A device cross-condition degradation evaluation method based on working condition potential field and health conservation constraint

Abstract

The application relates to the technical field of predictive maintenance, and discloses a device cross-working condition degradation evaluation method based on a working condition potential field and health conservation constraint. The method collects multi-source monitoring data and multi-source working condition parameters of the device, constructs a high-dimensional working condition space after time synchronization and standardization; a nonlinear mapping relationship between monitoring features and working condition variables is established according to health baseline historical data, and a working condition potential field is formed; static and transient working condition disturbance contribution amounts are determined in combination with a current working condition position, a change track and a working condition potential field gradient, and a preliminary health state component is separated from original monitoring features; dynamic correction is carried out by taking the continuous evolution of health states at adjacent moments as a constraint, the correction result is mapped to a unified health reference space, and a cross-working condition health state sequence is formed; and a device degradation state, a degradation trend and a residual life prediction result are output based on the sequence.

Description

Technical Field

[0001] This application relates to the field of predictive maintenance technology, and more specifically, to a method for assessing equipment degradation across operating conditions based on operating potential field and health conservation constraints. Background Technology

[0002] As electromechanical equipment such as wind turbines, aero engines, rail transit equipment, and industrial robots develop towards larger scale and continuous operation, monitoring data such as vibration, temperature, torque, and oil condition are typically used to assess the degradation status and remaining life of equipment components. During actual operation, the load, speed, power, ambient temperature, and control parameters of these devices continuously change with production tasks, the external environment, and control commands. Therefore, the monitoring characteristics are simultaneously affected by both the equipment's health status and its operating conditions.

[0003] Existing equipment degradation assessment methods typically employ fixed thresholds, single-condition models, or pre-divide operating conditions into multiple intervals and set characteristic benchmarks or assessment models for each interval. For equipment with relatively stable operating conditions and sufficient historical sample coverage, these methods can reduce the impact of operating condition differences on monitoring characteristics to some extent. However, in actual equipment, multiple operating parameters usually change simultaneously and continuously, with no clear physical boundaries between adjacent operating conditions. After discretizing continuous operating conditions, data from adjacent time points may invoke different characteristic benchmarks or assessment models due to being in different intervals, causing the processed health status representation to exhibit jumps at interval boundaries that are unrelated to actual equipment degradation. For combinations of operating conditions not fully covered by historical samples, it may also be impossible to obtain an effective benchmark that matches the current operating conditions.

[0004] Furthermore, monitoring characteristics are not only related to the current operating conditions but may also be affected by the process of reaching those conditions. For example, the vibration response of equipment when rapidly increasing from low to high load is not entirely the same as the vibration response when the equipment operates stably at the same high load for a long period. Rapid acceleration, rapid loading, start-up and shutdown, or sudden changes in control parameters may create peaks or offsets in the monitoring signal that last for several sampling periods. Existing processing methods that normalize or subtract from the benchmark based on the current operating point typically process data reaching the same operating point according to the same benchmark, thus making it difficult to eliminate transient responses caused by changes in operating conditions.

[0005] Even after the above processing, the resulting health status sequence may still contain baseline shifts caused by operating condition switching, short-term spikes caused by transient changes in operating conditions, and continuous changes caused by actual equipment degradation. If this sequence is directly input into the degradation assessment model, feature changes that appear in a short time and disappear as the operating condition recovers may be identified as sudden changes in equipment condition; conversely, slowly accumulating actual degradation changes may also be covered by feature shifts that repeatedly occur under different operating conditions. Therefore, the instability of subsequent assessment results is not solely caused by the assessment model itself, but is related to the fact that the health status sequence input has not fully excluded the influence of operating conditions.

[0006] For example, when a gearbox bearing is operating stably under low load, its root mean square (RMS) vibration value is within the normal fluctuation range. When the unit rapidly increases its power, even if the bearing does not suffer new mechanical damage, the RMS vibration value and the sideband energy of the meshing frequency may increase briefly and then gradually decrease after the power stabilizes. If degradation is judged directly based on the increased values, this recoverable change may easily be identified as bearing damage. If the bearing also has gradually developing fatigue spalling, the slow changes resulting from the actual degradation may be masked by the amplitude fluctuations under different loads.

[0007] Therefore, the technical problem that the existing technology needs to solve is: when the equipment is under multi-dimensional continuous variable operating conditions and transient changes in operating conditions, how to form a health status sequence that is stable in time evolution, can retain continuous degradation information, and can be compared between different operating conditions from the original monitoring characteristics affected by the operating conditions, so as to reduce the risk that changes in operating conditions are misidentified as equipment degradation and that true degradation is masked by fluctuations in operating conditions. Summary of the Invention

[0008] To overcome the problems in existing technologies where changes in operating conditions cause continuous disturbances to equipment monitoring characteristics, making it difficult to separate the impact of operating conditions from actual degradation, easily misjudging transient processes of operating conditions as sudden changes in equipment state, lacking a unified comparison benchmark for health status under different operating conditions, and having insufficient generalization ability of degradation assessment results across operating conditions, this application provides a cross-operating-condition degradation assessment method for equipment based on operating potential field and health conservation constraints.

[0009] This application constructs a working condition potential field that characterizes the intensity and direction of disturbances caused by operating conditions to monitoring features, and determines the gradient of the working condition potential field by combining the current operating condition change trajectory. This separates static and transient operating condition disturbances from the original monitoring features. Furthermore, by utilizing the characteristic of continuous evolution of equipment health status at adjacent time points, the health status after separation of operating condition disturbances is dynamically corrected, and the health status under different operating conditions is mapped to a unified health reference space. This forms a cross-operating condition health status sequence that can be used for equipment degradation status identification, degradation trend judgment, and remaining life prediction.

[0010] This application provides a method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints, including: Collect multi-source monitoring data and multi-source operating parameters from the equipment, synchronize and standardize the multi-source monitoring data and multi-source operating parameters in time, and construct a high-dimensional operating space based on the multi-source operating parameters; Based on historical operating data, a nonlinear mapping relationship is established between monitoring features and operating condition variables in the high-dimensional operating condition space to obtain the operating condition potential field that characterizes the disturbance intensity and direction of the operating condition on the monitoring features. Based on the current operating condition's position and trajectory within the operating condition potential field, the gradient of the operating condition potential field is calculated; based on the operating condition potential field and the gradient of the operating condition potential field, the contribution of operating condition disturbance in the original monitoring features is determined; the contribution of operating condition disturbance is separated from the original monitoring features to obtain the preliminary health status component. By taking the continuous evolution of health status at adjacent time points as a constraint, the initial health status components are dynamically corrected to obtain the optimal health status representation. The optimal health status representation is mapped to a unified health reference space to form a cross-condition health status sequence. Based on the cross-operating condition health status sequence, an equipment degradation assessment model is constructed, and the equipment degradation status, degradation trend and remaining life prediction results are output.

[0011] In one possible implementation, the operating condition potential field is constructed as follows: using multidimensional operating condition variables from historical operating data confirmed by health status as input, and corresponding monitoring feature values ​​or feature distribution statistics as output, differentiable nonlinear regression models corresponding to each monitoring feature are trained respectively; each differentiable nonlinear regression model is determined as a scalar field of the corresponding monitoring feature in the high-dimensional operating condition space, the field value of the scalar field represents the static disturbance intensity, the sign of the partial derivative of the scalar field with respect to the corresponding operating condition variable represents the direction of increase or decrease of the monitoring feature, and the absolute value of the partial derivative represents the sensitivity to change; multiple scalar fields are combined into a vector field according to a fixed order of monitoring features to obtain the operating condition potential field.

[0012] Through the above implementation method, the influence relationship of multidimensional operating condition variables on each monitoring feature can be described in the form of a continuous field, avoiding the division of continuously changing operating conditions of machinery into several fixed operating condition intervals, and reducing feature jumps and evaluation biases generated at the boundaries of operating condition partitions.

[0013] In one possible implementation, the working condition potential field gradient is obtained by: determining the rate of change of the working condition variables based on the working condition vector at the current time, the working condition vector at the previous time, and the sampling time interval; determining the Jacobian matrix of the working condition potential field at the current working condition vector; combining the Jacobian matrix with the rate of change of the working condition variables to obtain the time rate of change of the working condition potential field along the current working condition change trajectory, and using the time rate of change as the working condition potential field gradient.

[0014] Through the above implementation methods, not only can the static impact of the current operating condition position on the monitoring characteristics be identified, but the transient impact of the rate and direction of change of the operating condition on the monitoring characteristics can also be reflected, thereby reducing the risk that monitoring characteristic spikes caused by rapid acceleration, sudden loading, start-up and shutdown or rapid changes in control parameters are misjudged as equipment degradation.

[0015] In one possible implementation, determining the preliminary health state component includes: determining a static operating condition disturbance expectation value based on the operating condition vector at the current moment; determining a transient disturbance compensation amount based on the operating condition potential field gradient, sampling time interval, and dimensionless transient compensation coefficient; combining the static operating condition disturbance expectation value and the transient disturbance compensation amount to form the operating condition disturbance contribution amount; and subtracting the operating condition disturbance contribution amount from the original monitoring features to obtain the preliminary health state component relative to the health baseline.

[0016] Through the above implementation method, it is possible to simultaneously separate the static disturbance caused by the current operating condition and the transient disturbance caused by the change of operating condition in the same processing link, so that the preliminary health status component can more represent the change of the equipment's own health status, rather than the change of operating condition.

[0017] In one possible implementation, the dynamic correction includes: constructing a health-conserving energy function comprising a health conservation constraint term and a data fidelity term, wherein the health conservation constraint term is used to constrain changes in the health state at adjacent time points, and the data fidelity term is used to limit the deviation of the health state to be solved from the initial health state component; and solving the health-conserving energy function based on a positive definite state transition covariance matrix and a positive definite observation noise covariance matrix to obtain the optimal health state representation that minimizes the health-conserving energy function.

[0018] The health conservation principle is used to characterize the continuous evolution of the equipment's health status over a selected short timescale, rather than characterizing the equipment's health status as constant over long-term operation. Health conservation constraints suppress anomalous jumps that occur only in a single or a few processing cycles, while data fidelity terms preserve true degradation information that maintains a consistent direction of change over multiple consecutive processing cycles, thus balancing transient disturbance suppression with the preservation of continuous degradation trends.

[0019] In one possible implementation, forming the cross-condition health status sequence includes: determining the condition points in the condition neighborhood where the proportion of historical samples reaches a preset frequency condition and the variance or coefficient of variation of each condition variable in the condition neighborhood is not higher than a preset fluctuation threshold as standard reference conditions; determining a health baseline function for outputting health baseline values ​​corresponding to different conditions based on the optimal health status representation under each condition in the health baseline stage; inputting the current condition and the standard reference condition into the health baseline function respectively to obtain the health baseline value corresponding to the current condition and the health baseline value corresponding to the standard reference condition; subtracting the health baseline value corresponding to the current condition from the optimal health status representation at each time point, and adding the health baseline value corresponding to the standard reference condition to obtain the cross-condition health status sequence.

[0020] Through the above implementation methods, the optimal health status representation formed under different loads, speeds, power, environmental parameters, or control states can be transformed into the same health reference space, so that the health status at different times and under different operating conditions has a unified comparison benchmark, reducing baseline drift of the health status as the operating conditions change.

[0021] In one possible implementation, the equipment degradation assessment model takes the time window of the cross-condition health status sequence as input and outputs the current health index of the equipment, the discrete degradation stage to which the equipment belongs, the rate of change and trend direction of the health index, and the remaining life prediction result with a probability distribution. The equipment degradation assessment model adopts at least one of deep learning sequence model, probabilistic statistical model or physical information driven model, and is trained using the equipment's historical operation data from operation to failure.

[0022] By enabling the equipment degradation assessment model to read the cross-condition health state sequence formed after decoupling of the operating condition potential field, optimization of health conservation constraints, and mapping of the unified health reference space, the interference of the original operating condition differences on the identification of equipment degradation status and prediction of remaining life can be reduced.

[0023] In one possible implementation, when the residual of the operating condition disturbance interpretation continuously exceeds the feedback threshold set based on the health baseline data, or when the distance between the cross-operating condition health state sequence and the known degradation template exceeds the template threshold set based on the distance distribution within the template, the corresponding data is placed into the unconfirmed sample pool; the unconfirmed samples that have been confirmed to be fault-free and belong to the new operating condition are used for incremental learning or retraining of the operating condition potential field, and the unconfirmed samples that have been confirmed to have degradation are used for updating the equipment degradation assessment model; the operating condition potential field and the equipment degradation assessment model before and after the update are verified using the health baseline validation set that does not participate in the update and the labeled degradation validation set, and the updated model is put into use when the update result meets the preset acceptance indicators.

[0024] In one possible implementation, the feedback threshold is determined based on the residual distribution of the operating condition disturbance interpretation in the health baseline verification data and the preset false trigger control conditions; the preset number of continuous processing cycles is determined based on the sampling cycle of the equipment monitoring data, the degradation feedback response requirements, and the continuous triggering results of the health baseline verification data; and the template threshold is determined based on the template distance distribution among similar known degradation templates and the preset template matching conditions.

[0025] In one possible implementation, based on alarm records, work orders, oil or vibration retest results, and maintenance results from the same period, the samples to be confirmed are divided into fault-free new operating condition samples, degraded samples, and uncertain samples; the fault-free new operating condition samples are added to the health baseline training set for incremental learning or retraining of the operating condition potential field; the degraded samples are used for updating the equipment degradation assessment model and are not added to the health baseline training set; the uncertain samples are retained in the sample pool to be confirmed.

[0026] Using the healthy baseline validation set (not included in the update) and the labeled degradation validation set, calculate the median of the residuals of the operating condition disturbance explanation, the accuracy of the degradation trend identification, and the mean absolute error of the remaining lifetime before and after the update. When the median of the residuals of the operating condition disturbance explanation and the mean absolute error of the remaining lifetime after the update are not greater than the corresponding values ​​before the update, and the accuracy of the degradation trend identification after the update is not lower than the corresponding values ​​before the update, the updated operating condition potential field and the equipment degradation assessment model are put into use; otherwise, the operating condition potential field and the equipment degradation assessment model before the update continue to be used.

[0027] Compared with related technologies, this application has the following advantages: This application synchronizes and standardizes multi-source monitoring data and multi-source operating parameters over time, and constructs a high-dimensional operating condition space based on these parameters. It establishes a nonlinear mapping relationship between each monitoring feature and continuous operating condition variables based on historical operating data under healthy conditions. This ensures that each monitoring feature has a disturbance reference corresponding to the current operating condition within the continuous operating condition space, thereby reducing feature reference jumps caused by switching discrete operating condition partition boundaries. Furthermore, it determines static operating condition disturbances based on the position of the current operating condition in the operating condition potential field, and determines transient disturbance compensation based on the change trajectory of the current operating condition and the gradient of the operating condition potential field at the current operating condition. This ensures that the operating condition disturbance contribution simultaneously reflects the impact of the current operating condition value and the operating condition change process on the monitoring features. Finally, it separates the operating condition disturbance contribution from the original monitoring features. This reduces the risk of short-term responses caused by rapid loading, sudden acceleration, start-up and shutdown, or abrupt changes in control parameters being mistaken for equipment degradation. It dynamically corrects the initial health state components by constraining the continuous evolution of health states at adjacent time points and limits the deviation of the correction results relative to the initial health state components. This suppresses non-physical jumps that disappear with the recovery of operating conditions while retaining true degradation information that changes in the same direction over multiple consecutive time points. Finally, by mapping the optimal health state representation to a unified health reference space, the impact of differences in health baselines under different operating conditions on state comparisons is reduced. This forms a health state sequence that considers the separation of operating condition disturbances, the stability of time evolution, and the comparability across operating conditions, thus providing a sequence input with lower operating condition impact and consistent benchmarks for predicting equipment degradation status, deterioration trends, and remaining life. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the working process of the equipment cross-operating condition deterioration assessment method in the embodiments of this application.

[0029] Figure 2 This is a flowchart of multi-source data acquisition and high-dimensional working space construction in the embodiments of this application.

[0030] Figure 3 This is a flowchart of the construction of the working condition potential field and the calculation of the working condition potential field gradient in the embodiments of this application.

[0031] Figure 4 This is a flowchart of the decoupling of operating condition influences and optimization of health conservation constraints in the embodiments of this application.

[0032] Figure 5 This is a flowchart of the unified health reference space mapping, degradation assessment and feedback update in the embodiments of this application. Detailed Implementation

[0033] The present application will be further described below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are used to illustrate the technical solution of the present application and are not intended to limit the scope of protection of the present application. Without departing from the technical concept of the present application, those skilled in the art can make adaptive adjustments to the data dimensions, model forms and parameter values ​​according to the equipment type, monitoring object and operating conditions.

[0034] like Figure 1 As shown, this application processes equipment operating data according to the following main methods: multi-source data acquisition and operating condition space construction—operating condition potential field construction—operating condition potential field gradient calculation—operating condition influence decoupling—health conservation constraint optimization—unified health reference space mapping—cross-operating condition degradation assessment. The output of each node participates in subsequent processing: the operating condition vector is used to construct the operating condition potential field and locate the current operating condition; the operating condition potential field and its gradient are used to determine the contribution of operating condition disturbances; the preliminary health state components are used for health conservation optimization; the optimal health state representation, after being mapped through the unified health reference space, forms a cross-operating condition health state sequence, which is used to output degradation state, degradation trend, and remaining life prediction results.

[0035] This embodiment uses the drivetrain of a 3 MW wind turbine in an offshore wind farm as an example. The drivetrain includes a main shaft, a gearbox with two planetary gear stages and one parallel shaft gear stage, and a generator. The monitoring object is the high-speed shaft drive end bearing of the gearbox. The monitoring system is equipped with vibration acceleration sensors, temperature sensors, speed sensors, power sensors, and environmental sensors. The vibration acceleration sensor has a sampling frequency of 12.8 kHz and is installed on the bearing housings and gear rings of each stage of the gearbox, as well as the drive and non-drive end bearings of the generator. The gearbox oil temperature is collected by a temperature sensor installed in the gearbox lubrication circuit, the high-speed shaft speed is collected by a shaft end speed sensor, the generator power is read by the wind turbine main control system, and the nacelle temperature, nacelle humidity, and wind speed are collected by the nacelle environmental sensor and an anemometer, respectively. Control parameters such as pitch angle, generator torque setpoint, yaw angle, and grid connection status are obtained through the wind turbine main control system interface.

[0036] Step S1: Multi-source data acquisition and high-dimensional working space construction.

[0037] like Figure 2 As shown, multi-source monitoring data and multi-source operating parameters are collected by the sensor network and the wind turbine main control system interface, respectively. After time synchronization, quality verification, missing data processing, and unified time axis mapping, one path is used to extract the original monitoring feature vector, and the other path is used to form the operating condition vector and construct a high-dimensional operating condition space. Both the original monitoring feature vector and the operating condition vector participate in subsequent operating condition potential field construction, operating condition influence decoupling, and health status assessment.

[0038] Simultaneously collect multi-source monitoring data reflecting the equipment's health status and multi-source operating parameters describing the equipment's operating conditions. Multi-source monitoring data includes at least one of vibration acceleration, temperature, torque, and oil abrasive particles; multi-source operating parameters include at least one of load, speed, power, temperature, ambient temperature and humidity, and control parameters. Vibration acceleration signals are collected at 12.8kHz; the update frequency of temperature, speed, power, environmental parameters, and control parameters is no less than 1Hz, and they are aggregated to the field-level data center with a unified processing cycle of 1 second.

[0039] Before data acquisition, range, zero bias, and clock checks are performed on each sensor. Device identifiers, sensor identifiers, sampling timestamps, and quality markers are retained in the acquired data. Using the main control system clock as a reference, the sending time of the main control system and the receiving time of the edge acquisition node are read from each synchronization message. The difference between the two is taken as the clock deviation corresponding to the current synchronization message. The sampling timestamp of the edge acquisition node is corrected using the median clock deviation of five consecutive synchronization messages. Then, each data point is mapped to a 1-second unified time axis. This clock deviation correction and unified time axis mapping process constitutes the time synchronization of multi-source monitoring data and multi-source operating parameters. Zero-order hold or linear interpolation can be used for low-frequency parameters. Candidate interpolation durations are set sequentially according to integer multiples of the original update cycle. Data segments of the same length are manually masked in the health baseline verification data. Interpolation is performed separately, and the absolute deviation of the monitoring characteristics before and after interpolation is calculated. The maximum candidate duration whose absolute deviation does not exceed the standard deviation of the corresponding monitoring characteristic's health baseline fluctuation is determined as the preset interpolation duration. When the consecutive missing duration exceeds the preset interpolation duration, an unusable marker is set for the corresponding data segment. Only data points that have all the required operating parameters and monitoring characteristics within the same processing cycle, have no sensor quality alarms, and do not contain out-of-limit interpolation are determined as valid data points.

[0040] The 12,800 sampled values ​​within each 1-second vibration window are converted into features. This is applied to the vibration acceleration sequence. Root mean square value and steepness They are respectively:

[0041]

[0042] in, This represents the number of sampled values ​​within the vibration window. This represents the mean vibration acceleration within the window. When the variance of the vibration acceleration within the window is zero, the kurtosis of the window is set to an unavailable value, and it is not input into subsequent models. A Fast Fourier Transform is performed on the vibration window, based on the number of gear teeth. Corresponding axis rotation frequency The product determines the meshing frequency. Using the spectral resolution as the minimum bandwidth, the energy of the meshing frequency sideband is obtained by accumulating the squares of the spectral amplitudes within the meshing frequency and its transition sideband.

[0043] Bandpass filtering and Hilbert transform covering the bearing resonance band were performed on the vibration signal, and spectral analysis was conducted on the resulting envelope signal. (Bearing outer ring fault frequency) and bearing inner ring failure frequency They are respectively:

[0044]

[0045] in, The number of rolling elements; The diameter of the rolling element; The bearing pitch circle diameter; Contact angle. Bearing geometric parameters are obtained from equipment design documents or bearing model parameter tables; respectively in and The maximum spectral amplitude is selected within a frequency band defined by two spectral resolutions on each side to obtain the bearing outer ring fault frequency amplitude and the bearing inner ring fault frequency amplitude. The bearing resonance band is determined by the frequency band containing the peak value of the spectrum obtained from the bearing housing impact response test and remains unchanged during the model's use. The oil temperature change rate is determined by the ratio of the oil temperature difference between adjacent uniform processing cycles to the sampling time interval.

[0046] The aforementioned Fast Fourier Transform, Hilbert Transform, and Bandpass Filter can be implemented by digital signal processing programs, digital signal processors, or field-programmable gate arrays on edge computing nodes; their input is the vibration window, the processing methods are spectrum transformation, envelope demodulation, and frequency band filtering, and the output is the monitoring characteristics listed in Table 2.

[0047] For the Original operating parameters Read the mean of this parameter from the health baseline data. and standard deviation Standardize according to the following formula:

[0048] Indicates the first The sampling time, the first Standardized values ​​for the first operating condition variable; for the first... Original monitoring features Read the mean of this feature from the health baseline data. and standard deviation Standardize according to the following formula:

[0049] Indicates the first The sampling time, the first The standardized values ​​of each monitoring feature are used; when any standard deviation is zero, the corresponding parameter or feature is not used as model input. The mean and standard deviation are saved as model parameters, and subsequent online data will use the same parameters to perform standardization.

[0050] Timekeeping working condition vector for:

[0051] in, This represents the total number of dimensions for the operating condition variables. For the first The sampling time, the first Standardized values ​​of each operating condition variable; Total number of sampling times; superscript This represents the transpose of a vector. The operating condition vectors at all times span [the vectors]. High-dimensional operating space. Let the preprocessed original monitoring feature vector be:

[0052] in, To monitor feature dimensions. The resulting operating condition vector. and the original monitoring feature vector As input for steps S2 to S4.

[0053] The 10 operating condition variables selected in this embodiment are shown in Table 1.

[0054]

[0055] Table 1 Operating condition variables in this embodiment For the high-speed shaft drive end bearing of the gearbox, six monitoring features were extracted from the vibration acceleration signal and temperature signal, as shown in Table 2.

[0056]

[0057] Table 2 Monitoring characteristics of this embodiment Step S2: Construction of the potential field model under continuous operating conditions.

[0058] like Figure 3As shown, a differentiable nonlinear regression model corresponding to each monitoring feature is trained using representative samples from historical health baseline data, and each scalar field is combined into a vector value working condition potential field. In the online processing stage, the expected value of static working condition disturbance is obtained according to the current working condition vector, the rate of change of working condition variables is obtained according to the current working condition vector and the previous working condition vector, and the Jacobian matrix at the current working condition is combined with the rate of change of working condition variables to obtain the gradient of the working condition potential field.

[0059] The expected value of static operating condition disturbance is obtained as follows: the original operating condition parameters at the current moment are read and standardized using the mean and standard deviation of the health baseline data to obtain the current operating condition vector; the current operating condition vector is input into the differentiable nonlinear regression model corresponding to each monitoring feature, and each model outputs the predicted value of the corresponding monitoring feature when the device is in the health baseline state and operating under the current operating condition, and the predicted value of static operating condition disturbance is obtained by combining the monitoring features in a fixed order.

[0060] Using historical operational data from the health baseline phase, nonlinear mapping relationships between the operating condition vector and each monitoring feature are established, and these nonlinear mapping relationships are characterized by a nonlinear regression model to obtain the operating condition potential field function.

[0061] For the One monitoring feature, Working condition vector scalar field value at the location; The scalar fields are combined to form a vector field, resulting in the vector potential field under operating conditions. The field value characterizes the intensity of static disturbances from the operating conditions under the healthy baseline state to the corresponding monitored characteristics. The scalar field pair of the first The sign of the partial derivative of the nth operating condition variable indicates that the nth operating condition variable is constant. An increase in the first operating condition variable leads to the first The monitoring feature increases or decreases; the absolute value of the partial derivative represents the local sensitivity of the monitoring feature to the corresponding operating condition variable.

[0062] The output of the nonlinear regression model can be a single monitored feature value or a feature distribution statistic calculated within the operating condition neighborhood; the feature distribution statistic includes at least one of mean, median, variance, or quantile. Before model training, the type of statistic to be used is determined, and the same statistic is used for all training samples; the regression result of the selected statistic serves as the field value of the corresponding scalar field, participating in the subsequent calculation of the contribution of operating condition disturbances.

[0063] This embodiment collects six consecutive months of historical operating data from the initial commissioning phase of the unit, confirming its healthy state through factory tests, early operation records, and fault-free diagnostic results. This data covers an operating range from a cut-in wind speed of 3 m / s to a cut-out wind speed of 25 m / s, containing approximately 2.6 million valid data points. To reduce the number of training samples, the standardized high-dimensional operating space is first divided into hierarchical grids or clusters. Then, representative samples are extracted according to the sample size proportion of each operating condition's neighborhood. The representative samples cover the observed operating condition boundaries and the distribution area of ​​historical samples. Using representative samples as training samples, Gaussian process regression models corresponding to each monitoring feature are constructed respectively.

[0064] The root mean square value of the vibration signal For example, the corresponding first feature:

[0065] The Gaussian process uses a zero-mean function, and the covariance function uses a squared exponential autocorrelation kernel.

[0066] in, and For the first The and the first One training condition vector; This indicates that the observed noise follows a pattern with a mean of 0 and a variance of . The normal distribution; This represents the kernel function value between two working condition vectors; and These represent the first two working condition vectors, respectively. Standardized values ​​of each operating condition variable; For the first Length scale parameter for each working condition variable; The signal variance; Let the noise variance be denoted. Define the training case matrix. , No. Feature observation vectors nuclear matrix The Line 1 Column elements are , For the kernel matrix Identity matrices of the same order. The set of hyperparameters is determined by maximizing the log-marginal likelihood. :

[0067] For any working condition vector to be calculated The posterior predicted mean of the working condition potential field corresponding to the first feature for:

[0068] in, This represents the training observation vector of the first monitoring feature; Represents the training condition matrix; This represents the kernel matrix calculated based on the training condition vectors; Represents the identity matrix of the same order as the kernel matrix; superscript Indicates transpose; superscript 1 indicates matrix inversion; Represents the working condition vector to be calculated A column vector consisting of kernel function values ​​between all training case vectors; The above processing was repeated for the six monitoring features to obtain... Since the squared exponential kernel function is continuously differentiable with respect to all operating condition variables, the Gaussian process regression model constructed in this embodiment is a differentiable nonlinear regression model. Optionally, a deep neural network can be used as the differentiable nonlinear regression model; or a gradient boosting tree can be used for modeling, and local smoothing and finite difference can be performed in the neighborhood of the current operating condition to obtain the local rate of change required thereafter.

[0069] The inspection method in step S2 includes: dividing the health baseline data into a training set and a validation set in chronological order; calculating the Euclidean distance between each operating condition vector in the training set and its nearest neighbor operating condition vector; and determining the 95th quantile of the Euclidean distance distribution as the modeling domain distance threshold. For the operating condition vector to be calculated, the minimum Euclidean distance between it and the training operating condition vector is calculated; when the minimum Euclidean distance is not greater than the modeling domain distance threshold, it is determined to be within the observed modeling domain and input into the operating condition potential field; otherwise, it is marked as outside the modeling domain and no automatic degradation assessment result is directly generated. For each monitoring feature, the absolute residual between the observed value in the validation set and the predicted value in the operating condition potential field is calculated, and its 95th quantile is saved for use when setting the feedback threshold in step S8.

[0070] Step S3: Calculation of the potential field gradient under operating conditions and characterization of transient disturbances.

[0071] Get the current time working condition vector and the working condition vector of the previous moment The rate of change of the operating condition variables is calculated using backward differencing. :

[0072] in, This is the timestamp difference between two adjacent valid data points. For the first valid data point, it will be... Set to zero; when there is an unavailable data segment between adjacent data points, the difference is not calculated across the unavailable data segment, but the first valid data point after the unavailable data segment is taken as the new starting data point and its rate of change of the operating condition variable is set to zero. For the Gaussian process regression model, the first... The characteristic potential field is paired with the first... The partial derivatives of the various operating conditions are:

[0073] The partial derivative of the squared exponential kernel function is:

[0074] Calculate the partial derivatives for all monitored features and operating condition variables, and form... Viacobi matrix And define the gradient of the working potential field. for:

[0075] Indicates the first Scalar field corresponding to each monitored feature; Indicates the first One standardized operating condition variable; Represents the current operating condition vector Represents the working condition vector to be calculated A vector consisting of kernel function values ​​between all training case vectors; Indicates the first The training observation vector of each monitoring feature; the gradient of the working condition potential field is used to represent the rate of change of the working condition potential field along the trajectory of the current working condition, which is the product of the Jacobian matrix and the rate of change of the working condition variable, rather than referring to the spatial gradient of a certain scalar field in the working condition space alone. The The dimension of the component is the . Each monitoring feature per unit time characterizes the current working condition trajectory for the first... The additional disturbance rate caused by each monitoring feature.

[0076] For example, by selecting a rapid pitch change process window from the measured data and constructing the operating condition vector difference based on adjacent 1-second intervals, we can obtain a result primarily based on changes in pitch angle and generator power. ; will Compared with the current working condition position Multiplying them together yields the values ​​corresponding to each monitoring feature. When the operating condition vector remains unchanged, When the operating conditions change rapidly, the gradient of the potential field under the operating conditions is also zero. It participates in subsequent transient perturbation compensation. For deep neural networks, automatic differentiation can be used to obtain... For a locally smoothed gradient boosting tree model, increments of 0.01 can be applied to the positive and negative directions of the standardized operating condition variables, and the local partial derivatives are obtained by dividing the difference between the two field values ​​by 0.02. Both automatic differentiation and symmetric finite difference take the current operating condition vector as input and the Jacobian matrix as output. The improvement in this application is not focused on the differentiation procedure itself, but on the Jacobian matrix formed by combining it with the rate of change of the operating condition variables. Continue to participate in the calculation of the contribution of operating condition disturbance.

[0077] Step S4: Decoupling of operating condition effects.

[0078] like Figure 4 As shown, the expected value of static operating condition disturbance is combined with the transient disturbance compensation amount determined according to the gradient of the operating condition potential field to form the operating condition disturbance contribution amount. The operating condition disturbance contribution amount is then separated from the original monitoring features to obtain the preliminary health state component. Subsequently, the preliminary health state component is input into the health conservation constraint optimization link, and the optimal health state representation is jointly determined by the health conservation energy function, the state transition covariance matrix, and the observation noise covariance matrix.

[0079] Construct a state-space model with health status and operating condition disturbances as latent variables, and its observation equations for:

[0080] in, This is the original monitoring feature vector; This refers to the degradation residual component relative to the healthy baseline, rather than the absolute health of the device; The operating condition disturbance function; To observe the noise, this embodiment uses the operating condition disturbance function. Set to:

[0081] in, This represents the expected value of the static operating condition disturbance. is a dimensionless transient compensation coefficient; and They have the same characteristic dimensions. In this embodiment, the calibration time set is constituted by the transient window of the healthy baseline confirmed by maintenance records and fault-free diagnostic results. The nonnegative transient compensation coefficient that minimizes the sum of squares of the decoupling residuals of the calibration set is determined according to the following formula:

[0082] Indicates time The original monitoring feature vector; This represents the expected value of the static disturbance corresponding to the current operating condition; This represents the theoretical characteristic change caused by changes in operating conditions during the sampling time interval; when the denominator is zero, it will be... Set to zero. The result is... It participates in the calculation of the contribution of operating condition disturbances at each subsequent time step. Optionally, the transient compensation coefficient can be set using the same calculation method for different monitoring characteristics.

[0083] Observation noise satisfies Calculate the values ​​at each time point from the health baseline data. The time interval not exceeding the 20th quantile of the norm distribution and with no fault records is defined as the stable operating time; the time interval of the stable operating time is calculated. Sample covariance matrix The observation noise covariance matrix is ​​obtained according to the following formula:

[0084] in, It represents the sample covariance matrix calculated based on the difference between the original monitoring feature vector and the expected value of the static operating condition disturbance at a healthy baseline and stable operating conditions. It is used to reflect the fluctuation magnitude of the residuals of each monitoring feature and the correlation between features. express The sum of the elements on the main diagonal; This represents the average residual variance of each monitored feature; Represents the regularization coefficient; for An identity matrix of order 1; for traces; when At that time, Set as Therefore, Maintain positive definiteness. Separate the contribution of operating condition disturbances from the original monitoring characteristics to obtain the preliminary health status component:

[0085] Indicates time Preliminary health status components relative to a healthy baseline; Represents the original monitoring feature vector; This represents the expected value of the static disturbance caused by the current operating condition; This represents the compensation amount for transient disturbances generated along the actual change trajectory under the current operating condition. For example, for the rapid pitch process window in step S3, first read... As the static operating condition disturbance at that moment, further calculation As a transient disturbance caused by rapid changes in pitch angle and power, both are considered from the same moment... Subtracting from the middle yields Therefore, static field values, trajectory change rate, and original monitoring characteristics form corresponding processing results within the same decoupled link.

[0086] Step S5: Health conservation constraint optimization.

[0087] Health conservation implies that the health state of equipment evolves continuously over a selected short timescale, rather than remaining constant over long-term operation. The actual degradation of gearbox bearings typically does not undergo reversible instantaneous jumps due to a single gust of wind, emergency pitch control, or start-up / shutdown. Therefore, a health conservation energy function is constructed that includes health conservation constraints and data fidelity terms. :

[0088] in, This represents the total number of sampling moments involved in the optimization. Indicates the moment to be solved ; health status; Indicates the health status at the previous moment; This represents the initial health state component obtained after separating the operating condition disturbance; the first term is used to penalize changes in the health state at adjacent time points. Here is the state transition covariance matrix; the second term is used to limit the deviation of the health state to be solved from the components of the initial health state. To observe the noise covariance matrix; Indicated by The squared Mahalanobis distance is the weight.

[0089] State transition covariance matrix Determined by preliminary health status difference samples from adjacent time points in healthy baseline data and known early degradation data. Calculate the difference vector. Sample covariance matrix And set:

[0090] The sample covariance matrix representing the difference vector is used to reflect the natural variation range of each monitoring feature at adjacent time points and the correlation between the changes of different features; express An identity matrix of order 1; Indicates the dimension of the monitored features; when At that time, Set as The results Used for labeled verification periods determined by operation records, fault diagnosis, or maintenance inspection results; when the direction of change of a known continuous degradation period is reversed after optimization, the value is increased by a factor of 10. The corresponding diagonal elements are then processed until the continuing degradation direction is preserved. Thus, and All are positive definite matrices.

[0091] This embodiment solves the health conservation energy function using a linear dynamic system with a state transition matrix of identity:

[0092]

[0093] This represents the permissible change in health status between adjacent time points, and follows a mean of 0 and a covariance matrix of... The normal distribution; Used to describe the permissible range of change and correlation of each health status component at adjacent time points; Indicates observational noise; the preliminary health state component at the first valid time point. Set as initial state estimate The observation noise covariance matrix Set the initial state covariance First, a forward Kalman filter is used to obtain the filter estimate. and its covariance Then, the RTS smoother is used to iterate from back to front:

[0094] in, Indicates using time 1 to time 2 The time obtained from all the observation data at time A smooth and healthy state; Indicates that only the deadline is used. The filtered health status is obtained from the observed data; The covariance represents the health status of the filter. Indicates based on time Information prediction of the time Health status; This represents the corresponding prediction covariance; Indicates the time obtained using all observation data. The smoothed healthy state; the superscript -1 indicates matrix inversion. Since the state transition matrix is ​​the identity matrix, the smoothing gain is... ; This is the predicted state value. (Definition) All moments This constitutes the optimal health state representation that minimizes the health conservation energy function. Optionally, for nonlinear or non-Gaussian state evolution processes, particle filtering or graph optimization can be used to solve for the health conservation energy function.

[0095] For example, the fast pitch process window obtained in step S4 The sequence before and after this window is input into the linear dynamic system described above. The health conservation constraint term calculates the constraint cost for peaks that appear only in the transient window and subside after the operating condition recovers, based on the squared Mahalanobis distance between the difference vectors of adjacent time points; the data fidelity term ensures that degradation information that changes in the same direction across multiple consecutive windows is retained. middle.

[0096] Step S6: Construction of a unified health reference space and cross-condition mapping.

[0097] like Figure 5 As shown, the optimal health status representation is additively corrected using the current operating condition, standard reference operating condition, and health baseline function to form a cross-operating condition health status sequence. The cross-operating condition health status sequence is then aggregated through a time window and input into the equipment degradation assessment model to output the health index, discrete degradation stage, health index change rate, trend direction, and remaining life prediction probability distribution. When the feedback trigger condition is met, it enters the pending confirmation sample processing link.

[0098] In the health baseline data, the operating condition space is divided into neighborhoods according to the Euclidean distance between standardized operating condition vectors. The modeling domain distance threshold determined in step S2 is used as the neighborhood radius, and the number of historical samples in each operating condition neighborhood is counted. Percentage of historical samples and the variance of each continuous operating condition variable ,in, This represents the total number of healthy baseline samples. It includes all non-empty operating condition neighborhoods. The 75th quantile was determined as the frequency threshold. and will and The preset frequency condition is determined; the variance of each continuous operating condition variable in the operating condition neighborhood that satisfies the preset frequency condition is... The 25th quantile was determined as the preset fluctuation threshold for the corresponding variable. Optionally, for continuous operating condition variables where the original physical quantity is always positive and the neighborhood mean is not zero, the ratio of the standard deviation to the absolute value of the neighborhood mean can be used as the coefficient of variation, and the corresponding preset fluctuation threshold can be set according to the same 25th quantile method. For discrete operating condition variables such as grid connection status, the corresponding values ​​within the same neighborhood must remain consistent. Only those that simultaneously meet the preset frequency condition and all The neighborhood center of the operating condition was determined as a candidate point for the standard reference operating condition.

[0099] From the candidate points of the standard reference operating condition, calculate the Euclidean distance between each candidate point and the rated operating condition vector of the equipment, and determine the candidate point with the smallest distance as the standard reference operating condition. When the distances are the same, select the number of samples. Larger candidate points. In this embodiment, the rated operating condition vector has a wind speed of 11 m / s, a generator power of 3 MW, a pitch angle of 0°, and a rotational speed of rated speed; the gearbox oil temperature, nacelle temperature, nacelle humidity, yaw angle, and generator torque setpoints are taken from the median of the corresponding variables in the healthy baseline sample that meets the aforementioned rated operating conditions, and the grid connection status is set to 1. All ten physical quantities in Table 1 are standardized according to the mean and standard deviation saved in step S1 to form a ten-dimensional standard reference condition. .

[0100] During the initial health baseline phase of unit commissioning, optimal health status representations optimized in step S5 are collected under different operating conditions. For the first... Based on the monitoring characteristics, the health baseline function is defined as follows:

[0101] in, This indicates that the equipment is in a healthy state and is operating under the condition vector. At that time, the first The expected value of an optimal health state representation; Represents the mathematical expectation; Indicates the first Each component represents an optimal health state; vertical lines indicate "under given conditions"; This represents the health baseline dataset confirmed by equipment delivery records, maintenance records, and fault-free diagnostic results. The health baseline function can be obtained through Gaussian process regression, locally weighted regression, or a domain-adaptive mapping network. For any given time... , will the The optimal health status characterization is mapped to the standard reference operating condition:

[0102] Indicates time The first one after mapping to the standard reference condition One health status component; This represents the optimal health state component before mapping; This represents the baseline health value under the current operating conditions; This represents the health baseline value under standard reference conditions; the above additive correction is independent of the denominator and can be used for standardized features with zero or negative values. For features that are physically guaranteed to be positive and whose health baseline value is far from zero, a proportional mapping with a positive lower limit can also be used, but this proportional mapping is not used for features that can take zero or negative values. (For all...) Each monitoring feature undergoes the above processing to obtain a cross-condition health status vector. All moments This constitutes a health status sequence across operating conditions.

[0103] For example, at the current moment when the system is operating at off-rated power, the data is read from the health baseline function. and The difference between the two is used for correction. The result It is in the same health reference space as the same characteristic under rated power conditions and is used as the sequence input for step S7.

[0104] Step S7: Deterioration assessment and remaining life prediction.

[0105] A device degradation assessment model is constructed using a cross-condition health status sequence as input. The model outputs a device health index, discrete degradation stages, health index change rate and trend direction, and a remaining life prediction result with a probability distribution.

[0106] This embodiment employs a sequence-to-point regression model based on Transformer, a deep learning sequence model. The model input is a sequence of cross-condition health status data from the past 200 hours; for a uniform 1-second processing cycle, this time window contains 720,000 original feature points. Using a continuous 10-minute interval as a convergence interval, the mean, standard deviation, and maximum value of each dimension of cross-condition health status are calculated and concatenated into a convergence vector according to a fixed order of "feature dimension - statistic type," resulting in 1200 time-series points. The Transformer includes two encoding layers, each with four attention heads, a hidden layer dimension of 64, and a dropout ratio of 0.1. The output layer reads the encoding vector corresponding to the last time-series point and outputs the predicted mean of remaining lifespan. and unconstrained scale And ensure that the prediction variance is positive in the following manner:

[0107] in, The lower bound of variance is greater than zero, and in this embodiment it is set to... The model training used historical full-life data from five identical turbine units in the same wind farm, from commissioning to the replacement of the high-speed shaft bearing in the gearbox, forming five complete degradation-to-failure trajectories. The time difference between each moment and the actual failure moment was used as the remaining life label, with the actual failure moment determined from maintenance records. Training loss function... for:

[0108] in, The number of training samples for the remaining lifespan; For the first The true remaining lifetime label for each sample; and These represent the mean and variance of the remaining lifetime prediction output by the model, respectively. For cases with a limited number of full-lifetime trajectories, a leave-one-unit cross-validation approach is adopted, using four units as the training set and one unit as the validation set in each iteration. The Adam optimizer iteratively updates the model parameters; training stops when the validation set loss does not decrease for 20 consecutive training rounds, and the model parameters with the minimum validation set loss are retained. Optionally, a Long Short-Term Memory (LSTM) network can be used; when full-lifetime data is limited, a Wiener process or Gamma process, or other probabilistic statistical models, can be used; when the wear or crack propagation patterns of components are known, the corresponding evolution equations can be used as loss constraints or state transition constraints to construct a physical information-driven model. The Transformer, LSM network, and probabilistic statistical models mentioned above all take a cross-condition health state sequence as input and output a health index, degradation stage, or remaining lifetime distribution. The improvement focus of this application is not on the single-sequence model structure itself, but on using the cross-condition health state sequence formed after decoupling from the operating condition potential field, health conservation optimization, and unified health reference space mapping as the model input.

[0109] Hidden Markov Models (HMMs) are used to divide the degradation process into four states: normal, sub-healthy, degraded, and near-failure. The HMM uses a health index and its time difference to define observations. The initial state probability is determined by the proportion of times each state in the labeled lifespan data is the first state in the sequence. The state transition probability is determined by dividing the number of transitions from one state to another by the total number of transitions occurring in the previous state at adjacent labeled times. The Gaussian observation distribution for each state is determined based on the sample mean and sample covariance of the observations. The cross-condition health state corresponding to the root mean square value of the vibration signal is the main feature, and the health index is:

[0110] in, express The component corresponding to the root mean square value of the vibration signal; The mean of the main characteristics during the healthy baseline phase; The failure threshold is determined by the median of the principal feature mean within the last valid window before the failure of the four training units in each cross-validation round, and is validated on the reserved units. Only... The above health index formula is used when the condition is not met; when the condition is not met, the feature that has been verified by training data to increase monotonically with degradation and whose failure threshold is greater than the health baseline value is selected from the monitoring features listed in Table 2 as the main feature. Indicates when Output 0 when Time output ,when Output 1 at that time. The fixed health index range is only used to assign semantic labels to hidden states and as a validation rule: normal corresponds to... Sub-health correspondence Degeneration corresponds to Near-failure correspondence The actual state is output by combining the observation distribution and state transition probability by the Hidden Markov Model, thereby avoiding the use of fixed intervals and state transition models as two sets of parallel decision rules.

[0111] For the past preset trend window Perform a time-weighted linear fit and calculate the resulting slope. As the rate of change of the health index, candidate trend windows of no more than 200 hours and not spanning maintenance events are constructed sequentially in multiples of 24 hours. For each candidate trend window, the number of trend direction identification errors in the labeled and verified data is calculated, and the candidate window with the fewest errors is determined as the preset trend window; if the number of errors is the same, the candidate window with the shorter duration is selected. Based on the health baseline data... The 95th quantile of the distribution is used to set the trend dead zone threshold. ;when Output degradation trend, when The output showed no significant trend of change. Output the recovery trend in real time, and output simultaneously. As a quantifiable result of the deterioration trend.

[0112] The main parameters used in this embodiment and their setting methods are shown in Table 3.

[0113]

[0114] Table 3. Main parameters and their setting methods used in this embodiment. Step S8: Feedback update and manual confirmation.

[0115] Figure 5The feedback update chain, as shown, writes the corresponding data into the sample pool to be confirmed after triggering the feedback threshold condition or template threshold condition. After the sample type is manually confirmed, the working condition potential field update or equipment degradation assessment model update is performed respectively, and the updated model is verified. If the acceptance condition is not met, the previous version continues to be used.

[0116] Read raw monitoring features Expected value of static operating condition disturbance Transient disturbance compensation amount and optimal health status characterization The residuals of the operating condition disturbance are calculated according to the following formula. and its Mahalanobis distance statistics :

[0117]

[0118] In the health baseline validation data The 99th quantile of the distribution is set as the feedback threshold. ; Represents the observation noise covariance matrix; Indicates its inverse matrix; superscript This represents the vector transpose. The number of consecutive processing cycles is selected sequentially from 1 to 60. Simulate continuous triggering in the health baseline validation data; determine the minimum candidate value that does not produce continuous false triggers as the preset number of continuous processing cycles; when all candidate values ​​produce false triggers, Set to 60. When continuous The number of processing cycles is greater than When the residual of the interpretation of the operating condition disturbance continuously exceeds the feedback threshold.

[0119] Each type of degradation process confirmed by maintenance records is truncated into a known degradation template by extracting the corresponding cross-condition health state subsequences with the same time length. Dynamic time warping is performed on the current cross-condition health state window and each known degradation template. The Euclidean distance between corresponding vectors on the optimal alignment path is accumulated and divided by the number of matching points contained in the optimal alignment path to obtain the normalized template distance between the current window and the corresponding known degradation template. The normalized template distances between similar degradation templates are used to form an intra-template distance distribution, and the 95th quantile of this intra-template distance distribution is set as the template threshold. When the distance between the current window and the normalized template of all known degenerate templates is greater than... When the distance between the cross-condition health status sequence and the known degradation template exceeds the template threshold, the corresponding data is placed in the unconfirmed sample pool, without directly updating any model, provided that either the interpreted residual continuously exceeds the feedback threshold or the template distance exceeds the template threshold.

[0120] Maintenance personnel or equipment diagnostic personnel, combining alarm records, work orders, oil or vibration retest results, and maintenance results from the same period, categorize the samples to be confirmed into three types: samples confirmed to be fault-free and belonging to the new operating condition, samples confirmed to show degradation, and uncertain samples. Only samples confirmed to be fault-free and belonging to the new operating condition are added to the health baseline training set in step S2 for incremental learning or retraining of the operating condition potential field; samples confirmed to show degradation are only used for updating the equipment degradation assessment model in step S7 and are not added to the health baseline training set; uncertain samples are retained in the sample pool to be confirmed and will be processed after obtaining new inspection results.

[0121] New versions of the updated operating condition potential field and equipment degradation assessment models were generated. Using a healthy baseline validation set (not included in the update) and a labeled degradation validation set, the median of the explained residuals of operating condition disturbances, the accuracy of degradation trend identification, and the mean absolute error of remaining lifetime were calculated for both the updated and new models. Pre-set acceptance criteria were set as follows: the median of the explained residuals of operating condition disturbances after the update is not greater than the value before the update; the accuracy of degradation trend identification after the update is not lower than the value before the update; and the mean absolute error of remaining lifetime after the update is not greater than the value before the update. The new version was only used when all three conditions were met; otherwise, the previous version was used. The version number, training data range, parameters, and validation results were saved to support rollback.

[0122] In the application results of this embodiment, the health index of the high-speed shaft bearing of the unit's gearbox decreased to 0.25 on the 180th day after startup, with the model predicting a remaining life of 32 days and a 95% confidence interval of 26 to 38 days. Subsequently, on the 35th day, endoscopy confirmed fatigue spalling of the bearing's outer ring, with a prediction error of 3 days. During the 180-day monitoring period, the unit experienced transient events such as high wind shedding, rapid pitch control, and grid disturbances. After employing operating condition potential field gradient compensation and health conservation constraints, the spurious degradation fluctuations caused by transient disturbances were suppressed, while the continuous degradation trend over multiple time windows was retained in the output sequence.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. A method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints, characterized in that, include: Collect multi-source monitoring data and multi-source operating parameters from the equipment, synchronize and standardize the multi-source monitoring data and multi-source operating parameters in time, and construct a high-dimensional operating space based on the multi-source operating parameters; Based on historical operating data, a nonlinear mapping relationship is established between monitoring features and operating condition variables in the high-dimensional operating condition space to obtain the operating condition potential field that characterizes the disturbance intensity and direction of the operating condition on the monitoring features. Calculate the gradient of the working condition potential field based on the current working condition's position and trajectory within the working condition potential field; The contribution of operating condition disturbance in the original monitoring features is determined based on the operating condition potential field and the gradient of the operating condition potential field. The contribution of operating condition disturbance is separated from the original monitoring features to obtain the preliminary health status component. By taking the continuous evolution of health status at adjacent time points as a constraint, the initial health status components are dynamically corrected to obtain the optimal health status representation. The optimal health status representation is mapped to a unified health reference space to form a cross-condition health status sequence. Based on the cross-operating condition health status sequence, an equipment degradation assessment model is constructed, and the equipment degradation status, degradation trend and remaining life prediction results are output.

2. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints as described in claim 1, characterized in that, The operating condition potential field is constructed in the following way: Using multidimensional operating condition variables from historical operational data confirmed by health status as input, and the corresponding monitoring feature values ​​or feature distribution statistics as output, differentiable nonlinear regression models corresponding to each monitoring feature are trained respectively. Each differentiable nonlinear regression model is defined as a scalar field of the corresponding monitoring feature in the high-dimensional operating space. The field value of the scalar field represents the intensity of static disturbance, the sign of the partial derivative of the scalar field with respect to the corresponding operating variable represents the direction of increase or decrease of the monitoring feature, and the absolute value of the partial derivative represents the sensitivity to change. The scalar fields are combined into a vector field according to a fixed order of the monitoring features to obtain the working condition potential field.

3. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints according to claim 2, characterized in that, The working condition potential field gradient is obtained in the following way: The rate of change of the operating condition variable is determined based on the current operating condition vector, the previous operating condition vector, and the sampling time interval. Determine the Jacobian matrix of the working condition potential field at the current working condition vector; The Jacobian matrix is ​​combined with the rate of change of the operating condition variable to obtain the time rate of change of the operating condition potential field along the current operating condition trajectory, and the time rate of change is used as the gradient of the operating condition potential field.

4. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints as described in claim 1, characterized in that, Determining the preliminary health status components includes: Determine the expected value of the static operating condition disturbance based on the operating condition vector at the current moment; The transient disturbance compensation amount is determined based on the working condition potential field gradient, sampling time interval, and dimensionless transient compensation coefficient. The expected value of the static operating condition disturbance is combined with the compensation amount of the transient disturbance to form the operating condition disturbance contribution. The operating condition disturbance contribution is then subtracted from the original monitoring characteristics to obtain the preliminary health status component relative to the health baseline.

5. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints according to claim 4, characterized in that, The dynamic correction includes: A health conservation energy function is constructed, which includes a health conservation constraint term and a data fidelity term. The health conservation constraint term is used to penalize changes in the health state at adjacent time points, and the data fidelity term is used to limit the deviation of the health state to be solved from the initial health state component. The health conservation energy function is solved by using the positive definite state transition covariance matrix and the positive definite observation noise covariance matrix to obtain the optimal health state representation that minimizes the health conservation energy function.

6. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints as described in claim 1, characterized in that, The formation of the cross-condition health status sequence includes: The working condition is defined as the standard reference working condition when the proportion of historical samples in the working condition neighborhood reaches the preset frequency condition and the variance or coefficient of variation of each working condition variable in the working condition neighborhood is not higher than the preset fluctuation threshold. Based on the optimal health status characterization under each working condition in the health baseline stage, a health baseline function is determined to output the corresponding health baseline values ​​for different working conditions. Input the current operating condition and the standard reference operating condition into the health baseline function to obtain the health baseline value corresponding to the current operating condition and the health baseline value corresponding to the standard reference operating condition; The health baseline value corresponding to the current operating condition is subtracted from the representation at each time point, and the health baseline value corresponding to the standard reference operating condition is added to obtain the cross-operating condition health status sequence.

7. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints as described in claim 1, characterized in that, The equipment degradation assessment model takes the time window of the cross-condition health status sequence as input and outputs the current health index of the equipment, the discrete degradation stage to which the equipment belongs, the rate of change and trend direction of the health index, and the remaining life prediction result with probability distribution. The equipment degradation assessment model adopts at least one of deep learning sequence model, probabilistic statistical model or physical information driven model, and is trained using the equipment's historical operation to failure full life data.

8. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints according to claim 7, characterized in that, When the residual of the interpretation of the operating condition disturbance continuously exceeds the feedback threshold set according to the health baseline data, or when the distance between the cross-operating condition health status sequence and the known degradation template exceeds the template threshold set according to the distance distribution within the template, the corresponding data will be placed into the sample pool to be confirmed. The samples to be confirmed that are confirmed to be fault-free and belong to the new operating conditions are used for incremental learning or retraining of the operating condition potential field, and the samples to be confirmed that are confirmed to be degraded are used for updating the equipment degradation assessment model. The working condition potential field and the equipment degradation assessment model before and after the update are verified using a healthy baseline validation set that does not participate in the update and a labeled degradation validation set, and the updated model is put into use when the update results meet the preset acceptance criteria.

9. The method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints as described in claim 8, characterized in that, The feedback threshold, the preset number of consecutive processing cycles, and the template threshold are set in the following ways: Read the original monitoring features, static operating condition disturbance expectation value, transient disturbance compensation amount and optimal health state characterization corresponding to the health baseline verification data, calculate the operating condition disturbance interpretation residual and the distance statistic of the operating condition disturbance interpretation residual corresponding to each processing cycle, and determine the feedback threshold based on the distribution of the distance statistic in the health baseline verification data and the preset false trigger control conditions. Multiple candidate continuous processing cycles are set according to the sampling period of equipment monitoring data and the requirements for degradation feedback response. The health baseline verification data is used to continuously trigger and verify each candidate continuous processing cycle, and the preset continuous processing cycle is determined from the candidate continuous processing cycle that has never generated continuous false triggers. The cross-condition health status subsequences corresponding to various degradation processes confirmed by maintenance records are extracted according to a unified time scale to form the known degradation template. Time series alignment is performed on the known degradation templates of the same type. The template distance is determined according to the distance between the corresponding health status representations on the alignment path. The template threshold is determined according to the template distance distribution between the known degradation templates of the same type and the preset template matching conditions.

10. A method for evaluating equipment degradation across operating conditions based on operating potential field and health conservation constraints, as described in claim 8 or 9, characterized in that... Based on alarm records, work orders, oil or vibration retest results and maintenance results during the same period, the samples to be confirmed are divided into fault-free new working condition samples, degradation samples and uncertain samples. The fault-free new operating condition samples are added to the healthy baseline training set for incremental learning or retraining of the operating condition potential field; the degraded samples are used for updating the equipment degradation assessment model and are not added to the healthy baseline training set. The uncertain samples are retained in the pool of samples to be confirmed. Using the healthy baseline validation set that is not involved in the update and the labeled degradation validation set, the median of the residuals explained by the operating condition disturbances before and after the update, the accuracy of the degradation trend identification, and the mean absolute error of the remaining lifetime are calculated. When the median of the residuals and the mean absolute error of the remaining lifetime after the updated operating condition disturbance interpretation are not greater than the corresponding values ​​before the update, and the accuracy of the updated degradation trend identification is not lower than the corresponding values ​​before the update, the updated operating condition potential field and the equipment degradation assessment model shall be put into use; otherwise, the original operating condition potential field and the equipment degradation assessment model shall continue to be used.

Images