ACM data analysis methods and data monitoring systems

Abstract

The present invention discloses an ACM data analysis method and a data monitoring system, belonging to the technical field of aircraft maintenance. The method comprises: constructing an ACM data set, setting a feature engineering processing method, setting a dynamic window, calculating a dynamic correlation coefficient to extract coupling features, dividing working conditions, eliminating fluctuation interference, and generating a structured time series data set; obtaining ACM benchmark data of aircraft of the same model, setting a benchmark fault judgment method, judging the latest data in the structured time series data set, screening abnormal parameters, and judging whether the ACM is faulty in combination with actual temperature control effects, thereby reducing single parameter misjudgment, setting a coupled fault location method, diagnosing faulty ACM, distinguishing true faults from false abnormalities, and improving the accuracy of coupled fault diagnosis; and setting a prediction scheduling method, performing fault prediction on fault-free ACM, performing health status prediction and life assessment, thereby reducing the probability of sudden failures and unplanned groundings.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft maintenance and relates to an ACM data analysis method and a data monitoring system. Background Art

[0002] The air cycle machine (ACM) in an aircraft's air conditioning system plays a key role in maintaining a suitable cabin environment and is a core component of aircraft environmental control. The ACM primarily consists of a compressor and a turbine cooler. The turbine cooler performs a crucial heat exchange task, exchanging heat with the outside air or other media, in conjunction with the compressor to cool and condition the cabin air.

[0003] An existing Chinese patent with authorization publication number CN107807628B discloses a method for assessing the performance degradation of a heat exchanger in a civil aircraft air conditioning system. The method includes: first, collecting key performance parameters; then establishing a heat exchanger monitoring parameter system based on the key performance parameters; and then using the monitoring parameter system in a fault-free state to establish a heat exchanger performance baseline model. When the system receives new monitoring parameters, the heat exchanger performance baseline model calculates an estimated value of the key performance parameter. The estimated value is then subtracted from the actual monitored value to obtain a deviation value of the monitoring parameter. The deviation value is monitored and analyzed, and a fault warning is issued when the characteristic value becomes abnormal.

[0004] While existing technologies address the difficulty of online monitoring heat exchangers in air conditioning systems, avoiding unplanned maintenance of air conditioning systems caused by heat exchanger failures, reducing unnecessary cleaning and replacement of heat exchangers, saving airlines maintenance costs, and improving aircraft utilization, they lack comprehensive consideration of the coupling effects of multiple components in aircraft air conditioning systems (especially ACMs) and their lifecycle health management, resulting in insufficient fault diagnosis accuracy and a lack of predictability in maintenance strategies. In aircraft ACMs, the compressor and turbine cooler operate in tandem. Failure to consider the dynamic interaction between the compressor and turbine coolers can lead to misdiagnosis of faults, resulting in inaccurate fault location and incorrect repairs, such as misdiagnosing a compressor problem as a heat exchanger failure. Furthermore, since aviation maintenance requires advance planning of spare parts and downtime, only providing fault warnings without considering the use of historical deviation data to predict remaining service life fails to meet proactive maintenance needs, preventing airlines from optimizing spare parts inventory and flight scheduling. Therefore, this application provides an ACM data analysis method and data monitoring system. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide an ACM data analysis method and data monitoring system, which uses a dynamic window to extract the coupling characteristics of the compressor and turbine cooler, and uses Granger causality analysis to accurately locate coupling faults to avoid misjudgment. At the same time, combined with a deep autoencoder and a survival analysis model, the remaining life is predicted based on historical data, providing support for proactive maintenance, optimizing spare parts inventory and flight scheduling, and realizing multi-component coupling and full life cycle health management.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] ACM data analysis methods, including:

[0008] An ACM dataset was constructed, and a feature engineering method was set. Dynamic windows were set, dynamic correlation coefficients were calculated, and coupled features were extracted. The operating conditions were divided to generate a structured time series dataset. An ACM is an air cycle machine, consisting of a compressor and a turbine cooler. The ACM dataset includes offline data related to the ACM, as well as flight parameter data, ACM maintenance records, and flight environmental data.

[0009] Obtain ACM benchmark data for aircraft of the same model, set a benchmark fault judgment method, judge the latest data in the structured time series data set, filter abnormal parameters, and determine whether the ACM is faulty based on the actual temperature control effect. Set a coupled fault location method to diagnose the faulty ACM and distinguish between true faults and false anomalies. At the same time, set a predictive scheduling method to predict faults for fault-free ACMs, perform health status prediction and lifespan assessment;

[0010] When generating a structured time series dataset, the feature engineering processing method includes:

[0011] Read the timestamps of the ACM data set, calculate the differences between adjacent timestamps, and define the mode of the differences as the time resolution;

[0012] Conduct response tests on the ACM to simulate operation under different operating conditions, monitor the fluctuation characteristics of key parameters, and determine the time range when the key parameters exceed the preset variation range;

[0013] Dynamically adjust the window size and sliding step size based on the preset scale factor;

[0014] Sliding the window along the time series, performing Pearson correlation analysis on the compressor parameters and the turbine cooler parameters within each window to generate a dynamic correlation coefficient;

[0015] Associating the dynamic correlation coefficient with the start timestamp of the corresponding window to form a coupled characteristic time series;

[0016] When diagnosing a faulty ACM, the coupling fault location method includes:

[0017] In the structured time series data set, extracting the data segment corresponding to the latest timestamp based on the timestamp information;

[0018] Granger causality analysis was used to construct a bivariate regression model based on compressor parameters and turbine cooler parameters to evaluate the lag prediction ability between variables.

[0019] Using the AIC criterion, calculate different lag orders AIC value, select the one that minimizes the AIC value as the optimal lag order;

[0020] estimating the parameters of the bivariate regression model based on the least squares method, calculating and minimizing the residual sum of squares to fit the data, and calculating the causal relationship strength between the compressor and turbine cooler parameters;

[0021] Granger causality test is used to determine whether there is a causal relationship between the parameters; if , then it is determined that the turbine cooler parameters are the Granger causes of the compressor parameters; if , then it is determined that the compressor parameters are the Granger causes of the turbine cooler parameters; among them, 、 are the parameters of the bivariate regression model.

[0022] Furthermore, the feature engineering processing method further includes:

[0023] Divide the flight parameters into climb, cruise, and descent phases, traverse the ACM dataset, and mark the flight phase for the data at each time point based on the preset flight phase judgment rules;

[0024] Based on the flight phase identification results and dynamic correlation coefficients, the ACM dataset and the data in the coupled feature time series are saved into three constructed subsets: climb, cruise, and descent.

[0025] Standardize the data of each subset;

[0026] Using timestamps as row indexes, key parameters and dynamic correlation coefficients are arranged into a two-dimensional matrix to generate a structured time series dataset.

[0027] Furthermore, when determining whether the ACM is faulty, the baseline fault determination method includes:

[0028] Obtaining trouble-free operation data of the same model aircraft at the initial stage of delivery, filtering cruise phase data based on the flight phase judgment rules, and generating a benchmark database;

[0029] Calculate the mean and standard deviation of key parameters in the benchmark database and construct the parameter benchmark interval using the 3σ principle;

[0030] Integrate the mean value and parameter benchmark interval of each key parameter to build a standard ACM performance model;

[0031] The latest data in the structured time series data set is read, input into the standard ACM performance model, compared with the parameter benchmark interval, abnormal parameters are identified, and an abnormal parameter set and a normal parameter set are constructed.

[0032] Furthermore, the reference fault judgment method further includes:

[0033] Reading normal parameters in the normal parameter set, sequentially calculating derivatives between adjacent data points in a continuous sliding window, and obtaining a derivative sequence in the window;

[0034] For any single sliding window, the mean and variance of the derivative sequence are calculated to perform dynamic trend detection;

[0035] If the mean of the derivative sequence is not within the preset normal mean range or the variance exceeds the preset normal variance range, it is marked that the normal parameter has an abnormal derivative change within the sliding window; otherwise, the normal parameter does not have an abnormal derivative change;

[0036] Perform abnormality judgment on the continuous sliding window; if there is abnormal derivative change in all the continuous sliding windows, update the normal parameters to abnormal parameters and migrate them to the abnormal parameter set; otherwise, the normal parameters do not have abnormalities;

[0037] The actual temperature control effect in the computer cabin and the temperature control difference are calculated , comprehensively judge whether there is a fault in ACM;

[0038] when or , it is determined that the ACM has a fault; when and , it is determined that there is no fault in ACM; among them, is the number of exception parameters, is the temperature difference threshold, is the abnormal threshold.

[0039] Furthermore, the coupling fault locating method further includes:

[0040] Based on the result of the Granger causality test, determining whether there is a causal relationship between the compressor and turbine cooler parameters;

[0041] If there is no causal relationship, the fault type is determined to be a false abnormality and false abnormality processing is performed;

[0042] If there is a causal relationship, determine whether the dynamic correlation coefficient exceeds the preset normal correlation range;

[0043] like , determine the fault type as a pseudo-anomaly, and perform pseudo-anomaly processing; where, is the Z-score value of the correlation coefficient within the sliding window, is the normal correlation threshold;

[0044] like , combined with the flight phase information, the fault type is judged; if the aircraft is in the cruise phase, the fault type is determined to be a true fault and true fault processing is performed; otherwise, the fault type is determined to be a false anomaly and false anomaly processing is performed;

[0045] Generate a maintenance report based on the fault type.

[0046] Furthermore, the true fault processing includes:

[0047] Based on the abnormal parameter set, obtaining components corresponding to the abnormal parameters in the abnormal parameter set and determining faulty component candidates;

[0048] Calling a historical maintenance case library to filter historical cases related to the candidate faulty component;

[0049] Calculate the cosine similarity between the current fault data and each historical case data;

[0050] Sort historical cases based on cosine similarity and output the top three fault causes with the highest similarity and their confidence levels;

[0051] Generates a fault location report that includes the faulty component, cause, and confidence level, where the confidence level is the cosine similarity value.

[0052] Furthermore, the pseudo exception handling includes:

[0053] Extracting sensor data from the structured time series data set, checking the validity and continuity of the data point by point, and calculating a continuity index;

[0054] Comparing the continuity index with a preset continuity threshold, if the continuity index is lower than the continuity threshold, marking the corresponding sensor as a suspected fault; otherwise, the sensor has no abnormality;

[0055] Performing noise level analysis on sensor data, comparing the calculated noise index with a preset noise threshold, and marking the sensor as pending calibration if the noise index exceeds the noise threshold; otherwise, the sensor is normal;

[0056] Build a pseudo-abnormal pattern library to store the characteristic data of historical pseudo-abnormal events and manage them by category;

[0057] Using a dynamic time warping algorithm to match the sensor data without anomalies with the patterns in the pseudo-anomaly pattern library, and calculating the matching distance;

[0058] If the matching distance is less than the preset distance threshold, the corresponding pseudo-anomaly type is output;

[0059] Otherwise, update the false exception to a true fault and perform true fault processing;

[0060] Generates a fault location report that includes sensor status and false anomaly type.

[0061] Furthermore, the ACM data analysis method further includes using a fault prediction method to determine whether the ACM has failed; the fault prediction method includes:

[0062] Constructing a deep autoencoder model, training and evaluating the deep autoencoder model using the structured time series dataset, and calculating a reconstruction error based on an output of the deep autoencoder model;

[0063] Based on historical fault data, a degradation weight is set for each component, and the reconstruction error of each component is weighted and summed according to its degradation weight to construct a health index;

[0064] Construct an indicator prediction model, train the indicator prediction model with historical health indicator data, and predict the health indicator trend in the future time step. The length of the input sequence is , the prediction step length is ;

[0065] The most recent The health indicator data points are input into the indicator prediction model, and the future The predicted value of health indicators at each time step;

[0066] Building a survival analysis model based on historical failure data, inputting the predicted health indicator values and key parameters into the survival analysis model, and calculating the predicted failure probability at different time points;

[0067] The predicted failure probability at different time points is compared with a preset failure threshold. Once it is detected that the predicted failure probability at a certain time point exceeds the failure threshold, it is determined that the ACM is about to fail and the remaining life of the ACM is output.

[0068] Data monitoring system, including: data reading module, diagnosis and analysis module, prediction module and feedback module;

[0069] The data reading module is used to construct an ACM data set, set a dynamic window, extract coupling features based on dynamic correlation coefficients, divide working conditions, and generate a structured time series data set;

[0070] The diagnostic analysis module is used to combine static thresholds and parameter change rate analysis to screen out abnormal parameters, and based on the actual temperature control effect, determine whether the ACM is faulty, and diagnose the faulty ACM to distinguish true faults from false abnormalities;

[0071] The prediction module is used to predict potential faults of the fault-free ACM, and based on the fault prediction results, determine whether the ACM has failed and calculate the remaining life of the ACM;

[0072] The feedback module is used to receive feedback results from the staff after performing maintenance operations to determine the accuracy of fault diagnosis and prediction and perform update operations.

[0073] Beneficial effects of the present invention:

[0074] Through dynamic window and flight phase division, the dynamic coupling relationship between parameters is effectively captured, the interference of operating condition fluctuations on data analysis is eliminated, and the accuracy of feature extraction is improved; baseline fault judgment combines static thresholds with dynamic trend detection, which can not only quickly identify obvious anomalies, but also capture short-term parameter fluctuations, improving the sensitivity and reliability of fault detection. In addition, coupled fault location quantifies the interaction between components through Granger causal analysis and decision trees, solving the misjudgment problem of traditional single-parameter threshold methods and achieving accurate positioning of coupled faults; the health indicator prediction model based on deep autoencoders and LSTM can detect potential fault trends in advance; combined with the survival analysis model, it can accurately estimate the remaining life and provide a basis for preventive maintenance; the maintenance feedback mechanism continuously optimizes model parameters and algorithms by comparing diagnosis and prediction results, forming a closed-loop optimization system, and continuously improving the accuracy and adaptability of the system. While ensuring flight safety, it significantly reduces maintenance costs and improves aircraft operating efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 This is a flowchart of the ACM data analysis method;

[0076] Figure 2 Flowchart of the feature engineering processing method of the present invention;

[0077] Figure 3 is a flow chart of the benchmark fault judgment method of the present invention;

[0078] Figure 4 This is a flow chart of the coupling fault location method of the present invention;

[0079] Figure 5is a flow chart of the fault prediction method of the present invention;

[0080] Figure 6 This is the structure diagram of the data monitoring system. DETAILED DESCRIPTION

[0081] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0082] Example 1

[0083] refer to Figures 1 to 5 As shown, this embodiment introduces the ACM data analysis method, including:

[0084] A dedicated data reading device is used to connect to the aircraft's flight parameter card, which is also known as the flight data recorder. It is a portable storage medium used to store aircraft flight parameter data in avionics systems. Based on a pre-defined data protocol, offline data related to the ACM and flight parameter data are extracted, such as compressor speed, inlet and outlet pressures, turbine cooler temperature, and flow. At the same time, with the help of database query statements, such as SQL queries, the ACM maintenance record data is filtered out from the aircraft's maintenance database, including historical fault types, maintenance time, and replacement parts. A data interface is established with the airport weather station to obtain flight environmental data, such as outside temperature, air pressure, and humidity, through GPS time or flight phase marks, such as take-off / landing time; using GPS time or flight phase marks (such as take-off / landing time) as a reference, the environmental data is matched with the timestamp of the offline data. For data with inconsistent timestamps, linear interpolation is used to align the two in time, generating ACM data with aligned timestamps. The data set is collected to ensure time alignment and standardization of the data, providing highly reliable input for subsequent analysis. The data in the ACM dataset is preprocessed, feature engineering processing methods are set, dynamic windows are set, and the dynamic correlation coefficients of the compressor parameters and turbine cooler parameters are calculated for coupled feature extraction. At the same time, the data is clustered according to the flight phase to divide the operating conditions, eliminate fluctuation interference, and finally generate a structured time series dataset. During flight, adjustments to the data collection methods of airborne equipment or the connection of new data collection equipment will undermine the stability and reliability of the original aircraft system, posing a serious threat to flight safety. At the same time, according to the strict requirements of airworthiness certification, real-time data collection is a major change to the original aircraft design and requires comprehensive airworthiness re-certification, which is time-consuming and costly. In addition, the interfaces and data formats of airborne equipment on different models of aircraft are different, making it difficult to develop compatible collection equipment and software. Data transmission also faces challenges such as real-time performance, stability, and anti-interference. Therefore, considering the above factors, it is more appropriate to adopt an offline data processing method.

[0085] Through multi-dimensional data fusion and intelligent algorithms, accurate identification and location of ACM faults are achieved; baseline data of ACMs of the same model aircraft is obtained, and a baseline fault judgment method is set up to judge the latest data in the structured time series data set. Abnormal parameters are screened out by combining static thresholds and parameter change rate analysis. Combined with the actual temperature control effect, it is determined whether the ACM is faulty, reducing single parameter misjudgment. A coupled fault location method is set up to diagnose the faulty ACM. Granger causal analysis and decision tree algorithms are used to quantify the interaction between components, distinguish between true faults and false anomalies, improve the accuracy of coupled fault diagnosis, and accurately locate the faulty component. The cause of the fault is deeply analyzed, and the fault is graded according to its severity. For faults of different levels, corresponding maintenance instructions are sent to notify maintenance personnel to perform maintenance.

[0086] Through health indicator modeling and multi-model fusion, potential failures are predicted for fault-free ACMs, and the remaining life of the faulty components is determined based on the fault prediction results. A fault-free ACM is obtained, a fault prediction method is set, and the ACM's predicted health indicator is calculated to determine whether the ACM has failed. The remaining life of the ACM is calculated, reducing the probability of sudden failures, extending the equipment's service life, reducing unplanned groundings, and improving flight operation efficiency.

[0087] Receive feedback from staff after performing maintenance operations to determine the accuracy of fault diagnosis and prediction; for fault diagnosis, compare the actual maintenance results with the diagnosis results. If not, identify the specific problems and make targeted updates and modifications; for fault prediction, compare the prediction with the actual fault time or flight. If the predicted fault time or flight is significantly different from the actual situation, update the model parameters, improve the algorithm, or add new influencing factor variables to improve the accuracy of fault diagnosis and prediction, continuously improve the performance of monitoring ACM, enhance the adaptability to complex fault modes, ensure the accuracy and robustness of the monitoring system, and form a closed-loop improvement mechanism.

[0088] Furthermore, when generating structured time series datasets, feature engineering methods include:

[0089] Different time scales can reflect the operational characteristics of ACM at different levels. Fixed windows cannot fully capture this information. Dynamic windows can flexibly adjust the window size and sliding step size based on the actual operation of the ACM, thereby more accurately capturing the dynamic relationship between compressor and turbine cooler parameters at different time scales. The window size determines the data range considered when calculating the dynamic correlation coefficient, and the window sliding step size determines the interval at which the window moves in the time series, affecting the frequency of dynamic correlation coefficient calculation and the degree of data coverage.

[0090] Read the timestamps of the data in the ACM dataset, calculate the difference between adjacent timestamps using the diff function, and use the mode function to obtain the mode of the difference to determine the data collection frequency, which is defined as the time resolution of the ACM dataset. At the same time, the response test of ACM is carried out. By simulating the operation of ACM under different working conditions, the changes of key parameters are monitored, the fluctuation characteristics of key parameters on different time scales are analyzed, and the time range of key parameters exceeding the preset change range is determined. Among them, key parameters include compressor parameters, turbine cooler parameters, and other comprehensive parameters (such as turbine inlet and outlet pressure ratio or compressor inlet and outlet pressure ratio, turbine cooler inlet and outlet temperature difference or temperature difference with other components). The values of key parameters are offline data in the ACM dataset.

[0091] Set the scaling factor based on ACM characteristics and experience 、 、 、 , and calculate the window size And the sliding step , the expression is as follows:

[0092] ;

[0093] ;

[0094] The window sliding operation starts from the starting position of the time series data in the ACM dataset, with the first data point as the starting point of the window, and the window size is calculated based on the window size. Select a data segment as the first window data for calculating the dynamic correlation coefficient, and then calculate the data segment according to the step size. Slide the window backward along the time series in sequence. After each slide, a new window data is obtained, and finally a series of window data for calculating the dynamic correlation coefficient is output;

[0095] For each window data, the dynamic correlation coefficient of the compressor parameters and turbine cooler parameters in the window is calculated based on Pearson correlation analysis. The dynamic correlation coefficient of each window is associated with the corresponding window start timestamp to generate associated data containing the dynamic correlation coefficient and the corresponding window start timestamp.

[0096] The dynamic correlation coefficient values in the correlation data are arranged in chronological order to generate a coupling feature time series, preserving the time order to reflect the dynamic changes of ACM;

[0097] Based on the characteristics and experience of the flight process, flight phase judgment rules are formulated. Flight phases include climb, cruise, and descent. The flight parameter data in the ACM dataset are traversed to identify the aircraft's flight phase. According to the judgment rules, the flight phase judgment is performed on the data at each time point. Once the flight phase corresponding to a certain time point is determined, the data record with the same timestamp is found in the coupled feature time series, and the corresponding flight phase tag is added. The coupled feature time series with the flight phase tag is output.

[0098] Three independent data subsets were created, named the climb dataset, cruise dataset, and descent dataset. Based on the flight phase identification results, the ACM dataset and all data in the coupling feature time series, including coupling features, were classified into the climb, cruise, and descent phases and stored in the corresponding data subsets. The coupling feature refers to the dynamic correlation coefficient, which reflects the collaborative working relationship between key ACM components. Clustering based on the coupling feature helps to group data with similar operating conditions in the same flight phase, making subsequent data analysis more targeted. The data in each subset has similar operating conditions, effectively reducing the interference of different operating condition fluctuations on data analysis.

[0099] Since the numerical ranges and dimensions of different parameters vary greatly, the Z-score normalization method is used to process the data in each data subset, and three standardized data subsets are output;

[0100] Using timestamps as row indexes, compressor parameters, turbine cooler parameters, dynamic correlation coefficients, and other comprehensive parameters are arranged in column order. In each row, the parameter values corresponding to the same time point are sequentially filled into the corresponding columns to construct a two-dimensional matrix. This ensures that each row in the matrix represents a data record at a time point, and each column represents a specific parameter, thereby generating a structured time series dataset.

[0101] Furthermore, the flight phase judgment rules are as follows:

[0102] Set the height change threshold to 、 , the duration time is 、 ,in, , ;

[0103] When the altitude change rate is greater than , and lasts longer than , the speed gradually increases (the first-order derivative of the speed is greater than 0), and the engine thrust is in a high power output state (greater than 80% of the cruise thrust), which is determined to be the climbing stage;

[0104] When the altitude change rate is range and lasts longer than When the speed standard deviation is less than a specific threshold and the engine thrust is stable within the error range of ±10% of the cruise thrust, it is determined to be in the cruise phase;

[0105] When the altitude change rate is less than , and lasts longer than The entire process from the aircraft landing to the end of the rollout is considered the descent phase, which includes the descent preparation, descent process, and landing rollout phase.

[0106] For special flight situations, such as circling in the air, encountering meteorological disasters or aircraft failures that cause abnormal changes in flight parameters, classification is carried out according to the main parameter trends; if the main parameter trends are in line with the characteristic range of the cruise phase, it is classified as the cruise phase; if the parameter change characteristics are closer to the climb or descent phase, they are classified accordingly.

[0107] Furthermore, when determining whether the ACM is faulty, the baseline fault determination method includes:

[0108] Through the flight parameter card, at least The complete operational data of a trouble-free flight. Since the flight conditions of the aircraft during the cruise phase are relatively stable, the ACM operating status is also relatively stable. The data during this phase can better reflect the parameter characteristics of the normal operation of the system, providing a reliable data basis for building an accurate standard model. Based on the judgment rules of different flight phases, the data of the cruise phase are screened out to generate a benchmark database.

[0109] For key parameters in the benchmark database , calculate the mean and standard deviation , and use the 3σ principle as the key parameter Constructing parameter benchmark intervals ;in, The first key parameters, and , is the number of key parameters in the benchmark database;

[0110] The mean and benchmark interval of each key parameter are integrated to construct a standard ACM performance model, which serves as a standard for subsequent judgment of whether the latest data is abnormal, so as to quickly identify parameters that significantly deviate from the normal range. The input of the model is the offline data of each key parameter, and the output is the preliminary judgment result of each key parameter, including abnormal and normal. The input parameter value is compared with the corresponding benchmark interval in the standard ACM performance model. If the parameter value is within the benchmark interval, the output result is normal, and the key parameter is defined as a normal parameter; if the parameter value exceeds the benchmark interval, the output result is abnormal, and the key parameter is defined as an abnormal parameter.

[0111] Read the latest data from the structured time series dataset and input it into the standard ACM performance model to perform preliminary anomaly judgment. Based on the output results, construct abnormal parameter sets and normal parameter sets. Use static threshold detection to quickly identify parameters that significantly deviate from the normal range, providing a basis for preliminary judgment of whether the system is abnormal. The abnormal parameter set is used to store abnormal parameters and corresponding offline data, and the normal parameter set is used to store normal parameters and corresponding offline data.

[0112] Read the normal parameters in the normal parameter set, and for each normal parameter, select the latest continuous A sliding window is formed. In each sliding window, the derivatives between adjacent data points in the window are calculated in sequence to reflect the rate of change of the parameter and obtain the derivative sequence in the window.

[0113] For each sliding window, the changes in the derivatives within the window are analyzed, and the mean and variance of the derivative sequence are calculated. The calculated derivative mean and variance are compared with the pre-set normal fluctuation range data. Through dynamic trend detection, the changes in the derivatives within the sliding window are analyzed, which helps to capture abnormal changes in parameters in the short term, make up for the shortcomings of static threshold detection, and improve the sensitivity and accuracy of anomaly detection;

[0114] If the mean of the derivative sequence is not within the normal mean range, or the variance of the derivative sequence exceeds the normal variance range, then the corresponding normal parameter is marked as having abnormal derivative changes within the window; otherwise, the normal parameter is not abnormal;

[0115] For continuous A sliding window is used to judge if the continuous If there are abnormal changes in derivatives within each sliding window, the corresponding normal parameters are judged to be abnormal, and the corresponding normal parameters are updated to abnormal parameters and migrated to the abnormal parameter set. Otherwise, the corresponding normal parameters are not abnormal.

[0116] The ultimate goal of the air conditioning system is to provide a suitable temperature environment for the aircraft cabin. Therefore, the actual temperature control effect is an important indicator to measure whether the ACM is operating normally. By calculating the temperature control difference, the cooling effect of the system can be verified from the perspective of practical application, avoiding misdiagnosis caused by parameter misjudgment; obtaining the actual temperature control effect in the aircraft And the expected temperature control effect , and calculate the temperature difference , the expression is as follows:

[0117] ;

[0118] In the formula, several temperature values are obtained by temperature sensors installed at various locations inside the aircraft cabin, and the actual temperature control effect is obtained by calculating the average value of these temperature values, and , is the number of temperature sensors, For the The temperature value collected by a temperature sensor;

[0119] Define the key parameters in the abnormal parameter set as abnormal parameters and obtain the number of abnormal parameters , set the temperature difference threshold to , the abnormal threshold is ,To determine whether the ACM is faulty, the number of parameter anomalies and the actual temperature control effect are combined to avoid misjudgment of a single indicator;

[0120] when or , it is determined that the ACM has a fault; when and , it is determined that there is no fault in the ACM.

[0121] Furthermore, when diagnosing a faulty ACM, the coupled fault location method includes:

[0122] When an ACM fault occurs, all data segments corresponding to the latest timestamp are extracted from the structured time series dataset based on the timestamp information to obtain complete operating status data at the time of the fault. This provides comprehensive information for subsequent in-depth analysis of the fault cause, including compressor parameters, turbine cooler parameters, other comprehensive parameters, dynamic correlation coefficients, and flight phase marker information.

[0123] By quantifying the interaction between components, the problem of easy misjudgment of the traditional single-parameter threshold method is solved, and the coupling fault can be accurately located. The traditional single-parameter threshold method only focuses on whether a single parameter exceeds the normal range and ignores the interaction between components. Granger causal analysis and decision tree construction comprehensively consider the relationship between multiple parameters, which helps to locate the coupling fault more accurately; using multivariate time series association rule mining methods, such as Granger causal analysis, to construct a two-variable regression model, evaluate the predictive ability of the past value of one variable on the current value of another variable, and introduce the Akaike Information Criterion (AIC) to determine the optimal lag order of the model. The AIC criterion can find a balance between model complexity and goodness of fit, ensure the accuracy of the model, and use the AIC criterion to determine the lag order , for different lag orders Calculate the AIC value and select the one that minimizes the AIC value. As the optimal lag order, the expression is as follows:

[0124] ;

[0125] ;

[0126] Where, for Compressor variables at each moment The value of for Time turbo cooler variable The value of 、 、 、 is the parameter to be estimated, 、 are mutually independent white noise error terms, is the number of parameters in the bivariate regression model, is the sample size, is the likelihood function, is the index variable, representing the lag order, , for exist The value of the moment also represents Lag Expect, for exist The value of the moment also represents Lag Expect;

[0127] Use the least squares method to estimate the parameters of the bivariate regression model and calculate the residual sum of squares (RSS). By minimizing the residual sum of squares, the bivariate regression model can accurately fit the data and determine the parameters to be estimated.

[0128] To quantify the degree of mutual influence between compressor and turbine parameters, the causal relationship strength is calculated. , the expression is as follows:

[0129] ;

[0130] In order to determine whether there is a causal relationship between the parameters of the compressor and turbine cooler components, and thus discover the potential fault coupling relationship, a Granger causality test is performed; if , indicating that the lag value of the turbine cooler variable has a significant contribution to the prediction of the current value of the compressor variable, then the turbine cooler variable is determined to be the Granger cause of the compressor variable; if , indicating that the lagged value of the compressor variable has a significant contribution to the prediction of the current value of the turbine cooler variable, and the compressor variable is determined to be the Granger cause of the turbine cooler variable;

[0131] Based on the Granger causality verification results, a preliminary judgment is made on the fault type to determine whether there is a Granger causality between the compressor and the turbine cooler. The causal relationship includes the turbine cooler variable being the Granger cause of the compressor variable and the compressor variable being the Granger cause of the turbine cooler variable.

[0132] If there is no causal relationship, indicating that there is no obvious mutual influence between the compressor and turbine cooler parameters, the fault occurs independently or is caused by other non-coupled factors, then the fault type is determined to be a false abnormality and false abnormality processing is performed;

[0133] If there is a causal relationship, determine whether the dynamic correlation coefficient exceeds the preset normal correlation range; if , indicating that the dynamic correlation coefficient exceeds the normal correlation range and is marked as abnormal correlation; if , indicating that the dynamic correlation coefficient is within the normal correlation range, the fault type is determined to be a false abnormality, and false abnormality processing is performed; among them, is the Z-score value of the correlation coefficient within the sliding window, is the normal correlation threshold;

[0134] When the dynamic correlation coefficient is determined to be abnormal correlation, the fault type is judged in combination with the flight phase information. Since the aircraft operates stably in the cruise phase and the parameter fluctuations are relatively small, the true fault is judged by the cruise phase data to avoid the influence of transient interference in the take-off and landing phases on the fault judgment. If the aircraft is in the cruise phase, the fault type is judged to be a true fault and true fault processing is performed. By matching historical maintenance cases, the fault component and cause are determined. If the aircraft is not in the cruise phase, considering that the operating environment of the aircraft in the take-off and landing phase is complex and the parameters fluctuate frequently, the abnormality at this time is caused by transient interference, and the fault type is judged to be a false abnormality and false abnormality processing is performed.

[0135] According to the preset fault classification rules, the fault handling results are classified and a maintenance report is generated, including the fault location report, level, and type; among them, fault handling includes true fault handling and false abnormality handling, and the output of fault handling is the fault location report.

[0136] Furthermore, true fault handling includes:

[0137] Based on the abnormal parameter set, the component corresponding to the abnormal parameter is obtained and defined as a candidate for the faulty component. The historical maintenance case library is then called to filter out historical maintenance cases related to the current candidate faulty component, significantly narrowing the data matching scope and improving processing efficiency. The historical maintenance case library stores various fault cases that have occurred in the past ACM. Each case contains detailed data at the time of the fault, such as component parameters, fault symptoms, maintenance measures, and repair results.

[0138] For the selected historical cases, the current fault data and historical case data are converted into vector form, and the cosine similarity between the current fault data and each case data is calculated. , sort the calculated cosine similarities, and output the top 3 fault causes and their corresponding confidence levels, where the confidence level is the cosine similarity value;

[0139] Output fault location report, including fault components, causes and fault confidence. The fault cause is to use historical maintenance experience to quickly locate the fault cause and improve the efficiency and accuracy of fault diagnosis.

[0140] Furthermore, pseudo exception handling includes:

[0141] Extract sensor data from structured time series datasets, check the validity and continuity of the data point by point, and calculate sensor data continuity indicators , to quantify the continuity of the data and determine whether the sensor is working normally; the expression is as follows:

[0142] ;

[0143] Where, is the total number of sampling points in the structured time series dataset, is the number of consecutive valid points in the structured time series dataset. Continuous valid points are data points that are detected as normal by the standard ACM performance model.

[0144] Set the continuous threshold to , judge whether the sensor is working normally; if , indicating that the continuity index exceeds the continuity threshold, the sensor data has many discontinuities or abnormal points, the sensor fails, and the sensor is marked as suspected failure; if , sensor continuity is normal;

[0145] Analyze the noise level of sensor data, use statistical analysis methods to process sensor data over a period of time, obtain the standard deviation and mean of the data, calculate the ratio of the standard deviation to the mean of the data, and obtain the noise index , the expression is as follows:

[0146] ;

[0147] Where, is the mean of the sensor data, is the standard deviation of the sensor data;

[0148] If the noise index exceeds the preset noise threshold, it means that the sensor data noise is too large and interferes with fault diagnosis. The sensor is defined as a sensor to be calibrated and the noise anomaly is recorded. Otherwise, the sensor noise is normal.

[0149] Acquire historical pseudo-anomaly events, extract key features of different pseudo-anomaly patterns, convert them into standard pattern data, and store them in the constructed pseudo-anomaly pattern library. During the storage process, each pseudo-anomaly pattern is classified and managed, assigned a unique identifier, and detailed records of the pattern's characteristic information, applicable sensor type, and possible causes are recorded to provide a reference for subsequent pattern matching;

[0150] Using the DTW (Dynamic Time Warping) algorithm, normal sensor data is matched with patterns in a library of pseudo-anomaly patterns. By constructing a two-dimensional matrix, the distance between the current data and the pattern data in the library is calculated, and the optimal matching path between the two time series is found to minimize the cumulative distance on the matching path. If the matching distance is less than a preset distance threshold, the current data is determined to match the pseudo-anomaly pattern in the library, and the corresponding pseudo-anomaly type is output. Otherwise, the pseudo-anomaly is updated to a true fault and true fault processing is performed. The matching distance is the cumulative distance on the optimal matching path.

[0151] Output fault location report, including sensor status judgment result, false abnormality type (if matched successfully) or true fault handling process start information.

[0152] Furthermore, the fault classification rules are as follows:

[0153] If only one abnormal parameter exists in the abnormal parameter set, it is determined to be a minor fault. This has little impact on the overall performance of the ACM. A manual inspection is scheduled before the aircraft's next flight to continuously monitor parameter changes and ensure that the fault does not worsen. The aircraft's next flight is the next complete flight planned after the current mission.

[0154] A moderate fault is considered a failure if multiple parameters are abnormal or if the abnormal parameters cause a significant degradation in ACM performance. This does not affect flight safety, but the parameters of multiple components deviate from the normal range, affecting the cooling or heating performance of the air conditioning system. The aircraft will be grounded for maintenance within three days, with flight altitude or speed restricted to reduce the potential risk to flight safety, and undergo a comprehensive overhaul.

[0155] Failure of key functions, such as complete failure of the compressor or turbine cooler, is considered a serious fault. At this time, the aircraft's air conditioning system cannot operate normally, threatening flight safety. The aircraft will be immediately grounded to replace parts and prohibited from continuing to fly to ensure flight safety.

[0156] Furthermore, when determining whether the ACM has failed, the fault prediction method includes:

[0157] When the ACM is not faulty, a deep autoencoder model is constructed, consisting of an encoder and a decoder. The encoder uses a multi-layer neural network to map high-dimensional raw data, such as compressor parameters and turbine cooler parameters in structured time series datasets, to a low-dimensional feature space and extract the key features of the data. The decoder reconstructs the low-dimensional features back to the original data dimensions. The number of layers and the number of neurons in each layer need to be adjusted based on the data characteristics and task requirements.

[0158] Based on the structured time series data set, the training set and test set are divided, and the deep autoencoder model is trained using the training set to reconstruct the error With minimization as the goal, the stochastic gradient descent algorithm is used to update the model parameters; the expression is as follows:

[0159] ;

[0160] Where, is the number of samples in the training set, is the original data, To reconstruct data; ≤ ;

[0161] Use the trained deep autoencoder model to reconstruct the test set data and calculate the reconstruction error to evaluate the deep autoencoder model's ability to reconstruct the data and reflect the degree of data abnormality;

[0162] Different components have different importance and degradation rates in ACM. By setting degradation weights and calculating health indicators, the health status of ACM can be assessed more accurately, providing an effective indicator basis for remaining life prediction. Based on historical failure data and expert experience, degradation weights are set for each component to determine the degree of influence of different components on the health status of ACM. The reconstruction errors of each component are weighted and summed according to their degradation weights to construct a health index. , to comprehensively reflect the overall health status of ACM and provide a basis for subsequent remaining life prediction; the expression is as follows:

[0163] ;

[0164] Where, is the total number of parts, For the The degradation weight of each component, the historical failure data includes the time, type and impact range of the failure; For the The reconstruction error of each component;

[0165] Based on the LSTM neural network, an indicator prediction model is constructed. The model structure includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The LSTM layer is used to handle long-term dependencies in time series data, and the fully connected layer is used to map the output of the LSTM layer to the prediction dimension.

[0166] Historical health indicators The data is arranged in chronological order and divided into indicator training set and indicator verification set to determine the length of the input sequence and prediction step length The indicator prediction model is trained using the indicator training set, with the goal of minimizing the mean square error. The Adam optimization algorithm is used to update the model parameters. During the training process, the model performance is evaluated through the validation set, and the model hyperparameters are adjusted to improve the prediction accuracy of the model.

[0167] The most recent indivual Data points are input into the trained indicator prediction model to output the future time steps Predicted value, through Trend prediction to detect potential system failures in advance and provide support for maintenance decisions;

[0168] Use a survival analysis model, such as the Cox proportional hazards model, to estimate the parameters of the survival analysis model using historical failure data. Maximize the likelihood function and use the maximum likelihood estimation method to update the model parameters.

[0169] Output of the indicator prediction model The predicted values and key parameters are input into the trained survival analysis model, which outputs the predicted failure probability of the ACM at different time points. The predicted failure probability at different time points is compared with the preset failure threshold. Once the predicted failure probability at a certain time point is detected to exceed the failure threshold, it is determined that the ACM is about to fail, thereby estimating the remaining life of the ACM.

[0170] Example 2

[0171] See also Figure 6 , another embodiment provided by the present invention: a data monitoring system, comprising: a data reading module, a diagnosis and analysis module, a prediction module and a feedback module;

[0172] The data reading module is used to obtain offline data related to the aircraft air conditioning system (ACM), construct an ACM dataset, preprocess the data in the ACM dataset, set feature engineering processing methods, set dynamic windows, calculate the dynamic correlation coefficients of compressor parameters and turbine cooler parameters for coupled feature extraction, and cluster the data by flight phase to divide the operating conditions and generate a structured time series dataset.

[0173] The diagnostic analysis module is used to obtain baseline data for ACMs on aircraft of the same model, set a baseline fault judgment method, and judge the latest data in the structured time series data set. It combines static thresholds with parameter change rate analysis to screen out abnormal parameters. Based on the actual temperature control effect, it determines whether the ACM is faulty. It also sets a coupled fault location method to diagnose the faulty ACM. It uses Granger causal analysis and decision tree algorithms to quantify the interaction between components, distinguish true faults from false anomalies, and accurately locate the faulty component. It then conducts in-depth analysis of the fault cause and classifies the fault according to its severity. For faults of different levels, it sends corresponding maintenance instructions to notify maintenance personnel to perform repairs.

[0174] The prediction module is used to predict potential faults of fault-free ACMs through health indicator modeling and multi-model fusion, and to determine the remaining life of the faulty component based on the fault prediction results. It obtains a fault-free ACM, sets a fault prediction method, and calculates the ACM's predicted health indicator to determine whether the ACM has failed and calculate the ACM's remaining life.

[0175] The feedback module is used to receive feedback results from staff after performing maintenance operations to determine the accuracy of fault diagnosis and prediction; for fault diagnosis, the actual maintenance results are compared with the diagnosis results to see if they are consistent. If not, the specific difference is identified and targeted updates and modifications are made; for fault prediction, the prediction is compared with the actual fault time or flight. If the predicted fault time or flight is significantly different from the actual situation, the model parameters are updated, the algorithm is improved, or new influencing factor variables are added to improve the accuracy of fault diagnosis and prediction, continuously improve the performance of the monitoring ACM, enhance the adaptability to complex fault modes, ensure the accuracy and robustness of the monitoring system, and form a closed-loop improvement mechanism.

[0176] Furthermore, the baseline fault judgment method includes:

[0177] Through the flight parameter card, at least The complete operational data of a trouble-free flight is collected, and based on the judgment rules of different flight phases, the data of the cruise phase is screened out to generate a benchmark database;

[0178] For the key parameters in the benchmark database, calculate the mean and standard deviation, and use the 3σ principle to calculate the key parameters. Construct parameter benchmark intervals;

[0179] Integrate the mean and benchmark interval of each key parameter to build a standard ACM performance model;

[0180] Read the latest data from the structured time series dataset, input it into the standard ACM performance model, perform preliminary anomaly judgment, and construct abnormal parameter sets and normal parameter sets based on the output results;

[0181] Read the normal parameters in the normal parameter set, and for each normal parameter, select the latest continuous A sliding window is formed. In each sliding window, the derivatives between adjacent data points in the window are calculated in sequence to reflect the rate of change of the parameter and obtain the derivative sequence in the window.

[0182] For each sliding window, analyze the changes in the derivative within the window, calculate the mean and variance of the derivative sequence, and compare the calculated derivative mean and variance with the pre-set normal fluctuation range data;

[0183] If the mean of the derivative sequence is not within the normal mean range, or the variance of the derivative sequence exceeds the normal variance range, then the corresponding normal parameter is marked as having abnormal derivative changes within the window; otherwise, the normal parameter is not abnormal;

[0184] For continuous A sliding window is used to judge if the continuous If there are abnormal changes in derivatives within each sliding window, the corresponding normal parameters are judged to be abnormal, and the corresponding normal parameters are updated to abnormal parameters and migrated to the abnormal parameter set. Otherwise, the corresponding normal parameters are not abnormal.

[0185] Get the actual temperature control effect in the aircraft And the expected temperature control effect , and calculate the temperature difference ;

[0186] Define the key parameters in the abnormal parameter set as abnormal parameters and obtain the number of abnormal parameters , set the temperature difference threshold to , the abnormal threshold is , determine whether the ACM is faulty;

[0187] when or , it is determined that the ACM has a fault; when and , it is determined that there is no fault in the ACM.

[0188] Furthermore, the coupling fault location method includes:

[0189] When an ACM fails, all data segments corresponding to the latest timestamp are extracted from the structured time series dataset based on the timestamp information.

[0190] Using multivariate time series association rule mining methods, such as Granger causality analysis, to build a bivariate regression model, we evaluate the ability of the past value of one variable to predict the current value of another variable. To determine the optimal lag order of the model, we introduce the Akaike Information Criterion (AIC). The AIC criterion can find a balance between model complexity and goodness of fit, ensuring the accuracy of the model. The AIC criterion is also used to determine the lag order. , for different lag orders Calculate the AIC value and select the one that minimizes the AIC value. as the optimal lag order;

[0191] Use the least squares method to estimate the parameters of the bivariate regression model and calculate the residual sum of squares (RSS). By minimizing the residual sum of squares, the bivariate regression model can accurately fit the data and determine the parameters to be estimated.

[0192] To quantify the degree of mutual influence between compressor and turbine parameters, the causal relationship strength is calculated. ;

[0193] Perform Granger causality test to determine whether there is a causal relationship between the parameters of the compressor and turbine cooler components; if , indicating that the lag value of the turbine cooler variable has a significant contribution to the prediction of the current value of the compressor variable, then the turbine cooler variable is determined to be the Granger cause of the compressor variable; if , indicating that the lag value of the compressor variable has a significant contribution to the prediction of the current value of the turbine cooler variable, and the compressor variable is determined to be the Granger cause of the turbine cooler variable;

[0194] Based on the Granger causality verification results, a preliminary judgment is made on the fault type to determine whether there is a Granger causality between the compressor and turbine cooler;

[0195] If there is no causal relationship, indicating that there is no obvious mutual influence between the compressor and turbine cooler parameters, the fault occurs independently or is caused by other non-coupled factors, then the fault type is determined to be a false abnormality and false abnormality processing is performed;

[0196] If there is a causal relationship, determine whether the dynamic correlation coefficient exceeds the preset normal correlation range; if , indicating that the dynamic correlation coefficient exceeds the normal correlation range and is marked as abnormal correlation; if , indicating that the dynamic correlation coefficient is within the normal correlation range, the fault type is determined to be a false abnormality, and false abnormality processing is performed; among them, is the Z-score value of the correlation coefficient within the sliding window, is the normal correlation threshold;

[0197] When the dynamic correlation coefficient is determined to be abnormal, the fault type is judged in combination with the flight phase information. If the aircraft is in the cruise phase, the fault type is determined to be a true fault and true fault processing is performed. By matching historical maintenance cases, the faulty component and cause are determined. If the aircraft is not in the cruise phase, considering the complex operating environment of the aircraft during takeoff and landing and the frequent parameter fluctuations, the abnormality at this time is caused by transient interference, and the fault type is determined to be a false abnormality and false abnormality processing is performed.

[0198] According to the preset fault classification rules, the fault handling results are classified and a maintenance report is generated, including the fault location report, level, and type; among them, fault handling includes true fault handling and false abnormality handling, and the output of fault handling is the fault location report.

[0199] Furthermore, the fault prediction method includes:

[0200] When there is no fault in ACM, a deep autoencoder model is constructed;

[0201] Based on the structured time series data set, the training set and test set are divided, and the deep autoencoder model is trained using the training set to reconstruct the error With minimization as the goal, the stochastic gradient descent algorithm is used to update the model parameters;

[0202] Use the trained deep autoencoder model to reconstruct the test set data and calculate the reconstruction error to evaluate the deep autoencoder model's ability to reconstruct the data and reflect the degree of data abnormality;

[0203] Based on historical failure data and expert experience, a degradation weight is set for each component to determine the degree of influence of different components on the health status of the ACM. The reconstruction error of each component is weighted and summed according to its degradation weight to construct a health index. , to comprehensively reflect the overall health status of ACM and provide a basis for subsequent remaining life prediction;

[0204] Based on the LSTM neural network, an indicator prediction model is constructed. The model structure includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The LSTM layer is used to handle long-term dependencies in time series data, and the fully connected layer is used to map the output of the LSTM layer to the prediction dimension.

[0205] Historical health indicators The data is arranged in chronological order and divided into indicator training set and indicator verification set to determine the length of the input sequence and prediction step length The indicator prediction model is trained using the indicator training set, with the goal of minimizing the mean square error. The Adam optimization algorithm is used to update the model parameters. During the training process, the model performance is evaluated through the validation set, and the model hyperparameters are adjusted to improve the prediction accuracy of the model.

[0206] The most recent indivual Data points are input into the trained indicator prediction model to output the future time steps Predicted value, through Trend prediction to detect potential system failures in advance and provide support for maintenance decisions;

[0207] Use a survival analysis model, such as the Cox proportional hazards model, to estimate the parameters of the survival analysis model using historical failure data. Maximize the likelihood function and use the maximum likelihood estimation method to update the model parameters.

[0208] Output of the indicator prediction model The predicted values and key parameters are input into the trained survival analysis model, which outputs the predicted failure probability of the ACM at different time points. The predicted failure probability at different time points is compared with the preset failure threshold. Once the predicted failure probability at a certain time point is detected to exceed the failure threshold, it is determined that the ACM is about to fail, thereby estimating the remaining life of the ACM.

[0209] In summary, the present invention connects to flight parameter cards to extract offline data, flight parameters, and maintenance records, combines them with environmental data to generate an ACM dataset, uses dynamic windows to extract dynamic correlation coefficients between compressor and turbine parameters, clusters and normalizes them by flight phase, and generates structured time series data. A benchmark database is constructed, and abnormal parameters are screened through static threshold detection and dynamic derivative analysis. Faults are verified and identified using temperature adjustment differences. Coupled faults are located using Granger causality analysis and decision trees, distinguishing true faults from false anomalies. True faults are matched with historical cases, while false anomalies are checked for sensor noise. A deep autoencoder is constructed to calculate reconstruction error, and health indicators are generated using degradation weights. An LSTM is used to predict future trends, and a survival analysis model is used to estimate remaining life. Diagnostic and prediction accuracy is verified based on maintenance results, and model parameters or algorithms are updated to continuously improve system performance.

[0210] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. ACM data analysis method, characterized in that include: An ACM dataset was constructed, and a feature engineering method was set. Dynamic windows were set, dynamic correlation coefficients were calculated, and coupled features were extracted. The operating conditions were divided to generate a structured time series dataset. An ACM is an air cycle machine, consisting of a compressor and a turbine cooler. The ACM dataset includes offline data related to the ACM, as well as flight parameter data, ACM maintenance records, and flight environmental data. Obtain ACM benchmark data for aircraft of the same model, set a benchmark fault judgment method, judge the latest data in the structured time series data set, filter abnormal parameters, and determine whether the ACM is faulty based on the actual temperature control effect. Set a coupled fault location method to diagnose the faulty ACM and distinguish between true faults and false anomalies. At the same time, set a predictive scheduling method to predict faults for fault-free ACMs, perform health status prediction and lifespan assessment; When generating a structured time series dataset, the feature engineering processing method includes: Read the timestamps of the ACM data set, calculate the differences between adjacent timestamps, and define the mode of the differences as the time resolution; Conduct response tests on the ACM to simulate operation under different operating conditions, monitor the fluctuation characteristics of key parameters, and determine the time range when the key parameters exceed the preset variation range; Dynamically adjust the window size and sliding step size based on the preset scale factor; Sliding the window along the time series, performing Pearson correlation analysis on the compressor parameters and the turbine cooler parameters within each window to generate a dynamic correlation coefficient; Associating the dynamic correlation coefficient with the start timestamp of the corresponding window to form a coupled characteristic time series; When diagnosing a faulty ACM, the coupling fault location method includes: In the structured time series data set, extracting the data segment corresponding to the latest timestamp based on the timestamp information; Granger causality analysis was used to construct a bivariate regression model based on compressor parameters and turbine cooler parameters to evaluate the lag prediction ability between variables. Using the AIC criterion, calculate different lag orders AIC value, select the one that minimizes the AIC value as the optimal lag order; estimating the parameters of the bivariate regression model based on the least squares method, calculating the residual sum of squares and minimizing the fitted data, and calculating the causal relationship strength between the compressor and turbine cooler parameters; Granger causality test is used to determine whether there is a causal relationship between the parameters; if , then it is determined that the turbine cooler parameters are the Granger causes of the compressor parameters; if , then it is determined that the compressor parameters are the Granger causes of the turbine cooler parameters; among them, 、 are the parameters of the bivariate regression model.

2. The ACM data analysis method according to claim 1, characterized in that: The feature engineering processing method further includes: Divide the flight parameters into climb, cruise, and descent phases, traverse the ACM dataset, and mark the flight phase for the data at each time point based on the preset flight phase judgment rules; Based on the flight phase identification results and dynamic correlation coefficients, the ACM dataset and the data in the coupled feature time series are saved into three constructed subsets: climb, cruise, and descent. Standardize the data of each subset; Using timestamps as row indexes, key parameters and dynamic correlation coefficients are arranged into a two-dimensional matrix to generate a structured time series dataset.

3. The ACM data analysis method according to claim 2, characterized in that: When determining whether the ACM is faulty, the baseline fault determination method includes: Obtaining trouble-free operation data of the same model aircraft at the initial stage of delivery, filtering cruise phase data based on the flight phase judgment rules, and generating a benchmark database; Calculate the mean and standard deviation of key parameters in the benchmark database and construct the parameter benchmark interval using the 3σ principle; Integrate the mean value and parameter benchmark interval of each key parameter to build a standard ACM performance model; The latest data in the structured time series data set is read, input into the standard ACM performance model, compared with the parameter benchmark interval, abnormal parameters are identified, and an abnormal parameter set and a normal parameter set are constructed.

4. The ACM data analysis method according to claim 3, characterized in that: The benchmark fault judgment method further includes: Reading normal parameters in the normal parameter set, sequentially calculating derivatives between adjacent data points in a continuous sliding window, and obtaining a derivative sequence in the window; For any single sliding window, the mean and variance of the derivative sequence are calculated to perform dynamic trend detection; If the mean of the derivative sequence is not within the preset normal mean range or the variance exceeds the preset normal variance range, it is marked that the normal parameter has an abnormal derivative change within the sliding window; otherwise, the normal parameter does not have an abnormal derivative change; Perform abnormality judgment on the continuous sliding window; if there is abnormal derivative change in all the continuous sliding windows, update the normal parameters to abnormal parameters and migrate them to the abnormal parameter set; otherwise, the normal parameters do not have abnormalities; The actual temperature control effect in the computer cabin and the temperature control difference are calculated , comprehensively judge whether there is a fault in ACM; when or , it is determined that the ACM has a fault; when and , it is determined that there is no fault in ACM; among them, is the number of exception parameters, is the temperature difference threshold, is the abnormal threshold.

5. The ACM data analysis method according to claim 4, characterized in that: The coupling fault locating method further includes: Based on the result of the Granger causality test, determining whether there is a causal relationship between the compressor and turbine cooler parameters; If there is no causal relationship, the fault type is determined to be a false abnormality and false abnormality processing is performed; If there is a causal relationship, determine whether the dynamic correlation coefficient exceeds the preset normal correlation range; like , determine the fault type as a pseudo-anomaly, and perform pseudo-anomaly processing; where, is the Z-score value of the correlation coefficient within the sliding window, is the normal correlation threshold; like , combined with the flight phase information, the fault type is judged; if the aircraft is in the cruise phase, the fault type is determined to be a true fault and true fault processing is performed; otherwise, the fault type is determined to be a false anomaly and false anomaly processing is performed; Generate a maintenance report based on the fault type.

6. The ACM data analysis method according to claim 5, characterized in that: The true fault processing includes: Based on the abnormal parameter set, obtaining components corresponding to the abnormal parameters in the abnormal parameter set and determining faulty component candidates; Calling a historical maintenance case library to screen historical cases related to the candidate faulty component; Calculate the cosine similarity between the current fault data and each historical case data; Sort historical cases based on cosine similarity and output the top three fault causes with the highest similarity and their confidence levels; Generates a fault location report that includes the faulty component, cause, and confidence level, where the confidence level is the cosine similarity value.

7. The ACM data analysis method according to claim 6, characterized in that: The pseudo exception processing includes: Extracting sensor data from the structured time series data set, checking the validity and continuity of the data point by point, and calculating a continuity index; Comparing the continuity index with a preset continuity threshold, if the continuity index is lower than the continuity threshold, marking the corresponding sensor as a suspected fault; otherwise, the sensor has no abnormality; Performing noise level analysis on sensor data, comparing the calculated noise index with a preset noise threshold, and marking the sensor as pending calibration if the noise index exceeds the noise threshold; otherwise, the sensor is normal; Build a pseudo-abnormal pattern library to store the characteristic data of historical pseudo-abnormal events and manage them by category; Using a dynamic time warping algorithm to match the sensor data without anomalies with the patterns in the pseudo-anomaly pattern library, and calculating the matching distance; If the matching distance is less than the preset distance threshold, the corresponding pseudo-anomaly type is output; Otherwise, update the false exception to a true fault and perform true fault processing; Generates a fault location report that includes sensor status and false anomaly type.

8. The ACM data analysis method according to claim 7, characterized in that: The method further includes determining whether the ACM has failed using a fault prediction method; The fault prediction method comprises: Constructing a deep autoencoder model, training and evaluating the deep autoencoder model using the structured time series dataset, and calculating a reconstruction error based on an output of the deep autoencoder model; Based on historical fault data, a degradation weight is set for each component, and the reconstruction error of each component is weighted and summed according to its degradation weight to construct a health index; Construct an indicator prediction model, train the indicator prediction model with historical health indicator data, and predict the health indicator trend in the future time step. The length of the input sequence is , the prediction step length is ; The most recent The health indicator data points are input into the indicator prediction model, and the future The predicted value of health indicators at each time step; Building a survival analysis model based on historical failure data, inputting the predicted health indicator values and key parameters into the survival analysis model, and calculating the predicted failure probability at different time points; The predicted failure probability at different time points is compared with a preset failure threshold. Once it is detected that the predicted failure probability at a certain time point exceeds the failure threshold, it is determined that the ACM is about to fail and the remaining life of the ACM is output.

9. A data monitoring system for implementing the ACM data analysis method according to any one of claims 1 to 8, characterized in that: include: Data reading module, diagnosis and analysis module, prediction module and feedback module; The data reading module is used to construct an ACM data set, set a dynamic window, extract coupling features based on dynamic correlation coefficients, divide working conditions, and generate a structured time series data set; The diagnostic analysis module is used to combine static thresholds and parameter change rate analysis to screen out abnormal parameters, and based on the actual temperature control effect, determine whether the ACM is faulty, and diagnose the faulty ACM to distinguish true faults from false abnormalities; The prediction module is used to predict potential faults of the fault-free ACM, and based on the fault prediction results, determine whether the ACM has failed and calculate the remaining life of the ACM; The feedback module is used to receive feedback results from the staff after performing maintenance operations to determine the accuracy of fault diagnosis and prediction and perform update operations.

Images