Online diagnostic method and system for water content abnormalities in aviation fuel pipelines
By constructing a Gaussian mixture model and dynamically adjusting the decision threshold, the accuracy and stability issues of abnormal water content in fuel pipeline diagnosis in existing technologies have been resolved. This enables efficient online diagnosis in complex flight environments and ensures the safety of aviation fuel systems.
Patent Information
- Application Number
- CN202611618263.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing diagnostic technologies for water content anomalies in aviation fuel pipelines cannot achieve accurate sensing and high-stability identification in complex flight environments, resulting in high false alarm rates or high false negative rates, and failing to meet the stringent requirements of modern aviation flight missions.
By acquiring real-time measurements of water content, temperature, and pressure in the fuel pipeline, a Gaussian mixture model is constructed. The decision threshold is dynamically adjusted in conjunction with flight phase parameters. The hypothesis model is adjusted using the cumulative log-likelihood ratio and the moving average, thereby enabling online diagnosis of abnormal water content in the fuel pipeline.
This improved the accuracy and stability of diagnosing abnormal water content in fuel pipelines, reduced the false alarm rate, and ensured the safe and reliable operation of aviation fuel systems.
Smart Images

Figure CN123116647A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of diagnostics, and in particular relates to an online diagnostic method and system for abnormal water content in aviation fuel pipelines. Background Technology
[0002] The aircraft engine is the core power source of an aircraft, and the safe, stable, and reliable operation of the aviation fuel system is a prerequisite for ensuring the normal operation of the engine and the safety of the entire flight mission. During fuel production, transportation, storage, refueling, and actual flight, a certain amount of moisture inevitably mixes into and accumulates in the fuel pipeline due to factors such as changes in external environmental humidity and alternating thermal stress. Especially during high-altitude flight, the ambient temperature drops sharply, and the fuel pipeline is in a state of extreme cold for a long time. The water dissolved in the fuel is very likely to precipitate out in large quantities due to the sudden drop in temperature, forming free water, which then quickly condenses into tiny ice crystals. These ice crystals, carried by the high-speed flow of fuel, can easily accumulate at bends in the pipeline and clog precision filters, pump components, or fuel metering valves in the system. This can lead to drastic fluctuations in fuel supply pressure and a decrease in engine thrust, or even catastrophic flight accidents such as engine failure in mid-air. Therefore, establishing a scientific mechanism for real-time online diagnosis and continuous monitoring of abnormal water content in aviation fuel pipelines, so as to take timely anti-icing or dewatering intervention measures before excessive water content causes substantial blockage, has extremely important engineering significance and practical application value for the flight safety of modern aircraft.
[0003] During different flight phases, including takeoff, climb, cruise, descent, and landing, the fluid pressure and temperature within the fuel piping system undergo drastic fluctuations. Furthermore, it is subject to a combination of non-steady-state interference from factors such as high-frequency airframe vibrations, flight attitude changes, and external airflow loads. These environmental conditions result in real-time data collected by piping sensors being mixed with significant amounts of non-stationary noise and random fluctuations. Traditional online diagnostic methods largely rely on pre-set static statistical distribution models or fixed decision thresholds for status assessment. They cannot proactively adapt to shifts in fluid properties caused by changes in fuel temperature and pressure, nor can they flexibly adjust warning standards based on the fluctuation trends of the current data sequence. This diagnostic approach, lacking an evolutionary mechanism, is highly prone to misinterpreting transient environmental disturbances as abnormal water content in the piping when faced with drastic changes in flight operation status or severe fluctuations in operating conditions, leading to a very high false alarm rate. Moreover, in the early stages of the gradual evolution of water content anomalies, the lack of tolerance in the decision thresholds can easily result in missed warnings. The aforementioned issues prevent existing diagnostic technologies from accurately detecting and reliably identifying abnormal water content in fuel lines, thus failing to meet the stringent requirements of modern high-reliability aviation flight missions for fuel line fault diagnosis systems. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention provides an online diagnostic method and system for abnormal water content in aviation fuel pipelines, which solves the technical problems in the prior art.
[0005] In a first aspect, the present invention provides the following technical solution: an online diagnostic method for abnormal water content in aviation fuel pipelines, comprising the following steps:
[0006] The system acquires real-time measurements of water content, fuel temperature, and fuel pressure from aviation fuel pipelines. Based on these measurements, it queries a multidimensional lookup table to determine a Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null and alternative hypotheses. When the cumulative log-likelihood ratio of the previous time step is greater than zero, it approximates the comprehensive mean of the alternative hypotheses towards the moving average of water content within a preset time window and shifts the mean of each Gaussian component to construct the hypothesis model. Based on the flight phase operation parameters, the baseline Type I and Type II error probabilities are selected to calculate the baseline upper decision threshold and the baseline lower decision threshold. The first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence at the previous moment is calculated. The frequency of symbol inversion is statistically analyzed based on the first-order difference symbol sequence. The corresponding amplitude features are extracted from the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion. The dynamic margin is determined based on the frequency and amplitude features. The dynamic margin is superimposed on the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold. The single-step log-likelihood ratio is calculated using the assumed model and added to the cumulative log-likelihood ratio of the previous time step to obtain the current cumulative log-likelihood ratio. If the current cumulative log-likelihood ratio is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero; if it is less than the lower decision threshold, it is determined to be normal and cleared to zero; if it is between the two, the conclusion is maintained and the diagnostic process proceeds to the next time step.
[0007] Preferably, the acquisition of real-time water content, fuel temperature, and fuel pressure measurements from the aviation fuel pipeline includes: Configure the timer in the embedded microprocessor module of the system to generate an interrupt pulse of the reference period to trigger a synchronous sampling event; Send synchronous data acquisition commands to sensor nodes via the system communication bus; The system receives and parses the returned data packets with timestamps, extracts the measured values of water content, fuel temperature, and fuel pressure, and stores them in the specified memory buffer.
[0008] Preferably, the step of determining the baseline probability density function of the Gaussian mixture model representing normal and abnormal states as the null and alternative hypotheses based on the fuel temperature and fuel pressure measurements by querying a multidimensional lookup table includes: Within the full-range temperature and pressure range, a two-dimensional grid is divided according to a preset step size. For each grid node, the corresponding null hypothesis and alternative hypothesis Gaussian mixture model parameter set under the calibration state is stored in advance. The parameter set includes the weight, mean and variance of each Gaussian component to form an offline mapping dictionary. During diagnosis, the collected fuel temperature and fuel pressure measurements are matched to the nearest two-dimensional grid node; The memory address offset is calculated based on the coordinates of the matching node, and the corresponding parameter set is read from the offline mapping dictionary by a pointer as the baseline probability density function.
[0009] Preferably, the step of approximating the composite mean of the alternative hypotheses towards the moving average of moisture content within a preset time window and shifting the mean of each Gaussian component to construct the hypothesis model when the cumulative log-likelihood ratio of the previous time moment is greater than zero includes: If the cumulative log-likelihood ratio of the previous time step is greater than zero, calculate the moving average of the moisture content measurements within the preset time window; The adjusted weighting coefficient is calculated as one minus the exponent of the product of the negative preset constant with the natural constant as the base and the cumulative log-likelihood ratio of the previous time step. Calculate the weighted composite benchmark mean of each Gaussian component of the alternative hypothesis, and update the benchmark mean of each Gaussian component of the alternative hypothesis to a dynamic mean, wherein the dynamic mean is the benchmark mean plus the product of the adjustment weight coefficient and the moving average minus the weighted composite benchmark mean; When the cumulative log-likelihood ratio of the previous time step is not greater than zero, the weight coefficients are adjusted to zero, and the baseline mean of each Gaussian component is kept unchanged so that the alternative hypothesis model smoothly tracks the drift characteristics of the actual moisture content based on the confidence level of the anomaly.
[0010] Preferably, the step of selecting the baseline Type I and Type II error probabilities based on flight phase operation parameters and calculating the baseline upper decision threshold and the baseline lower decision threshold includes: The current flight characteristic parameters are analyzed to identify the flight stage, and the baseline Type I error probability and baseline Type II error probability corresponding to the flight stage are retrieved from the preset dictionary. Calculate the ratio of the difference between 1 and the baseline Type II error probability to the baseline Type I error probability, and use the natural logarithm of the ratio as the decision threshold on the baseline. The quotient obtained by dividing the baseline Type II error probability by 1 and subtracting the baseline Type I error probability is used as the decision threshold under the baseline. The calculated upper and lower decision thresholds are stored in the decision register as the base reference values for applying dynamic margins subsequently.
[0011] Preferably, the step of calculating the first-order difference symbol sequence after removing the zero-transition from the historical sequence of the cumulative log-likelihood ratio at the previous time step, statistically analyzing the frequency of symbol inversion based on the first-order difference symbol sequence, extracting the corresponding amplitude features from the historical sequence of the cumulative log-likelihood ratio according to the position index of the symbol inversion, determining a dynamic margin based on the frequency and amplitude features, and superimposing the dynamic margin onto the benchmark upper decision threshold and the benchmark lower decision threshold to obtain the upper decision threshold and the lower decision threshold, includes: Extract the cumulative log-likelihood ratio history sequence of a preset length, calculate the difference between adjacent sampling points, remove the difference caused by the zeroing operation, and then extract the sign function to generate a first-order difference sign sequence composed of +1 and -1; The total number of symbol inversions in the first-order difference symbol sequence is used as the frequency feature. Based on the position index of the symbol inversion, the cumulative log-likelihood ratio historical sequence is traced back, and the average absolute value of the difference between the local extrema before and after all inversion points in the historical sequence is calculated as the amplitude feature. Calculate the first product of the first preset positive weight coefficient and the frequency feature, and the second product of the second preset positive weight coefficient and the amplitude feature, respectively. The sum of the first product and the second product is used as the dynamic margin. The upper decision threshold is obtained by adding dynamic margin to the upper decision threshold, and the lower decision threshold is obtained by subtracting dynamic margin from the lower decision threshold. This is to suppress misjudgments caused by high-frequency oscillations in the log-likelihood ratio due to measurement noise.
[0012] Secondly, the present invention provides the following technical solution: an online diagnostic system for abnormal water content in aviation fuel pipelines, comprising the following modules: The module is used to acquire real-time measurements of water content, fuel temperature, and fuel pressure from aviation fuel pipelines. Based on the fuel temperature and fuel pressure measurements, a multidimensional lookup table is queried to determine a Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null hypothesis and alternative hypothesis. When the cumulative log-likelihood ratio of the previous time step is greater than zero, the comprehensive mean of the alternative hypothesis is approximated towards the moving average of water content within a preset time window, and the mean of each Gaussian component is shifted to construct the hypothesis model. The calculation module is used to select the baseline Type I and Type II error probabilities based on the flight phase operation parameters, calculate the baseline upper decision threshold and the baseline lower decision threshold, calculate the first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence at the previous moment, count the frequency of symbol inversion based on the first-order difference symbol sequence, extract the corresponding amplitude features in the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion, determine the dynamic margin based on the frequency and amplitude features, and superimpose the dynamic margin onto the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold. The determination module is used to calculate the single-step log-likelihood ratio using the assumed model and add it to the cumulative log-likelihood ratio of the previous time step to obtain the cumulative log-likelihood ratio of the current time step. If the cumulative log-likelihood ratio of the current time step is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero. If it is less than the lower decision threshold, it is determined to be normal and cleared to zero. If it is between the two, the conclusion is maintained and the next time step diagnosis process is entered.
[0013] Preferably, the acquisition of real-time water content, fuel temperature, and fuel pressure measurements from the aviation fuel pipeline includes: Configure the timer in the embedded microprocessor module of the system to generate an interrupt pulse of the reference period to trigger a synchronous sampling event; Send synchronous data acquisition commands to sensor nodes via the system communication bus; The system receives and parses the returned data packets with timestamps, extracts the measured values of water content, fuel temperature, and fuel pressure, and stores them in the specified memory buffer.
[0014] Preferably, the step of determining the baseline probability density function of the Gaussian mixture model representing the normal and abnormal states as the null and alternative hypotheses based on the fuel temperature and fuel pressure measurements by querying a multidimensional lookup table includes: Within the full-range temperature and pressure range, a two-dimensional grid is divided according to a preset step size. For each grid node, the corresponding null hypothesis and alternative hypothesis Gaussian mixture model parameter set under the calibration state is stored in advance. The parameter set includes the weight, mean and variance of each Gaussian component to form an offline mapping dictionary. During diagnosis, the collected fuel temperature and fuel pressure measurements are matched to the nearest two-dimensional grid node; The memory address offset is calculated based on the coordinates of the matching node, and the corresponding parameter set is read from the offline mapping dictionary by a pointer as the baseline probability density function.
[0015] Preferably, the step of approximating the composite mean of the alternative hypotheses towards the moving average of moisture content within a preset time window and shifting the mean of each Gaussian component to construct the hypothesis model when the cumulative log-likelihood ratio of the previous time moment is greater than zero includes: If the cumulative log-likelihood ratio of the previous time step is greater than zero, calculate the moving average of the moisture content measurements within the preset time window; The adjusted weighting coefficient is calculated as one minus the exponent of the product of the negative preset constant with the natural constant as the base and the cumulative log-likelihood ratio of the previous time step. Calculate the weighted composite benchmark mean of each Gaussian component of the alternative hypothesis, and update the benchmark mean of each Gaussian component of the alternative hypothesis to a dynamic mean, wherein the dynamic mean is the benchmark mean plus the product of the adjustment weight coefficient and the moving average minus the weighted composite benchmark mean; When the cumulative log-likelihood ratio of the previous time step is not greater than zero, the weight coefficients are adjusted to zero, and the baseline mean of each Gaussian component is kept unchanged so that the alternative hypothesis model smoothly tracks the drift characteristics of the actual moisture content based on the confidence level of the anomaly.
[0016] Preferably, the step of selecting the baseline Type I and Type II error probabilities based on flight phase operation parameters and calculating the baseline upper decision threshold and the baseline lower decision threshold includes: The current flight characteristic parameters are analyzed to identify the flight stage, and the baseline Type I error probability and baseline Type II error probability corresponding to the flight stage are retrieved from the preset dictionary. Calculate the ratio of the difference between 1 and the baseline Type II error probability to the baseline Type I error probability, and use the natural logarithm of the ratio as the decision threshold on the baseline. The quotient obtained by dividing the baseline Type II error probability by 1 and subtracting the baseline Type I error probability is used as the decision threshold under the baseline. The calculated upper and lower decision thresholds are stored in the decision register as the base reference values for applying dynamic margins subsequently.
[0017] Preferably, the step of calculating the cumulative log-likelihood ratio (CRL) history sequence after removing the zero-transition first-order difference symbol sequence, calculating the frequency of symbol inversion based on the first-order difference symbol sequence, extracting the corresponding amplitude features in the CRL history sequence according to the position index of the symbol inversion, determining a dynamic margin based on the frequency and amplitude features, and superimposing the dynamic margin onto the benchmark upper decision threshold and benchmark lower decision threshold to obtain the upper decision threshold and lower decision threshold includes: Extract the cumulative log-likelihood ratio history sequence of a preset length, calculate the difference between adjacent sampling points, remove the difference caused by the zeroing operation, and then extract the sign function to generate a first-order difference sign sequence composed of +1 and -1; The total number of symbol inversions in the first-order difference symbol sequence is used as the frequency feature. Based on the position index of the symbol inversion, the cumulative log-likelihood ratio historical sequence is traced back, and the average absolute value of the difference between the local extrema before and after all inversion points in the historical sequence is calculated as the amplitude feature. Calculate the first product of the first preset positive weight coefficient and the frequency feature, and the second product of the second preset positive weight coefficient and the amplitude feature, respectively. The sum of the first product and the second product is used as the dynamic margin. The upper decision threshold is obtained by adding dynamic margin to the upper decision threshold, and the lower decision threshold is obtained by subtracting dynamic margin from the lower decision threshold. This is to suppress misjudgments caused by high-frequency oscillations in the log-likelihood ratio due to measurement noise.
[0018] This invention acquires real-time measurements of water content, temperature, and pressure in pipelines, establishes a baseline probability density function using a multidimensional lookup table, and adjusts the alternative hypothesis mean in real-time based on the cumulative log-likelihood ratio and the moving average of water content from the previous time step. This adapts the diagnostic model to the actual conditions of different pipeline operating states, improving the detection accuracy of abnormal features. A baseline threshold is selected by combining flight operation parameters, and the fluctuation patterns and sign reversal characteristics of the likelihood ratio historical sequence are analyzed to calculate the compensation margin, obtaining more reasonable upper and lower decision thresholds in real time. This mechanism eliminates the shortcomings of traditional fixed thresholds, which are prone to false alarms and false negatives in complex environments, enhancing the anti-interference capability and stability of the inspection process. By comparing the current likelihood ratio with the real-time threshold, online diagnosis of water content anomalies in aviation fuel pipelines is achieved with an extremely low false positive rate. Attached Figure Description
[0019] Figure 1 A flowchart of an online diagnostic method for abnormal water content in aviation fuel pipelines; Figure 2 This is a schematic diagram of the probability density distribution of a Gaussian mixture model. Figure 3 This is a schematic diagram comparing the results of the ablation experiment. Detailed Implementation
[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.
[0021] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.
[0022] This specification provides one or more embodiments of an online diagnostic method for abnormal water content in aviation fuel pipelines, such as... Figure 1 As shown, it includes the following steps: S1. Obtain real-time measurements of water content, fuel temperature, and fuel pressure from the aviation fuel pipeline. Based on the fuel temperature and fuel pressure measurements, query a multidimensional lookup table to determine the Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null hypothesis and alternative hypothesis. Adjust the mean of the alternative hypothesis using the cumulative log-likelihood ratio of the previous time step. When the cumulative log-likelihood ratio of the previous time step is greater than zero, approximate the comprehensive mean of the alternative hypothesis towards the moving average of water content within a preset time window and shift the mean of each Gaussian component to construct the hypothesis model.
[0023] Continuous analog signals are acquired by a capacitive moisture content sensor, along with high-precision temperature and pressure transmitters installed on the aviation fuel pipeline. These analog signals are then converted to digital signals by a 16-bit analog-to-digital converter and input into a microprocessor to construct a real-time data array. One-dimensional median filtering is applied to eliminate high-frequency noise during hardware acquisition, yielding the cleaned moisture content, fuel temperature, and fuel pressure measurements. Prior to this, the Gaussian Mixture algorithm, based on the expectation-maximization mechanism, is used offline to cluster and train historical normal and abnormal datasets under different temperature and pressure conditions. The mean and covariance matrices of the Gaussian mixture model under each operating condition are extracted, and a multidimensional lookup table is constructed with temperature and pressure as discrete index dimensions. The structure of the dimension lookup table is a two-dimensional grid array with fuel temperature and fuel pressure as discrete index dimensions. The two-dimensional grid consists of 1250 grid nodes formed by orthogonal combination of 25 discrete nodes for full-range temperature (e.g., -50℃ to +70℃, step size 5℃) and 50 discrete nodes for full-range pressure (e.g., 0.1MPa to 5.0MPa, step size 0.1MPa). Each grid node contains a Gaussian mixture model parameter set for the null and alternative hypotheses under that specific temperature and pressure condition. For example, when the Gaussian model is set to have 3 components, the parameter set contains 18 32-bit single-precision floating-point numbers, including the weight coefficients, expected mean, and distribution variance of each component. Each node occupies 72 bytes and can be mapped to a memory address offset through two-dimensional coordinate indexing.
[0024] Real-time filtered temperature and pressure measurements are input into a multidimensional lookup table. A bilinear interpolation algorithm is used to calculate the normal state Gaussian mixture model parameters matching the current operating conditions, serving as the baseline probability density function for the null hypothesis. Interpolation is then used to calculate the abnormal state Gaussian mixture model parameters, serving as the baseline probability density function for the alternative hypothesis. The baseline probability density function provides initial statistical reference for online diagnostics of aviation fuel pipelines under specific temperature and pressure conditions. By inputting real-time acquired temperature and pressure data into the multidimensional lookup table and performing bilinear interpolation, Gaussian mixture model parameters matching the current operating conditions are calculated. These parameters constitute the initial basic probability distribution patterns representing the "normal" and "abnormal" states of the pipeline. The baseline function is not only used to calculate the total probability under the null hypothesis but also serves as the starting point for the dynamic evolution of the alternative hypothesis parameters, forming the basis for calculating the single-step log-likelihood ratio and constructing the adaptive diagnostic model. The null hypothesis and the alternative hypothesis provide two opposing statistical classification criteria for the sequential probability ratio (SPR) test of the system, enabling quantitative discrimination of the water content state of the fuel pipeline. The null hypothesis represents the pipeline being in a normal state of no water or with water content within acceptable limits, while the alternative hypothesis represents the pipeline being in an abnormal state with excessive water content. During real-time diagnosis, the currently collected water content data is substituted into the models for both hypotheses to calculate the probability of occurrence. The natural logarithm of the quotient of the alternative hypothesis probability and the null hypothesis probability, i.e., the log-likelihood ratio, is used to continuously assess the confidence level of the current data in favoring an abnormal state.
[0025] A fixed-length first-in-first-out queue is created to store historical moisture content data within a preset time window, and the moving average of moisture content within this time window is calculated. The cumulative log-likelihood ratio state variable from the previous moment is extracted from memory, and conditional judgment logic is executed. When the value of this variable is greater than zero, the difference between the moving average of moisture content and the comprehensive mean calculated by the weighted average of the Gaussian components of the current alternative hypothesis is calculated. This difference is multiplied by a preset learning rate to update the comprehensive mean of the alternative hypothesis, making the mean approximate the moving average. The Euclidean offset distance between the updated comprehensive mean and the original comprehensive mean is extracted, and this offset distance is proportionally added to the mean parameter of each Gaussian component of the alternative hypothesis Gaussian mixture model to complete the mean shift, constructing a hypothesis model adapted to the drift trend of the current operating conditions.
[0026] The hypothetical model is a Gaussian mixture model (GMM) with dynamic mean adjustment, composed of a weighted combination of multiple, for example, three Gaussian components. The structural parameters include fixed weight coefficients and covariance matrices for each Gaussian component, as well as dynamically iteratively updated expected mean. Based on the cumulative log-likelihood ratio state of the previous time step, the initial baseline mean of each Gaussian component of the alternative hypothesis is approximated and shifted towards the recent moving average of water content.
[0027] In one embodiment, acquiring real-time measurements of water content, fuel temperature, and fuel pressure from the aviation fuel pipeline includes: Configure the timer in the embedded microprocessor module of the system to generate an interrupt pulse of the reference period to trigger a synchronous sampling event; Send synchronous data acquisition commands to sensor nodes via the system communication bus; The system receives and parses the returned data packets with timestamps, extracts the measured values of water content, fuel temperature, and fuel pressure, and stores them in the specified memory buffer.
[0028] A timer is used to generate a reference interrupt pulse with a period of 50ms. Each interrupt triggers a synchronous sampling event. The controller sends a synchronous data acquisition command with a specific command word to the distributed sensor nodes via a CAN2.0B industrial bus or RS485 communication bus, configured with a baud rate of 500kbps, in broadcast frame format. Each sensor node synchronously completes the AD conversion within a limited window time (e.g., 2ms) after receiving the command and encapsulates the result into a data packet, returning it to the main controller.
[0029] The bus controller receives the returned data packets in a non-blocking manner via the DMA channel and performs CRC-16 hardware verification. After successful verification, the microprocessor parses the 64-bit Unix timestamp identifier contained in the data header and extracts the measured values of water content (range 0 to 500 ppm), fuel temperature (range -60°C to 80°C), and fuel pressure (range 0 to 5.0 MPa) in floating-point format. All extracted measurement parameters and their corresponding timestamps are pushed sequentially into a pre-allocated circular memory buffer in RAM with a depth of 1024 nodes. This buffer uses head and tail pointers for queue update management to ensure the consistency of continuous observation time series, reduce data latency, and enable traceability.
[0030] In one embodiment, determining the baseline probability density function of the Gaussian mixture model representing normal and abnormal states as the null and alternative hypotheses based on the fuel temperature and fuel pressure measurements by querying a multidimensional lookup table includes: Within the full-range temperature and pressure range, a two-dimensional grid is divided according to a preset step size. For each grid node, the corresponding null hypothesis and alternative hypothesis Gaussian mixture model parameter set under the calibration state is stored in advance. The parameter set includes the weight, mean and variance of each Gaussian component to form an offline mapping dictionary. During diagnosis, the collected fuel temperature and fuel pressure measurements are matched to the nearest two-dimensional grid node; The memory address offset is calculated based on the coordinates of the matching node, and the corresponding parameter set is read from the offline mapping dictionary by a pointer as the baseline probability density function.
[0031] During the calibration phase, the full-range temperature range of aviation fuel was set to -50℃ to +70℃, with a preset temperature step size of 5℃ and a total of 25 temperature discrete nodes. The full-range pressure range was set to 0.1MPa to 5.0MPa, with a preset pressure step size of 0.1MPa and a total of 50 pressure discrete nodes. A two-dimensional discrete mesh containing 1250 grid nodes was constructed through orthogonal combination. For each independent grid node, the parameter set of the Gaussian mixture model for the null hypothesis of anhydrous state and the alternative hypothesis of aqueous state, obtained from the experimental calibration, was pre-programmed into the non-volatile Flash memory. If the number of Gaussian model components K=3, the parameter set contains the weight coefficients of each of the three components, such as... , , The offline mapping dictionary contains 18 32-bit single-precision floating-point numbers, including the expected mean and distribution variance, and occupies approximately 90KB of contiguous memory address blocks.
[0032] In the real-time online diagnostic process, the current real-time measurement parameters are obtained, such as the current temperature of 21.3℃ and the pressure of 1.04MPa. The rounding algorithm is called to match the measurement parameters to the grid point with the closest Euclidean distance in space, namely the temperature node of 20℃, corresponding to the first-dimensional coordinate index i=14 and the pressure node of 1.0MPa, corresponding to the second-dimensional coordinate index j=9. Using the obtained index, according to the mapping formula Offset=(i×50+j)×72 bytes, the absolute address offset after the base address of the parameter storage memory is calculated. The calculated address is used to perform O(1) complexity memory access on the mapping dictionary, and the read parameter set is loaded into the cache area as the input variable for constructing the baseline probability density function of the null hypothesis and alternative hypothesis required for diagnostic testing, replacing the online model fitting, such as Figure 2 As shown in the figure, the solid line corresponds to the null hypothesis probability density curve under normal operating conditions, and the dashed line corresponds to the alternative hypothesis probability density curve under abnormal water content conditions. Samples with normal water content cluster around the 0 ppm area, while abnormal samples are dispersed in the 20–60 ppm range.
[0033] In one embodiment, adjusting the mean of the alternative hypothesis using the cumulative log-likelihood ratio of the previous time step, and when the cumulative log-likelihood ratio of the previous time step is greater than zero, approximating the overall mean of the alternative hypothesis towards the moving average of moisture content within a preset time window and shifting the mean of each Gaussian component to construct the hypothesis model, includes: If the cumulative log-likelihood ratio of the previous time step is greater than zero, calculate the moving average of the moisture content measurements within the preset time window; The adjusted weighting coefficient is calculated as one minus the exponent of the product of the negative preset constant with the natural constant as the base and the cumulative log-likelihood ratio of the previous time step. Calculate the weighted composite benchmark mean of each Gaussian component of the alternative hypothesis, and update the benchmark mean of each Gaussian component of the alternative hypothesis to a dynamic mean, wherein the dynamic mean is the benchmark mean plus the product of the adjustment weight coefficient and the moving average minus the weighted composite benchmark mean; When the cumulative log-likelihood ratio of the previous time step is not greater than zero, the weight coefficients are adjusted to zero, and the baseline mean of each Gaussian component is kept unchanged so that the alternative hypothesis model smoothly tracks the drift characteristics of the actual moisture content based on the confidence level of the anomaly.
[0034] The diagnostic algorithm sets a sliding time window with a sampling length N of 20 diagnostic cycles in its internal cache. At each diagnostic time, the arithmetic mean of the water content measurement observation sequence within the above 20 cycles is calculated using FIFO queue data to serve as the current sliding average, for example, an average result of 35.5 ppm. The cumulative log-likelihood ratio output from the previous time step is read. If detected ,For example This indicates that the current state exhibits characteristics of excessive moisture content. Adjustment weighting coefficients are obtained using exponential function calculations. The mathematical expression is The attenuation adjustment preset constant C ranges from 0.1 to 0.5; in this example, it is set to 0.2. The calculated adjustment weight coefficient is then... The index construction will It is limited to the range of 0 to 1.
[0035] Based on this calculation, the mean values are extracted from the three alternative hypotheses Gaussian components obtained from the table and set as follows: , , With the preset prior weight probabilities of each component The weighted average benchmark mean is calculated using the full expectation formula. For example, the calculated value is 30.0 ppm. The initial baseline mean for each Gaussian component is updated according to the difference scaling formula: if the original baseline mean for the first component is set to 28.0 ppm, then the dynamic mean is updated to... ppm. If the detection determines... This indicates that the current state is normal, and the program logic adjusts the weighting coefficients. Setting the values to zero stops the model parameter evolution, locking the model in its initial standard baseline state. This dynamic mean iteration mechanism enables the diagnostic model to adjust its mean cluster centers according to the confidence level, shifting the parameters that have deviated from the average value of the pipeline operating parameters within the current time window.
[0036] S2, based on the flight phase operation parameters, select the baseline Type I and Type II error probabilities, calculate the baseline upper decision threshold and the baseline lower decision threshold, calculate the first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence of the previous moment, statistically analyze the frequency of symbol inversion based on the first-order difference symbol sequence, and extract the corresponding amplitude features in the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion, determine the dynamic margin based on the frequency and amplitude features, and superimpose the dynamic margin onto the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold.
[0037] By parsing the airborne ARINC429 data bus, the system reads the flight phase status codes of the aircraft, such as takeoff, climb, cruise, or landing, and retrieves the baseline Type I error probability corresponding to the current flight phase from the internally preset fault-tolerant mapping dictionary. Type II error probability Perform the Wald approximate logarithm calculation, and the specific value of the decision threshold on the calculation basis is 1 minus Difference divided by Then, the natural logarithm is calculated, and the specific value of the decision threshold under the given benchmark is determined. Divide by 1 and subtract The natural logarithm is calculated after the difference. The cumulative log-likelihood ratio historical time series stored in the cache queue is extracted. The backward difference values of the sequence values at adjacent sampling times are calculated. An absolute value threshold for abnormal jumps is set. The `where` function of the NumPy library is used to locate the array index positions where the absolute difference value is greater than the threshold and the current state value is zero, identifying these as zero-clear jump points. Data corresponding to zero-clear jump points in the sequence is removed. The sign features are extracted from the continuous difference sequence after removing jump interference, generating a first-order difference sign sequence containing only +1, -1, and 0 values.
[0038] The index of the non-zero result in the first-order difference symbol sequence is found. The total number of sign reversals within a set historical time window is used as the frequency feature. The absolute value of the peak-to-trough difference of the cumulative log-likelihood ratio before and after each sign reversal event is extracted based on the reversal index. The average of these absolute values is calculated as the amplitude feature. A linear weighted summation algorithm is used to multiply the sign reversal frequency by a first weight constant and the amplitude by a second weight constant to calculate the dynamic margin value. The dynamic margin value is added to the benchmark upper decision threshold as a penalty term using an addition operator to obtain the upper decision threshold. The same dynamic margin value is subtracted from the benchmark lower decision threshold to obtain the lower decision threshold, thus realizing a mechanism to widen the decision interval under conditions of severe data fluctuation.
[0039] In one embodiment, the step of selecting the baseline Type I and Type II error probabilities based on flight phase operation parameters and calculating the baseline upper decision threshold and the baseline lower decision threshold includes: The current flight characteristic parameters are analyzed to identify the flight stage, and the baseline Type I error probability and baseline Type II error probability corresponding to the flight stage are retrieved from the preset dictionary. Calculate the ratio of the difference between 1 and the baseline Type II error probability to the baseline Type I error probability, and use the natural logarithm of the ratio as the decision threshold on the baseline. The quotient obtained by dividing the baseline Type II error probability by 1 and subtracting the baseline Type I error probability is used as the decision threshold under the baseline. The calculated upper and lower decision thresholds are stored in the decision register as the base reference values for applying dynamic margins subsequently.
[0040] The peripheral interface periodically parses and reads dynamic flight characteristic parameters such as barometric altitude, air Mach number, and current gas turbine engine speed from the flight control and management system. The diagnostic program internally nests a decision tree classification algorithm to identify the aircraft's current operational phase, such as taxiing / takeoff, cruise, and descent / landing. A pre-stored dictionary is configured with different fault tolerance threshold pairs for different flight phases. Taking the cruise phase as an example, the corresponding baseline Type I error probability and false alarm rate are retrieved from this dictionary. And the benchmark Type II error probability and false alarm rate .
[0041] After obtaining the stage-specific confidence parameters, the calculation is performed according to the threshold calculation rule of the sequential probability ratio test theory: 1 is subtracted from 0.05, divided by the baseline Type I error probability of 0.01 to obtain a scalar of 95. The scalar is then subjected to a natural logarithm function call based on Taylor series expansion or the CORDIC algorithm to obtain an upper bound ln(95)≈4.553, which serves as the baseline upper decision threshold for determining an abnormal situation. A similar process is used to obtain the natural logarithm of 0.0505, resulting in a lower bound ln(0.0505)≈-2.985, which serves as the baseline lower decision threshold for determining normality. The calculated floating-point results 4.553 and -2.985 are assigned to the designated global shared decision register, serving as the baseline for judgment boundaries and adding error tolerance bands in the subsequent real-time sequence discrimination stage.
[0042] In one embodiment, the step of calculating the first-order difference symbol sequence after removing the zero-transition from the historical sequence of the cumulative log-likelihood ratio at the previous time step, calculating the frequency of symbol inversion based on the first-order difference symbol sequence, extracting the corresponding amplitude features from the historical sequence of the cumulative log-likelihood ratio based on the position index of the symbol inversion, determining a dynamic margin based on the frequency and amplitude features, and superimposing the dynamic margin onto the benchmark upper decision threshold and the benchmark lower decision threshold to obtain the upper decision threshold and the lower decision threshold, includes: Extract the cumulative log-likelihood ratio history sequence of a preset length, calculate the difference between adjacent sampling points, remove the difference caused by the zeroing operation, and then extract the sign function to generate a first-order difference sign sequence composed of +1 and -1; The total number of symbol inversions in the first-order difference symbol sequence is used as the frequency feature. Based on the position index of the symbol inversion, the cumulative log-likelihood ratio historical sequence is traced back, and the average absolute value of the difference between the local extrema before and after all inversion points in the historical sequence is calculated as the amplitude feature. Calculate the first product of the first preset positive weight coefficient and the frequency feature, and the second product of the second preset positive weight coefficient and the amplitude feature, respectively. The sum of the first product and the second product is used as the dynamic margin. The upper decision threshold is obtained by adding dynamic margin to the upper decision threshold, and the lower decision threshold is obtained by subtracting dynamic margin from the lower decision threshold. This is to suppress misjudgments caused by high-frequency oscillations in the log-likelihood ratio due to measurement noise.
[0043] The embedded control system maintains a FIFO (First-In, First-Out) data history queue sequence in real time, with a length N and a cumulative log-likelihood ratio configured to be 100. During each sampling period, the program kernel performs first-order forward difference operations in parallel between two adjacent sampling points in the sequence. To eliminate jump interference caused by system reset operations, such as a step difference from +4.5 to 0, a data cleaning subroutine is executed to identify and filter out such jump difference points by comparing thresholds. The remaining continuous differences are mapped to 1 to represent an increasing trend and -1 to represent a decreasing trend sequence using a built-in sign identification operator, generating a first-order difference sign array sequence representing trend fluctuations.
[0044] The data statistics calculator traverses the symbol array sequence to identify feature points where the product of two adjacent symbol elements equals -1, and records the total number of positive and negative fluctuations within the current monitoring time window. As a frequency characteristic indicating the frequency of oscillations, for example, there were a total of 18 reversals; by tracing back the difference between the peak and trough of each reversal point, accumulating the absolute values of the differences, and then calculating the average, we obtain the result. As a characteristic of the average amplitude of the fluctuation, for example, the average amplitude is 0.8. Based on the measured model of pipeline disturbance, a first preset positive weighting coefficient is set. The range is 0.05 to 0.15; in this example, it is set to 0.05. This is the second preset positive weighting coefficient. The range is 0.3 to 0.8; in this example, it is set to 0.5, and the dynamic margin value is calculated. The arithmetic logic unit is used to determine the threshold on the baseline. For example, adding a dynamic margin of 1.3 to 4.553 updates the upper threshold to 5.853. Subtracting 1.3 from the lower threshold, such as -2.985, yields the lower threshold of -4.285. This widens the region to be determined and suppresses the triggering of the likelihood ratio test quantity and false alarms caused by microbubbles in the gas-liquid two-phase system or mechanical vibrations of the pipe wall.
[0045] S3. Calculate the single-step log-likelihood ratio using the assumed model and add it to the cumulative log-likelihood ratio of the previous time step to obtain the current cumulative log-likelihood ratio. If the current cumulative log-likelihood ratio is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero. If it is less than the lower decision threshold, it is determined to be normal and cleared to zero. If it is between the two, the conclusion is maintained and the diagnostic process proceeds to the next time step.
[0046] Substituting the current moisture content measurement into the probability density function of the dynamic hypothesis model (after mean shift) and the unadjusted null hypothesis baseline probability density function, the probability density value of each Gaussian component is calculated. These values are then weighted and summed with their corresponding mixing weights to obtain the total probability of the current observation under both the alternative and null hypotheses. The natural logarithm of the quotient of the total probability of the alternative hypothesis divided by the total probability of the null hypothesis is used to calculate the single-step log-likelihood ratio. This single-step log-likelihood ratio is then algebraically added to the cumulative log-likelihood ratio variable read from memory at the previous time step in the accumulator register to obtain the current cumulative log-likelihood ratio. This value is then used to update the state variables in memory.
[0047] The multi-branch conditional judgment logic compares the current cumulative log-likelihood ratio with the upper and lower decision thresholds. When the current cumulative log-likelihood ratio is greater than the upper decision threshold, a high-level abnormal alarm signal is sent to the pilot's control console via a general-purpose input / output pin to determine that the fuel line is in a water-containing abnormal state. The current cumulative log-likelihood ratio variable in memory is set to zero to reset the sequential test accumulator. When the current cumulative log-likelihood ratio is less than the lower decision threshold, a normal status confirmation data packet is output via the serial communication interface to determine that the fuel line system is normal. The current cumulative log-likelihood ratio variable in memory is set to zero to re-accumulate new observation data evidence. When the current cumulative log-likelihood ratio is greater than or equal to the lower decision threshold and less than or equal to the upper decision threshold, the system's previous diagnostic status output flag remains unchanged, and no alarm or confirmation command is triggered. The current cumulative log-likelihood ratio value is pushed into the historical sequence buffer queue, and a hardware timer interrupt triggers the instruction to enter the cyclic data acquisition and diagnostic process for the next sampling time. Optionally, after the decision to clear or maintain the state ends, the current cumulative log-likelihood ratio (including the 0 value after clearing) is pushed into the historical sequence cache queue, and then the next diagnostic process begins.
[0048] The test data used was a 500-hour fuel line sensor dataset from actual flight records of a certain type of passenger aircraft. Three schemes were compared in the experiment: the basic scheme used a traditional fixed Gaussian mixture model and a static decision threshold for state determination; the ablation scheme added a dynamic adjustment mechanism for the alternative hypothesis mean based on the cumulative log-likelihood ratio; and the complete scheme added a first-order difference symbol sequence to calculate the dynamic margin and update the decision threshold in real time. All three schemes were tested online for anomaly detection under the same 50ms sampling period and consistent baseline error probability.
[0049] After 500 hours of continuous testing, the basic solution achieved a false alarm rate of 4.5% and a false alarm rate of 2.1%, with an average inference latency of 12ms per diagnostic test. After incorporating a mean adjustment mechanism, the ablation solution reduced the false alarm rate to 1.8%, the false alarm rate to 2.3%, and the average inference latency to 14ms. With the addition of a dynamic threshold margin adjustment mechanism, the complete solution achieved a false alarm rate of 1.6%, a false alarm rate reduced to 0.4%, and an average inference latency of 16ms. Figure 3 As shown in the chart, the false alarm rate, false alarm rate and single diagnosis delay of the three sets of comparison schemes are statistically analyzed simultaneously. The indicators of the three sets of schemes form an orderly gradient. The complete scheme can reduce the false alarm level and the overall performance trade-off effect is clear.
[0050] Test data results show that the complete scheme achieves optimized overall detection performance while maintaining millisecond-level low latency response. The reduction in the false alarm rate of the ablation scheme is attributed to the ability of the adjusted alternative hypothesis mean mechanism to track the slow drift characteristics of water content based on confidence level, thereby improving the detection capability of concealed gradual faults. The reduction in the false alarm rate in the complete scheme is attributed to the dynamic margin mechanism, which widens the decision boundary based on the high-frequency differential characteristics of signal fluctuations, suppressing false judgments caused by likelihood ratio oscillations due to pipeline gas-liquid two-phase microbubbles or mechanical vibration noise.
[0051] This specification also provides one or more embodiments of an online diagnostic system for water content anomalies in aviation fuel pipelines, including the following modules: The module is used to acquire real-time measurements of water content, fuel temperature, and fuel pressure from aviation fuel pipelines. Based on the fuel temperature and fuel pressure measurements, a multidimensional lookup table is used to determine a Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null and alternative hypotheses. The mean of the alternative hypothesis is adjusted using the cumulative log-likelihood ratio of the previous time step. When the cumulative log-likelihood ratio of the previous time step is greater than zero, the comprehensive mean of the alternative hypothesis is approximated towards the moving average of water content within a preset time window, and the mean of each Gaussian component is shifted to construct the hypothesis model. The calculation module is used to select the baseline Type I and Type II error probabilities based on the flight phase operation parameters, calculate the baseline upper decision threshold and the baseline lower decision threshold, calculate the first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence at the previous moment, count the frequency of symbol inversion based on the first-order difference symbol sequence, extract the corresponding amplitude features in the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion, determine the dynamic margin based on the frequency and amplitude features, and superimpose the dynamic margin onto the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold. The determination module is used to calculate the single-step log-likelihood ratio using the assumed model and add it to the cumulative log-likelihood ratio of the previous time step to obtain the cumulative log-likelihood ratio of the current time step. If the cumulative log-likelihood ratio of the current time step is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero. If it is less than the lower decision threshold, it is determined to be normal and cleared to zero. If it is between the two, the conclusion is maintained and the next time step diagnosis process is entered.
[0052] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.
Claims
1. An on-line method of diagnosing water contamination in an aviation fuel line, the method comprising: Includes the following steps: The system acquires real-time measurements of water content, fuel temperature, and fuel pressure from aviation fuel pipelines. Based on these measurements, it queries a multidimensional lookup table to determine a Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null and alternative hypotheses. When the cumulative log-likelihood ratio of the previous time step is greater than zero, it approximates the comprehensive mean of the alternative hypotheses towards the moving average of water content within a preset time window and shifts the mean of each Gaussian component to construct the hypothesis model. Based on the flight phase operation parameters, the baseline Type I and Type II error probabilities are selected to calculate the baseline upper decision threshold and the baseline lower decision threshold. The first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence at the previous moment is calculated. The frequency of symbol inversion is statistically analyzed based on the first-order difference symbol sequence. The corresponding amplitude features are extracted from the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion. The dynamic margin is determined based on the frequency and amplitude features. The dynamic margin is superimposed on the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold. The single-step log-likelihood ratio is calculated using the assumed model and added to the cumulative log-likelihood ratio of the previous time step to obtain the current cumulative log-likelihood ratio. If the current cumulative log-likelihood ratio is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero; if it is less than the lower decision threshold, it is determined to be normal and cleared to zero; if it is between the two, the conclusion is maintained and the diagnostic process proceeds to the next time step.
2. The method of claim 1, wherein, The acquisition of real-time water content, fuel temperature, and fuel pressure measurements from the aviation fuel pipeline includes: Configure the timer in the embedded microprocessor module of the system to generate an interrupt pulse of the reference period to trigger a synchronous sampling event; Send synchronous data acquisition commands to sensor nodes via the system communication bus; The system receives and parses the returned data packets with timestamps, extracts the measured values of water content, fuel temperature, and fuel pressure, and stores them in the specified memory buffer.
3. The method of claim 1, wherein, The step of determining the baseline probability density function of the Gaussian mixture model representing normal and abnormal states as the null and alternative hypotheses by querying a multidimensional lookup table based on the measured fuel temperature and fuel pressure values includes: Within the full-range temperature and pressure range, a two-dimensional grid is divided according to a preset step size. For each grid node, the corresponding null hypothesis and alternative hypothesis Gaussian mixture model parameter set under the calibration state is stored in advance. The parameter set includes the weight, mean and variance of each Gaussian component to form an offline mapping dictionary. During diagnosis, the collected fuel temperature and fuel pressure measurements are matched to the nearest two-dimensional grid node; The memory address offset is calculated based on the coordinates of the matching node, and the corresponding parameter set is read from the offline mapping dictionary by a pointer as the baseline probability density function.
4. The method of claim 1, wherein, When the cumulative log-likelihood ratio of the previous time step is greater than zero, the composite mean of the alternative hypotheses is approximated towards the moving average of the moisture content within a preset time window, and the mean of each Gaussian component is shifted to construct the hypothesis model, including: If the cumulative log-likelihood ratio of the previous time step is greater than zero, calculate the moving average of the moisture content measurements within the preset time window; The adjusted weighting coefficient is calculated as one minus the exponent of the product of the negative preset constant with the natural constant as the base and the cumulative log-likelihood ratio of the previous time step. Calculate the weighted composite benchmark mean of each Gaussian component of the alternative hypothesis, and update the benchmark mean of each Gaussian component of the alternative hypothesis to a dynamic mean, wherein the dynamic mean is the benchmark mean plus the product of the adjustment weight coefficient and the moving average minus the weighted composite benchmark mean; When the cumulative log-likelihood ratio of the previous time step is not greater than zero, the weight coefficients are adjusted to zero, and the baseline mean of each Gaussian component is kept unchanged so that the alternative hypothesis model smoothly tracks the drift characteristics of the actual moisture content based on the confidence level of the anomaly.
5. The method of claim 1, wherein, The step of selecting the baseline Type I and Type II error probabilities based on flight phase operation parameters and calculating the baseline upper and lower decision thresholds includes: The current flight characteristic parameters are analyzed to identify the flight stage, and the baseline Type I error probability and baseline Type II error probability corresponding to the flight stage are retrieved from the preset dictionary. Calculate the ratio of the difference between 1 and the baseline Type II error probability to the baseline Type I error probability, and use the natural logarithm of the ratio as the decision threshold on the baseline. The quotient obtained by dividing the baseline Type II error probability by 1 and subtracting the baseline Type I error probability is used as the decision threshold under the baseline. The calculated upper and lower decision thresholds are stored in the decision register as the base reference values for applying dynamic margins subsequently.
6. The method of claim 1, wherein, The calculation of the cumulative log-likelihood ratio (CRL) history sequence after removing the zero-transition first-order difference symbol sequence, statistically analyzing the frequency of symbol inversion based on the first-order difference symbol sequence, extracting corresponding amplitude features from the CRL history sequence based on the position index of symbol inversion, determining a dynamic margin based on the frequency and amplitude features, and superimposing the dynamic margin onto the benchmark upper decision threshold and benchmark lower decision threshold to obtain the upper decision threshold and lower decision threshold, includes: Extract the cumulative log-likelihood ratio history sequence of a preset length, calculate the difference between adjacent sampling points, remove the difference caused by the zeroing operation, and then extract the sign function to generate a first-order difference sign sequence composed of +1 and -1; The total number of symbol inversions in the first-order difference symbol sequence is used as the frequency feature. Based on the position index of the symbol inversion, the cumulative log-likelihood ratio historical sequence is traced back, and the average absolute value of the difference between the local extrema before and after all inversion points in the historical sequence is calculated as the amplitude feature. Calculate the first product of the first preset positive weight coefficient and the frequency feature, and the second product of the second preset positive weight coefficient and the amplitude feature, respectively. The sum of the first product and the second product is used as the dynamic margin. The upper decision threshold is obtained by adding dynamic margin to the upper decision threshold, and the lower decision threshold is obtained by subtracting dynamic margin from the lower decision threshold. This is to suppress misjudgments caused by high-frequency oscillations in the log-likelihood ratio due to measurement noise.
7. An on-line diagnostic system for water contamination in an aviation fuel line, comprising: Includes the following modules: The module is used to acquire real-time measurements of water content, fuel temperature, and fuel pressure from aviation fuel pipelines. Based on the fuel temperature and fuel pressure measurements, a multidimensional lookup table is queried to determine a Gaussian mixture model representing normal and abnormal states as the baseline probability density function for the null hypothesis and alternative hypothesis. When the cumulative log-likelihood ratio of the previous time step is greater than zero, the comprehensive mean of the alternative hypothesis is approximated towards the moving average of water content within a preset time window, and the mean of each Gaussian component is shifted to construct the hypothesis model. The calculation module is used to select the baseline Type I and Type II error probabilities based on the flight phase operation parameters, calculate the baseline upper decision threshold and the baseline lower decision threshold, calculate the first-order difference symbol sequence after removing the zero-transition in the cumulative log-likelihood ratio history sequence at the previous moment, count the frequency of symbol inversion based on the first-order difference symbol sequence, extract the corresponding amplitude features in the cumulative log-likelihood ratio history sequence according to the position index of the symbol inversion, determine the dynamic margin based on the frequency and amplitude features, and superimpose the dynamic margin onto the baseline upper decision threshold and the baseline lower decision threshold to obtain the upper decision threshold and the lower decision threshold. The determination module is used to calculate the single-step log-likelihood ratio using the assumed model and add it to the cumulative log-likelihood ratio of the previous time step to obtain the cumulative log-likelihood ratio of the current time step. If the cumulative log-likelihood ratio of the current time step is greater than the upper decision threshold, it is determined to be abnormal and cleared to zero. If it is less than the lower decision threshold, it is determined to be normal and cleared to zero. If it is between the two, the conclusion is maintained and the next time step diagnosis process is entered.
8. The system of claim 7, wherein, The acquisition of real-time water content, fuel temperature, and fuel pressure measurements from the aviation fuel pipeline includes: Configure the timer in the embedded microprocessor module of the system to generate an interrupt pulse of the reference period to trigger a synchronous sampling event; Send synchronous data acquisition commands to sensor nodes via the system communication bus; The system receives and parses the returned data packets with timestamps, extracts the measured values of water content, fuel temperature, and fuel pressure, and stores them in the specified memory buffer.
9. The system of claim 7, wherein, The step of determining the baseline probability density function of the Gaussian mixture model representing normal and abnormal states as the null and alternative hypotheses by querying a multidimensional lookup table based on the measured fuel temperature and fuel pressure values includes: Within the full-range temperature and pressure range, a two-dimensional grid is divided according to a preset step size. For each grid node, the corresponding null hypothesis and alternative hypothesis Gaussian mixture model parameter set under the calibration state is stored in advance. The parameter set includes the weight, mean and variance of each Gaussian component to form an offline mapping dictionary. During diagnosis, the collected fuel temperature and fuel pressure measurements are matched to the nearest two-dimensional grid node; The memory address offset is calculated based on the coordinates of the matching node, and the corresponding parameter set is read from the offline mapping dictionary by a pointer as the baseline probability density function.
10. The system of claim 7, wherein, When the cumulative log-likelihood ratio of the previous time step is greater than zero, the composite mean of the alternative hypotheses is approximated towards the moving average of the moisture content within a preset time window, and the mean of each Gaussian component is shifted to construct the hypothesis model, including: If the cumulative log-likelihood ratio of the previous time step is greater than zero, calculate the moving average of the moisture content measurements within the preset time window; The adjusted weighting coefficient is calculated as one minus the exponent of the product of the negative preset constant with the natural constant as the base and the cumulative log-likelihood ratio of the previous time step. Calculate the weighted composite benchmark mean of each Gaussian component of the alternative hypothesis, and update the benchmark mean of each Gaussian component of the alternative hypothesis to a dynamic mean, wherein the dynamic mean is the benchmark mean plus the product of the adjustment weight coefficient and the moving average minus the weighted composite benchmark mean; When the cumulative log-likelihood ratio of the previous time step is not greater than zero, the weight coefficients are adjusted to zero, and the baseline mean of each Gaussian component is kept unchanged so that the alternative hypothesis model smoothly tracks the drift characteristics of the actual moisture content based on the confidence level of the anomaly.
