the coupling mode of the oxygen sensor 201 drift and the NOx model deviation is mapped to a hyperplane in the tensor space, and the output fault type is determined by the symbolic function.

[0100] Preferably, based on the conservation law residual verification and the parameter elastic network dynamic parameter adjustment mechanism, the physical rationality and robustness of the algorithm output are ensured, including the following steps:

[0101] The residual is calculated in real time based on the law of conservation of mass, and the judgment threshold is dynamically adjusted in combination with the load change rate. If the residual exceeds the threshold, the model parameters are reset. Specifically, the residual of mass-energy-momentum conservation is checked as Among them, when NCF>NCF th Trigger reset of model parameters;

[0102] A non-convex optimization rule with an exponential decay factor is used to update feature weights, constraining the absolute value of the weights to not exceed a fixed value, and automatically expanding the constraint boundary every 24 hours to maintain algorithm flexibility. Specifically, the feature weights are adjusted dynamically. Non-convex optimization avoids local minima and improves dual fault robustness, where η is the learning rate and β is the exponential decay parameter that controls weight updates.

[0103] In diesel generator sets, the combined failure of fuel oxidation and sensor drift will destroy the physical conservation relationship of the system and cause the traditional static model to fail due to changes in operating conditions. Based on the above method steps, the physical rationality and dynamic adaptability of the detection are ensured through conservation law verification and elastic network dynamic parameter adjustment, and a high combined fault detection rate is maintained under dynamic conditions, providing highly reliable operation and maintenance guarantees for diesel units.

[0104] The mass flow residual, energy conversion efficiency deviation and momentum conservation error in the combustion process are calculated in real time. The mass flow residual reflects the balance between fuel consumption and exhaust mass, the energy conversion efficiency deviation reflects the ratio of theoretical calorific value to actual output power, and the momentum conservation error reflects the integral of cylinder pressure fluctuation. The threshold is set dynamically, for example, the mass residual threshold is adjusted linearly with the load change rate. If the residual exceeds the limit continuously, the model is judged to be inaccurate and the parameters are reset to avoid cumulative errors caused by sensor drift or oxidation. For example, the drift of oxygen sensor 201 falsely increases the combustion efficiency, but the energy residual often triggers correction.

[0105] Then, a non-convex optimization rule is used to update the feature weights, preferentially suppressing the contribution of noise-sensitive features, constraining the absolute value of the weights and controlling the model complexity to prevent overfitting. At the same time, an exponential decay factor is introduced to make the weight update amplitude adaptively adjusted with the gradient strength, balancing the convergence speed and stability. The constraint boundaries are automatically expanded every 24 hours to adapt to the feature distribution drift caused by unit aging.

[0106] Finally, when the conservation law verification conflicts with the elastic network judgment, for example, oxidation causes the mass residual to increase but the feature weight does not trigger an alarm, the manual review process is initiated, and the source of the fault is located in combination with tensor decomposition visualization.

[0107] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A diesel generator set pre-start detection and control system, characterized in that: It comprises a data processing module (5), and a fuel system detection module (1), a combustion system detection module (2), an emission system detection module (3), and a control system information extraction module (4) which are communicatively connected thereto; The fuel system detection module (1), the combustion system detection module (2), the emission system detection module (3), and the control system information extraction module (4) collect parameters and transmit them to the data processing module (5); The data processing module (5) receives the parameters transmitted by each detection module, and performs consistency alignment on the spatiotemporal data of multiple sensors through the Lie group differential method to eliminate the spatiotemporal mismatch in the sensor network; The nonlinear dynamic characteristics caused by fuel oxidation are extracted using high-order recursive quantitative analysis, and the fuel viscoelastic degradation process is quantified through multi-scale Cauchy-Green tensor analysis. A fast approximation algorithm based on the curvature change of the Fisher information matrix and the Wasserstein distance is used to detect sensor signal drift. Using high-order tensor decomposition and tensor singular value entropy, multi-physical domain features are integrated and a joint feature tensor is generated. The decision boundary is optimized by diffeomorphic mapping and the composite fault classification is realized by combining nonlinear support tensor machine. Based on the conservation law residual verification and parameter elastic network dynamic parameter adjustment mechanism, the physical rationality and robustness of the algorithm output are ensured.

2. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The fuel system detection module (1) comprises a fuel flow meter (102) and a fuel pressure sensor (101), the combustion system detection module (2) comprises an oxygen sensor (201) and an exhaust temperature sensor (202), the emission system detection module (3) comprises a cylinder pressure sensor (301) and a NOx sensor (302), the control system information extraction module (4) extracts an ECU air-fuel ratio correction value and an injection pulse width instruction, and the data processing module (5) comprises a processor.

3. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The data processing module (5) receives the parameters transmitted by each detection module, performs consistency alignment on the spatiotemporal data of multiple sensors by using the Lie group differential method, and eliminates the spatiotemporal mismatch in the sensor network, including the following steps: The sensor network is defined as a Lie group structure, which consists of a three-dimensional rigid motion group and a sensor signal space, where the three-dimensional rigid motion group contains a rotation matrix and a translation vector; The spatiotemporal registration objective function is constructed to achieve alignment by minimizing the weighted distance of all sensor signals on the Riemann manifold, where the weighting coefficient is determined according to the signal-to-noise ratio of each sensor and the Riemann distance is calculated based on the logarithmic difference integral of the signal power spectral density. The Lie algebraic gradient descent method is used to iteratively optimize the group parameters. In each iteration, the group action parameters are updated according to the gradient direction of the loss function. The learning rate is fixed at 0.01, and the iteration termination condition is that the change in the loss function value for two consecutive times is less than the set threshold.

4. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The method of extracting the nonlinear dynamic characteristics caused by fuel oxidation by high-order recursive quantitative analysis and quantifying the fuel viscoelastic degradation process by multi-scale Cauchy-Green tensor analysis includes the following steps: The phase space of the fuel pressure signal is reconstructed, the time delay is determined by the mutual information method, and a multi-dimensional recursive matrix is ​​constructed. The recursive entropy spectrum is calculated by statistically analyzing the probability distribution of the diagonal length in the recursive matrix. When the embedding dimension is 5, if the entropy value drops by more than 20%, it is determined that the fuel is oxidized. The velocity field gradient is extracted from the fuel flow signal, and the Cauchy-Green tensor of the flow mapping Jacobian matrix is ​​calculated. The damage index is constructed by comparing the relative deviation between the real-time eigenvalue and the benchmark value. If the index exceeds 30% for more than 5 sampling cycles, the fuel viscoelastic degradation is confirmed.

5. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The fast approximation algorithm based on the curvature change of the Fisher information matrix and the Wasserstein distance detects the drift of the sensor signal, comprising the following steps: Assuming that the oxygen sensor (201) signal obeys Gaussian distribution, the curvature tensor of its Fisher information matrix is ​​calculated, a normal operating condition reference curvature is established through Monte Carlo simulation, the Frobenius norm of the curvature change is detected in real time, and a drift warning is triggered when it exceeds a preset threshold; An empirical distribution is constructed for the NOx sensor (302) signal, and a Sinkhorn iterative algorithm is used to quickly and approximately calculate the Wasserstein distance between the signal and the theoretical distribution. Combined with a dynamically adjusted statistical threshold, it is determined whether the sensor is abnormal.

6. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The method of utilizing high-order tensor decomposition and tensor singular value entropy to fuse multiple physical domain features and generate a joint feature tensor includes the following steps: Construct a three-dimensional joint feature tensor, where the dimensions correspond to the number of sensors, feature type, and number of sliding time windows, and each element is generated by a windowed feature extraction function; Tucker decomposition is used to decompose high-order tensors into the product of core tensors and factor matrices. The system health status is evaluated by counting the proportion of non-zero elements in the core tensor and calculating the entropy of the tensor singular value distribution. If the entropy is lower than the preset level, a compound fault is confirmed.

7. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The method of optimizing the decision boundary by differential homeomorphism mapping and implementing compound fault classification by combining a nonlinear support tensor machine includes the following steps: A nonlinear state space mapping is constructed, and the decision boundary is optimized by balancing the mapping smoothness and classification accuracy in the objective function, and the spectral method with a cutoff frequency of 10 is used to solve it. An exponential kernel function based on tensor Frobenius distance is defined, supporting tensors are selected from the historical fault library, and the classifier weights are solved through the sequential minimum optimization algorithm. The final decision function is the symbolic function of weighted kernel similarity.

8. A diesel generator set pre-start detection and control system according to claim 1, characterized in that: The residual verification based on conservation law and the dynamic parameter adjustment mechanism of parameter elastic network ensure the physical rationality and robustness of the algorithm output, including the following steps: The residual is calculated in real time based on the law of conservation of mass, and the judgment threshold is dynamically adjusted in combination with the load change rate. If the residual exceeds the threshold, the model parameters are reset; A non-convex optimization rule with an exponential decay factor is used to update the feature weights, constraining the absolute value of the weights not to exceed a fixed value, and automatically expanding the constraint boundaries every 24 hours to maintain the flexibility of the algorithm.

Images