Fault diagnosis method for helicopter velocity vector sensor

By unifying and standardizing fault form data and training with the K-nearest neighbor algorithm, the problems of accuracy and multi-factor reasoning in fault diagnosis of helicopter velocity vector sensors were solved, achieving more efficient and reliable fault identification.

CN117708687BActive Publication Date: 2026-11-10CHINA AERO POLYTECH ESTAB
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202311817307.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-27
Publication Date
2026-11-10
Estimated Expiration
2043-12-27

AI Technical Summary

Technical Problem

Existing technologies for fault diagnosis of velocity vector sensors used in helicopters are not accurate enough, especially in multi-factor fault reasoning, and fail to make full use of product usage data, resulting in highly subjective diagnostic results that are difficult to meet the specific needs of helicopters.

Method used

Data is collected using fault forms, standardized and normalized, feature sets are determined, and the model is trained using the K-nearest neighbor algorithm to automatically recommend fault causes.

Benefits of technology

It improves the reliability and efficiency of fault diagnosis, reduces the problem of incorrect manual input, optimizes the fault diagnosis dataset, and enhances the ability to identify multi-factor faults.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117708687B_ABST
    Figure CN117708687B_ABST
Patent Text Reader

Abstract

This invention provides a fault diagnosis method for helicopter velocity vector sensors, comprising the following steps: S1, collecting sensor fault forms; S2, standardizing and normalizing the field data in the fault forms; S3, determining the feature set for fault diagnosis of helicopter velocity vector sensors; S4, vectorizing the selected features; S5, determining the fault diagnosis dataset; S6, training the K-nearest neighbor algorithm model using the fault diagnosis dataset; and S7, automatically recommending fault causes when a new fault form is entered. This invention solves the problem of non-standardized data through standardization and normalization, fills in fault forms with missing values, vectorizes the features after selecting the feature set, and adjusts the test data or features, thereby optimizing the fault diagnosis dataset and improving the reliability and efficiency of fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a fault diagnosis method for a helicopter velocity vector sensor. Background Technology

[0002] The issues of fault analysis, troubleshooting, maintenance, and corrective improvement for aviation product failures are becoming increasingly prominent. Rapidly identifying the causes of failures, conducting in-depth analysis of the root causes, and implementing subsequent treatments are all hot research topics.

[0003] Currently, diagnosing aviation product malfunctions primarily relies on human experience, failing to fully utilize accumulated usage and malfunction-related data, resulting in highly subjective and inaccurate results. Current research on assisted diagnosis of malfunction causes mainly focuses on expert knowledge reasoning using FMEA data. While these studies have made significant progress, they still have shortcomings. First, regarding data sources, existing research largely uses FMEA data as input. However, FMEA data is part of product design documents, and based on practical experience, its content often differs from actual product usage, making it difficult to guarantee the accuracy and comprehensiveness of conclusions drawn from FMEA-based knowledge reasoning. Furthermore, FMEA-based malfunction correlation reasoning rules often directly utilize the hierarchical transmission relationship of malfunction impact-malfunction mode-malfunction cause in FMEA. While easy to obtain and understand, this approach can only express single-factor malfunction transmission.

[0004] Existing methods for diagnosing aviation product faults, particularly for helicopter velocity vector sensors, often fall short when dealing with multi-factor fault reasoning. This is because the logical rules governing multi-factor fault cause reasoning are not readily intuitive, making the development of such rules extremely difficult. Furthermore, current technologies prioritize general applicability while neglecting the unique characteristics of each component. Therefore, a novel fault diagnosis method is needed for helicopter velocity vector sensors. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a fault diagnosis method for a helicopter velocity vector sensor, which includes the following steps: S1, collect sensor fault forms; Fault data of helicopter velocity vector sensors are collected using a fault form. The fault form should include at least the following fields: reporting time, user type, unit name, fault part serial number, system to which it belongs, fault part location, fault part manufacturer, fault part manufacturing date, fault part operating time, fault phenomenon / mode, fault cause, and handling measures. S2, standardize and normalize the field data in the fault form, as follows: S21, standardize the fields; The reporting time and the manufacturing date of the faulty part should be standardized according to the date category requirements; the user type, unit name, system, and manufacturer of the faulty part should be standardized according to the coding category requirements; the number of times the faulty part was used should be standardized according to the numeric category; and the manufacturing number of the faulty part should be standardized according to the character category. S22, Normalize string description fields, including: (a) Determine the event library for the corresponding field; (b) Text similarity calculation; The read string field is treated as text A, and an item from the corresponding field's event library is treated as text B. Then, a text similarity comparison is performed. If the similarity score... If the similarity is greater than the set threshold, it means that text A and text B match. Then, the text A string in the field is normalized to text B. If the similarity between text A and multiple text Bs in the event library is greater than the threshold, then the text B with the highest similarity is selected. If text A is not similar to any text B in the event library, it is marked, and the subsequent selection of text B is done manually. S23, Fill in the missing values ​​in the fault form; S3, determine the feature set for fault diagnosis of helicopter velocity vector sensor; The fault phenomenon / mode is combined with the reporting time, user type, number of times the faulty part was used, and the manufacturing date of the faulty part to obtain a feature set for determining the cause of the fault. S4, vectorize the selected features; S5, determine the fault diagnosis dataset; Based on the fault forms in S2, the filling time, user type, number of times the faulty part was used, manufacturing date of the faulty part, and fault phenomenon / mode of each fault form are used as input features, and the fault cause is used as the output result to form a training data. The training data of all fault forms form the initial dataset. The training data with empty features in the initial set are deleted, as well as the duplicate training data are deleted, to obtain the fault diagnosis dataset. S6, Use the fault diagnosis dataset to train the K-nearest neighbor algorithm model; S7: When a new fault form is entered, the cause of the fault is automatically recommended; After a new fault form is entered, the field data in the fault form is unified and standardized according to step S2. Then, the fields used as features are vectorized according to step S4. The input features are input into the K-nearest neighbor algorithm model, and finally the cause of the fault is obtained.

[0006] Preferably, the requirement for uniformity of date class, encoding class, numeric class, and character class in step S21 is as follows: Dates must be in the format "YYYY-MM-DD", meaning the year is represented by a full four-digit number, and the month and day are represented by a full two-digit number with zeros added. For numeric classes, the units and precision of the fields must be standardized; Character classes require the specification and standardization of allowed characters or character lengths in the field; The encoding class requires that the field values ​​in the field must be within a predefined range.

[0007] Preferably, in step S22, the text similarity calculation involves comparing the texts using two methods, and then weighting the similarity scores to obtain the final similarity score between text A and text B. The formula is:

[0008] in, and These are the weights of the text similarity scores for the two methods, respectively. and These are the text similarity scores for the two methods, respectively.

[0009] Preferably, the missing value imputation in step S23 is performed in the following manner: (a) When a faulty form has only one missing field, check if there are other forms with the same fields. If so, fill in the missing field directly. (b) The reporting time has a logical relationship with the manufacturing date of the faulty part and the number of times the faulty part was used. The other field is derived from any two of these fields and filled in accordingly. (c) The manufacturer of the faulty part can be identified based on the part's serial number; (d) When only one field is missing in the cause of failure and the handling measures, it shall be filled in according to the other field, and the fault phenomenon / mode shall be filled in according to the cause of failure or the handling measures.

[0010] Preferably, the vectorization of features in step S4 specifically involves: The reporting time is processed into a seasonal classification code, using four Boolean vectors; When a faulty component is in operation, the number of working hours is directly used as the numerical vector. The manufacturing date of the faulty part is divided into two parts based on the year span of the manufacturing date of the faulty part in all fault forms. The year in the manufacturing date of the faulty part is 0 in the first part and 1 in the second part. Then, it is directly combined with the month to form a numerical vector. User type, fault phenomenon / mode, and fault cause are all directly converted into Boolean vectors using classification coding.

[0011] Preferably, the step S5 of deleting duplicate training data specifically involves the following steps: (1) When training data A and training data B are completely identical, delete one of the training data directly; (2) When the input features of training data A and training data B are exactly the same, but the causes of failure are different, record the different causes of failure, and find all training data in the dataset that are exactly the same as the input features of training data A and B. Count the number of different causes of failure for training data with exactly the same input features. When the number of at least two causes of failure reaches the specified threshold TH of the total number of causes of failure, return to step S3 to add features to distinguish the training data. If only one cause of failure reaches the specified threshold TH, delete the training data with causes of failure less than the specified threshold TH.

[0012] Preferably, in step S6, the k value of the K-nearest neighbor algorithm model is 3, and the distance metric used is Minkowski distance.

[0013] Compared with the prior art, the present invention has the following advantages: The focus of this invention is to improve the fault diagnosis dataset used for model training, which is the foundation of the fault diagnosis model. This application first addresses the issue of inconsistent, manually entered fault data in helicopter velocity vector sensor fault data through standardization and unification. Then, considering the limited test data and the inability to arbitrarily delete data, efforts are made to fill in fault forms with missing values. Next, a feature set is selected based on the characteristics of the helicopter velocity vector sensor. To facilitate model training, the features are vectorized. Finally, while deleting duplicate test data, the features are also adjusted. This invention optimizes the fault diagnosis dataset for helicopter velocity vector sensors, improving the reliability and efficiency of fault diagnosis. Attached Figure Description

[0014] Figure 1 This is a flowchart of the fault diagnosis method for helicopter velocity vector sensor of the present invention; Figure 2 This is a flowchart of the text similarity algorithm for fault phenomena / fault modes in this invention; Figure 3 This is a flowchart of the K-nearest neighbor algorithm model of the present invention; Figure 4 This is an example diagram showing the output results of the K-nearest neighbor algorithm model of this invention. Detailed Implementation

[0015] To better understand the technical solution, structural features, achieved objectives and effects of the present invention, the specific implementation methods of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0016] This invention relates to a fault diagnosis method for a helicopter velocity vector sensor, such as... Figure 1 As shown, the specific implementation steps are as follows: S1, collect sensor fault forms.

[0017] Fault data of helicopter velocity vector sensors are collected using a fault form. The fault form should include at least the following fields: reporting time, user type, unit name, fault part serial number, system to which it belongs, fault part location, fault part manufacturer, fault part manufacturing date, fault part operating time, fault phenomenon / mode, fault cause, and handling measures.

[0018] S2. Standardize and normalize the field data in the fault form. The specific method is as follows: S21. Standardize the fields.

[0019] The reporting time and the manufacturing date of the faulty part should be standardized according to the date category requirements; the user type, unit name, system, and manufacturer of the faulty part should be standardized according to the coding category requirements; the number of times the faulty part was used should be standardized according to the numeric category; and the manufacturing number of the faulty part should be standardized according to the character category.

[0020] In this embodiment, the date format is required to be uniformly "YYYY-MM-DD", meaning the year is represented by a complete four-digit number, and the month and day are represented by two complete digits with zeros padded. For example, "2023-01-04" meets the uniformity requirement. Therefore, the reporting time and the manufacturing date of the faulty part must be uniformly formatted as "YYYY-MM-DD". The following are the main reasons why dates may not meet the requirements: The year, month, and day are separated by a " / "; the year, month, and day are separated by a "."; the year is represented by two digits; the month or day is represented by a single digit; the order of the year, month, and day is incorrect; or other incorrect data is entered. For example, 2023 / 01 / 04, 2023.01.04, 23-01-04, 2023-1-04, 01-04-2023, or even the Chinese characters "quality problem" are all considered non-compliant manufacturing dates for faulty parts. Furthermore, multiple issues may exist simultaneously for the same date, such as "1 / 4 / 23".

[0021] To address the above issues, the following checks and corrections are required when standardizing dates: First, discard any erroneous Chinese characters in the date field. Next, replace " / " and "." in the time field with "-". Then, split the time field using "-" as the separator and check if the order of year, month, and day matches. If not, rearrange the order. Finally, check if the year, month, and day are in complete numerical format. If not, pad the year with "19" or "20" and the month and day with "0" as needed. Ultimately, process all date field data into a unified "YYYY-MM-DD" format for later matching and calculation.

[0022] For numerical fields, the units and precision must be standardized. Units that do not meet the requirements must be converted and the decimal point must be truncated according to the precision. For example, in this application, the unit for the faulty component is standardized to "hours" and the precision is one decimal place.

[0023] Character classes require the specification and standardization of the characters allowed in the field or the character length.

[0024] The encoding class requires that the field values ​​in the field must be within a predefined range.

[0025] S22, standardize the string description field.

[0026] In fault forms, fields such as fault location, fault phenomenon / mode, fault cause, and handling measures are typically described as strings. Although manuals usually provide regulations and requirements for the descriptions of each field, due to differences in personal habits, descriptions of the same event cannot be completely identical, and some slightly different descriptions need to be categorized into the same event. Therefore, it is necessary to standardize the string descriptions and form unified labels. For example, for fault phenomena / modes such as "no airspeed indication," "no airspeed display," and "airspeed stuck," they should all be standardized as "airspeed indication stuck / no display." The steps and methods of the standardization algorithm are as follows: (1) Determine the event library for the corresponding field.

[0027] The event library for the corresponding fields is preset before fault diagnosis. It is usually determined based on manuals, specifications and statistical data. For example, for helicopter speed vector sensors, the event library for fault phenomena / modes is preset to include no airspeed indication or inaccurate indication, air leakage in the total pressure and static pressure pipeline, sideslip angle not following, universal joint jamming, and heating function failure. The event library for fault causes is preset to include relative rotation between the bevel gear and the shaft, rotary transformer failure, and angle of attack signal error exceeding tolerance.

[0028] (2) Text similarity calculation, such as Figure 2 As shown, specifically: If the read string field is taken as text A, and an item in the corresponding field event library is taken as text B, then the text similarity comparison will begin.

[0029] To ensure the reliability of text comparison, this invention employs two methods to compare the texts, and then uses a weighted sum of the similarity scores to obtain the final similarity score between text A and text B. The formula is:

[0030] in, and These are the weights of the text similarity scores for the two methods, respectively. and These are the text similarity scores for the two methods, respectively.

[0031] like If the similarity is greater than the set threshold, it means that text A and text B match. Then, the text A string in the field is normalized to text B. If the similarity between text A and multiple text Bs in the event library is greater than the threshold, then the text B with the highest similarity is selected. If text A is not similar to any text B in the event library, it is marked, and the subsequent selection of text B is done manually.

[0032] In this embodiment, the first method is a text comparison method, which compares each word in text A with each word in text B pairwise. If the similarity score of the two words is greater than a set threshold, the comparison is successful. If the two words are related, then they are considered related words. Finally, the ratio of the number of related words in text A and text B to the total number of words (taking the smaller number of words in text A and text B) is calculated as the similarity score between text A and text B. .

[0033]

[0034]

[0035] in, This represents a word in text A. This represents a word in text B. This represents a sequence similarity metric, and its value is a floating-point number ranging from 0 to 1. For the set threshold, These represent the number of words in text A and text B, respectively.

[0036] The second method uses a vector comparison approach based on the model. For text A and text B, a pre-trained model is used to generate vectors to obtain... and The similarity between the two vectors is calculated to obtain a score. .

[0037]

[0038] in, and Let A and B be vectors, respectively. A model for calculating the similarity of vectors.

[0039] S23, fill in the missing values ​​for the unified and standardized fault forms.

[0040] Each value in the fault form does not exist in isolation. For example, the cause of the fault and the handling measures are one-to-one. When only one item is missing, it can be filled in based on the other item.

[0041] In this embodiment, missing value imputation is performed in the following manner: (1) When a fault form has only one missing field, check if there are other forms with the same fields. If so, fill in the missing field directly.

[0042] (2) There is a logical relationship between the reporting time and the manufacturing date and working time of the faulty part. The other field can be derived from any two fields and filled in, such as: (manufacturing date of faulty part - reporting time) * average daily usage time = working time of faulty part.

[0043] (3) The manufacturer of the faulty part can be filled in based on the faulty part's serial number.

[0044] (4) When only one field of fault cause and treatment measures is missing, it can be filled in according to the other field, and the fault phenomenon / mode can be filled in according to the fault cause or treatment measures.

[0045] S3, Establish a fault diagnosis dataset for helicopter velocity vector sensors.

[0046] S31, determine the mapping relationship between fault phenomena / modes and fault causes and handling measures.

[0047] Fault phenomena / modes are the features most relevant to fault causes and handling measures. Therefore, based on all the fault forms of helicopter speed vector sensors collected, a mapping table of fault phenomena / modes and fault causes and handling measures was obtained.

[0048] If the fault phenomenon / mode in the helicopter speed vector sensor is no airspeed indication or inaccurate indication, the mapping relationship between the fault phenomenon / mode and the fault cause and handling measures is shown in Table 1 (partial).

[0049] Table 1

[0050] The table above clearly shows that relying solely on fault symptoms / modes is insufficient to accurately determine the cause of a fault and the corresponding corrective measures. However, identifying the cause of a fault inevitably leads to the appropriate corrective measures. Therefore, to arrive at a relatively accurate diagnosis based on fault symptoms / modes, additional characteristics are needed for the determination.

[0051] S32, Determine the feature set used to determine the cause of the fault.

[0052] For helicopter velocity vector sensors, this embodiment selects the reporting time, user type, number of times the faulty component operated, and the manufacturing date of the faulty component as new features for determining the cause of the fault. The specific reasons for selecting them are as follows: The timing of the report is important. Some faults may be related to seasonal factors, usually involving the influence of environmental stresses such as temperature and humidity.

[0053] User type: Some faults may be related to user type factors, and usually also involve the influence of special environment or usage. For example, analysis shows that speed vector sensors at coastal airports may be affected by special corrosion effects from marine salt spray and acidic atmospheric environment, and the frame pipeline corrosion failure mode is significantly more frequent.

[0054] When a component fails, the causes of many product failures are related to the time interval in which it operates. Different failure modes tend to be concentrated in different time domains, and the operating conditions or environmental conditions corresponding to many failure causes have the characteristic of cumulative effects over time. For example, the failure of electronic components in a velocity vector sensor, and the wear and tear failure of non-electronic components.

[0055] The manufacturing date of a faulty part may be related to the batch time of the faulty part for causes of failure with certain batch characteristics. In addition, the difference between the reporting date and the manufacturing date also reflects the product's life cycle calendar time, and the operating conditions or environmental stresses corresponding to some failure phenomena and causes also have the characteristic of cumulative effects over time.

[0056] The fault phenomenon / mode is combined with the reporting time, user type, number of times the faulty part was used, and the manufacturing date of the faulty part to obtain a feature set for determining the cause of the fault.

[0057] There are no strict rules on which fields to select for features related to the cause of the fault; you can choose as needed. However, different combinations of feature fields will affect the accuracy of subsequent fault diagnosis.

[0058] S4, vectorize the selected features.

[0059] To facilitate model training, the feature set is vectorized. The specific processing method is as follows: The reporting time is processed into a seasonal classification code: four Boolean vectors are used to represent the four seasons of spring, summer, autumn and winter. Usually, March to May is defined as spring, June to August as summer, September to November as autumn, and December to February as winter. The specific separation dates of the seasons can be determined according to actual needs. For example, when only analyzing the causes of helicopter speed vector sensor failures in Northeast China, the winter time can be set to November to March, while the summer time is only July to August.

[0060] When a faulty component is in operation, the number of working hours is used directly as the numerical vector.

[0061] The manufacturing date of a faulty part is divided into two parts based on the year span of the manufacturing dates in all fault reports. The year in the manufacturing date is 0 for the first part and 1 for the second part. These parts are then combined with the month, ignoring the day, to form a numerical vector. For example, if the year span of the manufacturing date of a faulty part in the fault report is 2018-2022, then the year span is divided by 2020. Years from 2018 to 2019 are 0, and years from 2020 to 2022 are 1. When the manufacturing date of a faulty part is 2021-01-04, the manufacturing date of the faulty part is converted into a numerical vector of 101.

[0062] User type, fault phenomenon / mode, and fault cause are all classified and coded, and then converted into Boolean vectors.

[0063] S5, determine the fault diagnosis dataset.

[0064] Based on the fault forms in S2, the filling time, user type, number of times the faulty part was used, manufacturing date of the faulty part, and fault phenomenon / mode of each fault form are used as input features, and the fault cause is used as the output result to form a training data. The training data of all fault forms form the initial fault diagnosis dataset.

[0065] First, delete training data in the fault diagnosis dataset that has no features.

[0066] Then, duplicate training data is removed. The specific steps are as follows: (1) When training data A and training data B are completely identical, delete one of the training data directly.

[0067] (2) When the input features of training data A and training data B are exactly the same, but the causes of failure are different, record the different causes of failure, and find all training data in the dataset that are exactly the same as the input features of training data A and B. Count the number of different causes of failure for training data with exactly the same input features. When the number of at least two causes of failure reaches the specified threshold TH of the total number of causes of failure (usually 20%), it is necessary to return to step S32 to add features to distinguish the training data. If only one cause of failure reaches the specified threshold TH, delete the training data with causes of failure less than the specified threshold TH.

[0068] This step yields a fault diagnosis dataset that does not contain duplicate training data.

[0069] S6, use the fault diagnosis dataset to train the K-nearest neighbor algorithm model, such as Figure 3 As shown: The K-Nearest Neighbors (K-NN) model is trained using the fault diagnosis dataset (which does not contain duplicate training data) from step S4. The K-Nearest Neighbors algorithm is a basic classification and regression method. In this embodiment, k is chosen as 3, and the distance metric is the Minkowski distance.

[0070] S7 automatically recommends the cause of the fault when a new fault form is entered.

[0071] When a new fault form is entered, the K-nearest neighbor algorithm model simultaneously detects the form. Once all feature fields have been entered, it first unifies and standardizes the field data in the fault form according to step S2. Then, according to step S4, it vectorizes the fields used as input features and uses these input features as input to the K-nearest neighbor algorithm model obtained in step S6. Finally, it obtains the cause of the fault. The specific steps of the K-nearest neighbor algorithm are as follows: a. Calculate the Minkowski distance between the current point and a point in the known category dataset; the formula for the Minkowski distance is as follows:

[0072] in, The distance between vector sample i and vector sample j is represented by q, where q represents the order and p represents the dimension of the vector sample. This represents the k-th dimension of vector sample i. The formula for the minkowski distance in the k-th dimension of vector sample j.

[0073] b. Sort the points in ascending order of distance; select the K points with the smallest distance to the current point.

[0074] c. Determine the frequency of occurrence of the category of the first K points.

[0075] d. Return the category with the highest frequency among the first K points as the classification result for the current point.

[0076] This allows us to obtain the causes of failures under conditions that allow for multi-dimensional feature input, and to develop a correlation analysis model for handling measures based on the causes of failures.

[0077] Example: After the records of a certain fault form have been standardized and normalized, the fields used as input features are shown in Table 2: Table 2 Examples of Input Feature Fields for Fault Forms

[0078] Inputting it into the trained K-nearest neighbor algorithm model yields the following output: Figure 4 As shown, the output can be explained using the following example: For the fault mode "no airspeed indication or inaccurate indication", under the above characteristic constraints, the most likely fault cause is "angle of attack signal error exceeding tolerance", with a probability of 0.42; followed by "relative rotation between bevel gear and shaft", with a probability of 0.30; and then "heating temperature too high, exceeding the tolerance range of ferromagnetic fluid", with a probability of 0.28.

[0079] Analysis and evaluation of results: Based on the above algorithm, this invention uses a total of 600 fault records as the model dataset, of which 400 are used as the training set, 100 as the test set, and 100 as the validation set. The validation results are analyzed, as shown in Table 3. Table 3. Accuracy of Prediction Results Based on Test Data

[0080] The analysis of the above prediction results shows that the algorithm achieves high accuracy for fault modes with a large sample size (no or inaccurate air velocity indication, leakage in total pressure and static pressure pipelines); it also achieves high accuracy for fault modes with a small sample size but relatively concentrated causes (sideslip angle not following); however, the algorithm's accuracy is lower for fault modes with a small sample size and relatively dispersed causes (universal head jamming, heating function failure). This phenomenon is related to the small sample size and imbalance of the fault case data, as well as the effectiveness of feature selection and the choice of model parameters. Therefore, further optimization of the model can be achieved by accumulating more fault case data to increase the sample size, and by rationally selecting features and adjusting model parameters, thereby further improving the accuracy of the model's prediction results.

[0081] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A fault diagnosis method for a helicopter velocity vector sensor, characterized in that, It includes the following steps: S1, collect sensor fault forms; Fault data of helicopter velocity vector sensors are collected using a fault form. The fault form should include at least the following fields: reporting time, user type, unit name, fault part serial number, system to which it belongs, fault part location, fault part manufacturer, fault part manufacturing date, fault part operating time, fault phenomenon / mode, fault cause, and handling measures. S2, standardize and normalize the field data in the fault form, as follows: S21, standardize the fields; The reporting time and the manufacturing date of the faulty part should be standardized according to the date category requirements; the user type, unit name, system, and manufacturer of the faulty part should be standardized according to the coding category requirements; the number of times the faulty part was used should be standardized according to the numeric category; and the manufacturing number of the faulty part should be standardized according to the character category. S22, Normalize string description fields, including: (a) Determine the event library for the corresponding field; (b) Text similarity calculation; The read string field is treated as text A, and an item from the corresponding field's event library is treated as text B. Then, a text similarity comparison is performed. If the similarity score... If the similarity is greater than the set threshold, it means that text A and text B match. Then, the text A string in the field is normalized to text B. If the similarity between text A and multiple text Bs in the event library is greater than the threshold, then the text B with the highest similarity is selected. If text A is not similar to any text B in the event library, it is marked, and the subsequent selection of text B is done manually. S23, Fill in the missing values ​​in the fault form; S3, determine the feature set for fault diagnosis of helicopter velocity vector sensor; The fault phenomenon / mode is combined with the reporting time, user type, number of times the faulty part was used, and the manufacturing date of the faulty part to obtain a feature set for determining the cause of the fault. S4, vectorize the selected features; S5, determine the fault diagnosis dataset; Based on the fault forms in S2, the filling time, user type, number of times the faulty part was used, manufacturing date of the faulty part, and fault phenomenon / mode of each fault form are used as input features, and the fault cause is used as the output result to form a training data. The training data of all fault forms form the initial dataset. The fault diagnosis dataset is obtained by deleting training data with empty features from the initial set and deleting duplicate training data. To remove duplicate training data, the specific steps are as follows: (1) When training data A and training data B are completely identical, delete one of the training data directly; (2) When the input features of training data A and training data B are exactly the same, but the causes of failure are different, record the different causes of failure, and find all training data in the dataset that are exactly the same as the input features of training data A and B. Count the number of different causes of failure for training data with exactly the same input features. When the number of at least two causes of failure reaches the specified threshold TH of the total number of causes of failure, return to step S3 to add features to distinguish the training data. If only one cause of failure reaches the specified threshold TH, delete the training data with causes of failure less than the specified threshold TH. S6, Use the fault diagnosis dataset to train the K-nearest neighbor algorithm model; S7: When a new fault form is entered, the cause of the fault is automatically recommended; After a new fault form is entered, the field data in the fault form is unified and standardized according to step S2. Then, the fields used as features are vectorized according to step S4. The input features are input into the K-nearest neighbor algorithm model, and finally the cause of the fault is obtained.

2. The fault diagnosis method for a helicopter velocity vector sensor according to claim 1, characterized in that, The unified requirements for date, encoding, numeric, and character classes in step S21 are as follows: Dates must be in the format "YYYY-MM-DD", meaning the year is represented by a full four-digit number, and the month and day are represented by a full two-digit number with zeros added. For numeric classes, the units and precision of the fields must be standardized; Character classes require the specification and standardization of allowed characters or character lengths in the field; The encoding class requires that the field values ​​in the field must be within a predefined range.

3. The fault diagnosis method for a helicopter velocity vector sensor according to claim 1, characterized in that, In step S22, the text similarity calculation involves comparing the texts using two methods, and then weighting the similarity scores to obtain the final similarity score between text A and text B. The formula is: in, and These are the weights of the text similarity scores for the two methods, respectively. and These are the text similarity scores for the two methods, respectively.

4. The fault diagnosis method for a helicopter velocity vector sensor according to claim 1, characterized in that, Missing value imputation in step S23 is performed as follows: (a) When a faulty form has only one missing field, check if there are other forms with the same fields. If so, fill in the missing field directly. (b) The reporting time has a logical relationship with the manufacturing date of the faulty part and the number of times the faulty part was used. The other field is derived from any two of these fields and filled in accordingly. (c) The manufacturer of the faulty part can be identified based on the part's serial number; (d) When only one field of fault cause and handling measures is missing, it can be filled in based on the other field, and the fault phenomenon / mode can be filled in based on the fault cause or handling measures.

5. The fault diagnosis method for a helicopter velocity vector sensor according to claim 1, characterized in that, The vectorization of features in step S4 specifically involves: The reporting time is processed into a seasonal classification code, using four Boolean vectors; When a faulty component is in operation, the number of working hours is directly used as the numerical vector. The manufacturing date of the faulty part is divided into two parts based on the year span of the manufacturing date of the faulty part in all fault forms. The year in the manufacturing date of the faulty part is 0 in the first part and 1 in the second part. Then, it is directly combined with the month to form a numerical vector. User type, fault phenomenon / mode, and fault cause are all directly converted into Boolean vectors using a classification coding method.

6. The fault diagnosis method for a helicopter velocity vector sensor according to claim 1, characterized in that, In step S6, the k value of the K-nearest neighbor algorithm model is 3, and the distance metric used is Minkowski distance.

Citation Information

Patent Citations

  • Method and system for managing database containing record with missing value

    CN104516879A

  • Chinese semantic recognition method based on cosine similarity and minimum editing distance

    CN110674378A

  • Text classification-based textile equipment fault auxiliary processing method

    CN116484262A