Fault prediction and health management method for super-high pressure high-power diesel generator set

Abstract

This invention relates to the field of diesel generator set condition monitoring and fault diagnosis technology, specifically disclosing a method for fault prediction and health management of ultra-high voltage high-power diesel generator sets. The method involves collecting multi-source sensor data to construct a high-dimensional state space matrix; mapping data points to a low-dimensional manifold space to form a manifold distribution; extracting the operating condition cycle reference manifold and calculating the geodesic distance to obtain the manifold deviation; decomposing the manifold deviation into operating condition change components along the operating condition cycle trajectory and fault degradation components perpendicular to the operating condition cycle trajectory, extracting the fault degradation components as fault feature indicators; comparing the fault feature indicators with an adaptive early warning threshold to generate early warning information and updating the operating condition cycle reference manifold. This invention decouples the essence of operating condition changes and fault degradation through manifold space decomposition, enabling robust extraction of early minor fault features under conditions of severe load fluctuations, achieving adaptive dynamic early warning, and significantly improving the accuracy and reliability of fault diagnosis under varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of diesel generator set condition monitoring and fault diagnosis technology, specifically to a method for fault prediction and health management of ultra-high voltage high-power diesel generator sets. Background Technology

[0002] Ultra-high voltage, high-power diesel generator sets typically refer to large-scale power generation equipment with fuel injection pressures exceeding 200 MPa and single-unit power outputs exceeding several thousand kilowatts. They are widely used in critical applications such as data centers, nuclear power plant emergency power supplies, and electric propulsion for large ships. Their operational reliability directly impacts the power supply security of downstream loads. In actual operation, these units face complex conditions such as high altitudes, frigid environments, and frequent, drastic load fluctuations. Traditional periodic maintenance and reactive repair methods are insufficient to meet high reliability requirements. Therefore, fault prediction and health management technologies have become crucial means to ensure the safe operation of these units.

[0003] When ultra-high voltage high-power diesel generator sets face severe transient conditions with load change rates exceeding 20%, how can we fundamentally decouple and separate early micro-fault characteristics such as micron-level wear of the plunger pair from numerical disturbances caused by operating condition fluctuations in strongly nonlinearly coupled sensing signals? In other words, how can we solve the problem of characteristic isomorphism in traditional residual analysis, where operating condition changes and fault degradation both deviate from the baseline and cannot be distinguished in Euclidean space? Summary of the Invention

[0004] The purpose of this invention is to provide a method for fault prediction and health management of ultra-high voltage high-power diesel generator sets, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets includes the following steps:

[0007] S1: Collect multi-source sensor data of ultra-high voltage high-power diesel generator set during operation and construct a high-dimensional state space matrix containing time series.

[0008] S2: Map the data points in the high-dimensional state space matrix to the low-dimensional manifold space to form a manifold distribution that represents the unit's operating state. Different operating conditions correspond to different regions in the manifold space, and transient processes correspond to continuous transition trajectories in the manifold space.

[0009] S3: Extract the reference manifold for the operating cycle based on the manifold distribution, map the real-time data points to the manifold space, calculate the geodesic distance between the real-time data points and the reference manifold for the operating cycle, and obtain the manifold deviation.

[0010] S4: Decompose the manifold deviation into the condition change component along the condition cycle trajectory and the fault degradation component perpendicular to the condition cycle trajectory, and extract the fault degradation component as a fault characteristic index.

[0011] S5: Compare the fault characteristic indicators with the adaptive warning threshold. When the fault characteristic indicators exceed the adaptive warning threshold, generate warning information and feed the warning information and the current data point back to the operating condition cyclic reference manifold to update the operating condition cyclic reference manifold.

[0012] As a further aspect of the present invention: the generation process of the high-dimensional state space matrix is ​​as follows:

[0013] The multi-source sensor data includes: ultra-high pressure common rail pressure, injector return flow rate, cylinder head vibration acceleration, transient speed fluctuation, and combustion knock pressure;

[0014] The ultra-high pressure common rail pressure, injector return oil flow rate and combustion explosion pressure are encapsulated into independent time series vectors according to their original sampling frequencies;

[0015] Using crankshaft rotation angle as a unified reference, linear interpolation is used to synchronize each time series vector to the timing node corresponding to the highest sampling frequency, forming a multi-dimensional synchronization matrix.

[0016] Cross-correlation analysis was performed on cylinder head vibration acceleration and transient speed fluctuations. The delay time corresponding to the maximum correlation coefficient between the two in the time domain was extracted. Phase compensation was performed on the multidimensional synchronization matrix based on the delay time to obtain the high-dimensional state space matrix.

[0017] As a further aspect of the present invention: S2 specifically includes:

[0018] Using each data point in the high-dimensional state space matrix as a node, calculate the Euclidean distance between each data point and other data points in its neighborhood, select the nearest preset number of data points as adjacent points, and construct a local neighborhood graph.

[0019] Within each local neighborhood graph, the optimal weight coefficients for linear reconstruction of each data point from its neighboring points are solved to minimize the reconstruction error.

[0020] Keeping the optimal weight coefficients unchanged, we solve for the embedding coordinates of each data point in the low-dimensional space to minimize the local reconstruction error of each data point in the low-dimensional space, and use the solved embedding coordinates as the manifold distribution.

[0021] As a further aspect of the present invention: the construction of the local neighborhood graph specifically includes:

[0022] For each data point in the high-dimensional state space matrix, the operating condition interval is divided according to the transient speed fluctuation value corresponding to the data point. The Euclidean distance between the corresponding data point and other data points in the same interval is calculated in each operating condition interval.

[0023] The Euclidean distances obtained for each data point under different operating conditions are weighted and fused.

[0024] Adjacent points are selected based on the weighted fusion distance, and a local neighborhood graph is constructed using these adjacent points.

[0025] As a further aspect of the present invention: S3 specifically includes:

[0026] Based on the transient speed fluctuation values ​​corresponding to each data point in the manifold distribution, the manifold distribution is divided into multiple operating cycle segments, and each operating cycle segment corresponds to a complete load cycle process.

[0027] Within each working condition cycle segment, the local curvature of each data point is calculated, and continuous data points with local curvature less than a preset curvature threshold are selected to form candidate cyclic trajectory segments.

[0028] The candidate cyclic trajectory segments corresponding to each working condition cycle segment are weighted and superimposed, and the trajectory line formed by the weighted superposition is used as the working condition cycle reference manifold.

[0029] As a further aspect of the present invention: S4 specifically includes:

[0030] Based on the transient speed fluctuation value corresponding to the real-time data point, a reference point corresponding to the transient speed fluctuation value is determined on the working condition cycle trajectory in the manifold space, and a local orthogonal basis is established with the tangential direction of the reference point.

[0031] The manifold deviation is projected onto the tangential and normal directions of the local orthogonal basis to obtain the tangential component of the real-time data point in the tangential direction and the normal component in the normal direction.

[0032] Extract the normal component along the normal direction, and accumulate the normal component within a preset time window. Use the accumulated result as the fault characteristic value.

[0033] As a further aspect of the present invention: the establishment of a local orthogonal basis specifically includes:

[0034] Obtain the historical transient speed fluctuation value corresponding to each trajectory point on the working condition cycle trajectory, and use the reciprocal of the difference between the historical transient speed fluctuation value of each trajectory point and the transient speed fluctuation value of the real-time data point as the weight to weight the trajectory points on the working condition cycle trajectory.

[0035] A set of neighborhood points is formed by selecting a preset number of trajectory points with the largest weight after weighting. Local principal component analysis is performed on the neighborhood point set to extract the direction of the first principal component as the tangential direction and the direction of the second principal component as the normal direction.

[0036] Using the geometric center of each trajectory point in the neighborhood point set as the reference point, a local orthogonal basis is established with the tangential and normal directions.

[0037] As a further aspect of the present invention: S5 specifically includes:

[0038] A statistical distribution is constructed based on historical fault characteristic indicators. The quantile of the statistical distribution is used as the basic threshold. The basic threshold is dynamically corrected according to the transient speed fluctuation value at the current moment. The amount of dynamic correction is positively correlated with the difference between the current transient speed fluctuation value and the historical average transient speed fluctuation value, thus obtaining an adaptive warning threshold.

[0039] The fault characteristic indicators are compared with the adaptive warning threshold. When the fault characteristic indicators exceed the adaptive warning threshold for three consecutive times, a warning message is generated.

[0040] Mark the current data point corresponding to the generation of early warning information as an anomaly point, mark the remaining current data points that have not generated early warning information as normal points, incorporate the normal points into the construction dataset of the working condition cyclic baseline manifold, and re-execute the step of extracting the working condition cyclic baseline manifold based on the manifold distribution to complete the update of the working condition cyclic baseline manifold.

[0041] The beneficial effects of this invention are:

[0042] (1) This invention maps high-dimensional sensing data to a low-dimensional manifold space. Utilizing the essential difference between the ergodicity of operating condition changes and the unidirectionality of fault degradation in the manifold distribution, it employs local orthogonal basis decomposition to decouple the manifold deviation into operating condition change components along the operating condition cycle trajectory and fault degradation components perpendicular to the operating condition cycle trajectory. The normal component is then extracted as a fault characteristic index. Compared to traditional residual analysis methods, this invention solves the technical problem of small fault characteristic signals being submerged by operating condition fluctuations in existing technologies, enabling reliable identification and early warning of early faults.

[0043] (2) This invention continuously incorporates normal data points that do not trigger warnings into the constructed dataset, dynamically updating the operating condition cycle reference manifold. This allows the reference manifold to adaptively track the slow performance changes caused by normal wear and tear from long-term operation. Simultaneously, a basic threshold is constructed based on the 95th percentile of historical fault characteristic indicators, and dynamically corrected according to the difference between the current transient speed fluctuation value and the historical average value to obtain an adaptive warning threshold. This mechanism avoids false alarms caused by misjudging the normal aging process of the unit as a fault, and also prevents missed fault detection due to reference drift. While ensuring high warning accuracy, it extends the effective warning window period, providing a reliable decision-making basis for condition-based predictive maintenance. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Figure 1 This is a flowchart of the method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to the present invention.

[0046] Figure 2 This is a flowchart of the process of establishing a local orthogonal basis in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Please see Figure 1 As shown, this invention is a method for fault prediction and health management of ultra-high voltage high-power diesel generator sets, comprising the following steps:

[0049] S1: Collect multi-source sensor data of ultra-high voltage high-power diesel generator set during operation and construct a high-dimensional state space matrix containing time series.

[0050] S2: Map the data points in the high-dimensional state space matrix to the low-dimensional manifold space to form a manifold distribution that represents the unit's operating state. Different operating conditions correspond to different regions in the manifold space, and transient processes correspond to continuous transition trajectories in the manifold space.

[0051] S3: Extract the reference manifold for the operating cycle based on the manifold distribution, map the real-time data points to the manifold space, calculate the geodesic distance between the real-time data points and the reference manifold for the operating cycle, and obtain the manifold deviation.

[0052] S4: Decompose the manifold deviation into the condition change component along the condition cycle trajectory and the fault degradation component perpendicular to the condition cycle trajectory, and extract the fault degradation component as a fault characteristic index.

[0053] S5: Compare the fault characteristic indicators with the adaptive warning threshold. When the fault characteristic indicators exceed the adaptive warning threshold, generate warning information and feed the warning information and the current data point back to the operating condition cyclic reference manifold to update the operating condition cyclic reference manifold.

[0054] In S1, multi-source sensor data of the ultra-high voltage high-power diesel generator set during operation are collected to construct a high-dimensional state-space matrix containing time series data, specifically including:

[0055] A piezoresistive pressure sensor is installed on the ultra-high pressure common rail line of the ultra-high pressure high-power diesel generator set to collect the ultra-high pressure common rail pressure; a flow sensor is installed on the return line of each injector to collect the injector return flow rate; an acceleration sensor is installed on the cylinder head of each cylinder to collect cylinder head vibration acceleration; a magnetoelectric speed sensor is installed at the crankshaft flywheel end to collect transient speed fluctuations; and a piezoelectric cylinder pressure sensor is installed in the cylinder pressure sensor mounting hole of each cylinder to collect combustion knock pressure. All of the above sensors are triggered and acquired through a synchronous data acquisition card to ensure that the time reference of each data stream is consistent.

[0056] The collected ultra-high pressure common rail pressure, injector return oil flow rate, and combustion explosion pressure are encapsulated into independent time series vectors according to their original sampling frequencies. Each time series vector contains multiple sample values ​​arranged in chronological order and their corresponding timestamps.

[0057] Using crankshaft angle as a unified time reference, the three time series vectors mentioned above, along with the cylinder head vibration acceleration and transient speed fluctuation time series vectors, are synchronized. Specifically, the highest sampling frequency among all sensors is determined, and the sampling period corresponding to this frequency is used as the reference period. For time series vectors with sampling frequencies lower than the highest sampling frequency per second, the time series node of this reference period is taken as the target position. Two adjacent actual sampling points are taken before and after each target position. The weight is calculated based on the time difference between the timestamps of the two actual sampling points and the target position. The values ​​of the two actual sampling points are then weighted and averaged according to this weight. The weighted average result is used as the synchronization interpolation at the target position. After the above linear interpolation processing, all time series vectors are synchronized to the unified time series node with the highest sampling frequency, forming a multi-dimensional synchronization matrix. Each row of this multi-dimensional synchronization matrix corresponds to a time series node, and each column corresponds to a type of sensor data.

[0058] Cross-correlation analysis is performed on the cylinder head vibration acceleration column vector and the transient speed fluctuation column vector in the multidimensional synchronization matrix to eliminate phase deviations caused by differences in sensor installation positions or signal transmission paths. The specific calculation process is as follows: The cylinder head vibration acceleration column vector is kept fixed, while the transient speed fluctuation column vector is moved step-by-step according to time displacement. The displacement step size is set to the time interval between adjacent time nodes, i.e., the period corresponding to the highest sampling frequency. For each displacement, the value of each time node in the cylinder head vibration acceleration column vector is multiplied by the value of the corresponding time node in the moved transient speed fluctuation column vector, and all products are summed to obtain the cross-correlation value for that displacement. After traversing all preset displacement ranges, the displacement corresponding to the maximum cross-correlation value is selected as the delay time.

[0059] Phase compensation is performed on the multidimensional synchronization matrix based on the aforementioned delay time. Specifically: when the delay time is positive, it indicates that the transient speed fluctuation signal lags behind the cylinder head vibration acceleration signal. The value corresponding to each time-series node of the transient speed fluctuation column vector in the multidimensional synchronization matrix is ​​replaced with the value corresponding to that time-series node shifted backward by the delay time. When the delay time is negative, it indicates that the transient speed fluctuation signal leads the cylinder head vibration acceleration signal. The value corresponding to each time-series node of the transient speed fluctuation column vector is replaced with the value corresponding to that time-series node shifted forward by the absolute value of the delay time. After phase compensation, the cylinder head vibration acceleration and transient speed fluctuation at the same time-series node represent the state of the unit at the same physical moment. This compensated multidimensional synchronization matrix is ​​output as a high-dimensional state space matrix for subsequent analysis and processing.

[0060] In S2, data points in the high-dimensional state-space matrix are mapped to a low-dimensional manifold space, forming a manifold distribution representing the unit's operating state. Different operating conditions correspond to different regions in the manifold space, and transient processes correspond to continuous transition trajectories in the manifold space, specifically including:

[0061] When performing the mapping from the high-dimensional state space matrix to the low-dimensional manifold space, a local neighborhood graph needs to be constructed first to determine the topological relationship between each data point and its surrounding data points. The specific construction method is as follows: Each data point in the high-dimensional state space matrix is ​​taken as the current data point, and its operating range is determined based on the transient speed fluctuation value corresponding to that data point. The operating range is divided as follows: the entire range of transient speed fluctuation values, from the minimum to the maximum value, is divided into 10 equal intervals, each interval having a width of one-tenth of the range of transient speed fluctuation values. For the current data point, the Euclidean distance between the data point and other data points within each operating range is calculated. The Euclidean distance is calculated as follows: the numerical difference between two data points in each dimension is taken, the squares of the differences in each dimension are summed, and the square root is taken. The Euclidean distances obtained by the current data point under different operating conditions are weighted and fused. The weights for this weighting are determined as follows: The number of data points in each operating condition interval is counted, and the percentage of data points in that interval to the total number of data points is used as the density value of that interval. The reciprocal of this density value is then used as the weight of that interval, ensuring that intervals with dense data point distribution have smaller weights and intervals with sparse data point distribution have larger weights. The Euclidean distances calculated between the current data point and each other data point under different operating conditions are multiplied by the corresponding weights of the operating conditions, and then summed to obtain the weighted fused distance. Based on the weighted fused distance, a preset number of other data points with the closest distance are selected as the neighbors of the current data point. The preset number is set to 12. A local neighborhood graph is constructed using each data point as a node and the lines connecting each data point to its neighbors as edges.

[0062] After constructing the local neighborhood graph, the optimal weight coefficients for linear reconstruction of each data point from its neighboring points are determined. Specifically, for each data point, all its neighboring points in the local neighborhood graph are taken, forming the neighboring point set. A set of weight coefficients is defined, with each weight coefficient corresponding to one neighboring point, and the sum of all weight coefficients is 1. The difference between the numerical vector of the data point and the vector obtained by weighting and summing the numerical vectors of all neighboring points according to the weight coefficients is calculated, and the square of this difference is taken as the reconstruction error. The set of weight coefficients that minimizes the reconstruction error is then determined using the Lagrange multiplier method, i.e., by solving a system of linear equations to obtain the numerical solutions for each weight coefficient. The obtained weight coefficients are then used as the local reconstruction weights for that data point.

[0063] Keeping the local reconstruction weights of each data point obtained above unchanged, the embedding coordinates of each data point are solved in a low-dimensional space. The dimension of the low-dimensional space is set to 3. Initial embedding coordinates of each data point in the low-dimensional space are set. For each data point, its set of neighboring points is taken, and the embedding coordinates of each neighboring point are weighted and summed according to the determined local reconstruction weights of the data point. The difference between the embedding coordinates of the data point and the weighted summation result is calculated, and the square of this difference is taken as the reconstruction error in the low-dimensional space. The low-dimensional reconstruction errors of all data points are summed to obtain the total reconstruction error. The embedding coordinates of each data point are adjusted using an iterative optimization algorithm to minimize the total reconstruction error. The iterative optimization algorithm adopts the conjugate gradient method, with 500 iterations. The iteration termination condition is set to the relative change in the total reconstruction error between two adjacent iterations being less than one-thousandth. The coordinates of each data point obtained in the final solution in 3D space are used as the embedded coordinates of the data point. The embedded coordinates of all data points together constitute the manifold distribution. Each data point in the manifold distribution corresponds to the operating state of the unit at a sampling time. Data points corresponding to different steady-state conditions are clustered into different regions in the manifold distribution. Data points corresponding to transient processes are connected to different regions in the manifold distribution to form a continuous transition trajectory.

[0064] In S3, the reference manifold for the operating cycle is extracted based on the manifold distribution. After mapping real-time data points to the manifold space, the geodesic distance between the real-time data points and the reference manifold for the operating cycle is calculated to obtain the manifold deviation, which specifically includes:

[0065] When extracting the manifold distribution into a reference manifold for the load cycle, the load cycle segments are first divided based on the transient speed fluctuation values ​​corresponding to each data point in the manifold distribution. Specifically, the segmentation method involves peak detection of the sequence of transient speed fluctuation values ​​over time to identify all local maxima and minima. Starting from any local minima and ending at the next local minima, this continuous transient speed fluctuation process is defined as a complete load cycle. This process is repeated for all data points, dividing the manifold distribution into multiple load cycle segments. Each load cycle segment corresponds to a complete load cycle, and each segment contains multiple data points arranged in chronological order and their embedded coordinates within the manifold distribution.

[0066] Within each operating cycle segment, the local curvature of each data point is calculated. Specifically, for each data point within the operating cycle segment, its two adjacent data points within that segment are taken as the midpoint, and the two adjacent data points are designated as the preceding and following points, respectively. The Euclidean distances between the midpoint and the preceding point, between the midpoint and the following point, and between the preceding and following points are calculated separately. The Euclidean distance between the preceding and following points is taken as the chord length. Based on the geometric relationship of triangles, the angle of change of the tangential direction at the midpoint is calculated from the above three distances. The ratio obtained by dividing this angle by the chord length is taken as the local curvature of that midpoint. For the first and last data points within the operating cycle segment, the local curvature is calculated in the same way for their two adjacent data points. A preset curvature threshold is set to 0.5, and continuous data points with local curvatures less than 0.5 are selected to form candidate cyclic trajectory segments.

[0067] The candidate cyclic trajectory segments corresponding to each operating cycle segment are weighted and superimposed to form the operating cycle reference manifold. Specifically, the superposition method is as follows: all candidate cyclic trajectory segments in all operating cycle segments are aligned according to their positions in the manifold space, with the geometric center of each candidate cyclic trajectory segment as the alignment reference. For each position point in the manifold space, the number of trajectory segments containing data points within a preset radius of that position point is counted, and this number is used as the superposition count for that position point. The weight of each candidate cyclic trajectory segment is set as the ratio of the number of data points within that trajectory segment to the total number of data points in all candidate cyclic trajectory segments, and this ratio is used as the contribution weight of that trajectory segment. For each position point in the manifold space, the embedded coordinates of all candidate cyclic trajectory segments within the preset radius of that position point are weighted and averaged according to the contribution weights of each trajectory segment. The resulting coordinate points are used as trajectory points on the operating cycle reference manifold, and all trajectory points are connected sequentially to form the operating cycle reference manifold.

[0068] The preset radius is defined as: a value obtained by multiplying the arithmetic mean of the Euclidean distances between adjacent trajectory points in all candidate cyclic trajectory segments by 1.5 to 2.5.

[0069] After obtaining the operating condition cycle reference manifold, real-time data points acquired in real time and processed using the same mapping method are mapped to the manifold space to obtain the embedded coordinates of the real-time data points in the manifold space. The Euclidean distances between the embedded coordinates of the real-time data points and all trajectory points on the operating condition cycle reference manifold are calculated. The smallest Euclidean distance is selected as the geodesic distance between the real-time data point and the operating condition cycle reference manifold. This geodesic distance is output as the manifold deviation, which characterizes the degree to which the unit's operating state deviates from the normal cycle trajectory at the current moment.

[0070] Please seeFigure 2 As shown, in S4, the manifold deviation is decomposed into a condition variation component along the operating cycle trajectory and a fault degradation component perpendicular to the operating cycle trajectory. The fault degradation component is extracted as a fault characteristic index, specifically including:

[0071] When performing the decomposition of the manifold deviation and the extraction of fault characteristic indicators, it is first necessary to determine a reference point corresponding to the real-time data point on the operating condition cycle trajectory in the manifold space based on the transient speed fluctuation value corresponding to the real-time data point, and then establish a local orthogonal basis using the tangential and normal directions of the reference point. The specific implementation method is as follows:

[0072] The historical transient speed fluctuation value corresponding to each trajectory point on the operating condition cycle trajectory is obtained. This historical transient speed fluctuation value is recorded and stored when the operating condition cycle trajectory is constructed. For the current real-time data point, let its real-time transient speed fluctuation value be... For any trajectory point on the operating cycle trajectory, let its historical transient speed fluctuation value be... subscript This represents the sequence number of the trajectory point. The absolute value of the difference between the historical transient speed fluctuation value and the real-time transient speed fluctuation value for each trajectory point is calculated. The reciprocal of the difference is taken as the initial weight of the trajectory point. To prevent the weight from becoming too large due to a small difference, leading to numerical instability, a minimum constant of 0.001 is superimposed on the reciprocal of the difference. The initial weight of each trajectory point is then normalized by dividing the initial weight of each trajectory point by the sum of the initial weights of all trajectory points, resulting in the normalized weight, denoted as . .

[0073] According to the normalized weights The points are sorted from largest to smallest, and the 15 trajectory points with the highest weights are selected to form a neighborhood point set. Local principal component analysis (LPC) is performed on this neighborhood point set to extract the tangential and normal directions. The specific calculation process of LPC is as follows: Calculate the mean of the three-dimensional coordinates of all trajectory points in the neighborhood point set in the manifold space to obtain the coordinates of the center point. Subtract the coordinates of the center point from the three-dimensional coordinates of each trajectory point in the neighborhood point set to obtain a decentralized coordinate vector. Multiply the decentralized coordinate vector of each trajectory point by its own transpose to obtain a 3x3 matrix. Sum the matrices obtained from all trajectory points and divide by the number of trajectory points to obtain the covariance matrix. Solve for the three eigenvalues ​​and corresponding three eigenvectors of this covariance matrix. The eigenvector corresponding to the largest eigenvalue is taken as the first principal component direction, and the eigenvector corresponding to the second largest eigenvalue is taken as the second principal component direction. The first principal component direction is taken as the tangential direction, and the second principal component direction is taken as the normal direction. The tangential and normal directions are orthogonal to each other. Using the center point as a reference point, a local orthogonal basis is established with the reference point, the tangential direction, and the normal direction.

[0074] After establishing the local orthogonal basis, the manifold deviation is projected onto the tangential and normal directions of the local orthogonal basis. Specifically, the projection method is as follows: subtract the coordinates of the reference point from the embedded coordinates of the real-time data point in the manifold space to obtain the displacement vector from the reference point to the real-time data point. The dot product of this displacement vector and the unit vector in the tangential direction is calculated; this is the projected length of the displacement vector in the tangential direction, and this projected length is taken as the tangential component of the real-time data point in the tangential direction. Similarly, the dot product of this displacement vector and the unit vector in the normal direction is calculated; this projected length is taken as the normal component of the real-time data point in the normal direction.

[0075] The calculated normal component is used as the instantaneous deviation value at the current moment. A preset time window is set, with a length of 60 sampling moments. At each sampling moment, the normal component at the current moment is summed with the normal components from the past 59 moments, using the following summation formula:

[0076] ;

[0077] in, This represents the total number of sampling moments within the preset time window, with a value of 60. Indicates the number of times within the preset time window The normal component values ​​at each sampling time; the summation symbol indicates that... All from 1 to n Perform accumulation; divide the accumulated result by The mean value of the normal component within the window is obtained, and this mean value is output as the fault characteristic index at the current moment. This fault characteristic index characterizes the average degree to which the unit deviates from the normal operating condition cycle trajectory in a recent period of time, and the direction of this deviation is perpendicular to the tangent direction of the operating condition cycle trajectory, thereby effectively separating the irreversible drift caused by fault degradation from the periodic fluctuations caused by changes in operating conditions.

[0078] In S5, fault characteristic indicators are compared with adaptive warning thresholds. When a fault characteristic indicator exceeds the adaptive warning threshold, a warning message is generated and fed back to the operating condition cyclic reference manifold along with the current data point, updating the operating condition cyclic reference manifold. Specifically, this includes:

[0079] When performing the comparison and early warning generation of the aforementioned fault characteristic indicators, an adaptive early warning threshold needs to be constructed first. The specific construction method is as follows: Collect historical fault characteristic indicators from the past 1000 sampling times, and sort these 1000 historical fault characteristic indicators in ascending order of value. Take the 950th value after sorting as the 95th percentile, and use this percentile as the base threshold. Calculate the arithmetic mean of the transient speed fluctuation values ​​from the past 1000 sampling times to obtain the historical average transient speed fluctuation value. For the current moment, obtain the real-time transient speed fluctuation value, and calculate the difference between this real-time transient speed fluctuation value and the historical average transient speed fluctuation value. Set a correction coefficient, with a value of 0.05. Multiply the above difference by the correction coefficient to obtain the correction amount, which is directly proportional to the difference. Add the base threshold and the correction amount to obtain the adaptive early warning threshold for the current moment.

[0080] After obtaining the adaptive warning threshold, the fault characteristic indicator at the current moment is compared with the adaptive warning threshold. When the fault characteristic indicator is greater than the adaptive warning threshold, the moment is marked as an out-of-limit moment; when the fault characteristic indicator is less than or equal to the adaptive warning threshold, the moment is marked as a normal moment. The number of consecutive out-of-limit moments is recorded. When the number of consecutive out-of-limit moments reaches 3, a warning message is generated. This warning message includes the timestamp of the current moment, the value of the fault characteristic indicator, and the value of the adaptive warning threshold.

[0081] After generating an early warning message, the current data point corresponding to the time of message generation is marked as an anomaly. This anomaly includes the high-dimensional state space matrix data at that moment, the embedded coordinates in the manifold distribution, and the values ​​of fault characteristic indicators. For the remaining current data points that did not generate early warning messages, they are marked as normal points. All data points marked as normal points are incorporated into the construction dataset of the operating condition cyclic reference manifold. This construction dataset is merged with the original dataset used to construct the operating condition cyclic reference manifold, forming an updated construction dataset. Based on this updated construction dataset, the step of extracting the operating condition cyclic reference manifold based on the manifold distribution is re-executed, i.e., the process of re-dividing the operating condition cycle segments, calculating local curvature, selecting candidate cyclic trajectory segments, and weighted superposition is performed. The re-extracted trajectory line is used as the updated operating condition cyclic reference manifold. Through this update mechanism, the operating condition cyclic reference manifold can adapt to the slow performance changes caused by normal wear and tear after long-term unit operation, avoiding misjudging the normal aging process of the unit as a fault.

[0082] The working principle of this invention is as follows: First, multi-source sensor data, including ultra-high pressure common rail pressure, injector return flow, cylinder head vibration acceleration, transient speed fluctuation, and combustion knock pressure, are collected during unit operation. Time synchronization is achieved through linear interpolation using crankshaft angle as a unified benchmark, and phase compensation is performed using cross-correlation analysis of cylinder head vibration acceleration and transient speed fluctuation to construct a high-dimensional state space matrix. Second, the data points in this high-dimensional state space matrix are divided into operating condition intervals based on transient speed fluctuation values. Euclidean distances are calculated within each operating condition interval, and weighted fusion is performed based on the density of the operating condition intervals. Adjacent points are selected to construct a local neighborhood graph. Through local linear reconstruction and global optimization, the data points are mapped to a three-dimensional low-dimensional manifold space, forming a manifold distribution representing the unit's operating state. Then, operating condition cycle segments are divided based on the transient speed fluctuation values ​​corresponding to each data point in the manifold distribution. The local curvature of the data points is calculated within each operating condition cycle segment, and continuous points with curvature less than a preset threshold are selected to form candidate cyclic trajectory segments. The candidate cyclic trajectory segments are weighted and superimposed to extract the operating condition cycle benchmark manifold. The process involves mapping real-time data points to manifold space and calculating the geodesic distance between the data points and the baseline manifold for the operating cycle as the manifold deviation. Next, based on the transient speed fluctuations of the real-time data points, a neighborhood point set is selected using the reciprocal of the difference between the historical transient speed fluctuations and the real-time values ​​of each trajectory point on the operating cycle trajectory as the weight. Local principal component analysis is used to extract the tangential and normal directions to establish a local orthogonal basis. The manifold deviation is then decomposed into operating condition variation components along the operating cycle trajectory and fault degradation components perpendicular to the operating cycle trajectory. The normal components are accumulated and averaged within a preset time window to obtain fault characteristic indicators. Finally, a basic threshold is constructed based on the 95th percentile of the historical fault characteristic indicators. This basic threshold is dynamically adjusted based on the difference between the current transient speed fluctuation value and the historical average value to obtain an adaptive warning threshold. When the fault characteristic indicator exceeds this threshold for three consecutive moments, a warning message is generated. Simultaneously, normal data points that have not triggered a warning are included in the constructed dataset, and the extraction steps of the baseline manifold for the operating cycle are re-executed to achieve dynamic updates to the baseline manifold.

[0083] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for fault prediction and health management of ultra-high voltage high-power diesel generator sets, characterized in that, Includes the following steps: S1: Collect multi-source sensor data of ultra-high voltage high-power diesel generator set during operation and construct a high-dimensional state space matrix containing time series. S2: Map the data points in the high-dimensional state space matrix to the low-dimensional manifold space to form a manifold distribution that represents the unit's operating state. Different operating conditions correspond to different regions in the manifold space, and transient processes correspond to continuous transition trajectories in the manifold space. S3: Extract the reference manifold for the operating cycle based on the manifold distribution, map the real-time data points to the manifold space, calculate the geodesic distance between the real-time data points and the reference manifold for the operating cycle, and obtain the manifold deviation. S4: Decompose the manifold deviation into the condition change component along the condition cycle trajectory and the fault degradation component perpendicular to the condition cycle trajectory, and extract the fault degradation component as a fault characteristic index. S5: Compare the fault characteristic indicators with the adaptive warning threshold. When the fault characteristic indicators exceed the adaptive warning threshold, generate warning information and feed the warning information and the current data point back to the operating condition cyclic reference manifold to update the operating condition cyclic reference manifold.

2. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 1, characterized in that, The process of generating the high-dimensional state space matrix is ​​as follows: The multi-source sensor data includes: ultra-high pressure common rail pressure, injector return flow rate, cylinder head vibration acceleration, transient speed fluctuation and combustion knock pressure; The ultra-high pressure common rail pressure, injector return oil flow rate and combustion explosion pressure are encapsulated into independent time series vectors according to their original sampling frequencies; Using crankshaft rotation angle as a unified reference, linear interpolation is used to synchronize each time series vector to the timing node corresponding to the highest sampling frequency, forming a multi-dimensional synchronization matrix. Cross-correlation analysis was performed on cylinder head vibration acceleration and transient speed fluctuations. The delay time corresponding to the maximum correlation coefficient between the two in the time domain was extracted. Phase compensation was performed on the multidimensional synchronization matrix based on the delay time to obtain the high-dimensional state space matrix.

3. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 1, characterized in that, S2 specifically includes: Using each data point in the high-dimensional state space matrix as a node, calculate the Euclidean distance between each data point and other data points in its neighborhood, select the nearest preset number of data points as adjacent points, and construct a local neighborhood graph. Within each local neighborhood graph, the optimal weight coefficients for linear reconstruction of each data point from its neighboring points are solved to minimize the reconstruction error. Keeping the optimal weight coefficients unchanged, we solve for the embedding coordinates of each data point in the low-dimensional space to minimize the local reconstruction error of each data point in the low-dimensional space, and use the solved embedding coordinates as the manifold distribution.

4. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 3, characterized in that, The construction of the local neighborhood graph specifically includes: For each data point in the high-dimensional state space matrix, the operating condition interval is divided according to the transient speed fluctuation value corresponding to the data point. The Euclidean distance between the corresponding data point and other data points in the same interval is calculated in each operating condition interval. The Euclidean distances obtained for each data point under different operating conditions are weighted and fused. Adjacent points are selected based on the weighted fusion distance, and a local neighborhood graph is constructed using these adjacent points.

5. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 1, characterized in that, S3 specifically includes: Based on the transient speed fluctuation values ​​corresponding to each data point in the manifold distribution, the manifold distribution is divided into multiple operating cycle segments, and each operating cycle segment corresponds to a complete load cycle process. Within each working condition cycle segment, the local curvature of each data point is calculated, and continuous data points with local curvature less than a preset curvature threshold are selected to form candidate cyclic trajectory segments. The candidate cyclic trajectory segments corresponding to each working condition cycle segment are weighted and superimposed, and the trajectory line formed by the weighted superposition is used as the working condition cycle reference manifold.

6. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 1, characterized in that, S4 specifically includes: Based on the transient speed fluctuation value corresponding to the real-time data point, a reference point corresponding to the transient speed fluctuation value is determined on the working condition cycle trajectory in the manifold space, and a local orthogonal basis is established with the tangential direction of the reference point. The manifold deviation is projected onto the tangential and normal directions of the local orthogonal basis to obtain the tangential component of the real-time data point in the tangential direction and the normal component in the normal direction. Extract the normal component along the normal direction and accumulate the normal component within a preset time window. Use the accumulated result as a fault characteristic indicator.

7. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 6, characterized in that, The establishment of a local orthogonal basis specifically includes: Obtain the historical transient speed fluctuation value corresponding to each trajectory point on the working condition cycle trajectory, and use the reciprocal of the difference between the historical transient speed fluctuation value of each trajectory point and the transient speed fluctuation value of the real-time data point as the weight to weight the trajectory points on the working condition cycle trajectory. A set of neighborhood points is formed by selecting a preset number of trajectory points with the largest weight after weighting. Local principal component analysis is performed on the neighborhood point set to extract the direction of the first principal component as the tangential direction and the direction of the second principal component as the normal direction. Using the geometric center of each trajectory point in the neighborhood point set as the reference point, a local orthogonal basis is established with the tangential and normal directions.

8. The method for fault prediction and health management of ultra-high voltage high-power diesel generator sets according to claim 1, characterized in that, S5 specifically includes: A statistical distribution is constructed based on historical fault characteristic indicators. The quantile of the statistical distribution is used as the basic threshold. The basic threshold is dynamically corrected according to the transient speed fluctuation value at the current moment. The amount of dynamic correction is positively correlated with the difference between the current transient speed fluctuation value and the historical average transient speed fluctuation value, thus obtaining an adaptive warning threshold. The fault characteristic indicators are compared with the adaptive warning threshold. When the fault characteristic indicators exceed the adaptive warning threshold for three consecutive times, a warning message is generated. Mark the current data point corresponding to the generation of early warning information as an anomaly point, mark the remaining current data points that have not generated early warning information as normal points, incorporate the normal points into the construction dataset of the working condition cyclic baseline manifold, and re-execute the step of extracting the working condition cyclic baseline manifold based on the manifold distribution to complete the update of the working condition cyclic baseline manifold.

Images