Aero-engine batch component anomaly detection method based on multi-distribution probability fusion
By fitting multiple data distribution functions using a multi-distribution probability fusion method, the problems of insufficient fitting and interpretability in the batch component inspection of aero-engines are solved, and higher accuracy and reliability of anomaly detection are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH AT WEIHAI
- Filing Date
- 2024-01-09
- Publication Date
- 2026-07-21
AI Technical Summary
In the detection of anomalies in batch components of aero-engines, existing technologies suffer from insufficient fitting ability of mathematical and statistical methods and a lack of interpretability of machine learning methods, leading to missed detections, false detections, and difficulties in judgment.
Mathematical statistics methods are used to fit probability density functions of various distribution types to batch data. The BA value is used as a voting weight for anomaly detection. The final anomaly detection result is obtained by selecting the distribution type with the highest BA value and performing weighted voting.
It improves the accuracy of anomaly detection, reduces the probability of unknown faults, provides a theoretical explanation, avoids the bias of fitting a single distribution, and enhances the reliability of detection.
Smart Images
Figure CN117909899B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of aero-engine testing technology, and more specifically, it relates to a method for detecting anomalies in batch aero-engine components based on multi-distribution probability fusion. Background Technology
[0002] As the heart of an aircraft, the safety and reliability of aero engines are of paramount importance. Under conditions of high temperature, high pressure, and high-speed rotation, potentially defective components in a batch are more prone to failure, posing serious threats to flight safety. Therefore, conducting inspections for potentially defective components in a batch is an effective way to prevent flight accidents at their source.
[0003] To date, methods for detecting anomalies in high-end equipment can be broadly categorized into two types: methods based on mathematical statistics and methods based on machine learning.
[0004] Statistical methods typically employ probability density estimation based on a known distribution function, classifying samples with low probability densities as anomalies. For instance, based on hypothesis testing concepts, a dynamic threshold boundary that changes with operational status is constructed by fitting a probability density function, thereby enabling anomaly detection. In engineering applications, a normal distribution function is used to fit the wear data of bearings after stable operation, and the 3σ criterion is used to set an alarm threshold for bearing wear, achieving anomaly detection during bearing operation. However, research has revealed that due to supplier factors, the characteristics of mass-produced aero-engine components often do not conform to common distribution functions. When faced with real-world aero-engine data with complex distributions, statistical methods, such as traditional distribution function fitting, suffer from insufficient fitting capability. Using a specific traditional distribution function for fitting in such cases will result in biased boundary thresholds, leading to more missed and incorrect detections.
[0005] Machine learning-based methods attempt to train a classifier between normal and abnormal samples using data with known labels. For example, to address the problem of imbalanced data classes, an autoencoder combining a Long Short-Term Memory network and a convolutional neural network was built for feature extraction, enabling anomaly detection in coal mining machines with imbalanced samples. To address the difficulties in extracting acoustic features and identifying temporal anomalies in mechanical equipment under complex industrial environments, a bidirectional gated recurrent unit was added to the autoencoder to form a temporal feature extraction channel, and a self-supervised classifier was designed to achieve sound anomaly detection. However, machine learning-based methods face the challenge of lacking interpretability; for example, when determining a component as a potential abnormal component, sufficient theoretical explanation cannot be provided. Summary of the Invention
[0006] To achieve the above objectives, the technical solution adopted in this application is as follows: a method for detecting anomalies in batch components of aero-engines based on multi-distribution probability fusion is provided. Mathematical statistics methods are used to fit probability density functions of multiple distribution types to batch data, and BA values are used as voting weights to detect anomalies in batch data and determine whether the data is abnormal.
[0007] Optionally, the method for detecting anomalies in batch components of aero-engines based on multi-distribution probability fusion is characterized by: using mathematical statistics to fit probability density functions of multiple distribution types to batch data, using BA values as voting weights to detect anomalies in batch data, and determining whether the data is abnormal.
[0008] Optionally, the following steps are included: Step 1: Fit the probability density function of the training data to various distribution types to obtain the probability density function of each distribution. Step 2: For each probability density function distribution, anomaly detection is performed, and the confusion matrix and BA value are obtained; Step 3: Use a filter to select the three distribution types with the highest BA values; Step 4: Based on the anomaly detection boundaries of the three selected distribution types, perform preliminary anomaly detection on the test samples to obtain three preliminary anomaly detection results; Step 5: Sort the distribution types in descending order of BA value, take the BA values of the top three distribution types as voting weights, and perform weighted voting on the preliminary anomaly detection results of the test samples to obtain the final anomaly detection results of the test sample set. Step 6: Calculate the BA value of the final anomaly detection result.
[0009] Optionally, in step two, anomaly detection is performed for the distribution of each probability density function, as follows: a) Combining the cumulative probability distribution image, the anomaly detection boundary of the data distribution is obtained through the single distribution probability density function, including the upper boundary and the lower boundary; b) Using the above-mentioned anomaly detection boundary, each training sample is checked for anomalies; c) Obtain the confusion matrix and BA value by comparing the training sample labels and anomaly detection results.
[0010] Optionally, the BA value can be used as the evaluation metric for the anomaly detection model, representing the proportion of data detected as anomalous out of the actual anomalous data and the proportion of data detected as normal out of the actual normal data, as shown in the following formula: (6) TP indicates that the actual label is normal and the detection result is normal; FP indicates that the actual label is abnormal and the detection result is normal; FN indicates that the actual label is normal and the detection result is abnormal; TN indicates that the actual label is abnormal and the detection result is abnormal.
[0011] Optionally, the probability density function for various distribution types includes the normal distribution, whose probability density function is shown below: (1) in, For random variables The mean, i.e. The average of all possible values; for The variance reflects the degree to which the sample data deviates from the mean.
[0012] Optionally, the probability density function for various distribution types includes the Weibull distribution, whose probability density function is shown below: (2) in, It is a random variable. It is a proportional parameter. It is a shape parameter; when When the Weibull distribution is exponential, then the Weibull distribution follows an exponential distribution; if Then it is a minimized Weibull distribution.
[0013] Optionally, the probability density function for various distribution types includes the gamma distribution, whose probability density function is shown below: (3) In the formula, This is called the shape parameter, which determines the shape of the distribution curve; This is called the inverse scaling parameter, which determines the steepness of the curve. The functional expression for the gamma distribution is given by the formula: .
[0014] Optionally, the probability density function for various distribution types includes the Laplace distribution, whose probability density function is shown below: (4) In the formula, For position parameters; It is a scale parameter, requiring ; It is a natural number.
[0015] Optionally, the probability density function for various distribution types includes the Johnson-SU distribution, whose probability density function is shown below: (5) In the formula, The location parameter determines the center location of the distribution; The scale parameter determines the size of the distribution; and The shape parameter determines the shape of the distribution.
[0016] Optionally, the probability density functions for various distribution types include the uniform distribution and the Johnson-SB distribution.
[0017] This application proposes an anomaly detection method for batch aviation components based on multi-distribution probability fusion. By fitting multiple data distribution functions, the upper and lower boundaries of the data distribution anomalies are obtained, enabling effective monitoring of batch components with anomalies, thereby reducing the probability of unknown failures. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart of the batch component anomaly detection method for aero-engines based on multi-distribution probability fusion in this application; Figure 2 This is a schematic diagram of a normal distribution; Figure 3 A schematic diagram of the Weibull distribution; Figure 4 A schematic diagram of gamma distribution; Figure 5 A schematic diagram of the Laplace distribution; Figure 6 A schematic diagram of the Johnson-SU distribution; Figure 7 A histogram of the probability distribution of the simulated data; Figure 8 The probability density plot fitted to each distribution function; Figure 9 This is a schematic diagram of the decision results of multi-distribution probability fusion decision. Detailed Implementation
[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.
[0021] The present application provides a method for detecting batch anomalies in aero-engine components based on multi-distribution probability fusion. See also... Figure 1 The method for detecting anomalies in batch components of aero-engines based on multi-distribution probability fusion uses mathematical statistics to fit multiple distributions to batch data, and uses the BA value as the voting weight to detect anomalies in the batch data and determine whether the data is abnormal.
[0022] Includes the following steps: Step 1: Fit the probability density function of the training data to various distribution types to obtain the probability density function of each distribution.
[0023] The probability density functions for various distribution types include the normal distribution, Weibull distribution, gamma distribution, Laplace distribution, uniform distribution, Johnson-SU distribution, and Johnson-SB distribution.
[0024] The Weibull distribution includes the Weibull_min distribution and the Weibull_max distribution.
[0025] The gamma distribution includes the exponential distribution.
[0026] Step two, for each probability density function distribution, perform the following operations: a) Combining the cumulative probability distribution image, the anomaly detection boundary of the data distribution is obtained through the single distribution probability density function, including the upper boundary and the lower boundary; b) Using the above-mentioned anomaly detection boundary, each training sample is checked for anomalies; c) Obtain the confusion matrix and BA value by comparing the training sample labels and anomaly detection results; Step 3: Use a filter to select the three distribution types with the highest BA values; Step 4: Based on the anomaly detection boundaries of the three selected distribution types, perform preliminary anomaly detection on the test samples to obtain three preliminary anomaly detection results; Step 5: Sort the distribution types in descending order of BA value, take the BA values of the top three distribution types as voting weights, and perform weighted voting on the preliminary anomaly detection results of the test samples to obtain the final anomaly detection results of the test sample set. Step 6: Calculate the BA value of the final anomaly detection result.
[0027] In step one, the distribution is normal, such as Figure 2 As shown, the probability density function of the normal distribution is as follows: (1) in, For random variables The mean, i.e. The average of all possible values; for The variance is used to reflect the degree to which sample data deviates from the mean.
[0028] The method for detecting anomalies in batch components of aero-engines is based on normal distribution fitting, and through... Criteria for identifying outlier data. When training data follows a normal distribution, it exhibits a distribution pattern of high values in the middle and low values at both ends, with most data concentrated around the mean. Nearby, only a small amount of data is distributed in Outside of the area. Set in Data outside of these categories is considered abnormal and will be subject to close monitoring.
[0029] Weibull distribution, see Figure 3 The probability density function of the Weibull distribution is shown below: (2) in, It is a random variable. It is a proportional parameter. It is a shape parameter. When When the Weibull distribution is exponential, then the Weibull distribution follows an exponential distribution; if The Weibull distribution is minimized.
[0030] The Weibull distribution can more flexibly fit data with many complex distribution forms. The cumulative distribution function image is used to identify regions with a cumulative probability of 2% on both sides of the data distribution as outlier regions. Assuming the data distribution conforms to the Weibull distribution, most data is concentrated within a certain area of the data center, with only a small amount of data distributed at the extremes. Therefore, data in the extreme regions on both sides of the data distribution can be designated as outliers and given priority monitoring.
[0031] Gamma distribution, see Figure 4 The exponential distribution, commonly used in mathematical statistics, is a special case of the gamma distribution.
[0032] The probability density function of the gamma distribution is shown below: (3) In the formula, This is called the shape parameter, which determines the shape of the distribution curve; It is called the inverse scaling parameter and determines the steepness of the curve; The functional expression for the gamma distribution is given by the formula: .
[0033] The gamma distribution uses the cumulative distribution function graph to identify regions with a cumulative probability of 2% on both sides of the data distribution as outlier regions. Assuming the data distribution conforms to a gamma distribution, most data is concentrated within a certain area of the data center, with only a small amount of data distributed at the extremes. Therefore, data in the extreme regions on both sides of the data distribution can be designated as outlier data and given priority monitoring.
[0034] Laplace distribution, see Figure 4 The Laplace distribution, also known as the double exponential distribution, has the following probability density function: (4) In the formula, For position parameters; It is a scale parameter, requiring ; It is a natural number.
[0035] The Laplace distribution uses the cumulative distribution function graph to identify regions with a cumulative probability of 2% on both sides of the data distribution as outlier regions. Assuming the data distribution conforms to a Laplace distribution, most data is concentrated within a certain area of the data center, with only a small amount of data distributed at the extremes. Therefore, data in the extreme regions on both sides of the data distribution can be designated as outlier data and given priority monitoring.
[0036] Johnson-SU distribution, see Figure 5 The probability density function of the Johnson-SU distribution is shown below: (5) In the formula, The location parameter determines the center location of the distribution; The scale parameter determines the size of the distribution; and The shape parameter determines the shape of the distribution.
[0037] The Johnson-SU distribution uses the cumulative distribution function graph to identify regions with a cumulative probability of 2% on both sides of the data distribution as outlier regions. Assuming the data distribution conforms to the Johnson-SU distribution, most data is concentrated within a certain area of the data center, with only a small amount of data distributed at the extremes. Therefore, data in the extreme regions on both sides of the data distribution can be designated as outliers and given priority monitoring.
[0038] In step two, the anomaly detection of batch components of the aero-engine is a binary classification problem. Therefore, the detection results are mixed pairwise with all the results of the true labels to obtain four cases: normal components are judged as normal, normal components are judged as abnormal, abnormal components are judged as normal, and abnormal components are judged as abnormal. These four cases constitute the confusion matrix shown in Table 1.
[0039] Table 1. Schematic diagram of the confusion matrix
[0040] In Table 1, TP indicates that the actual label is normal and the detection result is normal; FP indicates that the actual label is abnormal and the detection result is normal; FN indicates that the actual label is normal and the detection result is abnormal; TN indicates that the actual label is abnormal and the detection result is abnormal.
[0041] In the anomaly detection of batch components for aero-engines, normal data constitutes the vast majority of the total data, while abnormal data is relatively scarce. Therefore, the BA value is used as the evaluation index for the anomaly detection model. The proportion of data detected as abnormal out of the actual abnormal data and the proportion of data detected as normal out of the actual normal data are shown in the following formulas: (6) This application uses simulated data for verification experiments. The data generation rules are as follows: Sub-datasets conforming to normal, exponential, uniform, and gamma distributions are randomly generated. These sub-datasets are then merged to obtain a simulated dataset, simulating complex data where the distribution type is difficult to determine in actual production processes. Figure 7 The figure shows the probability density distribution histogram of the simulated data. To determine the effectiveness of the method, the simulated data was labeled. The process is as follows: the top 1% and bottom 1% of the data were considered outliers and labeled -1; the middle 98% of the data were considered normal data and labeled 1.
[0042] Fitting algorithms for normal distribution, uniform distribution, exponential distribution, Laplace distribution, gamma distribution, Weibull_min distribution, Weibull_max distribution, JohnsonSB distribution, and JohnsonSU distribution are used as individual learners to obtain the probability density function curves of the data distribution and the data outlier boundaries.
[0043] The probability density of each distribution function fit is as follows Figure 8 As shown, through Figure 8 It is evident that most distribution types generally perform poorly on this data; among them, the exponential distribution performs particularly poorly. When using a single distribution with a very poor fit for anomaly detection, the upper and lower boundaries will deviate significantly from the true anomaly threshold, resulting in numerous false positives and false negatives.
[0044] After obtaining the probability density histogram of the data distribution and the data anomaly boundaries, the BA value of each distribution function is calculated. Then, the three highest-performing distribution functions are selected through a filter to participate in a voting process. The BA value is used as the voting weight for the distribution function fit to reduce the impact of poorly performing distribution functions on the fusion anomaly judgment.
[0045] See Figure 9 Using normal, exponential, and Laplace distributions as examples, this paper demonstrates the process of multi-distribution probability fusion decision-making. For each sample point in the sample sequence, different distribution types yield different anomaly detection results. The multi-distribution probability fusion decision-making method weights and fuses the anomaly detection results of the three distributions with the highest BA values to obtain a fused anomaly detection result.
[0046] The BA value obtained by the single-distribution fitting method is compared with that obtained by the multi-distribution probability fusion method to verify the effectiveness of the method. The comparative experiment was conducted by taking the mean of multiple experiments, and Table 2 shows the results of multiple experiments.
[0047] The experimental results in Table 2 clearly show that, in multiple experiments, the BA value of multi-distribution probability fusion is larger than that of the anomaly detection method based on single-distribution data fitting. Therefore, multi-distribution probability fusion can achieve higher TP and TN values, meaning it can classify more normal data as normal and more abnormal data as abnormal. Compared to single-distribution fitting methods, the anomaly detection method using multi-distribution probability fusion has a higher accuracy in identifying positive and negative samples; the decision-level probability fusion method has better anomaly detection performance for class imbalance problems.
[0048] Table 2. BA values (%) from multiple experiments
[0049] Experiments show that the average BA value of anomaly detection results based on decision-level probabilistic fusion is 2.18%, 1.57%, and 1.03% higher than that of the normal distribution, Weibull_min distribution, and Johnson SU distribution, respectively.
[0050] In this application, training data refers to the training dataset, and training samples are sample points within the dataset. Batch data is the total data to be used for anomaly detection, while training data is the data extracted from the batch data used to train the algorithm.
[0051] This application addresses the challenge of complex data distributions in batches of aero-engine components, making it difficult to achieve anomaly detection through single-distribution fitting. It proposes a multi-distribution probability fusion-based anomaly detection method for aero-engine batch components. This method obtains the upper and lower boundaries of anomalies in the data distribution by fitting multiple data distribution functions, enabling effective monitoring of anomalous batch components and reducing the probability of unknown faults. Compared to general neural network anomaly detection methods, this method is supported by mathematical principles, effectively avoiding the black-box problem of neural networks and providing theoretical support for anomaly data identification. Furthermore, the method integrates multiple data distribution types, effectively ensuring anomaly detection capabilities even with complex data distributions.
[0052] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for detecting anomalies in batch components of aero-engines based on multi-distribution probability fusion, characterized in that: Mathematical statistics methods are used to fit probability density functions of various distribution types to batch data, and BA values are used as voting weights to detect anomalies in batch data and determine whether the data is abnormal. Step 1: Fit the probability density function of the training data to various distribution types to obtain the probability density function of each distribution. Step 2: For each probability density function distribution, anomaly detection is performed, and the confusion matrix and BA value are obtained; Step 3: Use a filter to select the three distribution types with the highest BA values; Step 4: Based on the anomaly detection boundaries of the three selected distribution types, perform preliminary anomaly detection on the test samples to obtain three preliminary anomaly detection results; Step 5: Sort the distribution types in descending order of BA value, take the BA values of the top three distribution types as voting weights, and perform weighted voting on the preliminary anomaly detection results of the test samples to obtain the final anomaly detection results of the test sample set. Step 6: Calculate the BA value of the final anomaly detection result.
2. The method for detecting batch anomalies in aero-engine components based on multi-distribution probability fusion as described in claim 1, characterized in that: Step two: For each probability density function distribution, anomaly detection is performed as follows: a) Combining the cumulative probability distribution image, the anomaly detection boundary of the data distribution is obtained through the single distribution probability density function, including the upper boundary and the lower boundary; b) Using the above-mentioned anomaly detection boundary, each training sample is checked for anomalies; c) Obtain the confusion matrix and BA value by comparing the training sample labels and anomaly detection results.
3. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The BA value is used as the evaluation metric for the anomaly detection model. The proportion of data detected as anomalous out of the actual anomalous data, and the proportion of data detected as normal out of the actual normal data, are shown in the following formula: （6） TP indicates that the actual label is normal and the detection result is normal; FP indicates that the actual label is abnormal and the detection result is normal; FN indicates that the actual label is normal and the detection result is abnormal; TN indicates that the actual label is abnormal and the detection result is abnormal.
4. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density function for various distribution types includes the normal distribution, whose probability density function is shown below: (1) in, For random variables The mean, i.e. The average of all possible values; for The variance reflects the degree to which the sample data deviates from the mean.
5. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density function for various distribution types includes the Weibull distribution, whose probability density function is shown below: （2） in, It is a random variable. It is a proportional parameter. It is a shape parameter; when When the Weibull distribution is exponential, then the Weibull distribution follows an exponential distribution; if Then it is a minimized Weibull distribution.
6. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density function for various distribution types includes the gamma distribution, whose probability density function is shown below: （3） In the formula, This is called the shape parameter, which determines the shape of the distribution curve; This is called the inverse scaling parameter, which determines the steepness of the curve. The functional expression for the gamma distribution is given by the formula: .
7. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density function for various distribution types includes the Laplace distribution, whose probability density function is shown below: （4） In the formula, For position parameters; It is a scale parameter, requiring ; It is a natural number.
8. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density functions for various distribution types include the Johnson-SU distribution, whose probability density function is shown below: （5） In the formula, The location parameter determines the center location of the distribution; The scale parameter determines the size of the distribution; and The shape parameter determines the shape of the distribution.
9. The method for detecting batch anomalies of aero-engine components based on multi-distribution probability fusion as described in any one of claims 1 to 2, characterized in that: The probability density functions for various distribution types include the uniform distribution and the Johnson-SB distribution.