Deep learning based angular velocity sensor fault detection method

Abstract

The application provides a kind of angular velocity sensor fault detection method based on deep learning, belongs to the technical field of fault detection;The scheme adopted by the application is to obtain angular velocity data and voltage data of angular velocity sensor, various environmental data in real time;The whole acquisition process is divided into multiple periods, according to the relationship between various environmental data and angular velocity data in each period, and the dispersion degree of various environmental data, the overall disturbance index of various environmental data on angular velocity in each period is obtained, and the predicted deviation characteristics between angular velocity and various environmental data in the next period, the proportional relationship between angular velocity and voltage in each period, the heading angle difference are combined to obtain the degradation confidence in the next period, and then whether the angular velocity sensor fails in the next period is judged.The application provides analysis of the influence of various environmental factors on angular velocity sensor, improves the accuracy of angular velocity sensor fault detection.

Description

Technical Field

[0001] This application belongs to the field of fault detection technology, specifically relating to a fault detection method and system for angular velocity sensors based on deep learning. Background Technology

[0002] Angular velocity sensors are core components of aircraft fly-by-wire control systems, primarily used to measure the aircraft's angular velocities along the pitch, roll, and yaw axes. Their output voltage signal is proportional to the input angular velocity, providing crucial feedback signals to the fly-by-wire control system and thus improving flight stability. However, angular velocity sensors are prone to degradation and failure over long-term operation; once a failure occurs, it can seriously threaten flight safety. Therefore, developing a fault detection method for angular velocity sensors is of great significance for improving flight safety.

[0003] While existing technologies for fault analysis of angular velocity sensors have considered the impact of environmental factors on angular velocity data and implemented interference processing to improve detection accuracy to some extent, they only analyze the interference level of a single environmental factor and do not fully consider the coupled effects between multiple factors. This can lead to an inability to accurately distinguish between external environmental interference and sensor degradation faults, easily causing misjudgments or missed faults, resulting in poor fault detection performance. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, reduce misjudgments and omissions in angular velocity sensor fault analysis, and improve the accuracy of fault detection, this invention provides a deep learning-based angular velocity sensor fault detection method.

[0005] A deep learning-based fault detection method for angular velocity sensors includes the following steps: Step S1: Real-time continuous acquisition of angular velocity sensor data during aircraft flight; the sensor data includes angular velocity, voltage, aircraft GPS data, velocity heading angle, position heading angle, and environmental data; the environmental data includes vibration values, electromagnetic values, and temperature values ​​at the location of the angular velocity sensor; the acquired sensor data is divided into T periods; within the i-th period, the angular velocity at different times, the voltage at different times, the aircraft GPS data at different times, the velocity heading angle at different times, the position heading angle at different times, the vibration values ​​at different times, the electromagnetic values ​​at different times, and the temperature values ​​at different times respectively constitute the angular velocity sequence, voltage sequence, GPS sequence, velocity heading angle sequence, and position heading angle sequence. The dataset consists of angular velocity sequences, vibration value sequences, temperature sequences, and electromagnetic value sequences. Within all periods, the vibration value sequences, temperature sequences, and electromagnetic value sequences for each period are arranged according to time sequence to obtain vibration feature sequences, temperature feature sequences, and electromagnetic feature sequences. The angular velocity sequences for all periods are arranged according to time sequence to form a label sequence. Within all periods, the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight for each period are arranged according to time sequence to obtain average vibration interference weight sequences, average electromagnetic interference weight sequences, and average temperature interference weight sequences. The dataset comprises the vibration feature sequences, temperature feature sequences, electromagnetic feature sequences, label sequences, average vibration interference weight sequences, average electromagnetic interference weight sequences, and average temperature interference weight sequences. Simultaneously calculate the degradation extension coefficient and the overall vibration disturbance index for each cycle. Overall electromagnetic disturbance index and overall temperature disturbance index ; Step S2: Construct a prediction model; train the prediction model using the dataset, and obtain the optimal prediction model when the prediction model meets the convergence criterion; the input of the prediction model includes the vibration value sequence, temperature value sequence, electromagnetic value sequence, angular velocity sequence, vibration average disturbance weight, electromagnetic average disturbance weight, and temperature average disturbance weight of the i-th period; the output of the prediction model is the predicted vibration sequence, predicted temperature sequence, predicted electromagnetic sequence, and predicted angular velocity sequence of the (i+1)-th period; Step S3: Obtain the vibration sequence, temperature sequence, electromagnetic sequence, and angular velocity sequence of the angular velocity sensor during the current period of aircraft flight, and calculate the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight. Input the vibration sequence, temperature sequence, electromagnetic sequence, angular velocity sequence, average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight of the current period into the optimal prediction model to obtain the predicted vibration sequence, predicted temperature sequence, predicted electromagnetic sequence, and predicted angular velocity sequence for the next period. Step S4: Acquire the real-time vibration sequence, real-time temperature sequence, real-time electromagnetic sequence, and real-time angular velocity sequence of the angular velocity sensor for the next cycle during aircraft flight; calculate the degradation confidence level for the next cycle. If the confidence level degrades If the value is less than the preset fault threshold, it is determined that the angular velocity sensor has not malfunctioned in the next cycle; if the confidence level degrades... If the angular velocity sensor fails within the next cycle, it is determined that the failure is greater than or equal to the preset fault threshold.

[0006] Furthermore, the calculation of the degradation confidence level for the next cycle... for: ; In the formula, The confidence level for degradation in the next cycle. This is the degradation extension factor for the next cycle; It is the average of the overall vibration disturbance index, the overall temperature disturbance index, and the overall electromagnetic disturbance index in the next cycle; The relative deviation for the next period; This is the second average value for the next period; This is a preset constant; Furthermore, the angular velocity prediction deviation is the absolute value of the difference between the elements in the real-time angular velocity sequence and the elements in the predicted angular velocity sequence at the same time. The average value of all absolute values ​​is then calculated to obtain the angular velocity prediction deviation. The vibration prediction deviation, temperature prediction deviation, and electromagnetic prediction deviation are calculated in the same way as the angular velocity prediction deviation. The relative deviation is obtained by taking the maximum value among vibration prediction deviation, temperature prediction deviation and electromagnetic prediction deviation, and calculating the ratio of angular velocity prediction deviation to the maximum value, which is the relative deviation. The calculation process for the second mean is as follows: The prediction deviation of the angular velocity in the next cycle and all previous cycles is arranged in time sequence to obtain the prediction deviation sequence of the next cycle. The mean of the first difference sequence of the prediction deviation sequence of the next cycle is calculated, which is the second mean of the next cycle.

[0007] Furthermore, the calculation process for the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight for each cycle is as follows: The vibration average interference weight is the average of the vibration influence weights for all cycles; the electromagnetic average interference weight is the average of the electromagnetic influence weights for all cycles; and the temperature average interference weight is the average of the temperature influence weights for all cycles. The vibration influence weight for each cycle is the ratio of the vibration disturbance factor for each cycle to the sum of the vibration disturbance factors for all cycles; the electromagnetic influence weight for each cycle is the ratio of the electromagnetic disturbance factor for each cycle to the sum of the electromagnetic disturbance factors for all cycles; the temperature influence weight for each cycle is the ratio of the temperature disturbance factor for each cycle to the sum of the temperature disturbance factors for all cycles. The calculation process for the vibration disturbance factor, electromagnetic disturbance factor, and temperature disturbance factor for each cycle is as follows: In all cycles, the vibration disturbance factor is obtained by using the disturbance index calculation method based on the vibration value and angular velocity in the environmental data; the electromagnetic disturbance factor is obtained by using the disturbance index calculation method based on the electromagnetic value and angular velocity in the environmental data; and the temperature disturbance factor is obtained by using the disturbance index calculation method based on the temperature value and angular velocity in the environmental data. The process of calculating the disturbance index is as follows: The input to the perturbation index calculation method includes perturbation data and perturbation angular velocity for all periods; the perturbation data and perturbation angular velocity for each period are input into the Granger causality test algorithm to obtain the F value; wherein the perturbation data is the cause, the perturbation angular velocity is the effect, and the F value is the causal value of the perturbation data relative to the perturbation angular velocity.

[0008] Furthermore, the calculation process for the degradation extension coefficient is as follows: Based on the degradation factor for each cycle, calculate the degradation factor for each cycle divided by the degradation factor of the previous cycle and a preset constant. The sum of these values ​​yields the degradation extension coefficient; the degradation factor for each cycle is obtained as follows: Calculate the ratio of angular velocity to voltage at all times in each cycle to obtain the angular velocity-voltage ratio at all times in the cycle; then calculate the variance of the angular velocity-voltage ratio at all times in the cycle, and use the variance of the angular velocity-voltage ratio as the first discrete value for each cycle; in each cycle, use the spherical trigonometric formula to calculate the position heading angle at each time using GPS data at each time; calculate the absolute value of the difference between the velocity heading angle and the position heading angle at all times, and then calculate the mean of the absolute values ​​of the differences at all times, and use the mean as the first mean for each cycle; the product of the first discrete value and the first mean for each cycle is the degradation factor for each cycle.

[0009] Furthermore, the overall vibration disturbance index Overall electromagnetic disturbance index and overall temperature disturbance index The calculation process is as follows: The overall vibration disturbance index relative to the angular velocity during the i-th period for: ; Let be the variance of the vibration sequence in the i-th period; The influence factor of vibration on angular velocity in the i-th period; The overall electromagnetic disturbance index of the electromagnetic sub-angular velocity during the i-th period for: ; Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; The overall temperature disturbance index of temperature versus angular velocity during the i-th period for: ; Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; Among them, the influence factor of vibration on angular velocity during the i-th period The mean values ​​of the vibration disturbance factor and vibration regression coefficient after normalization; the influence factor of electromagnetic field on angular velocity. The mean values ​​of the electromagnetic disturbance factor and electromagnetic regression coefficient after normalization; the influence factor of temperature on angular velocity. This represents the mean value of the temperature disturbance factor and the temperature regression coefficient after normalization.

[0010] Furthermore, the solution process for the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient is as follows: Construct a regression model and solve for the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient for each cycle; The regression model for: In the formula, , , These are the angular velocity sequence, vibration sequence, temperature sequence, and electromagnetic sequence in the i-th period, respectively. This is the error term; , and These are the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient in the i-th period, respectively. , and It reflects the impact of each type of environmental data on angular velocity.

[0011] Furthermore, the methods for solving the regression model to obtain the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient include, but are not limited to, the least squares method, the ridge regression method, and the stepwise regression analysis method. Furthermore, the normalization methods include, but are not limited to, the Z-score method, the maximum value normalization method, and the maximum and minimum value normalization method.

[0012] Furthermore, the prediction model includes, but is not limited to, BP neural networks, LSTM neural networks, and GRU neural networks; The convergence criterion is a combination of any one or more of the following conditions: Condition A: The mean squared error loss function of the prediction model reaches the loss threshold; Condition B: The prediction model is trained for NP consecutive epochs, and the change in the mean squared error loss function value of the prediction model is less than the loss change threshold. Condition C: When the number of training rounds reaches the preset maximum number of iterations.

[0013] Furthermore, the preset fault threshold is ;in The mean of all degradation confidence levels. is the standard deviation of all degradation confidence levels.

[0014] The beneficial effects of this invention are as follows: by constructing a degradation extension coefficient, it is possible to preliminarily assess whether the angular velocity sensor has experienced performance degradation and the severity of the degradation; by constructing an overall disturbance index, it is possible to comprehensively reflect the interference intensity of environmental factors on the angular velocity sensor data; by constructing a degradation confidence level, it is possible to comprehensively assess the possibility of angular velocity sensor degradation failure after excluding external interference, thereby accurately distinguishing between environmental interference and sensor failure, improving the accuracy of angular velocity sensor fault detection, and ensuring flight safety. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a deep learning-based angular velocity sensor fault detection method provided in one embodiment of this application; Figure 2 A flowchart illustrating the process of obtaining the degradation confidence for the next cycle, as provided in one embodiment of this application. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] A deep learning-based fault detection method for angular velocity sensors includes the following steps: The sensor data of the angular velocity sensor is continuously collected in real time during the flight of the aircraft; the sensor data includes the angular velocity and voltage of the angular velocity sensor, the GPS data of the aircraft, the speed heading angle, the position heading angle, and environmental data; the environmental data includes the vibration value, the electromagnetic value, and the temperature value at the location of the angular velocity sensor. The collected sensor data is divided into T periods. Within the i-th period, the angular velocity, voltage, GPS data of the aircraft at different times, velocity heading angle, position heading angle, vibration value, electromagnetic value, and temperature value at different times constitute the angular velocity sequence, voltage sequence, GPS sequence, velocity heading angle sequence, position heading angle sequence, vibration value sequence, temperature sequence, and electromagnetic value sequence, respectively. The degradation factor and degradation extension coefficient for each period are calculated. The velocity heading angle is obtained by integrating the angular velocity of the heading axis; The position heading angle is obtained by calculating the GPS data using the spherical trigonometry formula; The method for obtaining the degradation factor for each cycle is as follows: Calculate the ratio of angular velocity to voltage at all times in each cycle to obtain the angular velocity-voltage ratio at all times in the cycle; then calculate the variance of the angular velocity-voltage ratio at all times in the cycle, and use the variance of the angular velocity-voltage ratio as the first discrete value for each cycle; in each cycle, use the spherical trigonometric formula to calculate the position heading angle at each time using GPS data at each time; calculate the absolute value of the difference between the velocity heading angle and the position heading angle at all times, and then calculate the mean of the absolute values ​​of the differences at all times, and use the mean as the first mean for each cycle; the product of the first discrete value and the first mean for each cycle is the degradation factor for each cycle; The degradation extension coefficient is calculated as follows: based on the degradation factor of each cycle, calculate the degradation factor of each cycle divided by the degradation factor of the previous cycle and a preset constant. The sum of these values ​​yields the degradation extension coefficient. In all cycles, the vibration disturbance factor is obtained by using the disturbance index calculation method based on the vibration value and angular velocity in the environmental data; the electromagnetic disturbance factor is obtained by using the disturbance index calculation method based on the electromagnetic value and angular velocity in the environmental data; and the temperature disturbance factor is obtained by using the disturbance index calculation method based on the temperature value and angular velocity in the environmental data. The process of calculating the disturbance index is as follows: The input to the disturbance index calculation method includes disturbance data and disturbance angular velocity within all periods; the disturbance data and disturbance angular velocity within the periods are input into the Granger causality test algorithm to obtain the F value; wherein the disturbance data is the cause, the disturbance angular velocity is the effect, and the F value is the causal value of the disturbance data relative to the disturbance angular velocity. Construct a regression model and solve for the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient for each cycle; The regression model for: In the formula, , , These are the angular velocity sequence, vibration sequence, temperature sequence, and electromagnetic sequence in the i-th period, respectively. This is the error term; , and These are the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient in the i-th period, respectively. , and This reflects the impact of each type of environmental data on angular velocity; The method for calculating the influence factor of the i-th period vibration data on the angular velocity is as follows: Calculate the influence factor of vibration on angular velocity during the i-th period. The mean values ​​of the vibration disturbance factor and the vibration regression coefficient after normalization are given respectively. Electromagnetic influence factors on angular velocity The mean values ​​of the electromagnetic disturbance factor and the electromagnetic regression coefficient after normalization are given respectively. The effect of temperature on angular velocity This represents the normalized mean values ​​of the temperature perturbation factor and the temperature regression coefficient. The overall vibration disturbance index relative to the angular velocity during the i-th period for: ; Let be the variance of the vibration sequence in the i-th period; The influence factor of vibration on angular velocity in the i-th period; The overall electromagnetic disturbance index of the electromagnetic sub-angular velocity during the i-th period for: ; Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; The overall temperature disturbance index of temperature versus angular velocity during the i-th period for: ; Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; The normalization methods include, but are not limited to, the sigmoid function and the tanh function; The methods for solving the regression model to obtain the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient include, but are not limited to, the least squares method, the ridge regression method, and the stepwise regression analysis method. The normalization methods include, but are not limited to, the Z-score method, the maximum value normalization method, and the maximum and minimum value normalization method.

[0018] Calculate the vibration influence weight, electromagnetic influence weight, temperature influence weight, average vibration disturbance weight, average electromagnetic disturbance weight, and average temperature disturbance weight for each cycle in all cycles. The vibration impact weight for each cycle is the ratio of the vibration disturbance factor for each cycle to the sum of the vibration disturbance factors for all cycles; the average vibration impact weight for all cycles is calculated to obtain the average vibration disturbance weight. The electromagnetic influence weight for each cycle is the ratio of the electromagnetic disturbance factor for each cycle to the sum of the electromagnetic disturbance factors for all cycles; the average electromagnetic influence weight for all cycles is calculated to obtain the average electromagnetic interference weight. The temperature influence weight for each cycle is the ratio of the temperature disturbance factor for each cycle to the sum of the temperature disturbance factors for all cycles; the average temperature influence weight for all cycles is calculated to obtain the average temperature disturbance weight. Build the dataset; Within all periods, the vibration value sequence, temperature sequence, and electromagnetic value sequence are arranged according to time sequence to obtain the vibration characteristic sequence, temperature characteristic sequence, and electromagnetic characteristic sequence. Arrange the angular velocity sequences within all periods according to time sequence to form a label sequence; The vibration average disturbance weight, electromagnetic average disturbance weight, and temperature average disturbance weight are arranged according to time sequence to obtain the vibration average disturbance weight sequence, electromagnetic average disturbance weight sequence, and temperature average disturbance weight sequence, respectively. The vibration feature sequence, temperature feature sequence, electromagnetic feature sequence, label sequence, vibration average disturbance weight sequence, electromagnetic average disturbance weight sequence, and temperature average disturbance weight sequence are used as the dataset; Build a predictive model; The inputs to the prediction model include the vibration value, temperature value, electromagnetic value, angular velocity, average vibration disturbance weight, average electromagnetic disturbance weight, and average temperature disturbance weight at time t of the i-th period. The output of the prediction model is the predicted vibration sequence, predicted temperature sequence, predicted electromagnetic sequence, and predicted angular velocity sequence for the next cycle. The prediction model is trained using the dataset (with supplementary training criteria) to obtain the optimal prediction model; The vibration, temperature, electromagnetic, and angular velocity sequences of the angular velocity sensor during the current period of aircraft flight are obtained, and the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight are calculated. The vibration, temperature, electromagnetic, and angular velocity sequences of the current period, the average vibration interference weight, the average electromagnetic interference weight, and the average temperature interference weight are input into the optimal prediction model to obtain the predicted vibration, temperature, electromagnetic, and angular velocity sequences for the next period. Collect the real-time vibration sequence, real-time temperature sequence, real-time electromagnetic sequence and real-time angular velocity sequence of the angular velocity sensor during the next cycle of the aircraft flight; Calculate the angular velocity prediction deviation: Calculate the absolute value of the difference between the elements in the real-time angular velocity sequence and the elements in the predicted angular velocity sequence at the same time, and then calculate the average of all absolute values ​​to obtain the angular velocity prediction deviation. The vibration prediction deviation, temperature prediction deviation, and electromagnetic prediction deviation were calculated using the same method as the angular velocity prediction deviation. The relative deviation is calculated as follows: obtain the maximum value among vibration prediction deviation, temperature prediction deviation and electromagnetic prediction deviation, and calculate the ratio of angular velocity prediction deviation to the maximum value, which is the relative deviation. Calculate the second mean for the next period: Arrange the angular velocity prediction deviations of the next cycle and all previous cycles in chronological order to obtain the prediction deviation sequence of the next cycle. Calculate the mean of the first difference sequence of the prediction deviation sequence of the next cycle, which is the second mean of the next cycle. The prediction models include, but are not limited to, BP neural networks, LSTM neural networks, and GRU neural networks; Calculate the degradation confidence level for the next period for: ; In the formula, The confidence level for degradation in the next cycle. This is the degradation extension coefficient for the next cycle; This represents the average of the overall vibration disturbance index, overall temperature disturbance index, and overall electromagnetic disturbance index for the next cycle. It is the ratio of the angular velocity prediction deviation to the maximum value among the vibration prediction deviation, temperature prediction deviation, and electromagnetic prediction deviation in the next cycle. This is the second average value for the next period; This is a preset constant; If the confidence level degrades If the value is less than the preset fault threshold, it is determined that the angular velocity sensor has not malfunctioned in the next cycle; otherwise, it is determined that the angular velocity sensor has malfunctioned in the next cycle. The convergence conditions of the prediction model include, but are not limited to, any one or more combinations of the following: (1): The loss function value is less than the preset convergence threshold (in this embodiment, MSE is used as the loss function, the convergence threshold is 0.001, and the actual value range is [0.0001, 0.01], which can be adjusted according to the data scale and accuracy requirements). (2): The change in the loss function value within N consecutive training epochs is less than the preset change threshold (N is 10 in this embodiment, and the change threshold is 1e-5). (3): The training rounds reach the preset maximum number of iterations (in this embodiment, the maximum number of iterations is 500); (4): The value of the validation set loss function no longer decreases within M consecutive training rounds (early stopping strategy, M is 20 in this embodiment).

[0019] If any of the above conditions are met, the model is considered to have converged. Training is stopped and the current model parameters are saved to obtain the optimal prediction model.

[0020] The following description, in conjunction with the accompanying drawings, details the specific scheme of the deep learning-based angular velocity sensor fault detection method and system provided in this application.

[0021] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based angular velocity sensor fault detection method according to an embodiment of this application. The method includes the following steps: Step 1: Acquire angular velocity and voltage data from the angular velocity sensor during aircraft flight, as well as the aircraft's GPS data; acquire vibration, electromagnetic, and temperature data at the location of the angular velocity sensor in real time, and record these three types of data as environmental data.

[0022] The flight phase from takeoff to a predetermined altitude, entering a stable flight state, until preparation for descent is defined as the steady flight phase. This application uses an aircraft on a fixed flight path in the steady flight phase as an example. It collects angular velocity data from the aircraft's angular velocity sensor along the heading axis, as well as the voltage data output by the angular velocity sensor during operation. Vibration data at the location of the angular velocity sensor is collected using a vibration acceleration sensor; electromagnetic data at the location of the angular velocity sensor is collected using an electromagnetic sensor; temperature data at the location of the angular velocity sensor is collected using a temperature sensor; and GPS data of the aircraft is collected using a GPS locator. For ease of description, the vibration data, electromagnetic data, and temperature data at the location of the angular velocity sensor are collectively referred to as environmental data.

[0023] The velocity heading angle at each moment is calculated by integrating the angular velocity data of the heading axis; the position heading angle at each moment is calculated by using spherical trigonometry formulas to calculate the GPS data at each moment. The steps of calculating the heading angle from angular velocity and GPS data are all performed in a geographic coordinate system. Calculating the heading angle from angular velocity and GPS signals is a well-known technique in the field and will not be elaborated upon here.

[0024] All data is collected synchronously and in real time. In this embodiment, the data collection time interval is 1 second, the collection duration of each cycle is 60 minutes, and the data is collected continuously between cycles.

[0025] Taking the i-th cycle as an example, based on the data collected within the i-th cycle, the angular velocity sequence, voltage sequence, vibration sequence, electromagnetic sequence, temperature sequence, velocity heading angle sequence, and position heading angle sequence for the i-th cycle are constructed respectively. Since it is difficult to reflect whether the angular velocity sensor has degraded or failed based on data from only a single cycle, data from T cycles (200 in this embodiment) are collected.

[0026] To eliminate the influence of different dimensions among the data, all data are normalized. The normalization method is not limited to Z-score, maximum value normalization, or maximum and minimum value normalization.

[0027] Step 2: Divide the entire acquisition process into multiple cycles. Based on the proportional relationship between angular velocity and voltage data in each cycle, and the difference between the heading angle calculated from the angular velocity data and the aircraft GPS data, obtain the degradation factor for each cycle. Combine the difference between the degradation factor and the degradation factor of the previous cycle to obtain the degradation extension coefficient for each cycle.

[0028] Degradation refers to the process by which the performance of a device gradually decreases over time. During degradation, the performance of an angular velocity sensor will gradually weaken, and the errors in the collected data will become increasingly larger.

[0029] Under normal operating conditions, the angular velocity signal acquired by the angular velocity sensor is directly proportional to the output voltage signal, and the degree of proportionality is relatively stable. However, if performance degradation occurs, the degree of proportionality between the angular velocity signal and the voltage signal will fluctuate.

[0030] Taking the i-th cycle as an example for analysis, the dispersion of the ratio of angular velocity data to voltage data at all times in the i-th cycle is denoted as the first discrete value of the i-th cycle. The first discrete value reflects whether the proportional relationship between angular velocity and voltage has fluctuated significantly in the i-th cycle; the greater the fluctuation, the greater the degradation of the angular velocity sensor in the i-th cycle. The calculation of dispersion is not limited to variance, coefficient of variation, and standard deviation; this embodiment uses variance for calculation.

[0031] Furthermore, the average of the absolute differences between the velocity heading angle and the position heading angle at all times in the i-th cycle is denoted as the first average of the i-th cycle. The first average reflects the overall deviation between the aircraft heading angles calculated by the angular velocity sensor and the GPS locator in the i-th cycle. The larger the value, the greater the directional difference, and thus the greater the possibility of errors in the data collected by the angular velocity sensor.

[0032] The product of the first discrete value and the first mean value in the i-th period is denoted as the degradation factor for the i-th period. The degradation factor comprehensively reflects the degree of error in the data collected by the angular velocity sensor in the i-th period. The larger the value, the greater the error, and thus the greater the possibility that the angular velocity sensor has suffered a serious degradation failure in the i-th period.

[0033] Since it is difficult to reflect the degradation fault of the angular velocity sensor in a single cycle, the degradation extension coefficient of each cycle is obtained based on the difference between the degradation factor of each cycle and the degradation factor of the previous cycle, which is used to characterize the degree of degradation of the angular velocity sensor in each cycle.

[0034] In this embodiment, the degradation extension coefficient of the i-th period is denoted as... Its specific expression is: In the formula, Let be the degradation extension coefficient for the i-th period. Let be the degradation factor for the i-th period; The degradation factor is the factor of the previous cycle in the i-th cycle; This is a preset constant. To avoid the denominator being 0, its value range is [0.005, 0.01]. The value has little impact on the calculation result and can be ignored. In this embodiment, it is 0.008.

[0035] The degradation extension coefficient can reflect the degradation trend of the angular velocity sensor in the i-th cycle; the larger the value, the greater the error in the data collected by the angular velocity sensor in the i-th cycle compared with the previous adjacent cycle, the more serious the performance degradation of the angular velocity sensor, and the greater the possibility of degradation failure.

[0036] Step 3: Based on the causal relationship between various types of environmental data and angular velocity data in each period, the degree of influence of various types of environmental data on angular velocity data, and the dispersion of various types of environmental data in each period, obtain the overall disturbance index of various types of environmental data on angular velocity in each period.

[0037] Furthermore, during flight, the aircraft is affected by high-altitude temperatures and airflow vibrations, which may cause data fluctuations and lead to errors in the degradation extension coefficient. Therefore, it is necessary to further analyze whether the abnormal changes in angular velocity data are caused by environmental factors, in order to accurately assess whether the angular velocity sensor has experienced degradation failure.

[0038] Take the vibration data in the i-th cycle as an example.

[0039] The vibration sequence and angular velocity sequence of the i-th period are used as inputs to the Granger causality test algorithm, with the vibration sequence as the cause and the angular velocity sequence as the effect. The F-value output by the Granger causality test algorithm is denoted as the causality value of the vibration data with respect to the angular velocity in the i-th period. The causality value reflects whether there is a strong causal relationship between the vibration sequence and the angular velocity sequence; the larger the value, the greater the influence of the vibration data on the angular velocity data.

[0040] Considering that environmental factors such as temperature, vibration, and electromagnetic fields can have a coupling effect on angular velocity sensors, it is necessary to further analyze the degree of influence of various environmental data on angular velocity data under the combined effect of multiple factors.

[0041] In this embodiment, a regression model is constructed to analyze the impact of various environmental data on angular velocity data. The constructed regression model is as follows: In the formula, , , These are the angular velocity sequence, vibration sequence, temperature sequence, and electromagnetic sequence in the i-th period, respectively. , , These are the regression coefficients for vibration data, temperature data, and electromagnetic data in the i-th cycle, respectively, to reflect the influence of each type of environmental data on angular velocity; This represents the error term. The regression coefficients for various environmental data in the i-th period can be obtained by solving for the regression coefficients. The methods for solving the regression coefficients are not limited to least squares, ridge regression, or stepwise regression analysis. The Granger causality test algorithm and the construction and solution of the regression model are well-known techniques and will not be elaborated upon here.

[0042] Regression coefficients of vibration data It can reflect the degree of influence of vibration data on angular velocity data under the combined effect of multiple factors; the larger the value, the more significant the interference of vibration changes on the angular velocity sensor, even under the interference of multiple environmental factors.

[0043] In the i-th cycle, the normalized values ​​of the causal relationship between the vibration data and the angular velocity, and the normalized values ​​of the regression coefficients of the vibration data are calculated respectively. Then, the mean between the two normalized values ​​is recorded as the influence factor of the vibration data on the angular velocity in the i-th cycle. The influence factor reflects the comprehensive degree of influence of the vibration data changes on the angular velocity data in the i-th cycle; the larger the value, the more significant the interference of the vibration on the angular velocity sensor.

[0044] It should be noted that normalization is performed to avoid the problem of inconsistent dimensions causing a certain parameter to have an effect that is too small or too large. Normalization methods are not limited to the sigmoid function or the tanh function.

[0045] As a preferred implementation, based on the causal relationship between various types of environmental data and angular velocity data in each period, the degree of influence of various types of environmental data on angular velocity data, and the dispersion of various types of environmental data in each period, the overall disturbance index of various types of environmental data on angular velocity in each period is obtained, which is used to characterize the degree of interference of various types of environmental data on angular velocity in each period.

[0046] In this embodiment, the overall perturbation index of the vibration data with respect to angular velocity in the i-th period is denoted as . Its specific expression is: In the formula, is the overall perturbation index of the vibration data with respect to angular velocity during the i-th period; Let be the variance of the vibration sequence in the i-th period; is the influence factor of vibration data on angular velocity in the i-th cycle.

[0047] The overall disturbance index reflects the comprehensive interference intensity caused by vibration to the angular velocity sensor in the i-th period. The smaller the value, the smaller the change in vibration data in the i-th period, and the smaller the impact on the angular velocity data. If the angular velocity data shows a large anomaly, it can reflect the greater possibility that the angular velocity sensor has failed.

[0048] Similarly, following the method of obtaining the overall perturbation index of vibration data with respect to angular velocity in the i-th period, the overall perturbation index of electromagnetic data and temperature data with respect to angular velocity in the i-th period is obtained.

[0049] Step 4: Based on the angular velocity data of the current cycle and all previous cycles, as well as various environmental data, predict the angular velocity data and various environmental data for the next cycle; based on the difference in prediction deviation between the angular velocity data and various environmental data in the next cycle, the changing trend of the prediction deviation of the angular velocity data, the degradation extension coefficient, and the average level of the overall disturbance index of all types of environmental data for angular velocity in the next cycle, obtain the degradation confidence level for the next cycle, and then determine whether the angular velocity sensor has failed in the next cycle.

[0050] The degradation of the angular velocity sensor can be further assessed by predicting the angular velocity data in the next cycle and comparing the residuals between the predicted data and the actual acquired data.

[0051] In the i-th cycle, following the steps described above for calculating the influence factor of vibration data on angular velocity, calculate the influence factors of temperature data and electromagnetic data on angular velocity in the i-th cycle respectively. Then, the ratio of the influence factor of vibration data on angular velocity to the sum of all influence factors is recorded as the influence weight of vibration data in the i-th cycle. Calculate the mean of the influence weights of vibration data across all cycles, and record it as the average disturbance weight of vibration data on angular velocity. Similarly, calculate the average disturbance weights of temperature data and electromagnetic data on angular velocity.

[0052] The angular velocity data of all cycles are sorted in temporal order and used as a label sequence. The environmental data of all cycles are arranged in temporal order and denoted as various feature sequences. The label sequence, various feature sequences, and the average interference weights of various environmental data on angular velocity are used as inputs to train the prediction model. The trained model is then used to predict the angular velocity data of the next cycle.

[0053] The neural networks used to construct the prediction model are not limited to BP, LSTM, and GRU. This embodiment uses a BP neural network model. The activation function used for model training is not limited to the Sigmoid function, tanh function, and ReLU function. The optimization algorithm is not limited to SGD and Adam. The loss function is not limited to MSE, RMSE, and MAE. The training of the prediction model is a well-known technique and will not be described in detail here.

[0054] Based on the calculation method of the average interference weight of vibration data, temperature data, and electromagnetic data on angular velocity, the average interference weight of angular velocity, temperature data, and electromagnetic data on vibration data can be calculated; then, the vibration data in the next cycle can be predicted according to the angular velocity prediction method; similarly, various environmental data in the next cycle can be predicted.

[0055] Following the data acquisition method in step one, collect angular velocity data and various environmental data for the next cycle; then calculate the absolute difference between the predicted angular velocity data and the actual angular velocity data at all sampling times in the next cycle, and record the average of all absolute differences as the prediction deviation of the angular velocity data in the next cycle.

[0056] Similarly, calculate the prediction deviation of various environmental data in the next cycle.

[0057] The ratio of the prediction deviation of the angular velocity data in the next cycle to the maximum value of the prediction deviations of various environmental data is denoted as the relative deviation for the next cycle. The relative deviation reflects the relative severity of the prediction deviation of the angular velocity sensor output data in the next cycle relative to the prediction deviations of other environmental factors.

[0058] If the angular velocity sensor degrades, the degree of degradation is unpredictable and becomes increasingly severe, which will lead to a greater and greater deviation in the prediction of angular velocity data.

[0059] The prediction deviation of the angular velocity data for the next cycle and all previous cycles is arranged chronologically and denoted as the prediction deviation sequence for the next cycle. The mean of the first-order difference sequence of the prediction deviation sequence for the next cycle is then denoted as the second mean for the next cycle. A positive and larger second mean indicates a gradually increasing prediction deviation, reflecting a greater likelihood of significant performance degradation in the angular velocity sensor. Conversely, a negative and smaller second mean indicates increasingly higher prediction accuracy of the angular velocity data, suggesting a greater likelihood of the angular velocity sensor being functioning normally.

[0060] As a preferred implementation, the degradation confidence level for the next period is obtained based on the difference in prediction deviation between the angular velocity data and various environmental data within the next period, the changing trend of the prediction deviation of the angular velocity data, the degradation extension coefficient, and the average level of the overall perturbation index of all types of environmental data for angular velocity within the next period. This confidence level is used to characterize the probability of angular velocity sensor failure within the next period. The flowchart for obtaining the degradation confidence level for the next period is as follows: Figure 2 As shown.

[0061] In this embodiment, the degradation confidence level for the next cycle is denoted as... Its specific expression is: In the formula, The confidence level for degradation in the next cycle. This is the degradation extension coefficient for the next cycle; This represents the average of the overall perturbation exponents of angular velocity for all environmental data types within the next period. It is the ratio of the prediction deviation of the angular velocity data in the next cycle to the maximum value of the prediction deviation of various environmental data. This is the second average value for the next period; This is a preset constant.

[0062] Degradation confidence This value comprehensively reflects the overall probability and severity of performance degradation failure of the angular velocity sensor in the next cycle. The larger the value, the greater the possibility that the data collected by the angular velocity sensor still contains errors after excluding interference from external environmental factors, and that the degree of error gradually increases, indicating a greater likelihood of angular velocity sensor degradation failure.

[0063] Since only sensor data was saved during the initial acquisition phase for all historical cycles, the degradation confidence score needs to be calculated retrospectively after the prediction model training is completed. Specifically, the trained prediction model is applied sequentially to each historical cycle (using data from previous cycles as input to predict the angular velocity sequence and various environmental sequences for that cycle), and then the degradation confidence score for each historical cycle is obtained using the same calculation method as the degradation confidence score for the next cycle. The degradation confidence scores of the next cycle and all previous cycles are then used as input to the outlier detection algorithm to obtain... Value, of which The mean of all degradation confidence levels. The standard deviation of all degradation confidence levels; the obtained The value is recorded as the preset fault threshold. If the degradation confidence of the next cycle is less than the preset fault threshold, it is determined that the angular velocity sensor has not failed in the next cycle and is in good condition, and can still be used. Otherwise, it is determined that the angular velocity sensor has failed in the next cycle. At this time, the error of the data collected by the angular velocity sensor will be larger, and the decision cannot be made entirely based on the angular velocity data. The angular velocity sensor needs to be processed in a timely manner.

[0064] Based on the same inventive concept as the above methods, this application also provides a deep learning-based angular velocity sensor fault detection system, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above deep learning-based angular velocity sensor fault detection methods.

Claims

1. A fault detection method for an angular velocity sensor based on deep learning, characterized in that, Includes the following steps: Step S1: Real-time continuous acquisition of angular velocity sensor data during aircraft flight; the sensor data includes angular velocity, voltage, aircraft GPS data, velocity heading angle, position heading angle, and environmental data; the environmental data includes vibration values, electromagnetic values, and temperature values ​​at the location of the angular velocity sensor; the acquired sensor data is divided into T periods; within the i-th period, the angular velocity at different times, the voltage at different times, the aircraft GPS data at different times, the velocity heading angle at different times, the position heading angle at different times, the vibration values ​​at different times, the electromagnetic values ​​at different times, and the temperature values ​​at different times respectively constitute the angular velocity sequence, voltage sequence, GPS sequence, velocity heading angle sequence, and position heading angle sequence. The dataset consists of angular velocity sequences, vibration value sequences, temperature sequences, and electromagnetic value sequences. Within all periods, the vibration value sequences, temperature sequences, and electromagnetic value sequences for each period are arranged according to time sequence to obtain vibration feature sequences, temperature feature sequences, and electromagnetic feature sequences. The angular velocity sequences for all periods are arranged according to time sequence to form a label sequence. Within all periods, the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight for each period are arranged according to time sequence to obtain average vibration interference weight sequences, average electromagnetic interference weight sequences, and average temperature interference weight sequences. The dataset comprises the vibration feature sequences, temperature feature sequences, electromagnetic feature sequences, label sequences, average vibration interference weight sequences, average electromagnetic interference weight sequences, and average temperature interference weight sequences. Simultaneously calculate the degradation extension coefficient and the overall vibration disturbance index for each cycle. Overall electromagnetic disturbance index and overall temperature disturbance index ; Step S2: Construct a prediction model; train the prediction model using the dataset, and obtain the optimal prediction model when the prediction model meets the convergence criterion; the input of the prediction model includes the vibration value sequence, temperature value sequence, electromagnetic value sequence, angular velocity sequence, vibration average disturbance weight, electromagnetic average disturbance weight, and temperature average disturbance weight of the i-th period; the output of the prediction model is the predicted vibration sequence, predicted temperature sequence, predicted electromagnetic sequence, and predicted angular velocity sequence of the (i+1)-th period; Step S3: Obtain the vibration sequence, temperature sequence, electromagnetic sequence, and angular velocity sequence of the angular velocity sensor during the current period of aircraft flight, and calculate the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight. Input the vibration sequence, temperature sequence, electromagnetic sequence, angular velocity sequence, average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight of the current period into the optimal prediction model to obtain the predicted vibration sequence, predicted temperature sequence, predicted electromagnetic sequence, and predicted angular velocity sequence for the next period. Step S4: Acquire the real-time vibration sequence, real-time temperature sequence, real-time electromagnetic sequence, and real-time angular velocity sequence of the angular velocity sensor for the next cycle during aircraft flight; calculate the degradation confidence level for the next cycle. If the confidence level degrades If the value is less than the preset fault threshold, it is determined that the angular velocity sensor has not malfunctioned in the next cycle; if the confidence level degrades... If the angular velocity sensor fails within the next cycle, it is determined that the failure is greater than or equal to the preset fault threshold.

2. The deep learning-based angular velocity sensor fault detection method according to claim 1, characterized in that, The calculation process for the average vibration interference weight, average electromagnetic interference weight, and average temperature interference weight for each cycle is as follows: The vibration average interference weight is the average of the vibration influence weights for all cycles; the electromagnetic average interference weight is the average of the electromagnetic influence weights for all cycles; and the temperature average interference weight is the average of the temperature influence weights for all cycles. The vibration influence weight for each cycle is the ratio of the vibration disturbance factor for each cycle to the sum of the vibration disturbance factors for all cycles; the electromagnetic influence weight for each cycle is the ratio of the electromagnetic disturbance factor for each cycle to the sum of the electromagnetic disturbance factors for all cycles; the temperature influence weight for each cycle is the ratio of the temperature disturbance factor for each cycle to the sum of the temperature disturbance factors for all cycles. The calculation process for the vibration disturbance factor, electromagnetic disturbance factor, and temperature disturbance factor for each cycle is as follows: In all cycles, the vibration disturbance factor is obtained by using the disturbance index calculation method based on the vibration value and angular velocity in the environmental data; the electromagnetic disturbance factor is obtained by using the disturbance index calculation method based on the electromagnetic value and angular velocity in the environmental data; and the temperature disturbance factor is obtained by using the disturbance index calculation method based on the temperature value and angular velocity in the environmental data. The process of calculating the disturbance index is as follows: The input to the perturbation index calculation method includes perturbation data and perturbation angular velocity for all periods; the perturbation data and perturbation angular velocity for each period are input into the Granger causality test algorithm to obtain the F value; wherein the perturbation data is the cause, the perturbation angular velocity is the effect, and the F value is the causal value of the perturbation data relative to the perturbation angular velocity.

3. The deep learning-based angular velocity sensor fault detection method according to claim 1, characterized in that, The calculation of the degradation confidence level for the next cycle for: ； In the formula, The confidence level for degradation in the next cycle. This is the degradation extension coefficient for the next cycle; It is the average of the overall vibration disturbance index, the overall temperature disturbance index, and the overall electromagnetic disturbance index in the next cycle; The relative deviation for the next period; This is the second average value for the next period; This is a preset constant.

4. The deep learning-based angular velocity sensor fault detection method according to claim 3, characterized in that, The calculation process for the second mean is as follows: The angular velocity prediction deviations of the next cycle and all previous cycles are arranged chronologically to obtain the prediction deviation sequence for the next cycle. The mean of the first-order difference sequence of the prediction deviation sequence for the next cycle is calculated, which is the second mean for the next cycle. The angular velocity prediction deviation is the absolute value of the difference between the elements in the real-time angular velocity sequence and the elements in the predicted angular velocity sequence at the same time. The average of all absolute values ​​is then calculated to obtain the angular velocity prediction deviation. The vibration prediction deviation, temperature prediction deviation, and electromagnetic prediction deviation are calculated using the same method as the angular velocity prediction deviation. The relative deviation is obtained by acquiring the maximum value among the vibration prediction deviation, temperature prediction deviation, and electromagnetic prediction deviation, and calculating the ratio of the angular velocity prediction deviation to the maximum value, which is the relative deviation.

5. The deep learning-based angular velocity sensor fault detection method according to claim 3, characterized in that, The calculation process for the degradation extension coefficient is as follows: Based on the degradation factor for each cycle, calculate the degradation factor for each cycle divided by the degradation factor of the previous cycle and a preset constant. The sum of these values ​​yields the degradation extension coefficient; the degradation factor for each cycle is obtained as follows: Calculate the ratio of angular velocity to voltage at all times in each cycle to obtain the angular velocity-voltage ratio at all times in the cycle; then calculate the variance of the angular velocity-voltage ratio at all times in the cycle, and use the variance of the angular velocity-voltage ratio as the first discrete value for each cycle; in each cycle, use the spherical trigonometric formula to calculate the position heading angle at each time using GPS data at each time; calculate the absolute value of the difference between the velocity heading angle and the position heading angle at all times, and then calculate the mean of the absolute values ​​of the differences at all times, and use the mean as the first mean for each cycle; the product of the first discrete value and the first mean for each cycle is the degradation factor for each cycle.

6. The deep learning-based angular velocity sensor fault detection method according to claim 3, characterized in that, The overall vibration disturbance index Overall electromagnetic disturbance index and overall temperature disturbance index The calculation process is as follows: The overall vibration disturbance index relative to the angular velocity during the i-th period for: ； Let be the variance of the vibration sequence in the i-th period; The influence factor of vibration on angular velocity in the i-th period; The overall electromagnetic disturbance index of the electromagnetic sub-angular velocity during the i-th period for: ； Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; The overall temperature disturbance index of temperature versus angular velocity during the i-th period for: ； Let be the variance of the electromagnetic sequence in the i-th period; is the influence factor of electromagnetic field on angular velocity in the i-th period; Among them, the influence factor of vibration on angular velocity during the i-th period The mean values ​​of the vibration disturbance factor and vibration regression coefficient after normalization; the influence factor of electromagnetic field on angular velocity. The mean values ​​of the electromagnetic disturbance factor and electromagnetic regression coefficient after normalization; the influence factor of temperature on angular velocity. This represents the mean value of the temperature disturbance factor and the temperature regression coefficient after normalization.

7. The deep learning-based angular velocity sensor fault detection method according to claim 6, characterized in that, The solution process for the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient is as follows: Construct a regression model and solve for the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient for each cycle; The regression model for: In the formula, , , These are the angular velocity sequence, vibration sequence, temperature sequence, and electromagnetic sequence in the i-th period, respectively. This is the error term; , and These are the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient in the i-th period, respectively. , and It reflects the impact of each type of environmental data on angular velocity.

8. The deep learning-based angular velocity sensor fault detection method according to claim 7, characterized in that, The methods for solving the regression model to obtain the vibration regression coefficient, temperature regression coefficient, and electromagnetic regression coefficient include, but are not limited to, the least squares method, the ridge regression method, and the stepwise regression analysis method; the normalization methods include, but are not limited to, the Z-score method, the maximum value normalization method, and the maximum and minimum value normalization method.

9. The deep learning-based angular velocity sensor fault detection method according to claim 1, characterized in that, The prediction models include, but are not limited to, BP neural networks, LSTM neural networks, and GRU neural networks; The convergence criterion is a combination of any one or more of the following conditions: Condition A: The mean squared error loss function of the prediction model reaches the loss threshold; Condition B: The prediction model is trained for NP consecutive epochs, and the change in the mean squared error loss function value of the prediction model is less than the loss change threshold. Condition C: When the number of training rounds reaches the preset maximum number of iterations.

10. The deep learning-based angular velocity sensor fault detection method according to claim 1, characterized in that, The preset fault threshold is: ;in The mean of all degradation confidence levels. is the standard deviation of all degradation confidence levels.

Images