An anomaly detection benchmark model adaptive updating method based on similarity-enhanced elastic weight consolidation
By employing a similarity-enhanced elastic weight integration method in aero-engines, combined with spatial DTW and similarity-enhanced EWC, online adaptive updating of the anomaly detection benchmark model was achieved. This solved the problem of model accuracy degradation under dynamic operating conditions, and improved detection accuracy and resource utilization efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-06-23
- Publication Date
- 2026-06-23
AI Technical Summary
Under dynamic operating conditions, the accuracy of the aero-engine anomaly detection benchmark model decreases due to changes in data distribution. Furthermore, with limited airborne resources, existing methods fail to effectively adapt and update, leading to reduced detection accuracy and model overfitting.
An anomaly detection method based on similarity-enhanced elastic weight integration (EWC) is adopted. By using spatial DTW multivariate similarity measurement and similarity-enhanced EWC model update module, the balance between new and old tasks is dynamically adjusted to achieve online adaptive updating of the model.
It improved the model's detection accuracy and generalization performance under dynamic conditions, reduced the false alarm rate, and significantly improved the Recall, Precision, and F1 scores.
Smart Images
Figure CN120724245B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aircraft condition monitoring technology, and is an adaptive update method for anomaly detection benchmark model based on similarity-enhanced elastic weight integration. Background Technology
[0002] Aero engines are extremely complex, multi-system coupled components that operate under harsh conditions such as high temperature, high pressure, and high speed. Therefore, even minor anomalies or malfunctions can develop into potential threats, leading to major aviation disasters and economic losses. Anomaly detection methods based on regression prediction can effectively assess the degree to which actual observed values of key engine parameters deviate from expectations, thereby identifying anomalies. However, under dynamic operating conditions, the distribution of flight parameter data changes dynamically with different environments, flight phases, and engine operating conditions. This change in data distribution leads to a mismatch in the anomaly detection baseline model, reducing detection accuracy. To address this issue, existing anomaly detection methods typically assume a known data distribution and update the model to adapt to new data. However, the data distribution of online flight parameter data is difficult to obtain during actual operation, resulting in the problem of unnoticed anomaly detection baseline model mismatch. Furthermore, the computing power and storage space of airborne computing platforms are limited, making it difficult to use large-scale data for online model updates. Therefore, the rational selection of update data and update timing is crucial for improving model performance. However, current research on online model update methods lacks effective quantitative analysis of adaptive boundary conditions, especially under dynamic operating conditions, where different types of new operating condition data have different levels of importance. Existing methods often fail to consider these differences, treating all new data as equally important and adopting a uniform update strategy. This approach can easily lead to model overfitting, reduced generalization performance, and even an inability to effectively adapt to critical operating conditions.
[0003] To address this, an adaptive update method for anomaly detection benchmark models based on Elastic Weight Consolidation (EWC) with enhanced similarity is proposed. This method comprises two parts: a multivariate similarity measurement module based on Dynamic Time Warping (DTW) and a model update module based on EWC. The DTW-based multivariate similarity measurement module combines spatial information extracted by Maximum Mean Discrepancy (MMD) with temporal information extracted by traditional DTW, accurately detecting distribution changes in multidimensional flight parameter data while improving computational efficiency compared to traditional DTW. When new operational data is detected, the EWC-based model update module dynamically adjusts the balance between old and new tasks based on the importance of the data, improving the model's online adaptability. Summary of the Invention
[0004] In dynamic operating conditions involving flight missions and changing environments, new operating condition data often emerges that differ from offline data, leading to a decrease in the accuracy of the anomaly detection benchmark model and causing model mismatch issues. To address the problem of unknowingly mismatched anomaly detection benchmark models under actual dynamic operating conditions, and the limited airborne resources preventing large-scale data updates, this invention proposes an adaptive update method for the anomaly detection benchmark model based on similarity-enhanced elastic weight integration, enabling online adaptive updates of the benchmark model.
[0005] This invention provides the following technical solutions:
[0006] An adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration, the method comprising the following steps:
[0007] Step 1: Deploy two identical models These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates;
[0008] Step 2: Cache the online flight parameter data, use spatial DTW to implement multivariate similarity measurement, and monitor changes in data distribution;
[0009] Step 3: When the similarity index exceeds the set threshold, it is determined that the current data distribution has changed significantly, the model is no longer suitable for the current data, and a model update based on similarity enhancement EWC is required.
[0010] Step 4: After the model is adaptively updated, store the new operating condition data in the flight parameter data standard library.
[0011] Preferably, step 2 specifically comprises:
[0012] Once the cached data reaches a predetermined length, a similarity measurement is performed to monitor changes in data distribution. The cache length is 1000 data points. During the similarity measurement, a similarity index needs to be calculated between the online flight parameter data and multiple historical flight parameter data sets to reflect differences under different operating conditions. The flight parameter data is divided into non-overlapping segments, and the length of each segment is set to [missing information]. Then online flight parameter data and historical flight data They were each divided into multiple non-overlapping subsequences. and :
[0013] (1)
[0014] (2)
[0015] The similarity index space DTW, namely MMD-DTW distance, is calculated between flight parameter data to reflect the similarity between flight parameter data.
[0016] Preferably, step 3 specifically comprises:
[0017] When the similarity index is detected to exceed the set threshold, additional auxiliary information needs to be introduced for judgment, namely, further judgment based on whether new control instructions or tasks have been issued:
[0018] If the current similarity index exceeds the threshold and a new control command or task appears, the current data is determined to be new operating condition data, and the benchmark model is updated online adaptively using the model update method based on similarity enhancement EWC.
[0019] If the current similarity index exceeds the threshold and no new control instructions or tasks are introduced, the model will not be updated to avoid the model learning abnormal data features.
[0020] Preferably, the model update method based on similarity enhancement EWC improves the model's learning ability for different working conditions by enhancing the model update process in EWC;
[0021] The specific implementation of the similarity enhancement EWC method is represented by equations (3) and (4):
[0022] (3)
[0023] (4)
[0024] in: Indicates dynamic weights, This represents the current similarity index, with 0.4 indicating the similarity threshold. EWC regularization loss is used to adjust the model's balance between old and new data. Representative and old working conditions The relevant Fisher information matrix, and These represent the parameter values of the model before and after learning the new task, respectively. These are hyperparameters, used to adjust the balance between old and new parameters. This represents the original loss for the current learning task, measuring only the model's performance on new data; This represents the total loss function for learning the current task, which balances the fit to new data with the preservation of knowledge from old data. for The sum of the losses from EWC regularization.
[0025] Preferably, the smaller the dynamic weight, the more important the current new operating condition data is, and the less the EWC constrains the old knowledge, so that the model pays more attention to the new operating condition data during the update process;
[0026] Conversely, the larger the dynamic weight, the greater the constraint of EWC on old knowledge, resulting in a smaller parameter update magnitude. The design of similarity-enhanced incremental learning enables the model to maintain the stability of old knowledge when adapting to new data, while flexibly adjusting the learning of new knowledge, thereby improving the detection accuracy and generalization.
[0027] Preferably, storing new operating condition data avoids the repeated identification of learned flight parameter data, thereby reducing duplicate detection of the same features and lowering the model's adaptive update frequency; once the anomaly detection baseline model is updated, the new model... It will be stored, replacing the old model. Continue performing anomaly detection and online adaptive update tasks.
[0028] Preferably, the flight parameter data collected during actual flight includes high-pressure rotor speed, low-pressure rotor speed, engine exhaust temperature, high-pressure compressor back static pressure, low-pressure compressor outlet temperature, high-pressure turbine exhaust temperature, high-pressure compressor outlet pressure, fuel consumption, high-pressure compressor outlet temperature, oil pressure, fan inlet temperature, and low-pressure compressor outlet pressure.
[0029] An adaptive update system for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration, the system comprising:
[0030] The deployment module deploys two identical models. These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates;
[0031] The data caching module caches online flight parameter data and uses spatial DTW to implement multivariate similarity measurement and monitor changes in data distribution.
[0032] The data analysis module determines that when the similarity index exceeds a set threshold, the current data distribution has changed significantly, the model is no longer suitable for the current data, and needs to be updated.
[0033] The update module stores the new operating condition data in the flight parameter data standard library after the model is adaptively updated.
[0034] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration.
[0035] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration.
[0036] The present invention has the following beneficial effects:
[0037] The anomaly detection method proposed in this invention significantly outperforms other comparative methods in terms of recall, precision, and F1 score. Due to the improved regression prediction accuracy of the proposed model, the false positive rate of anomaly detection is significantly lower than the other two methods, resulting in higher overall accuracy. Specifically, the proposed model improves the recall metric by an average of 3.71% to 6.52%, the precision metric by an average of 14.01% to 45.9%, and the F1 score by an average of 8.92% to 26.6%. Attached Figure Description
[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0039] Figure 1 The diagram shows a model update diagram for similarity enhancement EWC according to the present invention;
[0040] Figure 2 The flowchart shown is a flowchart of the online adaptive update process of the baseline model of the present invention.
[0041] Figure 3 Displayed as a baseline model prediction curve;
[0042] Figure 4 Displayed as a baseline model residual curve;
[0043] Figure 5 Displayed as a replay model prediction curve;
[0044] Figure 6 Displayed as a replay model residual curve;
[0045] Figure 7 Displayed as a prediction curve of the LwF model;
[0046] Figure 8 Displayed as the residual curve of the LwF model
[0047] Figure 9The graph displayed is the prediction curve of the model proposed in this invention.
[0048] Figure 10 The graph shown is the residual curve of the model proposed in this invention;
[0049] Figure 11 The results are displayed as an online data similarity measurement chart. Detailed Implementation
[0050] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] The present invention will be described in detail below with reference to specific embodiments. Specific Implementation Example 1:
[0053] according to Figures 1 to 11 As shown, the specific optimization technical solution adopted by the present invention to solve the above-mentioned technical problems is: The present invention relates to an adaptive update method for anomaly detection benchmark model based on similarity-enhanced elastic weight integration.
[0054] This invention provides an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration, the method comprising the following steps:
[0055] Step 1: Deploy two identical models These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates;
[0056] Step 2: Cache the online flight parameter data, use spatial DTW to implement multivariate similarity measurement, and monitor changes in data distribution;
[0057] Step 3: When the similarity index exceeds the set threshold, it is determined that the current data distribution has changed significantly, the model is no longer suitable for the current data, and needs to be updated;
[0058] Step 4: After the model is adaptively updated, store the new operating condition data in the flight parameter data standard library. Specific Implementation Example 2:
[0060] The only difference between Embodiment 2 and Embodiment 1 of this application is that:
[0061] Step 2 specifically involves:
[0062] Once the cached data reaches a predetermined length, a similarity measurement is performed to monitor changes in data distribution. The cache length is 1000 data points. During the similarity measurement, a similarity index needs to be calculated between the online flight parameter data and multiple historical flight parameter data sets to reflect differences under different operating conditions. The flight parameter data is divided into non-overlapping segments, and the length of each segment is set to [missing information]. Then online flight parameter data and historical flight data They were each divided into multiple non-overlapping subsequences. and :
[0063] (1)
[0064] (2)
[0065] The similarity index space DTW, namely MMD-DTW distance, is calculated between flight parameter data to reflect the similarity between flight parameter data. Specific Implementation Example 3:
[0067] The only difference between Embodiment 3 and Embodiment 2 of this application is that:
[0068] Step 3 specifically involves:
[0069] When the similarity index is detected to exceed the set threshold, additional auxiliary information needs to be introduced for judgment, namely, further judgment based on whether new control instructions or tasks have been issued:
[0070] If the current similarity index exceeds the threshold and a new control command or task appears, the current data is determined to be new operating condition data, and the benchmark model is updated online adaptively using the model update method based on similarity enhancement EWC.
[0071] If the current similarity index exceeds the threshold and no new control instructions or tasks are introduced, the model will not be updated to avoid the model learning abnormal data features. Specific Implementation Example 4:
[0073] The only difference between Embodiment 4 and Embodiment 3 of this application is that:
[0074] The model update method based on similarity enhancement EWC improves the model's ability to learn from different working conditions by enhancing the model update process in EWC.
[0075] The specific implementation of the similarity enhancement EWC method is represented by equations (3) and (4):
[0076] (3)
[0077] (4)
[0078] in: Indicates dynamic weights, This represents the current similarity index, with 0.4 indicating the similarity threshold. EWC regularization loss is used to adjust the model's balance between old and new data. Representative and old working conditions The relevant Fisher information matrix, and These represent the parameter values of the model before and after learning the new task, respectively. These are hyperparameters, used to adjust the balance between old and new parameters. This represents the original loss for the current learning task, measuring only the model's performance on new data; This represents the total loss function for learning the current task, which balances the fit to new data with the preservation of knowledge from old data. for The sum of the losses from EWC regularization. Specific Implementation Example 5:
[0080] The difference between Embodiment 5 and Embodiment 4 of the present invention lies only in:
[0081] The smaller the dynamic weight, the more important the current new operating condition data is, and the less the EWC constrains the old knowledge, making the model pay more attention to the new operating condition data during the update process.
[0082] Conversely, the larger the dynamic weight, the greater the constraint of EWC on old knowledge, resulting in a smaller parameter update magnitude. The design of similarity-enhanced incremental learning enables the model to maintain the stability of old knowledge when adapting to new data, while flexibly adjusting the learning of new knowledge, thereby improving the detection accuracy and generalization. Specific Implementation Example Six:
[0084] The difference between Embodiment Six and Embodiment Five of the present invention lies only in:
[0085] Storing new operating condition data avoids the repeated identification of learned flight parameter data, thereby reducing duplicate detection of the same features and lowering the frequency of model adaptive updates; once the anomaly detection baseline model has been updated, the new model... It will be stored, replacing the old model. Continue performing anomaly detection and online adaptive update tasks. Specific Implementation Example 7:
[0087] The difference between Embodiment Seven and Embodiment Six of the present invention lies only in:
[0088] Based on flight parameter data collected during actual flight, including high-pressure rotor speed, low-pressure rotor speed, engine exhaust temperature, high-pressure compressor back static pressure, low-pressure compressor outlet temperature, high-pressure turbine exhaust temperature, high-pressure compressor outlet pressure, fuel consumption, high-pressure compressor outlet temperature, oil pressure, fan inlet temperature, and low-pressure compressor outlet pressure. Specific Implementation Example 8:
[0090] The difference between Embodiment 8 and Embodiment 7 of the present invention lies only in:
[0091] This invention provides an adaptive update system for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration, the system comprising:
[0092] The deployment module deploys two identical models. These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates;
[0093] The data caching module caches online flight parameter data and uses spatial DTW to implement multivariate similarity measurement and monitor changes in data distribution.
[0094] The data analysis module determines that when the similarity index exceeds a set threshold, the current data distribution has changed significantly, the model is no longer suitable for the current data, and needs to be updated.
[0095] The update module stores the new operating condition data in the flight parameter data standard library after the model is adaptively updated. Specific Implementation Example Nine:
[0097] The difference between Embodiment Nine and Embodiment Eight of the present invention lies only in:
[0098] The present invention provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration. Specific Implementation Example 10:
[0100] The only difference between Embodiment 10 and Embodiment 9 of the present invention is that:
[0101] The present invention provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration. Specific Implementation Example Eleven:
[0103] The only difference between Embodiment Eleven and Embodiment Ten of this invention is that:
[0104] In dynamic operating conditions involving flight missions and changing environments, new operating condition data often emerges that differ from offline data, leading to a decrease in the accuracy of the anomaly detection benchmark model and causing model mismatch issues. To address the problem of unknowingly mismatched anomaly detection benchmark models under actual dynamic operating conditions, and the limited airborne resources preventing large-scale data updates, this invention proposes an adaptive update method for the anomaly detection benchmark model based on similarity enhancement EWC, enabling online adaptive updates of the benchmark model. The adaptive update method for the anomaly detection benchmark model based on similarity enhancement EWC specifically includes the following steps:
[0105] (1) The online adaptive update algorithm of the model and the flight parameter data anomaly detection are run in parallel. Therefore, two identical models are deployed first. These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates.
[0106] (2) Spatial DTW is used to implement multivariate similarity measurement to monitor changes in data distribution. Since online flight parameter data is transmitted in the form of a data stream, it is necessary to cache the online flight parameter data first. When the cached data reaches a predetermined length, similarity measurement is performed to monitor changes in data distribution. The cache length set in this invention is 1000 data points. When measuring similarity, it is necessary to calculate the similarity index between the online flight parameter data and multiple historical flight parameter data to reflect the differences under different operating conditions. In order to improve calculation efficiency, this invention divides the flight parameter data into non-overlapping segments. The length of each segment is set to be... Then online flight parameter data and historical flight data They were each divided into multiple non-overlapping subsequences. and :
[0107] (1)
[0108] (2)
[0109] The similarity index space DTW, namely MMD-DTW distance, is calculated between flight parameter data to reflect the similarity between flight parameter data.
[0110] (3) When the similarity index exceeds the set threshold, it is determined that the current data distribution has changed significantly, the model is no longer suitable for the current data, and an update is required. However, since abnormal data is also part of the new operating condition data in terms of data distribution, in order to prevent abnormal data from being used for model updates, when the similarity index exceeds the set threshold, additional auxiliary information needs to be introduced for judgment, that is, to further judge whether new control commands or tasks have appeared (such as new throttle openings or new combinations of known tasks and flight phases). If the current similarity index exceeds the threshold and new control commands or tasks appear, the current data is determined to be new operating condition data, and the model update method based on similarity enhancement EWC is used to perform online adaptive updates to the baseline model; if the current similarity index exceeds the threshold and no new control commands or tasks appear, the model is not updated to avoid the model learning abnormal data features. Similarity measurement and additional auxiliary information reduce the update frequency, avoid catastrophic forgetting and excessively long update time, and ensure that the model is efficient and stable under new operating conditions, preventing abnormal data from being mistakenly included.
[0111] The model update method based on similarity-enhanced EWC improves the model's learning ability under different working conditions by enhancing the model update process within EWC. A schematic diagram of the similarity-enhanced EWC model update method is shown below. Figure 1 As shown.
[0112] The specific implementation of the similarity enhancement EWC method is represented by equations (3) and (4):
[0113] (3)
[0114] (4)
[0115] in: Indicates dynamic weights, This represents the current similarity index, with 0.4 indicating the similarity threshold. The learning rate during model updates is dynamically controlled based on the importance of operating conditions, enabling the model to effectively learn the current data features. It is a value greater than or equal to 0.4. EWC regularization loss is used to adjust the model's balance between old and new data. Representative and old working conditions The relevant Fisher information matrix, and These represent the parameter values of the model before and after learning the new task, respectively. These are hyperparameters used to adjust the balance between old and new parameters. This represents the original loss for the current learning task, measuring only the model's performance on new data; This represents the total loss function for learning the current task, which balances the fit to new data with the preservation of knowledge from old data. for The sum of the losses from EWC regularization.
[0116] By assigning different dynamic weights to different new operating condition data, the model can effectively distinguish their relative importance during the learning process. Smaller dynamic weights indicate greater importance of the current new operating condition data, less constraint of EWC on old knowledge, and a greater emphasis on new operating condition data during model updates. Conversely, larger dynamic weights result in greater constraint of EWC on old knowledge, leading to smaller parameter update magnitudes. The similarity-enhanced incremental learning design allows the model to maintain stability with old knowledge while adapting to new data, while flexibly adjusting the learning of new knowledge, thereby improving detection accuracy and generalization. This design ensures that the model can be effectively updated and adjusted under dynamic operating conditions, enhancing its performance in practical applications. It is important to note that model updates only involve updating the anomaly detection baseline model and do not update the anomaly threshold.
[0117] (4) After the model adaptively updates, the new operating condition data will be stored in the flight parameter data standard library. Storing the new operating condition data can avoid the repeated identification of the learned flight parameter data, thereby reducing the repeated detection of the same features and reducing the frequency of model adaptive updates. Once the anomaly detection benchmark model has been updated, the new model... It will be stored, replacing the old model. Continue performing anomaly detection and online adaptive update tasks.
[0118] In the proposed adaptive update method for the model, the flight parameter data standard library plays a core role. This library provides data support for similarity measurement and stores newly detected operating condition data. The flight parameter data standard library consists of representative operating condition data segments from historical flight parameter data. During similarity measurement, the similarity index between these historical data and the online data is calculated to determine whether new operating condition data has appeared. The online adaptive update process of the baseline model in this invention is as follows: Figure 2 As shown. Detailed Implementation Method Twelve:
[0120] This experiment was conducted based on flight parameter data collected during actual flight. The target parameter was the exhaust temperature of the aero-engine, and 11 related parameters were selected as inputs after parameter selection. All parameters used in the experiment are shown in Table 1.
[0121] To verify the effectiveness of the proposed model update method on aero-engine flight parameter data, this invention uses real flight parameter data from two different operating conditions for experiments. This invention uses a pre-constructed anomaly detection benchmark model for the model update experiment, which is trained using flight parameter data from operating condition 1. This invention uses flight parameter data from operating condition 2 to simulate online data under dynamic operating conditions, and conducts online adaptive model update experiments. The flight parameter data information used in the model update experiments is shown in Table 2.
[0122] Table 1. Parameters used in the experiment
[0123]
[0124] Table 2 Flight parameter data used in the model update experiment
[0125]
[0126] This experiment uses sortie 2-1 as the training set to simulate online data for online adaptive model updates. During the experiment, the data from sortie 2-1 is transmitted as a data stream. After the model is updated, the data from sorties 2-2, 2-3, and 2-4 are used as the test set to verify the model's adaptability to the new operating conditions, and the data from sortie 2-5 is used to evaluate the impact of the model update on anomaly detection accuracy.
[0127] (2) To evaluate the effectiveness of the proposed model update method, this invention compares the baseline model (the unupdated anomaly detection benchmark model), the LwF model (a model update method based on non-forgetting learning), the replay model (a model update method based on replay), and the proposed model (a model update method based on similarity enhancement EWC). The model update method based on non-forgetting learning updates the model using only new working condition data. The model update method based on replay updates the model by concatenating historical data and new working condition data. The proposed model has different hyperparameters during the model update process. The value is set to , Numerical values influence the learning rate; therefore, in the proposed method, the learning rate changes with the similarity metric during model updates. Dynamic changes. Based on the adaptive update method proposed in this invention, similarity measurement is performed on flight 2-1. The data length of the similarity measurement sliding window is set to 1000 data points, containing a total of 51 sliding windows. The average time for similarity measurement in each sliding window is approximately 1.36 seconds. The results of the adaptive update of the model for flight 2-1 are as follows: Figure 3 As shown.
[0128] from Figures 3-10As can be seen, the proposed similarity measurement method can accurately identify new operating condition data, detecting changes in data distribution at data points 20,000 to 21,000 and 41,000 to 42,000, and adaptively updating the model accordingly. This adaptive update significantly improves the model's adaptability to online flight parameter data. Compared to the baseline model, the updated model better fits the online flight parameter data and reduces the prediction residual. The proposed method enhances the model update process, enabling the model to learn more important information. Compared to the replay model, the results show that the proposed model can capture key temporal information and achieve more accurate predictions, greatly improving the accuracy and sensitivity of anomaly detection. Furthermore, the flight parameter data standard library proposed in this invention significantly reduces redundant learning of the same type of new operating condition data by storing new operating condition data, thereby reducing the number of model updates. Figure 4 As shown, storing new operating condition data and using it for future similarity metrics can reduce redundant updates.
[0129] Figure 11 In this table, the "similarity index before online update" represents the result calculated between the online flight parameter data and the historical flight parameter data in the flight parameter data standard library. When the similarity index exceeds the threshold and a new control command or task occurs, the flight parameter data in the current sliding window is identified as new operating condition data and stored in the flight parameter data standard library. Simultaneously, the model is updated online adaptively. Because the new operating condition data is stored in the flight parameter data standard library during model updates, previously learned flight parameter data will not be repeatedly identified during subsequent similarity measurements. This reduces redundant model updates and lowers the risk of catastrophic forgetting caused by repeated model updates. The model update time for different methods is shown in Table 3.
[0130] Table 3. Time taken for different model update methods (s)
[0131]
[0132] As shown in Table 3, the method proposed in this invention is more efficient in terms of model update time, saving an average of 4.01 seconds per update compared to the playback-based method. This invention was validated on multiple test runs, and the prediction results for different model update methods are shown in Table 4.
[0133] Table 4. Prediction results of different model update methods
[0134]
[0135] According to the results in Table 4, the method proposed in this invention outperforms the other two methods in both MAE and RMSE indices, and achieves the best results in all test flights. This is because the proposed model can distinguish the importance of new operating condition data during the update process, which allows the model to better handle the ratio of new and old knowledge during the update, enhancing the model's adaptability to different operating conditions after the update. Specifically, the MAE index decreased by an average of 26.61% to 40.34%, and the RMSE index decreased by an average of 23.78% to 36.50%, proving the effectiveness of the proposed method. The results show that the proposed method can accurately monitor changes in the data distribution of online flight parameter data using similarity metrics and promptly identify new operating condition data. After regression prediction, to verify the accuracy of different model methods in anomaly detection, the anomaly detection results are shown in Table 5.
[0136] Table 5. Anomaly detection results of different model update methods in flights 2-5.
[0137]
[0138] As shown in Table 5, the anomaly detection method proposed in this invention significantly outperforms other comparative methods in terms of Recall, Precision, and F1 score. Due to the improved regression prediction accuracy of the proposed model, the false positive rate of anomaly detection is significantly lower than the other two methods, resulting in higher overall accuracy. Specifically, the proposed model improves the Recall metric by an average of 3.71% to 6.52%, the Precision metric by an average of 14.01% to 45.9%, and the F1 score by an average of 8.92% to 26.6%.
[0139] The above description is merely a preferred embodiment of an adaptive update method for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration. The scope of protection for such an adaptive update method is not limited to the above embodiments; all technical solutions falling within this framework are within the protection scope of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the protection scope of this invention.
Claims
1. An adaptive update method for anomaly detection benchmark models based on similarity-enhanced elastic weight integration, characterized by: The method includes the following steps: Step 1: Deploy two identical models These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates; Step 2: Cache the online flight parameter data, use spatial DTW to implement multivariate similarity measurement, and monitor changes in data distribution; Step 3: When the similarity index exceeds the set threshold, it is determined that the current data distribution has changed significantly, the model is no longer suitable for the current data, and a model update based on similarity-enhanced elastic weight integration is required. Step 4: After the model adaptively updates, store the new operating condition data in the flight parameter data standard library; Step 2 specifically involves: Once the cached data reaches a predetermined length, a similarity measurement is performed to monitor changes in data distribution. The cache length is 1000 data points. The similarity measurement involves calculating a similarity index between the online flight parameter data and multiple historical flight parameter data sets to reflect differences under different operating conditions. The flight parameter data is divided into non-overlapping segments, with each segment having a set length of [missing information]. Then online flight parameter data and historical flight data They were each divided into multiple non-overlapping subsequences. and : (1) (2) The similarity index space DTW, namely MMD-DTW distance, is calculated between flight parameter data to reflect the similarity between flight parameter data. Step 3 specifically involves: When the similarity index is detected to exceed the set threshold, additional auxiliary information needs to be introduced for judgment, namely, further judgment based on whether new control instructions or tasks have been issued: If the current similarity index exceeds the threshold and a new control command or task appears, the current data is determined to be new operating condition data, and the benchmark model is updated online adaptively using the model update method based on similarity enhancement EWC. If the current similarity index exceeds the threshold and no new control instructions or tasks are introduced, the model will not be updated to avoid the model learning abnormal data features. The model update method based on similarity enhancement EWC improves the model's ability to learn from different working conditions by enhancing the model update process in EWC. The specific implementation of the similarity enhancement EWC method is represented by equations (3) and (4): (3) (4) in: Indicates dynamic weights. This represents the current similarity index, with 0.4 indicating the similarity threshold. EWC regularization loss is used to adjust the model's balance between old and new data. Representative and old working conditions The relevant Fisher information matrix, and These represent the parameter values of the model before and after learning the new task, respectively. These are hyperparameters, used to adjust the balance between old and new parameters. This represents the original loss for the current learning task, measuring only the model's performance on new data; This represents the total loss function for learning the current task, which balances the fit to new data with the preservation of knowledge from old data. for The sum of the losses from EWC regularization.
2. The method according to claim 1, characterized in that: The smaller the dynamic weight, the more important the current new operating condition data is, and the less the EWC constrains the old knowledge, making the model pay more attention to the new operating condition data during the update process. Conversely, the larger the dynamic weight, the greater the constraint of EWC on old knowledge, resulting in a smaller parameter update magnitude. The design of similarity-enhanced incremental learning enables the model to maintain the stability of old knowledge when adapting to new data, while flexibly adjusting the learning of new knowledge, thereby improving the detection accuracy and generalization.
3. The method according to claim 2, characterized in that: Storing new operating condition data avoids the repeated identification of learned flight parameter data, thereby reducing duplicate detection of the same features and lowering the frequency of model adaptive updates; once the anomaly detection baseline model has been updated, the new model... It will be stored, replacing the old model. Continue performing anomaly detection and online adaptive update tasks.
4. The method according to claim 3, characterized in that: Based on flight parameter data collected during actual flight, including high-pressure rotor speed, low-pressure rotor speed, engine exhaust temperature, high-pressure compressor back static pressure, low-pressure compressor outlet temperature, high-pressure turbine exhaust temperature, high-pressure compressor outlet pressure, fuel consumption, high-pressure compressor outlet temperature, oil pressure, fan inlet temperature, and low-pressure compressor outlet pressure.
5. An adaptive update system for an anomaly detection benchmark model based on similarity-enhanced elastic weight integration, wherein the system operates based on the method of claim 1, characterized in that: The system includes; The deployment module deploys two identical models. These are used for anomaly detection and online model updates, respectively, to avoid omissions in anomaly detection due to model updates; The data caching module caches online flight parameter data and uses spatial DTW to implement multivariate similarity measurement and monitor changes in data distribution. The data analysis module determines that when the similarity index exceeds a set threshold, the current data distribution has changed significantly, the model is no longer suitable for the current data, and needs to be updated. The update module stores the new operating condition data in the flight parameter data standard library after the model is adaptively updated.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method as claimed in any one of claims 1-4.
7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the method of any one of claims 1-4.