A real-time lubricating oil state monitoring method based on machine learning

Abstract

The application discloses a kind of real-time lubricating oil state monitoring methods based on machine learning, specifically related to state monitoring technical field, including after lubricating oil system enters steady working condition, by collecting the static characteristic in oil tank as input, the output of oil tank fault grade probability distribution is fuzzy bayesian neural network model, construct fuzzy bayesian neural network model, according to the output of fuzzy bayesian neural network model in monitoring interval, the probability value of different fault grade in monitoring interval is comprehensively analyzed, the state of lubricating oil in oil tank is determined, based on the dynamic characteristic of lubricating oil in pipeline, the physical information and risk information at the preset position of pipeline are collected, the state of lubricating oil in pipeline is determined, based on the state of lubricating oil in quantization oil tank and pipeline, according to the trained neural network model, realize the early warning function to lubricating oil state, the application helps the comprehensiveness and accuracy of lubricating oil state monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of state monitoring, and more particularly to a real-time lubricating oil state monitoring method based on machine learning. BACKGROUND

[0002] With the increasing requirements of industrial equipment on reliability and maintenance efficiency, lubricating oil state monitoring and early warning has become an important means to ensure the safe operation of equipment. The existing technology usually collects a single index to evaluate the risk through a threshold alarm or a traditional mechanism model, often lacks comprehensive judgment of the degradation state of the oil itself in the tank and the pipeline, resulting in lack of grading judgment and uncertainty quantification of fault levels, and there is a maintenance blind area. Local alarm can only reflect the single risk of the pipe section or component, and cannot reveal the degradation evolution trend of the oil itself.

[0003] In order to solve the above defects, a technical scheme is provided. SUMMARY

[0004] In order to overcome the above defects of the prior art, the embodiments of the present application provide a real-time lubricating oil state monitoring method based on machine learning to solve the problems raised in the background art.

[0005] To achieve the above object, the present application provides the following technical scheme:

[0006] A real-time lubricating oil state monitoring method based on machine learning, comprising the following steps:

[0007] S1: After the lubricating oil system enters a steady state, static characteristics in the oil tank are collected as input, and the fault level probability distribution of the oil tank is taken as the output of the fuzzy Bayesian neural network model, and a fuzzy Bayesian neural network model is constructed, wherein the static characteristics in the oil tank include viscosity health index, temperature index and particle concentration of different particle sizes;

[0008] S2: According to the output of the fuzzy Bayesian neural network model in the monitoring interval, the probability values of different fault levels in the monitoring interval are comprehensively analyzed to determine the state of the lubricating oil in the oil tank;

[0009] S3: Based on the dynamic characteristics of the lubricating oil in the pipeline, the physical information and risk information at the preset position of the pipeline are collected, and the physical information and risk information at different preset positions of the pipeline are comprehensively analyzed to determine the state of the lubricating oil in the pipeline;

[0010] S4: Based on the state of the lubricating oil in the oil tank and the pipeline, the state of the lubricating oil in the oil tank and the pipeline is analyzed according to the trained neural network model to realize the early warning function of the lubricating oil state.

[0011] In a preferred embodiment, constructing a fuzzy Bayesian neural network model includes:

[0012] By using viscosity health index, temperature index, and particle concentration of different sizes as inputs to a fuzzy Bayesian neural network model, and the probability distribution of tank failure levels as the output, the input space is divided into fuzzy regions using fuzzy rules. The output of the fuzzy Bayesian neural network model is then divided into different failure levels: no failure, low failure, medium failure, and high failure. The output of the fuzzy Bayesian neural network model is represented as: P, , This represents the probability value for a fault-free rating. This represents the probability value for a low failure level. This represents the probability value for a medium-level fault. This represents the probability value for a high failure level.

[0013] The particle concentration for different particle sizes is determined by setting monitoring intervals, and the particle concentration of the lubricating oil in each particle size range within the monitoring interval of the oil tank is marked as follows: , where i = 1, 2, 3, ..., I, I is a positive integer, and i is the number of the particle size segment.

[0014] In a preferred embodiment, the logic for obtaining the viscosity health index is as follows:

[0015] Obtain the viscosity data within the monitoring range in the oil tank, and label the viscosity data within the monitoring range in the oil tank as follows: m = 1, 2, 3, ..., M, where M is a positive integer and m is the sampling number within the monitoring interval;

[0016] Offline GMM modeling is performed. In offline GMM modeling, different types of viscosity data in the oil tank are obtained based on the training data. The different viscosity data include the viscosity data of new oil, the viscosity data of slightly deteriorated oil, and the viscosity data of severely deteriorated oil. The number of components K is set, where K is a modeling parameter, which indicates how many different Gaussian sub-distributions are considered to constitute the reading distribution in a GMM fitting. The GMM is fitted by the EM algorithm to complete the offline GMM modeling.

[0017] The liability score is calculated using the trained GMM parameters. The viscosity data within the monitoring range of the oil tank is used to calculate the liability score for different implicit components based on the trained GMM parameters, and the liability scores for different implicit components are labeled as follows: ,in, , Let be the weight corresponding to the k-th hidden component, representing the proportion belonging to the k-th state. The viscosity average value under the k-th implicit component. The viscosity variance under the k-th implicit component;

[0018] The viscosity health index is calculated using the following formula: ;in, For viscosity health index, The health weights are for different components k.

[0019] In a preferred embodiment, the logic for obtaining the temperature index is as follows:

[0020] The temperature data within the monitoring interval of the fuel tank is obtained. This temperature data is then fitted with time to determine the variation function of the temperature data with time within the monitoring interval. This variation function is then labeled as follows: ;

[0021] The temperature index is calculated using the following formula: ;in, Temperature index This refers to the time period during operation when the temperature exceeds the set threshold.

[0022] In a preferred embodiment, determining the state of the lubricating oil in the tank includes:

[0023] Based on the output of the fuzzy Bayesian neural network model within the monitoring interval, the probability values ​​of different fault levels within the monitoring interval are weighted and summed to obtain the fuel tank fault rating coefficient. The formula for calculating the fuel tank fault rating coefficient is as follows: ;in, This is the fuel tank failure rating coefficient. This is the proportionality coefficient representing the probability value of the fault-free level. This is a proportionality coefficient representing the probability value of a low failure level. This is the proportionality coefficient for the probability value of a medium-level fault. These are the proportional coefficients for high failure levels. .

[0024] In a preferred embodiment, the physical information at different preset locations of the pipeline includes:

[0025] The physical information at different preset locations on the pipeline is represented by the differential pressure deviation amplitude coefficient and the mechanical acoustic spectrum measurement coefficient;

[0026] The logic for obtaining the differential pressure deviation amplitude coefficient is as follows: collect the time series of differential pressure data within the monitoring interval at a preset location in the pipeline, perform stationary processing on the differential pressure data time series, that is, perform multiple differences on the series until it is considered stationary after passing the unit root test, estimate the autoregression order and moving average order in ARIMA using the least squares method, and obtain an ARIMA model with fixed structure and parameters through model validation.

[0027] Using the ARIMA model at a preset location, the time series of differential pressure data obtained within the monitoring interval is input into the ARIMA model to obtain the differential pressure prediction value. By comparing the model prediction value with the actual observed value, the differential pressure deviation amplitude coefficient is calculated.

[0028] The formula for calculating the differential pressure deviation amplitude coefficient is: ;in, Here, q represents the differential pressure deviation amplitude coefficient, where q = 1, 2, 3, ..., Q, where Q is a positive integer and q represents the index of the differential pressure data sampled at different time points within the monitoring interval. To monitor the actual differential pressure data time series within the monitoring range, This represents the predicted pressure difference value of the ARIMA model at time point q.

[0029] In a preferred embodiment, the logic for obtaining the mechanical acoustic spectrum metric coefficients is as follows:

[0030] Acoustic signals within a monitoring interval at a preset location in the pipeline are collected, and the acoustic signals at the preset location are subjected to Fourier transform to obtain the spectrum S(f). Where f is the frequency, The acoustic signal at a predetermined location in the pipeline;

[0031] The mechanical spectral metric coefficient is calculated by dividing the low-frequency and high-frequency thresholds. The calculation formula is as follows: ;in, For mechanical acoustic spectrum measurement coefficients. , For low-frequency thresholds, For high-frequency band thresholds, The natural frequency of the structure, This refers to the abnormal frequency band generated by impurities or wear debris in the oil impacting the pipe wall and by friction of the lubricating oil.

[0032] In a preferred embodiment, risk information at different predetermined locations along the pipeline includes:

[0033] Risk information at different preset locations in the pipeline is represented by a particle deposition risk assessment coefficient.

[0034] The logic for obtaining the particle deposition risk assessment coefficient is as follows: based on the particles of different particle sizes of lubricating oil at preset positions in the pipeline within the monitoring interval, determine the impact of particles of different particle sizes on the operating status, and obtain the impact score of particles of different particle sizes based on the impact of particles of different particle sizes on the operating status.

[0035] Based on the particles of different sizes collected at preset locations in the pipeline within the monitoring interval, the impact scores and concentrations of particles of different sizes are used as input features of the logistic regression model, and the particle deposition risk assessment coefficient is used as the output feature of the logistic regression model to construct the logistic regression model and generate the particle deposition risk assessment coefficient.

[0036] The expression for the particle deposition risk assessment coefficient is as follows:

[0037] ;

[0038] in, This is the particle deposition risk assessment coefficient. The weights of particles of different sizes, The influence of particles of different size ranges is scored, with e as the base.

[0039] In a preferred embodiment, determining the state of the lubricating oil in the pipeline includes:

[0040] A comprehensive analysis of physical and risk information at predetermined locations within the pipeline is conducted. The pressure difference deviation amplitude coefficient, mechanical acoustic spectrum coefficient, and particle deposition risk assessment coefficient are weighted and summed to construct a pipeline location assessment model and generate pipeline location assessment coefficients. The formula for calculating the pipeline location assessment coefficients is as follows: ;in, Let be the pipeline position evaluation coefficient at the preset position j, where j = 1, 2, 3, ..., J, J is a positive integer, and j is the number of the preset position in the pipeline. , , These are the proportional coefficients for the differential pressure deviation amplitude coefficient, the mechanical acoustic spectrum measurement coefficient, and the particle deposition risk assessment coefficient, respectively. , , All are greater than 0;

[0041] Based on the number of preset positions in the pipeline and their impact on the overall operating state of the lubricating oil, the overall pipeline fault score is obtained by weighted summation of pipeline position evaluation coefficients at different preset positions. The formula for calculating the pipeline fault score coefficient is as follows: ;in, This is the pipeline fault rating coefficient. Weights at different preset positions.

[0042] In a preferred embodiment, the function of providing an early warning of the lubricating oil condition includes:

[0043] The fault warning coefficient is calculated using the following formula: ;in, This is the fault warning coefficient. It is a specific proportionality coefficient.

[0044] The tank fault rating coefficient, pipeline fault rating coefficient, and monitoring interval time length are used as inputs to train the neural network model, and the fault warning coefficient is used as the output of the neural network model. The model is trained by assigning a value of 0 or 1, where 0 indicates that the fault warning coefficient is less than or equal to the fault warning coefficient threshold, and 1 indicates that the fault warning coefficient is greater than the fault warning coefficient threshold.

[0045] The technical effects and advantages of this invention are as follows:

[0046] This invention considers both the static characteristics of the oil tank and the dynamic characteristics of the pipeline, enabling it to capture changes in the chemical and physical properties of the oil and reflect wear and risk accumulation during transportation. Specifically, it uses a fuzzy Bayesian neural network for static analysis of the oil tank and constructs a neural network model by analyzing dynamic parameters in the pipeline to provide early warning of the overall lubricating oil condition. This invention contributes to the comprehensiveness and accuracy of lubricating oil condition monitoring. Attached Figure Description

[0047] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0048] Figure 1 This is a flowchart illustrating a real-time lubricating oil condition monitoring method based on machine learning according to the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Example 1

[0051] Figure 1 This is a flowchart illustrating a real-time lubricating oil condition monitoring method based on machine learning according to the present invention, which specifically includes the following steps:

[0052] S1: After the lubricating oil system enters a stable working state, the static features in the oil tank are collected as input, and the probability distribution of oil tank failure level is used as the output of the fuzzy Bayesian neural network model to construct the fuzzy Bayesian neural network model. The static features in the oil tank include viscosity health index, temperature index and particle concentration of different particle sizes.

[0053] S2: Based on the output of the fuzzy Bayesian neural network model within the monitoring interval, the probability values ​​of different fault levels within the monitoring interval are comprehensively analyzed to determine the state of the lubricating oil in the tank.

[0054] S3: Based on the dynamic characteristics of the lubricating oil in the pipeline, collect physical information and risk information at preset locations in the pipeline, and conduct a comprehensive analysis of the physical information and risk information at different preset locations in the pipeline to determine the state of the lubricating oil in the pipeline;

[0055] S4: Based on the quantitative state of lubricating oil in the tank and pipeline, the state of lubricating oil in the tank and pipeline is analyzed according to the trained neural network model to realize the early warning function of lubricating oil state.

[0056] After the lubricating oil system enters a stable operating state, the collected data is evaluated. The criteria for judging a stable operating state include the system temperature reaching a stable range, small fluctuations in equipment load, stable pipeline flow, and the oil tank meeting static conditions.

[0057] Lubricating oil is forced to flow in pipelines at high velocities, often entering laminar or turbulent flow states. Pressure fluctuations, shear stress, eddies, and pulsations are very obvious. Therefore, the characteristics of lubricating oil collected in pipelines are usually dynamic. In contrast, the oil in the tank is basically static or has only very weak natural convection, without obvious flow shear and pulsations. The main focus is on static or slowly changing characteristics such as temperature stratification, oil-water interface, water sedimentation, and suspended particles.

[0058] In pipelines, the focus is on faults directly related to flow characteristics, such as abnormal flow resistance, cavitation / cavitation, and particulate blockage. In oil tanks, the emphasis is on detecting oil contamination, oil-water separation, thermal separation, and physicochemical changes after long periods of stagnation. These are manifested as slow drifts in concentration distribution, oil level changes, and temperature gradients.

[0059] Since the oil in the tank is basically static or only has very weak natural convection, and there is often ambiguity in actual annotation, the static features in the tank are collected as input and fed into the fuzzy Bayesian neural network.

[0060] It should be noted that fuzzy representation maps features in the fuel tank that cannot be precisely divided by a single threshold into a set of membership degrees that "belong to a certain semantic category to some extent" through fuzzy sets and membership functions. This preserves both numerical information and the ability of experts to make qualitative judgments such as "slightly high", "medium", and "low" in the neural network.

[0061] The static features in the fuel tank include viscosity health index, temperature index, and particle concentration of different sizes. By using viscosity health index, temperature index, and particle concentration of different sizes as input features of fuzzy Bayesian neural network, the probability distribution of fuel tank failure level is used as the output of fuzzy Bayesian neural network model.

[0062] Specifically, the particle concentrations of different sizes are measured using sensors with particle size resolution capabilities. This means the sensors can statistically analyze the particle concentration of the lubricating oil within each particle size range of the monitoring interval in the oil tank. By setting monitoring intervals, the particle concentration of the lubricating oil within each particle size range of the monitoring interval in the oil tank is labeled as follows: Where i = 1, 2, 3, ..., I, I is a positive integer, and i is the number of the particle size segment;

[0063] It should be noted that the monitoring interval is a specific time period set by professionals in the field. The monitoring interval can be set according to the specific operating status, that is, each operating cycle can be used as a monitoring interval.

[0064] The logic for obtaining the viscosity health index is as follows: obtain the viscosity data within the monitoring range in the oil tank, and mark the viscosity data within the monitoring range in the oil tank as: m = 1, 2, 3, ..., M, where M is a positive integer and m is the sampling number within the monitoring interval;

[0065] Offline GMM modeling is performed. In offline GMM modeling, different types of viscosity data in the oil tank are obtained based on the training data. The different viscosity data include the viscosity data of new oil, the viscosity data of slightly deteriorated oil, and the viscosity data of severely deteriorated oil. The number of components K is set, where K is a modeling parameter, indicating how many different Gaussian sub-distributions are considered to constitute the reading distribution in a GMM fitting. The GMM is fitted by the EM algorithm to complete the offline GMM modeling.

[0066] It should be noted that the different types of viscosity data are set by professionals in the field. The number of components K in the GMM can be preset to the number of types. By training the GMM model, the number of components K with the smallest error can be obtained.

[0067] The liability score is calculated using the trained GMM parameters. The viscosity data within the monitoring range of the oil tank is used to calculate the liability score for different implicit components based on the trained GMM parameters, and the liability scores for different implicit components are labeled as follows: ,in, , Let be the weight corresponding to the k-th hidden component, representing the proportion belonging to the k-th state. The viscosity average value under the k-th implicit component. The viscosity variance under the k-th implicit component;

[0068] The viscosity health index is calculated using the following formula: ;in, For viscosity health index, The health weights are for different components k.

[0069] It should be noted that, These are parameters obtained by fitting the EM algorithm during offline GMM training, and are used in online monitoring. It remains unchanged and is directly used for liability calculation;

[0070] Viscosity data are continuous numerical measurements, but they often do not come from a single, homogeneous oil state. Instead, they are a mixture of various "operating conditions" or "oil aging degrees." For example, in a sampling interval, some new oils have not yet begun to deteriorate, some oils have already deteriorated slightly, and some oils may have already deteriorated severely. These three situations each produce a slightly different set of viscosity reading distributions, but we cannot see their labels (new, slightly, severely) in the data. These are called "implicit states."

[0071] The analysis of hidden states helps to capture multimodal distributions. That is, if all viscosity values ​​are directly treated as the same mode for modeling, a single Gaussian or linear model often does not fit well when there are multiple oil states. By introducing hidden states, multiple Gaussian distributions can be used to fit the viscosity mode of each state, and the overall distribution can fit the "multi-peak" or "long-tail" distribution well.

[0072] The logic for obtaining the temperature index is as follows: Obtain the temperature data within the monitoring range of the fuel tank; fit the temperature data within the monitoring range of the fuel tank to time; determine the change function of the temperature data within the monitoring range over time; and label the change function of the temperature data within the monitoring range over time as: ;

[0073] The temperature index is calculated using the following formula: ;in, Temperature index This refers to the time period during operation when the temperature exceeds the set threshold.

[0074] It should be noted that after the equipment is started, the lubricating oil is heated by heat conduction from components such as the engine and gearbox. When the equipment stops running, the lubricating oil gradually cools down. Therefore, the temperature of the lubricating oil in the tank will change, and the lubricating oil will deteriorate rapidly in a high-temperature environment. If the temperature exceeds a certain threshold, it means that the condition has deteriorated and should be regarded as a negative indicator in the health index.

[0075] By using viscosity health index, temperature index, and particle concentration of different sizes as inputs to a fuzzy Bayesian neural network model, and the probability distribution of tank failure levels as the output, the input space is divided into fuzzy regions using fuzzy rules. The output of the fuzzy Bayesian neural network model is then divided into different failure levels: no failure, low failure, medium failure, and high failure. The output of the fuzzy Bayesian neural network model is represented as: P, , This represents the probability value for a fault-free rating. This represents the probability value for a low failure level. This represents the probability value for a medium-level fault. The probability value for a high fault level is represented by the sample dataset of the fuzzy Bayesian neural network model as follows: , .

[0076] Based on the output of the fuzzy Bayesian neural network model within the monitoring interval, the probability values ​​of different fault levels within the monitoring interval are weighted and summed to obtain the fuel tank fault rating coefficient. The formula for calculating the fuel tank fault rating coefficient is as follows: ;in, This is the fuel tank failure rating coefficient. This is the proportionality coefficient representing the probability value of the fault-free level. This is a proportionality coefficient representing the probability value of a low failure level. This is the proportionality coefficient for the probability value of a medium-level fault. These are the proportional coefficients for high failure levels. .

[0077] Since the lubricating oil in the pipeline is in a dynamic state, the dynamic characteristics in the pipeline are collected to assess the health status of the lubricating oil. The dynamic characteristics include physical information and risk information. The physical information is represented by the differential pressure deviation amplitude coefficient and the mechanical acoustic spectrum measurement coefficient, and the risk information is represented by the particle deposition risk assessment coefficient.

[0078] Points to note when using the differential pressure deviation amplitude coefficient include:

[0079] Different pressure differentials at different locations in a pipeline can be caused by factors such as different pipeline geometries, different lubricating oil flow rates, and increased local resistance due to wear, blockage, or contamination. Therefore, it is necessary to pre-install differential pressure sensors at key locations in the system, collect and model the differential pressure data at each location to reflect its unique dynamic behavior, and achieve multi-dimensional, distributed health monitoring of the lubrication system.

[0080] Under normal operating conditions, the differential pressure data at a fixed location in the pipeline will not fluctuate significantly or irregularly. Instead, it will typically fluctuate slightly around a stable mean, thus giving the differential pressure data at that location a time series stationarity or quasi-stationarity.

[0081] Pressure differential data from fixed locations exhibits time-series stationarity. By using the ARIMA model to model and predict pressure differential data from the same collection point and leveraging the regularities in historical data, abnormal fluctuations or trend changes can be effectively identified, thus aiding in health status assessment.

[0082] The logic for obtaining the differential pressure deviation amplitude coefficient is as follows: collect the time series of differential pressure data within the monitoring interval at a preset location in the pipeline, perform stationary processing on the differential pressure data time series, that is, perform multiple differences on the series until it is considered stationary after passing the unit root test, estimate the autoregression order and moving average order in ARIMA using the least squares method, and obtain an ARIMA model with fixed structure and parameters through model validation.

[0083] Using the ARIMA model at a preset location, the time series of differential pressure data obtained within the monitoring interval is input into the ARIMA model to obtain the differential pressure prediction value. By comparing the model prediction value with the actual observed value, the differential pressure deviation amplitude coefficient is calculated.

[0084] The formula for calculating the differential pressure deviation amplitude coefficient is: ;in, Here, q represents the differential pressure deviation amplitude coefficient, where q = 1, 2, 3, ..., Q, where Q is a positive integer and q represents the index of the differential pressure data sampled at different time points within the monitoring interval. To monitor the actual differential pressure data time series within the monitoring range, This represents the predicted pressure difference value of the ARIMA model at time point q.

[0085] As can be seen from the formula, the larger the differential pressure deviation amplitude coefficient, the more the current differential pressure deviates from the historical pattern, and there may be problems such as blockage, leakage, or abnormal wear. Conversely, the closer the differential pressure deviation amplitude coefficient is to 0, the closer the current differential pressure is to the predicted value, and the more stable the system state is.

[0086] Among the points to note regarding the mechanical acoustic spectrum metric coefficients are:

[0087] When the lubricating oil is in good condition, the moving parts are adequately lubricated, friction is low, and the high-frequency components in the vibration / sound spectrum are weak. When the lubricating oil deteriorates or impurities increase, friction increases, high-frequency noise and vibration energy increase significantly, and the proportion of high-frequency energy increases significantly.

[0088] High-frequency energy mainly comes from rapid transient phenomena such as impact of tiny particles, cavitation, and local faults in lubricating oil. Therefore, by collecting sound waves caused by particles, bubbles, or abnormal vibrations during the oil flow process, it can serve as an important indicator for lubricating oil performance degradation and system fault early warning.

[0089] The logic for obtaining the mechanical acoustic spectrum coefficients is as follows: Acoustic signals within a monitoring interval at a preset location in the pipeline are collected; the acoustic signals at the preset location are then subjected to Fourier transform to obtain the spectrum S(f). Where f is the frequency, The acoustic signal at a predetermined location in the pipeline;

[0090] The mechanical spectral metric coefficient is calculated by dividing the low-frequency and high-frequency thresholds. The calculation formula is as follows: ;in, For mechanical acoustic spectrum measurement coefficients. , For low-frequency thresholds, For high-frequency band thresholds, The natural frequency of the structure, This refers to the abnormal frequency band generated by impurities or wear debris in the oil impacting the pipe wall and by friction of the lubricating oil.

[0091] It should be noted that the acoustic signal is obtained by installing an acoustic sensor at a preset location. The spectrum is divided into "low frequency band" and "high frequency band" and corresponding thresholds are set. This is mainly to more accurately separate the components generated by different physical mechanisms in the signal, and then extract more targeted and reliable health monitoring features. The low frequency band mainly includes macroscopic motion features such as the overall mechanical resonance of the system, rotational speed-related components, and fluid pulsation. The high frequency band mainly includes transient impact signals such as microscopic friction of lubricating oil, particle impact, and cavitation bubble collapse.

[0092] As can be seen from the formula, the larger the mechanical acoustic spectrum metric coefficient, the higher the proportion of high-frequency components in the overall spectrum, indicating that there may be deterioration of lubricating oil, increased pollution, or minor faults in the system.

[0093] Among the points to note in the particle deposition risk assessment coefficient are:

[0094] Since particles of different sizes have different effects on the lubricating oil operating system, the impact of particles on different components is obtained based on the differences in the impact of particles on different parts of the pipeline. The risk assessment coefficient can comprehensively consider the particle concentration and its impact differences in different particle size segments, improve the ability to characterize the real risks. Different particle sizes have different damage mechanisms to equipment, and risk modeling based on segmented sensitivity is realized.

[0095] The logic for obtaining the particle deposition risk assessment coefficient is as follows: based on the particles of different particle sizes of lubricating oil at preset positions in the pipeline within the monitoring interval, determine the impact of particles of different particle sizes on the operating status, and obtain the impact score of particles of different particle sizes based on the impact of particles of different particle sizes on the operating status.

[0096] It should be noted that solid particles of different sizes have different mechanisms of influence on friction, wear, and flow resistance in a lubrication system. Therefore, their "degree of harm" or "indicative significance" to the operating state also varies. For example, particles with a diameter range of 1–4 μm undergo Brownian motion with the oil film and are relatively stable, so their impact is small. Particles with a diameter range greater than 14 μm move with the fluid impact and are prone to particle collision, causing impact wear. Therefore, segmented monitoring of the concentration of each particle size can not only quantitatively assess the overall contamination level, but also distinguish different types of wear and failure modes, and provide accurate information for targeted maintenance and lubrication management.

[0097] Based on the particles of different sizes collected at preset locations in the pipeline within the monitoring interval, the impact scores and concentrations of particles of different sizes are used as input features of the logistic regression model, and the particle deposition risk assessment coefficient is used as the output feature of the logistic regression model to construct the logistic regression model and generate the particle deposition risk assessment coefficient.

[0098] The expression for the particle deposition risk assessment coefficient is as follows:

[0099] ;

[0100] in, This is the particle deposition risk assessment coefficient. The weights of particles of different sizes, The influence of particles of different particle sizes is scored, with e as the base.

[0101] As can be seen from the formula, the larger the particle deposition risk assessment coefficient, the higher the concentration of particles of different sizes at the preset location in the pipeline, which may pose a potential hazard. Conversely, the smaller the particle deposition risk assessment coefficient, the lower the concentration of particles of different sizes at the preset location in the pipeline, which may have a smaller impact on the machinery.

[0102] Based on dynamic characteristics, a comprehensive analysis of physical and risk information at predetermined locations within the pipeline is performed. The pressure difference deviation amplitude coefficient, mechanical acoustic spectrum coefficient, and particle deposition risk assessment coefficient are weighted and summed to construct a pipeline location assessment model and generate pipeline location assessment coefficients. The formula for calculating the pipeline location assessment coefficients is as follows: ;in, Let be the pipeline position evaluation coefficient at the preset position j, where j = 1, 2, 3, ..., J, J is a positive integer, and j is the number of the preset position in the pipeline. , , These are the proportional coefficients for the differential pressure deviation amplitude coefficient, the mechanical acoustic spectrum measurement coefficient, and the particle deposition risk assessment coefficient, respectively. , , All are greater than 0.

[0103] As can be seen from the formula, the larger the differential pressure deviation amplitude coefficient, the mechanical acoustic spectrum measurement coefficient, and the particle deposition risk assessment coefficient, the larger the pipeline position assessment coefficient at the preset location in the pipeline, indicating that there is a greater possibility of a fault at the preset location in the pipeline. Conversely, the smaller the differential pressure deviation amplitude coefficient, the mechanical acoustic spectrum measurement coefficient, and the particle deposition risk assessment coefficient, the smaller the pipeline position assessment coefficient at the preset location in the pipeline, indicating that there is a less possibility of a fault at the preset location in the pipeline.

[0104] Based on the number of preset positions in the pipeline and their impact on the overall operating state of the lubricating oil, the overall pipeline fault score is obtained by weighted summation of pipeline position evaluation coefficients at different preset positions. The formula for calculating the pipeline fault score coefficient is as follows: ;in, This is the pipeline fault rating coefficient. Weights at different preset positions.

[0105] The fault warning coefficient is calculated using the following formula: ;in, This is the fault warning coefficient. It is a specific proportionality coefficient.

[0106] The tank fault rating coefficient, pipeline fault rating coefficient, and monitoring interval time length are used as inputs to train the neural network model, and the fault warning coefficient is used as the output of the neural network model. The model is trained by assigning a value of 0 or 1, where 0 indicates that the fault warning coefficient is less than or equal to the fault warning coefficient threshold, and 1 indicates that the fault warning coefficient is greater than the fault warning coefficient threshold.

[0107] It should be noted that the fault warning coefficient threshold is set by professionals in the field and needs to be comprehensively considered in conjunction with historical data, expert experience, and model output in order to binarize the fault warning coefficient.

[0108] This invention considers both the static characteristics of the oil tank and the dynamic characteristics of the pipeline, enabling it to capture changes in the chemical and physical properties of the oil and reflect wear and risk accumulation during transportation. Specifically, it uses a fuzzy Bayesian neural network for static analysis of the oil tank and constructs a neural network model by analyzing dynamic parameters in the pipeline to provide early warning of the overall lubricating oil condition. This invention contributes to the comprehensiveness and accuracy of lubricating oil condition monitoring.

[0109] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0110] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0111] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0112] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A real-time lubricating oil condition monitoring method based on machine learning, characterized in that, Includes the following steps: S1: After the lubricating oil system enters a stable working state, the static features in the oil tank are collected as input, and the probability distribution of oil tank failure level is used as the output of the fuzzy Bayesian neural network model to construct the fuzzy Bayesian neural network model. The static features in the oil tank include viscosity health index, temperature index and particle concentration of different particle sizes. S2: Based on the output of the fuzzy Bayesian neural network model within the monitoring interval, the probability values ​​of different fault levels within the monitoring interval are comprehensively analyzed to determine the state of the lubricating oil in the tank. S3: Based on the dynamic characteristics of the lubricating oil in the pipeline, collect physical information and risk information at preset locations in the pipeline, and conduct a comprehensive analysis of the physical information and risk information at different preset locations in the pipeline to determine the state of the lubricating oil in the pipeline; S4: Based on the quantitative state of lubricating oil in the tank and pipeline, the state of lubricating oil in the tank and pipeline is analyzed according to the trained neural network model to realize the early warning function of lubricating oil state; The physical information at different preset locations of the pipelines includes: The physical information at different preset locations on the pipeline is represented by the differential pressure deviation amplitude coefficient and the mechanical acoustic spectrum measurement coefficient; The logic for obtaining the differential pressure deviation amplitude coefficient is as follows: collect the time series of differential pressure data within the monitoring interval at a preset location in the pipeline, perform stationary processing on the differential pressure data time series, that is, perform multiple differences on the series until it is considered stationary after passing the unit root test, estimate the autoregression order and moving average order in ARIMA using the least squares method, and obtain an ARIMA model with fixed structure and parameters through model validation. Using the ARIMA model at a preset location, the time series of differential pressure data obtained within the monitoring interval is input into the ARIMA model to obtain the differential pressure prediction value. By comparing the model prediction value with the actual observed value, the differential pressure deviation amplitude coefficient is calculated. The formula for calculating the differential pressure deviation amplitude coefficient is: ;in, Here, q represents the differential pressure deviation amplitude coefficient, where q = 1, 2, 3, ..., Q, where Q is a positive integer and q represents the index of the differential pressure data sampled at different time points within the monitoring interval. To monitor the actual differential pressure data time series within the monitoring range, This represents the predicted pressure difference value of the ARIMA model at time point q. The logic for obtaining the mechanical acoustic spectrum metric coefficients is as follows: Acoustic signals within a monitoring interval at a preset location in the pipeline are collected, and the acoustic signals at the preset location are subjected to Fourier transform to obtain the spectrum S(f). Where f is the frequency, The acoustic signal at a predetermined location in the pipeline; The mechanical spectral metric coefficient is calculated by dividing the low-frequency and high-frequency thresholds. The calculation formula is as follows: ;in, For mechanical acoustic spectrum measurement coefficients. , For low-frequency thresholds, For high-frequency band thresholds, The natural frequency of the structure, This refers to the abnormal frequency band generated by impurities or wear debris in the oil impacting the pipe wall and by friction of the lubricating oil.

2. The real-time lubricating oil condition monitoring method based on machine learning according to claim 1, characterized in that, Constructing a fuzzy Bayesian neural network model includes: By using viscosity health index, temperature index, and particle concentration of different sizes as inputs to a fuzzy Bayesian neural network model, and the probability distribution of tank failure levels as the output, the input space is divided into fuzzy regions using fuzzy rules. The output of the fuzzy Bayesian neural network model is then divided into different failure levels: no failure, low failure, medium failure, and high failure. The output of the fuzzy Bayesian neural network model is represented as: P, , This represents the probability value for a fault-free rating. This represents the probability value for a low failure level. This represents the probability value for a medium-level fault. This represents the probability value for a high failure level. The particle concentration for different particle sizes is determined by setting monitoring intervals, and the particle concentration of the lubricating oil in each particle size range within the monitoring interval of the oil tank is marked as follows: , where i = 1, 2, 3, ..., I, I is a positive integer, and i is the number of the particle size segment.

3. The real-time lubricating oil condition monitoring method based on machine learning according to claim 2, characterized in that, The logic for obtaining the viscosity health index is as follows: Obtain the viscosity data within the monitoring range in the oil tank, and label the viscosity data within the monitoring range in the oil tank as follows: m = 1, 2, 3, ..., M, where M is a positive integer and m is the sampling number within the monitoring interval; Offline GMM modeling is performed. In offline GMM modeling, different types of viscosity data in the oil tank are obtained based on the training data. The different viscosity data include the viscosity data of new oil, the viscosity data of slightly deteriorated oil, and the viscosity data of severely deteriorated oil. The number of components K is set, where K is a modeling parameter, which indicates how many different Gaussian sub-distributions are considered to constitute the reading distribution in a GMM fitting. The GMM is fitted by the EM algorithm to complete the offline GMM modeling. The liability score is calculated using the trained GMM parameters. The viscosity data within the monitoring range of the oil tank is used to calculate the liability score for different implicit components based on the trained GMM parameters, and the liability scores for different implicit components are labeled as follows: ,in, , Let be the weight corresponding to the k-th hidden component, representing the proportion belonging to the k-th state. The viscosity average value under the k-th implicit component. The viscosity variance under the k-th implicit component; The viscosity health index is calculated using the following formula: ;in, For viscosity health index, The health weights are for different components k.

4. The real-time lubricating oil condition monitoring method based on machine learning according to claim 3, characterized in that, The logic for obtaining the temperature index is as follows: The temperature data within the monitoring interval of the fuel tank is obtained. This temperature data is then fitted with time to determine the variation function of the temperature data with time within the monitoring interval. This variation function is then labeled as follows: ; The temperature index is calculated using the following formula: ;in, Temperature index This refers to the time period during operation when the temperature exceeds the set threshold.

5. The real-time lubricating oil condition monitoring method based on machine learning according to claim 4, characterized in that, Determine the condition of the lubricating oil in the tank, including: Based on the output of the fuzzy Bayesian neural network model within the monitoring interval, the probability values ​​of different fault levels within the monitoring interval are weighted and summed to obtain the fuel tank fault rating coefficient. The formula for calculating the fuel tank fault rating coefficient is as follows: ;in, This is the fuel tank failure rating coefficient. This is the proportionality coefficient representing the probability value of the fault-free level. This is a proportionality coefficient representing the probability value of a low failure level. This is the proportionality coefficient for the probability value of a medium-level fault. These are the proportional coefficients for high failure levels. .

6. The real-time lubricating oil condition monitoring method based on machine learning according to claim 5, characterized in that, Risk information at different pre-defined locations on different pipelines includes: Risk information at different preset locations in the pipeline is represented by a particle deposition risk assessment coefficient. The logic for obtaining the particle deposition risk assessment coefficient is as follows: based on the particles of different particle sizes of lubricating oil at preset positions in the pipeline within the monitoring interval, determine the impact of particles of different particle sizes on the operating status, and obtain the impact score of particles of different particle sizes based on the impact of particles of different particle sizes on the operating status. Based on the particles of different sizes collected at preset locations in the pipeline within the monitoring interval, the impact scores and concentrations of particles of different sizes are used as input features of the logistic regression model, and the particle deposition risk assessment coefficient is used as the output feature of the logistic regression model to construct the logistic regression model and generate the particle deposition risk assessment coefficient. The expression for the particle deposition risk assessment coefficient is as follows: ; in, This is the particle deposition risk assessment coefficient. The weights of particles of different sizes, The influence of particles of different size ranges is scored, with e as the base.

7. The real-time lubricating oil condition monitoring method based on machine learning according to claim 6, characterized in that, Determine the state of the lubricating oil in the pipeline, including: A comprehensive analysis of physical and risk information at predetermined locations within the pipeline is conducted. The pressure difference deviation amplitude coefficient, mechanical acoustic spectrum coefficient, and particle deposition risk assessment coefficient are weighted and summed to construct a pipeline location assessment model and generate pipeline location assessment coefficients. The formula for calculating the pipeline location assessment coefficients is as follows: ;in, Let be the pipeline position evaluation coefficient at the preset position j, where j = 1, 2, 3, ..., J, J is a positive integer, and j is the number of the preset position in the pipeline. , , These are the proportional coefficients for the differential pressure deviation amplitude coefficient, the mechanical acoustic spectrum measurement coefficient, and the particle deposition risk assessment coefficient, respectively. , , All are greater than 0; Based on the number of preset positions in the pipeline and their impact on the overall operating state of the lubricating oil, the overall pipeline fault score is obtained by weighted summation of pipeline position evaluation coefficients at different preset positions. The formula for calculating the pipeline fault score coefficient is as follows: ;in, This is the pipeline fault rating coefficient. Weights at different preset positions.

8. The real-time lubricating oil condition monitoring method based on machine learning according to claim 7, characterized in that, Based on the trained neural network model, the state of lubricating oil in the tank and pipeline is analyzed to achieve an early warning function for the lubricating oil state, including: The fault warning coefficient is calculated using the following formula: ;in, This is the fault warning coefficient. It is a specific proportionality coefficient; The tank fault rating coefficient, pipeline fault rating coefficient, and monitoring interval time length are used as inputs to train the neural network model, and the fault warning coefficient is used as the output of the neural network model. The model is trained by assigning a value of 0 or 1, where 0 indicates that the fault warning coefficient is less than or equal to the fault warning coefficient threshold, and 1 indicates that the fault warning coefficient is greater than the fault warning coefficient threshold.

Images