Method and system for predictive maintenance of civil aircraft oil equipment

Abstract

The invention relates to a method and system for predictive maintenance of civil aircraft oil equipment. The method comprises the following steps: S1, collecting and preprocessing data of oil equipment, and performing data cleaning and preprocessing; s2, a model algorithm is adopted to process the preprocessed data, trend prediction is carried out on the working state of the oil equipment based on the processed data, association fusion is carried out on different types of data, and trend prediction is carried out in real time by means of a self-adaptive dynamically-adjusted threshold value; s3, a prediction result is stored and output, the steps S1-S3 are automatically scheduled and managed in a unified mode according to the triggering conditions of the steps S1-S3, and therefore the steps S1-S3 are executed in a coordinated and linkage mode. According to the invention, oil faults can be automatically and efficiently monitored, the accuracy, reliability and real-time performance of predictive maintenance are obviously improved, the complexity and variability of different equipment are effectively handled, and the operation and maintenance efficiency of oil equipment is obviously improved.

Description

Technical Field

[0001] The present invention relates to the field of predictive maintenance, and in particular to a method and system for predictive maintenance of oil equipment of civil aircraft. Background Art

[0002] In civil aircraft, oil, as a crucial medium for various devices, performs crucial tasks such as lubrication, temperature regulation, and power transmission. Its pressure and temperature can serve as early indicators of equipment wear, seal failure, or potential malfunctions. Therefore, by monitoring oil temperature, pressure, and other parameters, the health of related equipment and components can be indirectly inferred.

[0003] In order to obtain the temperature, pressure and other state parameters of the oil, corresponding sensors are usually used in the existing technology. However, the sampling frequency of existing sensors is low and the coverage parameters are limited. They can often only measure one or two parameters. Secondly, oil-related faults usually involve a variety of different equipment. For example, abnormal oil pressure in the aircraft hydraulic system may be caused by pump wear, pipeline leakage or valve failure, while insufficient oil pressure in the aircraft landing gear system may cause the retraction mechanism to stick, insufficient brake pressure output, etc. Traditional methods lack in-depth analysis of this complex correlation, resulting in limited accuracy of fault prediction. In addition, in the existing technology, maintenance information is scattered, data analysis task scheduling is not unified, there is a lack of effective linkage between systems, resource utilization efficiency between systems is low, management is inconvenient, and it is difficult to form a scientific and dynamic fault warning mechanism.

[0004] The above problems reduce the operational efficiency of aircraft and even affect their safe operation. Therefore, there is an urgent need for accurate and efficient predictive maintenance methods and systems for oil failures to achieve real-time monitoring and early warning of aircraft oil failures. Summary of the Invention

[0005] The object of the present invention is to provide a method for predictive maintenance of oil equipment, which solves the problems in the prior art at least to a certain extent.

[0006] The above object is achieved by a method for predictive maintenance of oil equipment according to the present invention. The method comprises the following steps:

[0007] S1: Collect and preprocess the data of oil equipment and perform data cleaning and preprocessing;

[0008] S2: Use model algorithms to process the pre-processed data and perform trend predictions on the operating status of the oil equipment based on the processed data. Different types of data are correlated and fused, and trend predictions are performed in real time using adaptively and dynamically adjusted thresholds.

[0009] S3: Store and output the prediction results, and take corresponding maintenance measures based on the prediction results.

[0010] The steps S1-S3 are uniformly and automatically scheduled and managed according to the triggering conditions of the steps S1-S3, so that the steps S1-S3 are executed in a coordinated and linked manner. Preferably, the triggering conditions include task priority and data processing requirements.

[0011] Through the present invention, oil faults can be automatically and efficiently monitored, the accuracy, reliability and real-time performance of predictive maintenance can be significantly improved, the complexity and variability of different equipment can be effectively coped with, and the operation and maintenance efficiency of oil equipment can be significantly improved.

[0012] In one embodiment of the present invention, step S1 further includes distributing data in real time using a message queue. This ensures the smooth execution of subsequent steps while ensuring high-quality input data and providing a foundation for the accuracy and reliability of subsequent predictive maintenance. The use of a message queue enables efficient and accurate data distribution.

[0013] In one embodiment of the present invention, the data cleaning step includes: standardizing the format of the collected data; identifying and removing noisy data using statistical methods; correcting missing values ​​using interpolation and / or mean filling; and marking anomalous data using an anomaly detection algorithm. Preferably, the format standardization step includes timestamp unification and unit conversion. This achieves unified data format, consistent units, and temporal continuity, improving data purity and ensuring data integrity, while also enabling the detection and timely removal of anomalous data.

[0014] In one embodiment of the present invention, the data preprocessing step includes normalizing the collected data; optimizing the data using smoothing and noise reduction algorithms; and aggregating the data by time window. Normalization facilitates subsequent data processing, while smoothing and noise reduction algorithms further eliminate noise. Aggregating the data by time window allows for the collection of various data within different time windows.

[0015] In one embodiment of the present invention, the step of distributing data in real time includes: shunting the data to corresponding processing modules based on the data attributes and the data requirements of different modules. By shunting different types of data to corresponding processing modules according to their respective attributes, efficient and reliable processing of various data with different attributes can be achieved.

[0016] In one embodiment of the present invention, step S2 includes: S21: identifying the interval within which the feature under investigation lies; S22: obtaining the model eigenvalue within the feature interval; S23: calculating a dynamic threshold based on the model eigenvalue and its changing pattern, wherein the threshold range is dynamically updated based on newly added data, and a dynamic threshold is established as a basis for determining whether the measured value is normal; and S24: predicting the future trend of the data based on the dynamic threshold and the measured value. Here, the data is deeply processed to discover its inherent changing patterns, obtain the correlations between different data, predict future trends, and dynamically adjust them in real time based on the latest situation, thereby improving the accuracy, reliability, and real-time performance of trend prediction.

[0017] In one embodiment of the present invention, step S21 includes: dividing the data into data segments using a sliding window or clustering method; and identifying and labeling characteristic segments within the data segments that meet conditions under different operating states or conditions. This step allows finding periods of stable data or distinct characteristics and their corresponding data segments, and identifying the intervals within the acquired data where specific characteristics exist.

[0018] In one embodiment of the present invention, step S22 includes: screening the high-dimensional data within the feature interval through principal component analysis and / or correlation analysis to identify preliminary potential relationships between the high-dimensional data; and based on the preliminary potential relationships, further screening feature data that meets key indicator requirements from the high-dimensional data, constructing a feature vector for the feature data, and reducing the feature dimension of the feature vector to obtain a final feature value. In this step, the model feature value of interest is obtained within the identified feature interval, and the data dimension is further reduced to prepare for subsequent trend prediction.

[0019] In one embodiment of the present invention, step S23 further includes: establishing an initial threshold model based on historical data and design standards; defining the normal fluctuation range of each feature as the corresponding threshold range based on the initial threshold determined by the initial threshold model; and comparing the dynamic threshold with the real-time feature value as the measured value to determine whether the deviation between the dynamic threshold and the real-time feature value exceeds the threshold range, and determining a confidence interval for the deviation. In this way, a threshold for trend prediction can be constructed, which references historical data and design specifications and, importantly, can reflect current data dynamics in real time.

[0020] In one embodiment of the present invention, the step of dynamically updating the threshold range includes: obtaining a prior distribution of the distribution of previous experience values, constructing a likelihood function for evaluating whether the current data conforms to the prior distribution, inferring a posterior distribution of the future distribution trend of the data based on the prior distribution and the likelihood function, making a threshold judgment based on the current data and the posterior distribution, and dynamically updating the threshold based on the threshold judgment. Here, an implementation method for dynamically updating the threshold range is provided, through which the threshold can be updated in real time to reflect the actual current changes in the data, thereby taking into account the differences or changes between different devices and different environments.

[0021] In one embodiment of the present invention, step S24 includes: constructing a deep learning network model, inputting time series feature data and threshold information; training the deep learning network model using historical data to learn the changing patterns of the oil equipment's operating status; optimizing the deep learning network model's parameters through cross-validation and error analysis; and predicting the future state of the oil equipment and outputting a fault risk assessment. This approach enables accurate, reliable, efficient, and real-time prediction of data trends, estimating the future operating state of the oil equipment, and promptly detecting anomalies.

[0022] In one embodiment of the present invention, the method further includes: automatically scheduling the execution of all steps based on task priority and data processing requirements; collecting and storing the data processing results of each step; monitoring the execution status and processing results of each step, logging them, and issuing exception alerts; and providing a visualization page displaying key indicators, forecast results, and maintenance recommendations. By utilizing unified, automated task scheduling management, data processing efficiency is significantly improved, coordination between tasks is deepened, overall resources are fully and rationally utilized, human intervention is reduced, and the efficiency, accuracy, stability, and reliability of trend forecasting are enhanced.

[0023] According to another aspect, the present invention also relates to a system for predictive maintenance of oil equipment, which is used to implement the method described above. The system comprises:

[0024] A data collection and preprocessing module, wherein the data collection and preprocessing module is designed to perform data cleaning, data preprocessing, and data distribution;

[0025] A model algorithm module, wherein the model algorithm module is designed to perform feature segment identification, feature value calculation, threshold construction, and trend prediction;

[0026] A comprehensive scheduling management module, which is designed to automatically and uniformly manage and coordinate the execution of data collection tasks, data preprocessing tasks, model algorithm tasks, and model result collection tasks based on trigger conditions;

[0027] A data storage and management module, wherein the data storage and management module is designed to centrally store and manage raw data, pre-processed data, and historical records; and

[0028] The model display and analysis module is designed to output fleet information, model results, abnormality alarms, and perform management functions.

[0029] The various embodiments and advantages described above for the method for predictive maintenance of oil equipment are also applicable to the system of the present invention and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] To better understand the above and other objects, features, advantages, and functions of the present invention, reference may be made to the preferred embodiments shown in the accompanying drawings. Like reference numerals in the accompanying drawings represent like components. Those skilled in the art should understand that the accompanying drawings are intended to illustrate preferred embodiments of the present invention by way of illustration and are not intended to limit the scope of the present invention. The components in the drawings are not drawn to scale.

[0031] Figure 1 is a general flow chart of a method for predictive maintenance of oil equipment according to the present invention;

[0032] Figure 2A is a flow chart of the steps for collecting and preprocessing data in a method according to the present invention;

[0033] Figure 2B is a flow chart of the steps for data processing and trend prediction in the method according to the present invention;

[0034] Figure 3 is an exemplary flow chart of steps for feature interval identification in the method according to the present invention;

[0035] Figure 4 is an exemplary flow chart of steps for trend prediction in a method according to the present invention; and

[0036] Figure 5 is a block diagram of a system for predictive maintenance of oil equipment according to the present invention. DETAILED DESCRIPTION

[0037] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. What is described here is only a preferred embodiment of the present invention. Those skilled in the art can conceive of other ways to implement the present invention based on the preferred embodiment, and other ways also fall within the scope of the present invention.

[0038] The following combination Figures 1 to 5 The present invention provides a method 100 for predictive maintenance of oil equipment in civil aircraft. Figure 1 According to the present invention, a method 100 for predictive maintenance of civil aircraft oil and fluid equipment includes step S1: collecting various oil and fluid equipment data and performing data cleaning S11 and preprocessing S12. This data includes, for example, historical operational data, design documents, maintenance records, time series data, current operating status, and environmental data of the aircraft's oil and fluid equipment. This raw data can be collected using, for example, the aircraft's flight data recorder, sensors, line maintenance records, and design materials. This raw data is then transmitted to a data receiver and stored in a structured, semi-structured, or unstructured format.

[0039] After collecting and storing the required raw data, it is cleaned and pre-processed, that is, step S1 includes: data cleaning S11 and data pre-processing S12. Step S1 also includes data distribution S13, that is, using the message queue to distribute the data in real time, see Figure 2A In this way, the high quality of input data is ensured, thereby guaranteeing the accuracy and reliability of subsequent predictive maintenance and providing the basis and prerequisite for the smooth execution of subsequent steps.

[0040] Specifically, the step S11 of cleaning data includes: standardizing the format of the collected data; identifying and eliminating noise data using statistical methods; repairing missing values ​​using interpolation and / or mean filling; and marking abnormal data using anomaly detection algorithms. Preferably, the step of format standardization includes timestamp unification and unit conversion. Here, the data cleaning algorithm is called to filter out noise data and outliers, unify the timestamps (i.e., the time format of the unified data generation) and unit format. For missing data, linear interpolation or mean filling, missing value filling and other methods are used to repair it. Suspected outliers are marked or eliminated by, for example, the statistical anomaly detection algorithm Z-score or IQR (interquartile range) method. Through the step of data cleaning S11, the format and unit consistency of the data can be achieved, the purity of the data can be improved, the quality and integrity of the data can be ensured, and abnormal data can be discovered and marked.

[0041] After the data cleaning step S11, the data preprocessing step S12 is performed, which includes normalizing the collected data; optimizing the data using smoothing and noise reduction algorithms; and aggregating the data by time window. The purpose of normalization is to facilitate subsequent data processing. Smoothing and noise reduction algorithms can further eliminate noise from the data, building on the initial removal of noise during data cleaning, thereby further improving data quality. Aggregating data by time window allows for the collection of various data corresponding to different time periods within different time windows, facilitating subsequent data processing.

[0042] Then, step S13 of real-time data distribution is executed, which includes: shunting data to corresponding processing modules based on data attributes and the data requirements of different modules. Here, model parameters and common parameter data required for analysis of the complete flight segment are extracted and distributed to downstream processing modules via the message queue described above. In the present invention, model data includes, for example, oil temperature, oil pressure, oil quantity, engine core speed (engine core speed refers to the corresponding engine's main speed, expressed as a percentage, such as 70%), and fuel temperature. Common data includes altitude, total temperature, total pressure, touchdown signal, Mach number, flight phase, etc. Utilizing message queues enables efficient, accurate, and real-time data distribution, effectively alleviating the pressure on data reception and data cleaning services when experiencing massive amounts of data, thereby ensuring service stability. Necessary statistical analysis is performed before data distribution and transmission, recording metadata such as data volume, time range, and key attributes. By shunting various data types to corresponding processing modules based on their respective attributes and the data requirements of different modules, efficient and reliable processing of diverse data with different attributes is achieved.

[0043] Next, step S2 is executed: various model algorithms are used to process the pre-processed data, and based on the processed data, trend prediction is performed on the working status of the oil equipment, wherein different types of data are correlated and fused, and trend prediction is performed in real time with the help of adaptively dynamically adjusted thresholds. This step S2 includes: identifying the interval where the examined feature is located, i.e., feature interval identification S21; obtaining the model feature value within the feature interval, i.e., model feature value obtaining S22; calculating the dynamic threshold value based on the model feature value and its change law, wherein the threshold range is dynamically updated according to the newly added data, and a dynamic threshold value is established as the basis for judging whether the measured value is normal or not, i.e., threshold construction S23; and predicting the future trend of data changes based on the dynamic threshold value and the measured value, i.e., trend prediction S24, see Figure 2B. Here, for example, after the eigenvalue is obtained, a dynamic threshold value is calculated based on the eigenvalue and its changing pattern. In an exemplary embodiment, the initial threshold value is first calculated based on the data generated at the previous time, and then a new eigenvalue is calculated based on the data generated at the current time, and then a new threshold value is calculated based on the new eigenvalue to replace the initial threshold value, thereby realizing the dynamic update of the threshold value. Through step S2, the data is deeply processed to explore its inherent changing patterns, obtain the correlation between different data, predict future changing trends, and dynamically make real-time adjustments based on the latest situation, thereby improving the accuracy, reliability, and real-time performance of trend prediction.

[0044] Specifically, the step S21 of identifying the feature interval includes: dividing the data into data segments by a sliding window or clustering method; and identifying and marking the feature segments that meet the conditions under different operating states or working conditions in the data segments. In this step S21, the time series data is segmented by, for example, a sliding window or clustering method. For different working conditions, such as different flight phases of the aircraft, system operating modes, etc., judgment conditions are set to identify relatively stable or time periods with obvious characteristics, namely the so-called feature segments. The start and end indexes of these feature segments and the corresponding working condition information are recorded and marked, and the feature segment indexes and related parameters that meet the conditions are stored in the stable segment index list or the feature segment index list for subsequent feature extraction. For example, in one embodiment, the flight segment data can first be preliminarily screened based on rough initial screening conditions, such as parameter A being equal to 7 and parameter B being greater than 0.3, and the corresponding segments that meet the initial screening conditions are identified as preliminary feature segments; then, a sliding window mechanism is introduced within the scope of the preliminary feature segments, and the data within each window is further screened using more stringent and complex judgment conditions to identify accurate feature intervals. It should be pointed out that it is usually necessary to identify multiple feature intervals in a complete flight segment, for example, feature interval 1 that meets screening condition 1, feature interval 2 that meets screening condition 2, feature interval 3 that meets screening condition 3, and so on. Generally speaking, these conditions are defined according to different operating conditions, that is, the so-called "identifying and marking feature segments that meet the conditions under different operating conditions in the data segment." In this way, the characteristic value obtained is not just one, but multiple characteristic values ​​that meet different conditions.

[0045] In an exemplary embodiment of the present invention, for example, a stable lubricating oil temperature segment is extracted as a characteristic segment. For time series data, a sliding window combined with logical judgment is employed. Specifically, a stable interval where the lubricating oil pressure is greater than 15 psi (pounds per square inch) and the engine core speed remains above 70% (i.e., the engine speed is above 70% of its rated speed for most of the time) is selected as the characteristic interval. For this characteristic interval, a key indicator is defined as, for example, a temperature change of less than 0.7 degrees Celsius within a 1-minute time window, and the duration exceeds 10 minutes. The corresponding data segment is denoted, for example, as work_indices. However, key indicators are not limited to this and can be determined based on specific operating conditions. A sliding window is defined: the unit time window is 1 minute, the step length is 1 second, and the total data length is at least 10 minutes. For example, a stable_list is used to store the logical value of each sliding window judgment. If the key indicator condition is met, a 1 is stored; if not, a 0 is stored, and the process continues one step forward. The loop ends when the sliding window reaches the end of work_indices. The data segments with consecutive 1s in stable_list are extracted and the number of 1s is calculated. If this number is greater than or equal to the threshold duration last_time, the corresponding list of first and last index pairs is returned. The index between the first and last indexes is completed, and the corresponding model parameters such as fuel temperature and lubricating oil temperature are stored in the stable segment index stable_indices.

[0046] In this embodiment, if Figure 3 As shown, a stable interval with oil pressure greater than 15 psi and engine core speed greater than 70% is selected, and the start and end points of this interval are recorded or annotated. The data is segmented using a sliding window. First, the sliding window's tail end is checked to see if it exceeds the aforementioned stable interval. If not, the sliding window remains within the stable interval. The data is then checked to see if the key indicator condition is met: whether the difference between the maximum and minimum temperature values ​​within 60 seconds from the start point is less than 0.7 degrees Celsius. If so, a 1 is recorded in the stable_list; otherwise, a 0 is recorded. The sliding window is then moved forward one second, and the start and end points of the interval are recorded again. This cycle repeats until the window's tail end exceeds the stable interval. At this point, the total duration that the stable_list value remains 1 is determined to be greater than the required minimum duration of stability, last_time. If so, the interval between the start and end points of the stable_list value plus the sliding window length is recorded and designated as the stable characteristic interval. If not, no interval in the current window meets the condition.

[0047] Then, step S22 of obtaining the model eigenvalues ​​is performed, which includes: filtering the high-dimensional data within the feature interval through principal component analysis (PCA) and / or correlation analysis to discover preliminary potential relationships between the high-dimensional data; and based on these preliminary potential relationships, further filtering the high-dimensional data to identify feature data that meets key indicator requirements, constructing feature vectors for these feature data, and reducing the feature dimensions of these feature vectors to obtain final eigenvalues. In this step, statistical analysis or probabilistic inference, such as probabilistic principal component analysis or probabilistic factor analysis, is used to discover the potential relationships between the multidimensional data. Bayesian regression modeling is employed, combining system principles with design documents and maintenance information, to obtain the theoretical eigenvalue distribution and uncertainty information such as its mean and variance. The difference between the actual observed value and the theoretical value is calculated, and the confidence interval or probability distribution of the difference can be quantified as the key eigenvalue. If high-dimensional data is available, core features can be selected through dimensionality reduction, such as PCA or correlation analysis, to output a feature vector containing multiple features.

[0048] In an exemplary embodiment for obtaining model eigenvalues, a correspondence is established between lubricating oil temperature, fuel temperature, and engine core speed. Based on design documents and historical data, a Bayesian regression algorithm is used to build a model, obtaining a "theoretical oil temperature" containing uncertainty information such as mean and variance (the theoretical oil temperature is obtained from design documents and historical data; the oil temperature here refers only to the lubricating oil temperature). The difference between the actual average oil temperature and the theoretical oil temperature distribution is calculated as the eigenvalue. Other methods, such as weighting, can also be used to obtain the eigenvalue. The degree of deviation is calculated, and the confidence interval is quantified. The obtained characteristic interval typically has multiple parameters, which are correlated with each other. Principal component analysis and / or correlation analysis are used to determine the potential relationship between them, resulting in some primary characteristic indicators (or primary eigenvalues, which are subsequently used for further calculations). Then determine the definition of key indicators, such as the standard deviation of the difference between D and E under condition C, the similarity P between M and N, the above-mentioned temperature change within a 1-minute unit time window is less than 0.7 degrees, and similar situations. Then, construct a feature vector based on these indicators, and further reduce the dimension of the feature vector to obtain, for example, the final eigenvalue.

[0049] Next, step S23 of threshold construction is executed, which includes: establishing an initial threshold model based on historical data and design standards; defining the normal fluctuation range of each feature as the corresponding threshold range based on the initial threshold determined by the initial threshold model; and comparing the dynamic threshold mentioned above with the real-time characteristic value as the measured value to determine whether the deviation between the dynamic threshold and the real-time characteristic value exceeds the threshold range, and determining the confidence interval of the deviation. In this step, an initial threshold model is established to determine the upper and lower limits of the basic threshold by combining historical data statistics with design and maintenance requirements. Bayesian inference (prior distribution, likelihood function, posterior distribution) or adaptive filtering is used to adjust the threshold interval in real time. The current characteristic value is compared with the dynamic threshold. If their deviation exceeds a certain confidence interval, it is determined to be abnormal. At the same time, the current posterior distribution is used as the prior distribution for the next time, forming a dynamic threshold adjustment mechanism for continuous iteration.

[0050] In an exemplary embodiment, for example, the value of a certain parameter should not fall below 100. This value, for example, corresponds to the line 0 in the model. Therefore, the defined characteristic value cannot fall below 0, otherwise it indicates an anomaly, such as an aircraft fault alarm. The line 0 represents the initial threshold model, which is typically the baseline. Based on the calculation logic of the model characteristic value, a reasonable threshold range is derived, such as 2-10. As long as the characteristic value is within the range of 2-10, it is likely normal. If the current characteristic value is in the range of 0-2, the aircraft may not report a fault, but the model determines that there is a potential risk. When a new characteristic value is generated, it is determined whether it is within the threshold range. If it is, it is likely normal. Otherwise, the component monitored by the model is likely abnormal (the probability is specifically inferred from the confidence interval). For example, if the new characteristic value is 4, which is within the range of 2-10, it is likely normal (90%-95% confidence interval). If the new eigenvalue is 1.5, there is a 65%-75% chance of an anomaly, and the confidence interval for the deviation is 65%-75%. If the new eigenvalue is 0.2, there is a 95%-99% chance of an anomaly, and the confidence interval for the deviation is 95%-99%.

[0051] In the present invention, the aforementioned dynamic updating of the threshold range includes: obtaining the distribution of prior empirical values, i.e., a prior distribution; constructing a likelihood function to assess whether the current data conforms to the prior distribution; inferring the future distribution trend of the data, i.e., a posterior distribution, based on the prior distribution and the likelihood function; performing a threshold determination based on the current data and the posterior distribution; and dynamically updating the threshold based on the threshold determination. In an exemplary embodiment for dynamically updating the threshold, a prior distribution is first determined. Specifically, for example, by statistically analyzing lubricating oil temperature characteristic values ​​from multiple voyages, the mean and variance of the distance characteristic value are constructed, and lower and upper thresholds based on the distance characteristic value are constructed using the three-sigma (3δ) principle. The "distance" here refers to the difference between the actual value and the standard value or theoretical value (or the value of a specific curve), i.e., the so-called characteristic value. Subsequent calculations are based on this distance, such as calculating the mean, variance, and likelihood function of the distance. The so-called prior distribution is equivalent to a probability distribution of the uncertainty or subjective perception of the existing data, that is, using a probability distribution to describe this uncertainty. For example, based on recent work, the lubricating oil temperature is estimated to have a mean of 105 degrees Celsius, with an amplitude fluctuation of approximately 5 degrees Celsius. Therefore, a normal distribution with a mean of 105 and a variance of 25 can be used as the prior distribution. A likelihood function is then constructed. For example, a window of 20 is set and the likelihood function is constructed using the 20 most recent historical eigenvalues ​​to obtain a characteristic sample distribution. The posterior distribution is then determined. For example, if the posterior distribution still conforms to a normal distribution, the posterior distribution function is calculated according to Bayes' theorem and used as the latest threshold reference. It should be noted that the prior distribution is derived from historical data, and the likelihood function is derived from the most recent data. The prior distribution and likelihood function are combined to calculate the posterior distribution according to the Bayesian formula. When both the prior distribution and the likelihood function are normal, the posterior distribution also conforms to a normal distribution. Next, a threshold determination is performed. For example, the characteristic value of the current flight segment is calculated using the latest flight segment data. If the latest characteristic value deviates significantly from the posterior distribution threshold, it is determined to be abnormal. Finally, a dynamic update is performed. Specifically, for example, the previous posterior distribution is used as the new prior distribution for the next iteration. Combined with statistical process control, the moving average and standard deviation of the characteristics of the last 20 flight segments are calculated, and the threshold is updated in real time. The "moving average" here actually calculates the likelihood function. The likelihood function is based on the value of the most recent period, which can be 20 flight segments, for example.

[0052] After establishing the threshold, the trend prediction step S24 can be executed. This step includes: constructing a deep learning network model, inputting time series feature data and threshold information; training the deep learning network model using historical data to learn the changing patterns of the oil equipment's operating status; optimizing the deep learning network model's parameters through cross-validation and error analysis; and predicting the future state of the oil equipment and outputting a failure risk assessment. In this step, an end-to-end time series prediction model is established, using deep learning structures such as LSTM and GRU (Gated Recurrent Unit). Uncertainty is quantified within the neural network through random dropouts (deactivation, discarding) or other methods. The model first filters and smoothes the feature value sequence, using the smoothed feature values ​​for training to learn the system's long- and short-term dependency patterns (similar to the changing patterns). In the prediction phase, the latest time series data and external features are input into the model to obtain predicted feature value distributions for multiple future time points. The previously established dynamic thresholds are also used to determine anomalies. If the predicted value is likely to exceed a safe range, an alert is triggered and recorded in the model's operation log.

[0053] In an exemplary embodiment for trend prediction, see Figure 4 In order to perform the trend prediction step, a prediction model network structure is adopted, which consists of two parts: the encoder-decoder framework and the prediction network. The encoder-decoder framework uses two layers of LSTM (Long Short-Term Memory) units to capture the inherent patterns in the input time series data during the pre-training phase and output the hidden state learned by the encoder (the inherent pattern and the hidden state are similar to the change law, Figure 4The "Z" in the code represents hidden features, and the decoder outputs pre-trained data. The first layer of LSTM units learns the underlying temporal patterns of the input sequence (such as short-term fluctuations). The second layer of LSTM units extracts higher-order global features (such as periodicity and trends) based on this. This allows the model to account for both short-term and long-term dependencies while reducing model complexity. During the prediction phase, the prediction network inputs the hidden states generated by the encoder-decoder framework and the latent external features (which include other influencing factors) into a multilayer perceptron. The multilayer perceptron outputs the prediction results and quantifies the uncertainty. The multilayer perceptron is used to perform the final regression or classification prediction after sequence feature extraction. The specific implementation steps are: smoothing the feature values ​​in the data using, for example, a Kalman filter; segmenting the smoothed feature values ​​into time series of fixed length; then training the end-to-end model using this time series data until convergence; using the trained model to predict feature trends at several future time points and determine whether there are anomalies based on a pre-set threshold range; if the predicted value remains within the threshold range, the feature is considered normal; otherwise, an anomaly alarm is triggered.

[0054] After the result is predicted, step S3 of storing and outputting the prediction result and taking corresponding maintenance measures according to the prediction result is executed. At this point, all the steps are completed. It should be emphasized here that in the present invention, all the above steps S1-S3 are uniformly and automatically scheduled and managed according to the triggering conditions of the steps S1-S3, so that the steps S1-S3 are coordinated and linked with each other. Preferably, the triggering conditions include task priority and data processing requirements. Regarding this unified and automatic scheduling management, the method of the present invention also includes: automatically scheduling the execution of all steps according to task priority and data processing requirements, including algorithms such as data cleaning, distribution, preprocessing, feature extraction, threshold construction and prediction; collecting and storing the data processing results of each step; monitoring the execution status and processing results of each step, and performing log records and abnormal alarms; and providing a visualization page to display key indicators, prediction results and maintenance suggestions. See the following introduction for details.

[0055] According to another aspect of the present invention, see Figure 5The present invention also relates to a system 200 for predictive maintenance of oil equipment, which is used to implement the method 100 described above. The system 200 includes a data collection and preprocessing module 1, a model algorithm module 2, a comprehensive scheduling and management module 3, a data storage and management module 4, and a model display and analysis module 5. The data collection and preprocessing module 1 is designed to perform data cleaning, data preprocessing, and data distribution. It is responsible for collecting raw flight data, maintenance data, etc. from the route and performing data cleaning, data preprocessing, and data distribution operations. The model algorithm module 2 is designed to perform feature segment identification, feature value calculation, dynamic threshold construction, and trend prediction. It conducts in-depth analysis of preprocessed data and outputs fault prediction results and related feature indicators. The comprehensive scheduling and management module 3 is designed to schedule data collection tasks, data preprocessing tasks, model algorithm tasks, and model result collection tasks. It is responsible for coordinating data flow and task execution between modules, automatically scheduling data collection, preprocessing, model calculation, storage, prediction, and feedback processes, and ensuring efficient system operation. The data storage and management module 4 is designed to centrally store and manage raw data, pre-processed data, and historical records, providing data query and backup interfaces to ensure data persistence and traceability. The model display and analysis module 5 is designed to output (e.g., via a front-end visualization platform) fleet information, real-time predicted model results, historical trends, anomaly alerts, and other information, as well as perform management functions, including model management and user management, to facilitate user monitoring and decision-making.

[0056] Specifically, the data collection and preprocessing module 1 automatically identifies incoming data, completes data collection, cleaning, and preprocessing, and records the data status, including, for example, aircraft flight segment information, flight segment data size, data status, and processing timestamp. After confirmation, the data distribution process is executed, sending only the parameters required for the lubricating oil temperature model, such as those required for the model algorithm module, to ensure sufficient data and minimize data loss.

[0057] In model algorithm module 2, whenever new data is received, the feature segment identification algorithm, feature extraction (i.e., eigenvalue calculation algorithm), threshold construction algorithm, and trend prediction algorithm are automatically scheduled for execution. The algorithm execution results are imported into the data storage module by the scheduler. These results include model eigenvalues, timestamps, feature segment numbers, eigenvalue status scalars, upper and lower thresholds, predicted values ​​(saved as vector values), and anomaly judgment scalars. Operational information is also recorded, including the amount of data executed, execution time, and execution effect evaluation.

[0058] In the integrated scheduling management module 3, all of the aforementioned steps, including data collection, preprocessing, and modeling, are managed and scheduled automatically and coordinated based on trigger conditions. Task processes and dependencies are established. All of these steps are scheduled and logged, and the operational status of each module is monitored in real time.

[0059] In the data storage and management module 4, data storage and indexing are performed. To this end, data generated by each link in the entire model workflow is received, and multiple databases are used in a mixed manner for hierarchical storage, where feature values ​​are stored in a relational database, original feature segments are stored in a time series database, and task execution status information is stored in a document database. The storage reading efficiency is guaranteed by efficient data indexing such as the LZ4 data compression algorithm and the B+ tree index. In addition, security management and access control are performed in the data storage and management module 4. To this end, strict authority management and identity authentication are implemented. Only the required relevant system parameters, such as those mentioned in the above-mentioned model algorithm module, are sent for the conversion valve model, avoiding full data access for a single model. In addition, data backup and fault tolerance are performed in the data storage and management module 4. To this end, clear data lineage and dependency relationships are formed through full-process data monitoring and logging, which facilitates data backtracking after an error in a certain step. A distributed storage architecture is used to achieve data security backup.

[0060] In Model Display and Analysis Module 5, the front-end display page receives data from the scheduling system related to the predictive maintenance analysis model, such as lubricating oil temperature, including chapter number, timestamp, component information, and all of the aforementioned model data. Using the system-component selection model, the page selects the specific unit by unit number to display the predictive maintenance analysis results. Furthermore, within Model Display and Analysis Module 5, the display page can filter intervals by time dimension, color-code eigenvalues, thresholds, and predicted values, and issue alerts for eigenvalues ​​exceeding thresholds, sending notifications to operations and maintenance personnel. These personnel can then combine historical model information, route maintenance information, and model predictions to comprehensively assess the current system status and assist maintenance personnel in making decisions.

[0061] Since the system 200 according to the present invention is used to implement the method 100 for predictive maintenance of oil equipment described above, the various embodiments and advantages described above for the method 100 according to the present invention are also applicable to the system 200 according to the present invention, and will not be repeated here.

[0062] Existing predictive maintenance model analysis methods and systems employ only a single model. In contrast, the present invention provides a universal predictive maintenance method and system for civil aircraft oil and fluid equipment. By utilizing multiple model algorithms, employing dynamic threshold adjustment algorithms, probabilistic inference, and other methods, the resulting model possesses enhanced interpretability, making prediction results more reliable and significantly improving fault prediction accuracy and system operational safety. Through unified management and automated scheduling, the entire process, from data cleaning, preprocessing, distribution, feature extraction, to threshold construction and trend prediction, is fully automated, ensuring clear processes, well-defined inputs and outputs, and orderly scheduling.

[0063] The advantages and effects of the present invention are mainly reflected in:

[0064] Process Automation: This invention automates the entire data processing process through unified management and automatic scheduling, seamlessly integrating all aspects of the process, from data cleaning, preprocessing, and distribution to feature extraction, threshold construction, and trend prediction. This automated process not only ensures real-time and efficient data transmission and processing, but also significantly reduces the errors and delays that can arise from manual operations, lowering the cost of human intervention. Furthermore, through automatic monitoring and scheduling, the system can dynamically adjust the collaborative work between modules, improving overall maintenance efficiency and prediction accuracy, and providing strong support for equipment maintenance and fault prevention.

[0065] Flexible Threshold Management: By establishing different threshold ranges based on each aircraft's historical data and extracted feature value information, we can reduce feature and threshold deviations caused by long-term operating environment differences. However, each model still needs to be based on fixed baseline thresholds defined based on design and maintenance requirements to ensure compliance with flight safety requirements and maintenance standards. On this basis, thresholds are then modified and optimized based on actual operating conditions, achieving both universality and specificity, thereby improving the accuracy of fault diagnosis.

[0066] Model interpretability: Using algorithms based on prior knowledge and probabilistic inference to extract feature values ​​from characteristic segments, each extracted feature is given clear credibility and interpretability, making the physical meaning and role of the feature more intuitive. Deep learning algorithms can handle complex nonlinear dynamic changes in trend forecasting, providing highly accurate predictions. This combination retains the interpretability of Bayesian methods while leveraging the strengths of machine learning in modeling complex trends, thereby achieving both transparent feature sources and precise predictive maintenance.

[0067] Data traceability: Through unified management and automated scheduling, the relationship between each piece of data is clearly displayed. When anomalies or failures occur, the data flow can be traced back to quickly locate the root cause. This also provides complete records for subsequent audits, ensuring compliance with security and compliance standards, thereby enhancing the system's compliance capabilities.

[0068] While some embodiments of the present invention have been described for illustrative purposes, the present invention is not limited to these embodiments. Numerous modifications and variations will occur to those skilled in the art. Therefore, these embodiments have been selected and described to better illustrate the principles and practical applications of the present invention and to enable those skilled in the art to understand the contents thereof. In other words, all modifications and variations made without departing from the spirit of the present invention will fall within the scope of protection of the present invention as defined by the appended claims.

Claims

1. A method (100) for predictive maintenance of oil equipment of a civil aircraft, the method comprising the following steps: (S1): collecting data of the oil equipment and performing data cleaning (S11) and preprocessing (S12); (S2): Processing the pre-processed data using a model algorithm, and performing trend prediction on the working status of the oil equipment based on the processed data, wherein different types of data are correlated and fused, and the trend prediction is performed in real time with the help of a threshold that is adaptively and dynamically adjusted; (S3): Storing and outputting the prediction results, and taking corresponding maintenance measures according to the prediction results, Wherein, the steps (S1)-(S3) are uniformly and automatically scheduled and managed according to the triggering conditions of the steps (S1)-(S3), so that the steps (S1)-(S3) are coordinated and linked to each other.

2. The method (100) according to claim 1, wherein: The triggering conditions include task priority and data processing requirements.

3. The method (100) according to claim 2, wherein: The step (S1) further includes: utilizing a message queue to distribute data in real time (S13).

4. The method (100) according to claim 3, wherein: The step of cleaning data (S11) includes: Standardizing the format of the collected data; Use statistical methods to identify and eliminate noisy data; Use interpolation and / or mean imputation to fix missing values; and Anomaly detection algorithms are used to mark abnormal data.

5. The method (100) according to claim 4, wherein: The format standardization steps include timestamp unification and unit conversion.

6. The method (100) according to claim 3, wherein: The data preprocessing step (S12) includes: Normalizing the collected data; Optimizing the data using smoothing and noise reduction algorithms; and Aggregate the data by time window.

7. The method (100) according to claim 3, wherein: The step of distributing data in real time (S13) includes: dividing the data into corresponding processing modules according to data attributes and different module data requirements.

8. The method (100) according to claim 2, wherein: The step (S2) comprises: (S21): Identify the interval where the feature under investigation is located; (S22): obtaining the model characteristic value within the characteristic interval; (S23): Calculating a dynamic threshold value based on the model characteristic value and its changing pattern, wherein the threshold range is dynamically updated according to the newly added data, and establishing a dynamic threshold value as a basis for judging whether the measurement value is normal; and (S24): Predicting future data change trends based on the dynamic threshold and the measurement value.

9. The method (100) according to claim 8, wherein: Step (S21) includes: Dividing the data into data segments by a sliding window or clustering method; and In the data segments, characteristic segments that meet conditions under different operating states or working conditions are identified and marked.

10. The method (100) according to claim 8, wherein: Step (S22) includes: Screening the high-dimensional data within the feature interval by principal component analysis and / or correlation analysis to mine preliminary potential relationships between the high-dimensional data; and Based on the preliminary potential relationship, feature data that meets the key indicator requirements are further screened out from the high-dimensional data, a feature vector of the feature data is constructed, and the feature dimension of the feature vector is reduced to obtain a final feature value.

11. The method (100) according to claim 8, wherein: Step (S23) further includes: Establish an initial threshold model based on historical data and design criteria; Based on the initial threshold value determined by the initial threshold model, defining the normal fluctuation interval of each feature as the corresponding threshold range; and The dynamic threshold is compared with the real-time characteristic value as the measurement value to determine whether the deviation between the dynamic threshold and the real-time characteristic value exceeds the threshold range, and a confidence interval of the deviation is determined.

12. The method (100) of claim 8, wherein: The steps of dynamically updating the threshold range include: obtaining a prior distribution of the distribution of previous experience values, constructing a likelihood function for evaluating whether the current data conforms to the prior distribution, inferring a posterior distribution of the future distribution trend of the data based on the prior distribution and the likelihood function, performing a threshold judgment based on the current data and the posterior distribution, and dynamically updating the threshold according to the threshold judgment.

13. The method (100) of claim 8, wherein: Step (S24) includes: Build a deep learning network model and input time series feature data and threshold information; Using historical data to train the deep learning network model to learn the changing patterns of the operating status of the oil equipment; Optimizing the parameters of the deep learning network model using cross-validation and error analysis; and Predict the future state of the oil equipment and output a failure risk assessment.

14. The method (100) according to any one of claims 2 to 13, wherein: The method further comprises: Automatically schedule and execute all steps according to the task priority and the data processing requirements; Collect and store the data processing results of each step; Monitor the execution status and processing results of each step, and perform logging and exception alerts; and Provides a visualization page to display key indicators, prediction results and maintenance recommendations.

15. A system (200) for predictive maintenance of oil equipment of a civil aircraft, for implementing the method (100) according to any one of the preceding claims 1-14, the system (200) comprising: A data collection and preprocessing module (1), wherein the data collection and preprocessing module (1) is designed to perform data cleaning, data preprocessing, and data distribution; A model algorithm module (2), wherein the model algorithm module (2) is designed to perform feature segment identification, feature value calculation, threshold construction, and trend prediction; A comprehensive scheduling management module (3), wherein the comprehensive scheduling management module (3) is designed to be able to uniformly and automatically schedule and manage according to trigger conditions and to coordinate and link the execution of data collection tasks, data preprocessing tasks, model algorithm tasks, and model result collection tasks; A data storage and management module (4), wherein the data storage and management module (4) is designed to centrally store and manage raw data, pre-processed data, and historical records; as well as A model display and analysis module (5) is designed to output fleet information, model results, abnormality alarms, and perform management functions.

16. The system (200) of claim 15, wherein: The triggering conditions include task priority and data processing requirements.

Images