A new energy unit degradation trend evaluation and risk early warning method and system

Abstract

The application discloses a new energy unit degradation trend evaluation and risk early warning method and system, belongs to the technical field of equipment degradation prediction, and aims at the problems of high false degradation false alarm rate under complex variable working conditions, lack of physical constraints of health indexes, and insufficient early warning reliability, and comprises the following steps: collecting multi-modal states and working condition parameters, and dividing three types of operation segments of steady state, transient impact and abnormal disturbance; a working condition-degradation decoupling model is established to strip working condition response and random disturbance and extract irreversible degradation characteristics; after multi-dimensional feature screening, three types of physical constraints of monotonic degradation, fatigue damage accumulation and degradation rate boundary are introduced to generate a degradation health index; a degradation digital twin model is constructed, health state and residual life probability distribution are output through online Bayesian updating; and a graded risk early warning result is generated by comprehensively considering multi-dimensional indexes. The application can effectively reduce the false alarm of variable working conditions, and improve the physical interpretability and early warning reliability of degradation evaluation.

Description

Technical Field

[0001] This invention belongs to the field of equipment degradation prediction technology, specifically relating to a method and system for assessing the degradation trend and providing early warning of risks in new energy power generation units. More specifically, this invention relates to condition monitoring, fault diagnosis, and predictive maintenance of key components of new energy power generation equipment. Background Technology

[0002] New energy power units such as wind turbine generators, photovoltaic tracking drive mechanisms, energy storage converters, and electric drive equipment typically operate under complex conditions of varying wind speeds, loads, temperatures, and frequent start-ups and shutdowns. During long-term operation, key components of these units are prone to degradation phenomena such as bearing wear, gear pitting, rotor imbalance, insulation aging, and lubrication deterioration. Failure to identify degradation trends and provide timely risk warnings can lead to unplanned outages, decreased power generation efficiency, increased maintenance costs, and even safety accidents.

[0003] In existing technologies, risk warning for new energy generating units often employs fixed thresholds, single-signal trend judgments, or ordinary data-driven models. Fixed threshold methods struggle to adapt to normal fluctuations in wind speed, load, and ambient temperature, easily misinterpreting changes in operating conditions as degradation anomalies. Single-signal methods fail to cover the complete degradation chain of early wear, mid-term structural damage, and late-stage thermal anomalies. While ordinary machine learning or deep learning methods can fit historical data, they typically lack constraints such as degradation monotonicity, fatigue damage accumulation, and physical rate boundaries, resulting in insufficient interpretability of the output results and poor stability when applied across generating units and operating conditions.

[0004] Furthermore, existing multi-source fusion methods often directly normalize or reduce the dimensionality of vibration, temperature, and oil characteristics, failing to clearly distinguish between reversible responses caused by operating conditions, random disturbances, and irreversible degradation caused by component damage. For new energy units, the same vibration amplitude or temperature rise may be a normal response under high load conditions, but may indicate an anomaly under low load conditions. Without operating condition-degradation decoupling, health indicators are prone to pseudo-degradation jumps, leading to false alarms or missed alarms.

[0005] When new energy power units operate under conditions of varying wind speed, load, temperature, and frequent start-stop cycles, measurements such as vibration amplitude, bearing temperature rise, current characteristics, and lubrication status will reversibly change with the operating conditions. Increased vibration and temperature caused by high loads do not necessarily indicate irreversible damage to components; if fixed thresholds are set or degradation models are trained directly based on the original measurements, sudden changes in operating conditions can easily be misidentified as a sudden drop in health status. On the other hand, health indicators that rely entirely on data fitting may experience short-term reverse recovery or jumps exceeding the physically permissible degradation rate. Therefore, it is necessary to simultaneously achieve operating condition response stripping, degradation physical constraints, and propagate health status uncertainties to remaining life and failure probability within the maintenance window.

[0006] Therefore, it is necessary to propose a method and system for assessing the degradation trend and providing early warning of risks of new energy generating units that can be oriented towards complex and variable operating conditions, distinguish between operating condition response and actual degradation, have physical consistency constraints, and output probabilistic risk levels, so as to predict and manage the equipment. Summary of the Invention

[0007] To address the aforementioned issues, this invention proposes a method and system for assessing the degradation trend and providing early warning of risks in new energy generating units. This method and system can effectively reduce false alarms under varying operating conditions and improve the physical interpretability of degradation assessments and the reliability of early warnings.

[0008] The data processed in this invention corresponds to the actual operating status of the target new energy unit. The status monitoring data is obtained by vibration, temperature, oil status and current detection units set at or near the key components of the unit. The operating condition parameters are obtained by the unit control system or the monitoring and data acquisition system SCADA. After time synchronization, all kinds of data are used to identify the operating condition response, irreversible degradation and subsequent remaining life and risk level, so that the calculation results directly serve the status assessment and maintenance decision of key components.

[0009] The present invention adopts the following technical solution:

[0010] On the one hand, a method for assessing the degradation trend and providing early warning of risks in new energy generating units includes the following steps:

[0011] S101, Collect multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data including vibration signals, temperature signals, oil state signals and current signals, extract multimodal state features. Based on the operating condition parameters including speed, load, output power, ambient temperature and wind speed, construct an operating condition vector.

[0012] S102, Based on the operating condition state vector, the multimodal state data is divided into operating segments to obtain steady-state operating segments, transient impact segments and abnormal disturbance segments;

[0013] S103. Establish a condition-degradation decoupling model for each operating segment, decompose the multimodal state features into condition response components, random disturbance components, and irreversible degradation components, and construct a condition-independent degradation feature set based on the irreversible degradation components. Among them, the condition response component is the condition response component caused by the condition state vector, which is obtained by nearest neighbor retrieval and kernel weighted estimation of historical normal samples; the random disturbance component is the zero-mean fluctuation of the multimodal state features relative to the local smooth trend within a preset local time window after deducting the condition response component; the irreversible degradation component is the remaining component after deducting the condition response component and random disturbance component from the multimodal state features, and is obtained by dynamic health baseline deviation and monotonic constraint correction.

[0014] S104, the set of degraded features unrelated to operating conditions is screened based on trend, monotonicity, fault separability, and weak correlation of operating conditions. The qualified degraded features unrelated to operating conditions are retained and fused to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, the fused health index is projected and corrected to generate a physical constraint degradation health index. Among them, fault separability is the degree of difference in the distribution of degraded features unrelated to operating conditions between normal samples and faulty or degraded samples, and weak correlation of operating conditions is the characteristic that the correlation between degraded features unrelated to operating conditions and the operating condition state vector is lower than a preset correlation threshold.

[0015] S105, establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly.

[0016] S106, using real-time acquired multimodal state data to update the parameters of the degraded digital twin model online, and performing uncertainty propagation based on the updated posterior state samples to obtain the health status prediction distribution, failure time distribution and remaining lifetime probability distribution.

[0017] S107 generates graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate, and degradation acceleration within the preset maintenance window.

[0018] Preferably, the multimodal state data is divided into operating segments based on the operating condition state vector to obtain steady-state operating segments, transient impact segments, and abnormal disturbance segments, specifically including:

[0019] By calculating the rate of change and amplitude fluctuation of the operating condition state vector within adjacent sampling windows, and calculating the distance between the current operating condition state vector and the historical normal operating condition cluster, the data is judged in the order of abnormal disturbance, steady-state operation, and transient impact: first, data whose distance from the historical normal operating condition cluster exceeds the third threshold is classified as abnormal disturbance segment; for data that are not classified as abnormal disturbance segment, data whose rate of change is less than the first threshold and whose amplitude fluctuation is less than the second threshold is classified as steady-state operation segment; and the remaining data that do not meet the steady-state operation conditions are classified as transient impact segment.

[0020] The operating condition state vector is represented as follows:

[0021] ;

[0022] in, For time; Here is the operating condition state vector; n(t) is the rotational speed; L(t) is the load; P(t) is the output power; T env v(t) represents the ambient temperature; v(t) represents the wind speed.

[0023] Preferably, the operating condition-degradation decoupling model satisfies:

[0024] ;

[0025] in, Let g be the j-th multimodal state feature; j is the multimodal state feature number; g j (c(t)) represents the condition response component caused by the condition state vector for the j-th multimodal state feature; r j (t) represents the random perturbation component of the j-th multimodal state feature; d j (t) represents the irreversible degradation component of the j-th multimodal state feature, which is obtained by monotonic constraint correction of the deviation between the original feature and the dynamic health baseline.

[0026] Preferably, the dynamic health baseline is determined based on a historical normal sample library and the current operating condition state vector. Specifically, the operating condition state vectors of historical normal samples in the historical normal sample library and the current operating condition state vector are normalized respectively. The Euclidean distance between the normalized operating condition state vectors is calculated, and the samples are sorted in ascending order of Euclidean distance. The 3 to 20 historical normal samples with the smallest distances are selected as historical normal samples similar to the current operating condition state. The corresponding weights are determined based on the Euclidean distance between the selected historical normal samples and the current operating condition state, where a smaller Euclidean distance corresponds to a larger weight, and the sum of all weights is 1. For the j-th multimodal state feature, a weighted estimate is performed on the j-th multimodal state feature corresponding to the selected historical normal samples to obtain the normal response value of the j-th multimodal state feature under the current operating condition, and this normal response value is used as the operating condition response component. The operating condition response components corresponding to each multimodal state feature together constitute the dynamic health baseline under the current operating condition; the normalized deviation between the real-time multimodal state features and their corresponding operating condition response components is used as the candidate features for degradation after operating condition compensation.

[0027] Preferably, the qualified condition-independent degradation features are specifically condition-independent degradation features with a comprehensive screening score greater than a preset score threshold; the comprehensive screening score is expressed as:

[0028] ;

[0029] in, The comprehensive screening score for the m-th condition-independent degradation feature; m is the number of the condition-independent degradation feature; is the monotonicity index of the degradation characteristics independent of the m-th operating condition; This is a trend indicator of the degradation characteristics unrelated to the m-th operating condition; The fault separability index is the feature of the m-th operating condition that is independent of degradation. The correlation penalty term between the m-th condition-independent degradation feature and the condition state vector is used; α, β, γ and η are weight coefficients; condition-independent degradation features with a comprehensive screening score greater than a preset score threshold are retained to participate in the construction of the physical constraint degradation health index.

[0030] Preferably, the physical constraint degradation health index generated in S104 satisfies the feasible degradation state space of physical consistency, as follows:

[0031] ;

[0032] in, for Physical constraints at any given moment degrade health index; For time intervals; for Physical constraints at any given moment leading to a decline in health index; v max (c(t)) represents the maximum permissible degradation rate determined by the current operating condition state vector; the feasible degradation state space is the range of health index values ​​jointly determined by the monotonic degradation constraint and the degradation rate boundary constraint; when the health index obtained by real-time feature fusion is greater than the physical constraint degradation health index of the previous moment, it is corrected to the physical constraint degradation health index of the previous moment; when the health index obtained by real-time feature fusion is lower than the lower limit of the health index determined by the maximum permissible degradation rate, it is corrected to the lower limit of the health index; when the real-time fusion result is within the range of values, the fusion result remains unchanged.

[0033] Preferably, the degenerate digital twin model uses discrete sampling time k as the modeling unit, and the state vector is represented as:

[0034] ;

[0035] Where k is the discrete sampling time; H is the state vector at time k; k D represents the health status at time k; k b is the cumulative damage at time k; k Let a be the degradation rate at time k; k Let θ be the degradation acceleration at time k; k For individual chemical conditions sensitive parameters at time k; It is the transpose of the vector;

[0036] The degradation state equation is expressed as:

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] Among them, Π [0,1] {·} is the interval projection operator; The health status value at time k-1; Let be the degradation rate at time k-1; The time interval between adjacent sampling times; Let be the degradation acceleration at time k-1; This is the mapping coefficient of cumulative damage to the decline in health status; This represents the cumulative damage at time k-1; As the health status quantity The observed values, and on Perform online calibration; , , , and These are random perturbation terms for health status parameters, cumulative damage, degradation rate, degradation acceleration, and individual chemical condition sensitive parameters, respectively. Let k be the working condition vector at time k; For individual chemical conditions sensitive parameters at time k-1; The cumulative fatigue damage intensity per unit time at time k; It is a minimum value function; This represents the maximum allowable degradation rate under current operating conditions. It is a function for maximizing the value;

[0043] The fatigue damage accumulation term is expressed as follows:

[0044] ;

[0045] in, , , , and These are the normalized rotational speed, load, output power, ambient temperature, and wind speed, respectively. to is the individualized sensitivity coefficient for the corresponding working condition; max{0,·} is used to ensure that the cumulative fatigue damage intensity is non-negative.

[0046] The random drift term is represented as:

[0047] , ;

[0048] in, For random drift terms; This is the process noise covariance matrix, used to characterize the impact of unmodeled disturbances, sensor fluctuations, and individual differences on the degradation process;

[0049] The observation equation is expressed as:

[0050] ;

[0051] ;

[0052] in, The observed values ​​of the physical constraint-induced health degradation index; To observe noise; To observe the noise covariance.

[0053] Preferably, S106 specifically includes:

[0054] The observed values ​​are calculated from the real-time acquired multimodal state data. Input the degraded digital twin model and update the degradation state and degradation rate online; when using Bayesian recursion, the update formula for the posterior distribution is:

[0055] ;

[0056] in, For a given observation sequence The posterior distribution of the state after that; Given the observation sequence up to time k, ={y1,y2,…,y k}; To observe the likelihood; This is the observation sequence up to time k-1; To predict the prior distribution;

[0057] When implemented using particle filtering, the weight update and observation likelihood of the i-th particle are calculated using the following formulas:

[0058] ;

[0059] ;

[0060] in, For the first Normalized weights of particles at time k; Number the particles; For the first The weights of each particle at time k-1; The total number of particles; For the first The predicted state of each particle; ℓ is the index for weighted normalized summation; Let be the weight of the ℓth particle at time k−1; Let be the state of the ℓth particle predicted from time k−1 to time k; For the first The health status quantity predicted by each particle.

[0061] In the health status prediction phase, the posterior particles are propagated forward h steps along the future operating condition sequence to obtain the future health status prediction distribution, calculated using the following formula:

[0062] ;

[0063] in, For a given observation sequence Predicted distribution of health status at time k+h; for Physical constraints at any given moment degrade health index; Let be the health status quantity obtained by the i-th particle when it propagates to time k+h; It is the Dirac function;

[0064] During the remaining lifetime calculation phase, a failure threshold is set, and the formula for calculating the failure time and remaining lifetime of the i-th particle is as follows:

[0065] ;

[0066] in, Let be the failure time of the i-th particle; The earliest time when the failure condition is met; H represents the time corresponding to the h-th step in the future; fai This is the failure threshold; This is the remaining lifetime sample of the i-th particle; Let k be the time interval.

[0067] The formulas for calculating the cumulative distribution function of remaining lifetime, the probability of failure within the maintenance window, and the confidence interval of remaining lifetime are as follows:

[0068] ;

[0069] ;

[0070] in, Let τ be the cumulative distribution function of remaining lifetime within the maintenance window τ; Let be the probability that the remaining lifetime is no greater than τ; τ is the maintenance window; I(·) is the indicator function; To maintain the failure probability within window τ; Q represents the remaining lifetime confidence interval at confidence level 1-α; α / 2 (RUL) and Q 1-α / 2 (RUL) is the weighted quantile of the RUL sample.

[0071] Preferably, the graded risk warning results include a concern level, a warning level, an alarm level, and an emergency level; when the degradation rate increases for multiple consecutive windows but the failure probability within the preset maintenance window does not exceed the first probability threshold, the concern level is output; when the lower bound of the remaining lifetime confidence is lower than the first maintenance threshold or the failure probability exceeds the first probability threshold, the warning level is output; when the failure probability exceeds the second probability threshold and the degradation acceleration exceeds the acceleration threshold, the alarm level is output; when the lower confidence bound of the health status prediction distribution is lower than the failure threshold and the alarm condition is met for multiple consecutive windows, the emergency level is output.

[0072] On the other hand, a degradation trend assessment and risk early warning system for new energy generating units includes:

[0073] The multimodal state data acquisition module is used to collect multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data, including vibration signals, temperature signals, oil state signals and current signals, multimodal state features are extracted. Based on the operating condition parameters, including speed, load, output power, ambient temperature and wind speed, an operating condition vector is constructed.

[0074] The segment division module is used to divide multimodal state data into segments based on the operating condition state vector, resulting in steady-state operating segments, transient impact segments, and abnormal disturbance segments.

[0075] The load condition-degradation decoupling model building module is used to build load condition-degradation decoupling models for each operating segment. It decomposes multimodal state features into load condition response components, random disturbance components, and irreversible degradation components, and constructs a load condition-independent degradation feature set based on the irreversible degradation components. Among them, the load condition response component is the load condition response component caused by the load condition state vector, which is obtained by nearest neighbor retrieval and kernel weighted estimation of historical normal samples; the random disturbance component is the zero-mean fluctuation of the multimodal state features relative to the local smooth trend within a preset local time window after deducting the load condition response component; the irreversible degradation component is the remaining component after deducting the load condition response component and random disturbance component from the multimodal state features, and is obtained by dynamic health baseline deviation and monotonic constraint correction.

[0076] The physical constraint degradation health index generation module is used to filter the set of condition-independent degradation features based on trend, monotonicity, fault separability, and weak correlation with operating conditions. The module retains the qualified condition-independent degradation features and fuses them to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, the fused health index is projected and corrected to generate the physical constraint degradation health index. Fault separability refers to the degree of difference in the distribution of condition-independent degradation features between normal samples and faulty or degraded samples, and weak correlation with operating conditions refers to the characteristic that the correlation between condition-independent degradation features and the operating condition state vector is lower than a preset correlation threshold.

[0077] The observation equation establishment module is used to establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly.

[0078] The state distribution prediction module is used to update the parameters of the degraded digital twin model online using real-time collected multimodal state data, and to perform uncertainty propagation based on the updated posterior state samples to obtain the health state prediction distribution, failure time distribution and remaining lifetime probability distribution.

[0079] The graded risk warning result generation module is used to generate graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate, and degradation acceleration within a preset maintenance window.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] (1) This invention divides the operating segment into operating state space and decomposes the multimodal features into operating response components, random disturbance components and irreversible degradation components by combining the operating condition-degradation decoupling model. It can accurately distinguish between normal fluctuations in operating conditions such as start-up and shutdown, load change, and gusts and actual damage to components, thus avoiding misjudging the operating condition response as degradation anomaly from the source and greatly reducing the false alarm rate.

[0082] (2) In the feature fusion stage, the present invention introduces monotonic degradation constraints, fatigue damage accumulation constraints and degradation rate boundary constraints, so that the generated physical constraint degradation health index conforms to the irreversible damage evolution mechanism of mechanical parts, avoids the problem of irregular jumps in pure data-driven models, and the output results have clear physical meaning and engineering rationality.

[0083] (3) The present invention constructs a degradation digital twin model containing fatigue damage accumulation terms, and with the online Bayesian update mechanism, the degradation state, degradation rate and individual working condition sensitive parameters can be continuously corrected with real-time operating data, and the application stability in cross-unit and cross-working condition scenarios is significantly improved.

[0084] (4) This invention abandons the single threshold alarm mode and integrates multiple indicators such as failure probability within the maintenance window, remaining lifetime confidence interval, degradation rate and degradation acceleration. It outputs four levels of risk results: attention, warning, alarm and emergency. It can directly guide the priority ranking of maintenance and the planning of maintenance windows, and improve the efficiency of operation and maintenance and the safety of equipment operation.

[0085] (5) This invention uses a joint scoring mechanism of trend, monotonicity, fault separability and weak correlation with operating conditions to screen features, eliminate redundant features that are highly sensitive to operating conditions but not sensitive to damage, ensure that the final health index is highly sensitive to the degradation process, and at the same time has strong resistance to operating condition interference. Attached Figure Description

[0086] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0087] Figure 1This is a flowchart of the method for assessing the degradation trend and providing early warning of risks of new energy generating units according to an embodiment of the present invention;

[0088] Figure 2 This is a time series diagram of representative daily wind speed and ambient temperature used in embodiments of the present invention;

[0089] Figure 3 The following are representative daily power output timing diagrams for different schemes used in embodiments of the present invention;

[0090] Figure 4 The following are timing diagrams of fatigue load proxy schemes used in the embodiments of the present invention;

[0091] Figure 5 The following are temperature timing diagrams of the gearbox using different schemes in the embodiments of the present invention;

[0092] Figure 6 These are vibration RMS timing diagrams for different schemes used in embodiments of the present invention;

[0093] Figure 7 These are timing diagrams of dimensionless reliability proxy schemes used in different embodiments of the present invention.

[0094] Figure 8 This is a structural block diagram of the new energy unit degradation trend assessment and risk warning system according to an embodiment of the present invention. Detailed Implementation

[0095] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0096] like Figure 1 As shown in the figure, this embodiment presents a method for assessing the degradation trend and providing early warning of risks of new energy power units, which includes the following steps.

[0097] S101, collect multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data including vibration signal, temperature signal, oil state signal and current signal, extract multimodal state features. Based on the operating condition parameters including speed, load, output power, ambient temperature and wind speed, construct an operating condition state vector.

[0098] Specifically, the status monitoring data is collected by corresponding detection units located on or near the key components of the target new energy unit, and the operating condition parameters are obtained by the unit control system or the monitoring and data acquisition system SCADA.

[0099] Furthermore, the state monitoring data is time-synchronized and windowed according to a preset feature extraction time window. Within each feature extraction time window, features are extracted from the vibration signal, temperature signal, oil state signal, and current signal to obtain vibration state feature sub-vectors, temperature state feature sub-vectors, oil state feature sub-vectors, and current state feature sub-vectors. These state feature sub-vectors are then combined according to their corresponding time windows to form a multimodal state feature vector. The vibration state features include at least one or more of vibration amplitude statistical features, frequency band energy features, and spectral features. The temperature state features include at least one or more of temperature statistical features, temperature rise features, and temperature change rate features. The oil state features include at least one or more of abrasive particle state features, oil contamination state features, and oil physicochemical state features. The current state features include at least one or more of current amplitude features, harmonic features, three-phase imbalance features, and characteristic sideband energy features.

[0100] It should be noted that, in this application, unless explicitly stated as "posterior state sample" or "particle," the term "sample" refers to a synchronous data record formed using a preset feature extraction time window as the basic unit. This data record includes at least the multimodal state feature vector and the operating condition state vector corresponding to the same time window; when there is maintenance or fault labeling, it also includes the state label corresponding to the time window. Historical normal samples refer to samples whose sampling time is earlier than the current time and whose target component fault or degradation anomaly has not been confirmed according to unit operation records, alarm records, and maintenance records; fault or degradation samples refer to samples that correspond temporally to confirmed fault events, maintenance conclusions, or labeled degradation stages; the historical normal sample library is a collection of multiple historical normal samples. The posterior state sample or particle in S106 is a random implementation of the degradation state vector during Bayesian recursion or particle filtering, used to characterize the uncertainty of state estimation, and does not belong to the aforementioned state monitoring data samples.

[0101] S102, Based on the operating condition state vector, the multimodal state data is divided into operating segments to obtain steady-state operating segments, transient impact segments and abnormal disturbance segments;

[0102] Specifically, the operating condition state vector is represented as follows:

[0103] ;

[0104] in, For time; Here is the operating condition state vector; n(t) is the rotational speed; L(t) is the load; P(t) is the output power; T env v(t) represents the ambient temperature; v(t) represents the wind speed.

[0105] By calculating the rate of change and amplitude fluctuation of the operating condition state vector within adjacent sampling windows, and calculating the distance between the current operating condition state vector and the historical normal operating condition cluster, the data is classified in the order of abnormal disturbance, steady-state operation, and transient impact: First, data whose distance from the historical normal operating condition cluster exceeds the third threshold is classified as abnormal disturbance segment; for data that are not classified as abnormal disturbance segment, data whose rate of change is less than the first threshold and whose amplitude fluctuation is less than the second threshold is classified as steady-state operation segment; the remaining data that do not meet the steady-state operation conditions are classified as transient impact segment, so that any sampling window is classified into one of the steady-state operation segment, transient impact segment, or abnormal disturbance segment.

[0106] S103. Establish a condition-degradation decoupling model for each operating segment, decompose the multimodal state features into condition response components, random disturbance components, and irreversible degradation components, and construct a condition-independent degradation feature set based on the irreversible degradation components.

[0107] Specifically, the operating condition-degradation decoupling model satisfies:

[0108] ;

[0109] in, Let be the j-th multimodal state feature; j is the multimodal state feature number; gj(c(t)) is the working condition response component of the j-th multimodal state feature caused by the working condition state vector, which is obtained through nearest neighbor retrieval and kernel weighted estimation of historical normal samples; rj(t) is the random disturbance component of the j-th multimodal state feature, which is the zero-mean fluctuation of the multimodal state feature relative to the local smooth trend within a preset local time window after deducting the working condition response component; dj(t) is the irreversible degradation component of the j-th multimodal state feature, which is the remaining component after deducting the working condition response component and the random disturbance component from the original feature, and is obtained by monotonic constraint correction of the deviation between the remaining component and the dynamic health baseline.

[0110] The process of determining the operating condition response component is as follows: Using the value ranges of various operating condition parameters (speed, load, output power, ambient temperature, and wind speed) in the historical normal sample library as a benchmark, the operating condition state vectors of historical normal samples in the historical normal sample library and the current operating condition state vector are normalized respectively. The Euclidean distance between the normalized operating condition state vectors is calculated, and the samples are sorted in ascending order of Euclidean distance. The 3 to 20 historical normal samples with the smallest distance are selected as historical normal samples similar to the current operating condition state. The corresponding weights are determined based on the Euclidean distance between the selected historical normal samples and the current operating condition state, where the smaller the Euclidean distance, the larger the corresponding weight, and the sum of all weights is 1. For the j-th multimodal state feature, a weighted estimation is performed on the j-th multimodal state feature corresponding to the selected historical normal samples to obtain the normal response value of the j-th multimodal state feature under the current operating condition, and this normal response value is used as the operating condition response component. .

[0111] Furthermore, the operating condition response components corresponding to each multimodal state feature together constitute the dynamic health baseline under the current operating condition; the normalized deviation between the real-time multimodal state features and their corresponding operating condition response components is used as the candidate features for degradation after operating condition compensation.

[0112] S104, the set of degradation features unrelated to operating conditions is screened based on trend, monotonicity, fault separability and weak correlation of operating conditions, and the qualified degradation features unrelated to operating conditions are retained and fused to obtain the fused health index; based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, the fused health index is projected and corrected to generate the physical constraint degradation health index.

[0113] Specifically, a comprehensive screening score is calculated for degradation characteristics unrelated to each operating condition. The calculation formula is as follows:

[0114] ;

[0115] in, The comprehensive screening score for the m-th condition-independent degradation feature; m is the number of the condition-independent degradation feature; is the monotonicity index of the degradation characteristics independent of the m-th operating condition; This is a trend indicator of the degradation characteristics unrelated to the m-th operating condition; The fault separability index is used to characterize the degree of distribution difference of the feature between normal samples and faulty or degraded samples for the m-th condition-independent degradation feature. The correlation penalty term between the m-th condition-independent degradation feature and the condition state vector is given, where the weak correlation indicates that the correlation between the feature and the condition state vector is lower than a preset correlation threshold; α, β, γ and η are weighting coefficients.

[0116] Condition-independent degradation features with a comprehensive screening score greater than a preset score threshold are retained as qualified condition-independent degradation features and fused to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current condition state vector, the fused health index is projected and corrected to generate a physical constraint degradation health index.

[0117] The feasible degenerate state space that satisfies the physical constraint degradation health index and physical consistency is as follows:

[0118] ;

[0119] in, for Physical constraints at any given moment degrade health index; For time intervals; for Physical constraints at any given moment leading to a decline in health index; v max (c(t)) represents the maximum permissible degradation rate determined by the current operating condition state vector; the feasible degradation state space is the range of health index values ​​jointly determined by the monotonic degradation constraint and the degradation rate boundary constraint; when the health index obtained by real-time feature fusion is greater than the physical constraint degradation health index of the previous moment, it is corrected to the physical constraint degradation health index of the previous moment; when the health index obtained by real-time feature fusion is lower than the lower limit of the health index determined by the maximum permissible degradation rate, it is corrected to the lower limit of the health index; when the real-time fusion result is within the range of values, the fusion result remains unchanged.

[0120] In this embodiment, the projection correction is not a new model calculation, but a physical consistency correction for the health index obtained by feature fusion. Based on the monotonic degradation constraint, the health index at the next time step should not be higher than the health index at the previous time step; based on the degradation rate boundary constraint, the decrease in the health index at the next time step should not exceed the product of the maximum allowable degradation rate corresponding to the current operating condition and the time interval. Therefore, the feasible value range of the health index is determined by the above two constraints. Fusion results exceeding this range are corrected to the corresponding boundary, while fusion results within the range remain unchanged.

[0121] S105, establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly.

[0122] Specifically, the degenerate digital twin model uses discrete sampling time k as the modeling unit, and the state vector is represented as:

[0123] ;

[0124] Where k is the discrete sampling time; H is the state vector at time k; k D represents the health status at time k; k b is the cumulative damage at time k; k Let a be the degradation rate at time k; k Let θ be the degradation acceleration at time k; k For individual chemical conditions sensitive parameters at time k; It is the transpose of the vector.

[0125] The degradation state equation is expressed as:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] Among them, Π [0,1] {·} is the interval projection operator; The health status value at time k-1; Let be the degradation rate at time k-1; The time interval between adjacent sampling times; Let be the degradation acceleration at time k-1; This is the mapping coefficient of cumulative damage to the decline in health status; This represents the cumulative damage at time k-1; , , , and These are random perturbation terms for health status parameters, cumulative damage, degradation rate, degradation acceleration, and individual chemical condition sensitive parameters, respectively. Let k be the working condition vector at time k; For individual chemical conditions sensitive parameters at time k-1; The cumulative fatigue damage intensity per unit time at time k; It is a minimum value function; This represents the maximum allowable degradation rate under current operating conditions. It is a function for maximizing the value.

[0132] The fatigue damage accumulation term is expressed as follows:

[0133] ;

[0134] in, , , , and These are the normalized rotational speed, load, output power, ambient temperature, and wind speed, respectively. to is the individualized sensitivity coefficient for the corresponding working condition; max{0,·} is used to ensure that the cumulative fatigue damage intensity is non-negative.

[0135] The random drift term is represented as:

[0136] , ;

[0137] in, For random drift terms; This is the process noise covariance matrix, used to characterize the impact of unmodeled disturbances, sensor fluctuations, and individual differences on the degradation process;

[0138] The observation equation is expressed as:

[0139] ;

[0140] ;

[0141] in, The observed value of the physical constraint degradation health index output in step S104; To observe noise; To observe the noise covariance.

[0142] S106 utilizes real-time acquired multimodal state data to update the parameters of the degraded digital twin model online, and performs uncertainty propagation based on the updated posterior state samples to obtain the health status prediction distribution, failure time distribution, and remaining lifetime probability distribution.

[0143] The specific implementation process of S106 is as follows.

[0144] The observed values ​​are calculated from the real-time acquired multimodal state data. Input the degraded digital twin model and update the degradation state and degradation rate online; when using Bayesian recursion, the update formula for the posterior distribution is:

[0145] ;

[0146] in, For a given observation sequence The posterior distribution of the state after that; Given the observation sequence up to time k, ={y1,y2,…,y k}; To observe the likelihood; This is the observation sequence up to time k-1; To predict the prior distribution.

[0147] When implemented using particle filtering, the weight update and observation likelihood of the i-th particle are calculated using the following formulas:

[0148] ;

[0149] ;

[0150] in, For the first Normalized weights of particles at time k; Number the particles; For the first The weights of each particle at time k-1; The total number of particles; For the first The predicted state of each particle; ℓ is the index for weighted normalized summation; Let be the weight of the ℓth particle at time k−1; Let be the state of the ℓth particle predicted from time k−1 to time k; For the first The health status quantity predicted by each particle.

[0151] In the health status prediction phase, the posterior particles are propagated forward h steps along the future operating condition sequence to obtain the future health status prediction distribution, calculated using the following formula:

[0152] ;

[0153] in, For a given observation sequence Predicted distribution of health status at time k+h; for Physical constraints at any given moment degrade health index; Let be the health status quantity obtained by the i-th particle when it propagates to time k+h; This is the Dirac function.

[0154] During the remaining lifetime calculation phase, a failure threshold is set, and the formula for calculating the failure time and remaining lifetime of the i-th particle is as follows:

[0155] ;

[0156] in, Let be the failure time of the i-th particle; The earliest time when the failure condition is met; H represents the time corresponding to the h-th step in the future; fai This is the failure threshold; This is the remaining lifetime sample of the i-th particle; Let k be the time at time k.

[0157] S107 generates graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate, and degradation acceleration within the preset maintenance window.

[0158] Specifically, the graded risk warning results include attention level, warning level, alarm level, and emergency level; when the degradation rate increases for multiple consecutive windows but the failure probability within the preset maintenance window does not exceed the first probability threshold, the attention level is output; when the lower bound of the remaining lifetime confidence is lower than the first maintenance threshold or the failure probability exceeds the first probability threshold, the warning level is output; when the failure probability exceeds the second probability threshold and the degradation acceleration exceeds the acceleration threshold, the alarm level is output; when the lower confidence bound of the health status prediction distribution is lower than the failure threshold and the alarm condition is met for multiple consecutive windows, the emergency level is output.

[0159] In this embodiment, the graded risk warning results are further associated with the target component identifier, health status, remaining lifespan, and recommended maintenance window, and output to the unit operation and maintenance terminal or status monitoring system for maintenance scheduling.

[0160] The following will be a detailed explanation using an example of multimodal state and risk indicators based on numerical data of a general 2.5 MW doubly fed wind turbine generator set.

[0161] To further illustrate the implementation of this invention under complex and variable operating conditions using specific data, this embodiment uses data and data graphs from a general 2.5 MW class doubly-fed wind turbine generator numerical case, without constructing separate health status, remaining life, or failure probability curves. This numerical case uses A0 as the original system and sets up blade replacement (A1), gearbox optimization (A2), and pitch system upgrade (A3). The four schemes are compared under identical environmental and operational boundaries. It should be noted that this data is example numerical data for the parameterized construction of a general 2.5 MW class doubly-fed wind turbine generator, not field, bench, or hardware-in-the-loop (HIL) measured data, and does not correspond to a specific manufacturer, commercial model, or proprietary controller.

[0162] In this embodiment, the rated power is 2.5 MW, the rotor diameter is 121 m, and the cut-in / rated / cut-out wind speeds are 3 / 11 / 25 m / s, respectively. Representative daily data consists of 96 consecutive 15-minute samples, totaling 24 hours; all four schemes use the same fixed environmental sequence. Wind speed ranges from 5.96 to 12.32 m / s, with a mean of 9.02 m / s and a standard deviation of 1.76 m / s; ambient temperature ranges from 10.3 to 22.0 ℃, with a mean of 15.54 ℃; average air density is 1.215 kg / m³, average turbulence intensity is 0.113, and yaw error is mainly distributed between -6.08° and 7.13°. Specific data boundaries are shown in Table 1.

[0163] Table 1. Main operating conditions and data boundaries of the numerical case:

[0164]

[0165] Figure 2 The figure shows representative daily wind speed and ambient temperature data used in this embodiment. It indicates that both wind speed and ambient temperature fluctuate continuously over 24 hours, corresponding to wind speed v(t) and ambient temperature Tenv(t) in S101 of this invention, and together with rotational speed, load, and output power, constitute the operating condition state vector. Figure 3 The power output timing of A0, A1, A2, and A3 under the same environmental input is shown, with average daily power of 1320.7, 1386.1, 1352.6, and 1328.2 kW, respectively. Therefore, the changes in state characteristics cannot be directly interpreted as irreversible degradation without considering real-time operating conditions and unit configuration. In this embodiment, operating segments are first divided and operating condition-degradation decoupling is performed in A2 and A3.

[0166] Before entering the degradation assessment model, the data undergoes unified time-base processing according to the preprocessing rules of the numerical case: a central Hampel window with 7 sampling points is used for key channels such as power, torque, speed, temperature, and vibration. When a sample deviates from the local median by more than 3 × 1.4826 times the absolute deviation of the median, it is marked as a candidate outlier and verified in conjunction with the engineering range; the actual transitions corresponding to the preset operating state switching are retained. Linear interpolation is used when no more than 2 sampling points are missing consecutively; missing segments with more than 2 sampling points are removed from the corresponding operating condition interval. The example data used in this embodiment is complete, so the missing data branch is not actually triggered.

[0167] Figure 4 The fatigue load proxy timing sequence used in this embodiment is shown. The average fatigue load proxies for A0, A1, A2, and A3 are 0.870, 0.924, 0.837, and 0.785, respectively; relative to A0, A1 increases by 6.11%, A2 decreases by 3.90%, and A3 decreases by 9.83%. The corresponding average peak load proxies are 1.674, 1.736, 1.636, and 1.465, respectively, with changes relative to A0 of 3.72% for A1, 2.25% for A2, and 12.49% for A3. The above load values ​​are comparative proxies at the response level, not the physical damage equivalent load DEL calculated based on the cyclic load history. In this embodiment, this type of load proxy can be used as an auxiliary input when setting fatigue damage accumulation constraints and degradation rate boundaries, but it is not directly equivalent to the healthy state or failure probability.

[0168] Figure 5 and Figure 6 The gearbox temperature timing and vibration RMS timing used in this embodiment are shown respectively. The average gearbox temperatures of A0, A1, A2, and A3 are 73.20, 74.95, 67.24, and 71.74 ℃, respectively, with corresponding maximum temperatures of 88.5, 91.0, 82.1, and 86.2 ℃; the average vibration RMS are 2.127, 2.260, 1.817, and 1.960 mm / s, respectively. Compared to A0, the average vibration RMS of A1 increases by 6.27%, while A2 and A3 decrease by 14.60% and 7.84%, respectively. Temperature and vibration are both important components of the multimodal state data described in step S101 of this invention, and they change together with the operating conditions in the high wind speed and high power range, further illustrating that using a dynamic health baseline and weakly correlated operating condition features for screening is more suitable for complex and variable operating condition scenarios than directly using a fixed threshold.

[0169] Figure 7The dimensionless reliability proxy timing diagram used in this embodiment is shown, with mean values ​​of 0.8065, 0.7704, 0.9125, and 0.8735 for A0, A1, A2, and A3, respectively. The changes relative to A0 are a 4.49% decrease for A1, a 13.13% increase for A2, and an 8.30% increase for A3. This reliability quantity is only a comparative proxy indicator and is not interpreted as a failure probability. In the risk warning process of this embodiment, it can be used together with temperature, vibration, and load proxy as auxiliary state information for health assessment; the actual failure probability within the maintenance window is still obtained according to S106 based on the propagation of the degraded digital twin posterior state samples.

[0170] Table 2. Comprehensive index results of numerical cases:

[0171]

[0172] As shown in Table 2, the relationship between a single power generation index and load, temperature, vibration, and reliability proxy is not always unidirectional. For example, A1 has the highest AEP gain at 3.32%, but its fatigue load proxy, peak load proxy, average temperature, and vibration RMS are all higher than A0. A2 has an AEP gain of 1.52%, while its fatigue load proxy decreases by 3.90%, peak load proxy decreases by 2.25%, average temperature drops to 67.24 ℃, vibration RMS drops to 1.817 mm / s, and reliability proxy reaches 0.9125. These results can be used to explain why this invention needs to separate the operating condition response from the multimodal state variables and introduce damage accumulation and physical consistency constraints in the construction of the health index, avoiding the direct judgment of a single performance gain or short-term fluctuation as health improvement or degradation.

[0173] Since the data used in this embodiment only covers a representative 24-hour day and the comparison results of different modification schemes, and does not provide the complete life trajectory of the same component evolving from a healthy state to failure, this embodiment does not artificially supplement values ​​such as the median remaining lifespan, confidence interval, posterior failure time, or failure probability within the maintenance window. After the target unit accumulates continuous degradation samples and determines the failure threshold, the temperature, vibration, load, and operating condition data shown in this embodiment are used to construct a physical constraint degradation health index according to steps S103 and S104. Then, based on the degradation digital twin and Bayesian recursive formula in steps S105 and S106, the predicted distribution of health status and the probability distribution of remaining lifespan are obtained. Finally, according to S107, the risk results of concern level, early warning level, alarm level, or emergency level are output.

[0174] In summary, in the numerical case study of a general 2.5 MW doubly-fed wind turbine generator, 96 consecutive 15-minute samples covered wind speeds ranging from 5.96 to 12.32 m / s and ambient temperatures ranging from 10.3 to 22.0 °C. Under the same environmental sequence, using A0 as the baseline of the original system, the fatigue load proxy and peak load proxy of the gearbox optimization scheme A2 decreased by 3.90% and 2.25%, respectively, the average gearbox temperature decreased from 73.20 °C to 67.24 °C, and the average vibration RMS decreased from 2.127 mm / s to 1.817 mm / s. While the blade replacement scheme A1 achieved a 3.32% increase in annual power generation, its fatigue load proxy and peak load proxy increased by 6.11% and 3.72%, respectively. The above numerical results indicate that relying solely on power or a single threshold cannot fully reflect the health risks of the unit; it is necessary to combine multimodal information such as operating conditions, loads, temperatures, and vibrations for operating condition compensation, degradation feature extraction, and comprehensive risk assessment.

[0175] like Figure 8 As shown in the figure, this embodiment also discloses a new energy unit degradation trend assessment and risk early warning system, including:

[0176] The multimodal state data acquisition module 801 is used to acquire multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data including vibration signals, temperature signals, oil state signals and current signals, multimodal state features are extracted. Based on the operating condition parameters including speed, load, output power, ambient temperature and wind speed, an operating condition vector is constructed.

[0177] The segment division module 802 is used to divide multimodal state data into segments based on the operating condition state vector to obtain steady-state operating segments, transient impact segments and abnormal disturbance segments.

[0178] The operating condition-degradation decoupling model establishment module 803 is used to establish an operating condition-degradation decoupling model for each operating segment. It decomposes the multimodal state features into operating condition response components, random disturbance components, and irreversible degradation components, and constructs an operating condition-independent degradation feature set based on the irreversible degradation components. Among them, the operating condition response component is the operating condition response component caused by the operating condition state vector, which is obtained through nearest neighbor retrieval and kernel weighted estimation of historical normal samples; the random disturbance component is the zero-mean fluctuation of the multimodal state features after deducting the operating condition response component, relative to the local smooth trend within a preset local time window; the irreversible degradation component is the remaining component after deducting the operating condition response component and random disturbance component from the multimodal state features, and is obtained by dynamic health baseline deviation and monotonic constraint correction.

[0179] The physical constraint degradation health index generation module 804 is used to filter the set of degradation features unrelated to operating conditions based on trend, monotonicity, fault separability, and weak correlation of operating conditions. It retains the qualified degradation features unrelated to operating conditions and fuses them to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, it projects and corrects the fused health index to generate the physical constraint degradation health index. Fault separability refers to the degree of difference in the distribution of degradation features unrelated to operating conditions between normal samples and faulty or degraded samples, and weak correlation of operating conditions refers to the characteristic that the correlation between degradation features unrelated to operating conditions and the operating condition state vector is lower than a preset correlation threshold.

[0180] The observation equation establishment module 805 is used to establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly.

[0181] The state distribution prediction module 806 is used to update the parameters of the degraded digital twin model online using real-time acquired multimodal state data, and to perform uncertainty propagation based on the updated posterior state samples to obtain the health state prediction distribution, failure time distribution and remaining lifetime probability distribution.

[0182] The graded risk warning result generation module 807 is used to generate graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate and degradation acceleration within a preset maintenance window.

[0183] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assessing the degradation trend and providing early warning of risks in new energy generating units, characterized in that, Includes the following steps: S101, Collect multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data including vibration signals, temperature signals, oil state signals and current signals, extract multimodal state features. Based on the operating condition parameters including speed, load, output power, ambient temperature and wind speed, construct an operating condition vector. S102, Based on the operating condition state vector, the multimodal state data is divided into operating segments to obtain steady-state operating segments, transient impact segments and abnormal disturbance segments; S103. Establish a condition-degradation decoupling model for each operating segment, decompose the multimodal state features into condition response components, random disturbance components, and irreversible degradation components, and construct a condition-independent degradation feature set based on the irreversible degradation components. Among them, the condition response component is the condition response component caused by the condition state vector, which is obtained by nearest neighbor retrieval and kernel weighted estimation of historical normal samples; the random disturbance component is the zero-mean fluctuation of the multimodal state features relative to the local smooth trend within a preset local time window after deducting the condition response component; the irreversible degradation component is the remaining component after deducting the condition response component and random disturbance component from the multimodal state features, and is obtained by dynamic health baseline deviation and monotonic constraint correction. S104, the set of degraded features unrelated to operating conditions is screened based on trend, monotonicity, fault separability, and weak correlation of operating conditions. The qualified degraded features unrelated to operating conditions are retained and fused to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, the fused health index is projected and corrected to generate a physical constraint degradation health index. Among them, fault separability is the degree of difference in the distribution of degraded features unrelated to operating conditions between normal samples and faulty or degraded samples, and weak correlation of operating conditions is the characteristic that the correlation between degraded features unrelated to operating conditions and the operating condition state vector is lower than a preset correlation threshold. S105, establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly. S106, using real-time acquired multimodal state data to update the parameters of the degraded digital twin model online, and performing uncertainty propagation based on the updated posterior state samples to obtain the health status prediction distribution, failure time distribution and remaining lifetime probability distribution. S107 generates graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate, and degradation acceleration within the preset maintenance window.

2. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 1, characterized in that, Based on the operating condition state vector, the multimodal state data is divided into operating segments, resulting in steady-state operating segments, transient impact segments, and abnormal disturbance segments, specifically including: By calculating the rate of change and amplitude fluctuation of the operating condition state vector within adjacent sampling windows, and calculating the distance between the current operating condition state vector and the historical normal operating condition cluster, the data is judged in the order of abnormal disturbance, steady-state operation, and transient impact: first, data whose distance from the historical normal operating condition cluster exceeds the third threshold is classified as abnormal disturbance segment; for data that are not classified as abnormal disturbance segment, data whose rate of change is less than the first threshold and whose amplitude fluctuation is less than the second threshold is classified as steady-state operation segment; and the remaining data that do not meet the steady-state operation conditions are classified as transient impact segment. The operating condition state vector is represented as follows: ； in, For time; Here is the operating condition state vector; n(t) is the rotational speed; L(t) is the load; P(t) is the output power; T env v(t) represents the ambient temperature; v(t) represents the wind speed.

3. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 2, characterized in that, The operating condition-degradation decoupling model satisfies: ； in, Let g be the j-th multimodal state feature; j is the multimodal state feature number; g j (c(t)) represents the condition response component caused by the condition state vector for the j-th multimodal state feature; r j (t) represents the random perturbation component of the j-th multimodal state feature; d j (t) represents the irreversible degradation component of the j-th multimodal state feature, which is obtained by monotonic constraint correction of the deviation between the original feature and the dynamic health baseline.

4. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 3, characterized in that, The dynamic health baseline is determined based on the historical normal sample library and the current operating condition vector. The specific process is as follows: the operating condition vectors of the historical normal samples in the historical normal sample library and the current operating condition vector are normalized respectively. The Euclidean distance between the normalized operating condition vectors is calculated and sorted in ascending order of Euclidean distance. The 3 to 20 historical normal samples with the smallest distance are selected as historical normal samples that are similar to the current operating condition. The corresponding weights are determined based on the Euclidean distance between the selected historical normal samples and the current operating condition. The smaller the Euclidean distance, the larger the corresponding weight, and the sum of all weights is 1. For the j-th multimodal state feature, a weighted estimate is performed on the corresponding j-th multimodal state feature in the selected historical normal samples to obtain the normal response value of the j-th multimodal state feature under the current operating condition, and the normal response value is used as the operating condition response component. The operating condition response components corresponding to each multimodal state feature together constitute the dynamic health baseline under the current operating condition; the normalized deviation between the real-time multimodal state features and their corresponding operating condition response components is used as the candidate features for degradation after operating condition compensation.

5. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 1, characterized in that, The qualified condition-independent degradation features are specifically those with a comprehensive screening score greater than a preset score threshold; the comprehensive screening score is expressed as: ； in, The comprehensive screening score for the m-th condition-independent degradation feature; m is the number of the condition-independent degradation feature; is the monotonicity index of the degradation characteristics independent of the m-th operating condition; This is a trend indicator of the degradation characteristics unrelated to the m-th operating condition; The fault separability index is the feature of the m-th operating condition that is independent of degradation. The correlation penalty term between the m-th condition-independent degradation feature and the condition state vector is used; α, β, γ and η are weight coefficients; condition-independent degradation features with a comprehensive screening score greater than a preset score threshold are retained to participate in the construction of the physical constraint degradation health index.

6. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 1, characterized in that, The physical constraint degradation health index generated in S104 satisfies the feasible degradation state space of physical consistency, as follows: ； in, for Physical constraints at any given moment degrade health index; For time intervals; for Physical constraints at any given moment leading to a decline in health index; v max (c(t)) represents the maximum permissible degradation rate determined by the current operating condition state vector; the feasible degradation state space is the range of health index values ​​jointly determined by the monotonic degradation constraint and the degradation rate boundary constraint; when the health index obtained by real-time feature fusion is greater than the physical constraint degradation health index of the previous moment, it is corrected to the physical constraint degradation health index of the previous moment; when the health index obtained by real-time feature fusion is lower than the lower limit of the health index determined by the maximum permissible degradation rate, it is corrected to the lower limit of the health index; when the real-time fusion result is within the range of values, the fusion result remains unchanged.

7. The method for assessing the degradation trend and providing early warning of risks of new energy generating units as described in claim 1, characterized in that, The degenerate digital twin model uses discrete sampling time k as the modeling unit, and its state vector is represented as: ； Where k is the discrete sampling time; H is the state vector at time k; k D represents the health status at time k; k b is the cumulative damage at time k; k Let a be the degradation rate at time k; k Let θ be the degradation acceleration at time k; k For individual chemical conditions sensitive parameters at time k; It is the transpose of the vector; The degradation state equation is expressed as: ； ； ； ； ； Among them, Π [0,1] {·} is the interval projection operator; The health status value at time k-1; Let be the degradation rate at time k-1; The time interval between adjacent sampling times; Let be the degradation acceleration at time k-1; This is the mapping coefficient of cumulative damage to the decline in health status; This represents the cumulative damage at time k-1; As the health status quantity The observed values, and on Perform online calibration; , , , and These are random perturbation terms for health status parameters, cumulative damage, degradation rate, degradation acceleration, and individual chemical condition sensitive parameters, respectively. Let k be the working condition vector at time k; For individual chemical conditions sensitive parameters at time k-1; The cumulative fatigue damage intensity per unit time at time k; It is a minimum value function; This represents the maximum allowable degradation rate under current operating conditions. It is a function for maximizing the value; The fatigue damage accumulation term is expressed as follows: ； in, , , , and These are the normalized rotational speed, load, output power, ambient temperature, and wind speed, respectively. to is the individualized sensitivity coefficient for the corresponding working condition; max{0,·} is used to ensure that the cumulative fatigue damage intensity is non-negative. The random drift term is represented as: ， ； in, For random drift terms; This is the process noise covariance matrix, used to characterize the impact of unmodeled disturbances, sensor fluctuations, and individual differences on the degradation process; The observation equation is expressed as: ； ； in, The observed values ​​of the physical constraint-induced health degradation index; To observe noise; To observe the noise covariance.

8. The method for assessing the degradation trend and providing early warning of risks for new energy generating units as described in claim 7, characterized in that, S106 specifically includes: The observed values ​​are calculated from the real-time acquired multimodal state data. Input the degraded digital twin model and update the degradation state and degradation rate online; when using Bayesian recursion, the update formula for the posterior distribution is: ； in, For a given observation sequence The posterior distribution of the state after that; Given the observation sequence up to time k, ={y1,y2,…,y k }; To observe the likelihood; This is the observation sequence up to time k-1; To predict the prior distribution; When implemented using particle filtering, the weight update and observation likelihood of the i-th particle are calculated using the following formulas: ； ； in, For the first Normalized weights of particles at time k; Number the particles; For the first The weights of each particle at time k-1; The total number of particles; For the first The predicted state of each particle; ℓ is the index for weighted normalized summation; Let be the weight of the ℓth particle at time k−1; Let be the state of the ℓth particle predicted from time k−1 to time k; For the first The health status quantity predicted by each particle. In the health status prediction phase, the posterior particles are propagated forward h steps along the future operating condition sequence to obtain the future health status prediction distribution, calculated using the following formula: ； in, For a given observation sequence Predicted distribution of health status at time k+h; for Physical constraints at any given moment degrade health index; Let be the health status quantity obtained by the i-th particle when it propagates to time k+h; It is the Dirac function; During the remaining lifetime calculation phase, a failure threshold is set, and the formula for calculating the failure time and remaining lifetime of the i-th particle is as follows: ； in, Let be the failure time of the i-th particle; The earliest time when the failure condition is met; H represents the time corresponding to the h-th step in the future; fai This is the failure threshold; This is the remaining lifetime sample of the i-th particle; Let k be the time interval. The formulas for calculating the cumulative distribution function of remaining lifetime, the probability of failure within the maintenance window, and the confidence interval of remaining lifetime are as follows: ； ； in, Let τ be the cumulative distribution function of remaining lifetime within the maintenance window τ; Let be the probability that the remaining lifetime is no greater than τ; τ is the maintenance window; I(·) is the indicator function; To maintain the failure probability within window τ; Q represents the remaining lifetime confidence interval at confidence level 1-α; α / 2 (RUL) and Q 1-α / 2 (RUL) is the weighted quantile of the RUL sample.

9. The method for assessing the degradation trend and providing early warning of risks of new energy generating units as described in claim 1, characterized in that: The graded risk warning results include attention level, warning level, alarm level, and emergency level. When the degradation rate increases for multiple consecutive windows but the failure probability within the preset maintenance window does not exceed the first probability threshold, the attention level is output. When the lower bound of the remaining lifetime confidence is lower than the first maintenance threshold or the failure probability exceeds the first probability threshold, the warning level is output. When the failure probability exceeds the second probability threshold and the degradation acceleration exceeds the acceleration threshold, the alarm level is output. When the lower confidence bound of the health status prediction distribution is lower than the failure threshold and the alarm condition is met for multiple consecutive windows, the emergency level is output.

10. A system for assessing the degradation trend and providing early warning of risks in new energy generating units, characterized in that, include: The multimodal state data acquisition module is used to collect multimodal state data during the operation of the new energy unit. The multimodal state data includes state monitoring data and operating condition parameters. Based on the state monitoring data, including vibration signals, temperature signals, oil state signals and current signals, multimodal state features are extracted. Based on the operating condition parameters, including speed, load, output power, ambient temperature and wind speed, an operating condition vector is constructed. The segment division module is used to divide multimodal state data into segments based on the operating condition state vector, resulting in steady-state operating segments, transient impact segments, and abnormal disturbance segments. The load condition-degradation decoupling model building module is used to build load condition-degradation decoupling models for each operating segment. It decomposes multimodal state features into load condition response components, random disturbance components, and irreversible degradation components, and constructs a load condition-independent degradation feature set based on the irreversible degradation components. Among them, the load condition response component is the load condition response component caused by the load condition state vector, which is obtained by nearest neighbor retrieval and kernel weighted estimation of historical normal samples; the random disturbance component is the zero-mean fluctuation of the multimodal state features relative to the local smooth trend within a preset local time window after deducting the load condition response component; the irreversible degradation component is the remaining component after deducting the load condition response component and random disturbance component from the multimodal state features, and is obtained by dynamic health baseline deviation and monotonic constraint correction. The physical constraint degradation health index generation module is used to filter the set of condition-independent degradation features based on trend, monotonicity, fault separability, and weak correlation with operating conditions. The module retains the qualified condition-independent degradation features and fuses them to obtain a fused health index. Based on the monotonic degradation constraint that the health index does not increase with the degradation process and the maximum allowable degradation rate determined by the current operating condition state vector, the fused health index is projected and corrected to generate the physical constraint degradation health index. Fault separability refers to the degree of difference in the distribution of condition-independent degradation features between normal samples and faulty or degraded samples, and weak correlation with operating conditions refers to the characteristic that the correlation between condition-independent degradation features and the operating condition state vector is lower than a preset correlation threshold. The observation equation establishment module is used to establish a degradation digital twin model of the new energy unit, determine the state vector and establish a degradation state equation containing fatigue damage accumulation term and random drift term, take the physical constraint degradation health index as the observed value of the health state quantity in the state vector, and establish the observation equation accordingly. The state distribution prediction module is used to update the parameters of the degraded digital twin model online using real-time collected multimodal state data, and to perform uncertainty propagation based on the updated posterior state samples to obtain the health state prediction distribution, failure time distribution and remaining lifetime probability distribution. The graded risk warning result generation module is used to generate graded risk warning results based on the failure probability, remaining lifetime confidence interval, degradation rate, and degradation acceleration within a preset maintenance window.

Images