Oil opening flash point online prediction method and system based on multiple physical quantities and temperature and oil age correction

Abstract

The invention discloses an oil flash point online prediction method based on multiple physical quantities and temperature and use duration correction, and belongs to the field of oil performance monitoring and equipment health management. The online prediction method comprises the following steps: collecting multi-physical-quantity sensing signals of oil in the operation process of target equipment, and executing timestamp synchronization, interpolation alignment, abnormal value elimination and sample matrix construction; performing benchmark processing on the temperature sensitive variable by using the temperature correction model; the oil temperature is monitored in real time, the accumulated high-temperature operation duration when the temperature exceeds a threshold value is calculated, time correction is conducted on a flash point according to an empirical formula and an exponential decay model, and the flash point decline effect caused by aging is compensated; taking the multi-physical quantity vector after temperature and oil age correction as input, constructing a multi-model fusion prediction framework, and generating a flash point prediction value through a weighted fusion or stacking strategy; and comparing the predicted flash point with a baseline value, executing hierarchical early warning, generating a structured evidence chain log, and recording input parameters, a model version, a prediction result and threshold setting.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and safety assessment of oil condition, and in particular to an online prediction method and system for the open flash point of oil based on multiple physical quantities and temperature and oil age correction. It is mainly applied to power equipment such as diesel engines and other industrial lubrication systems to realize real-time assessment, deterioration identification and risk warning of lubricating oil quality. Background Technology

[0002] Currently, the determination of oil flash point mainly relies on standard laboratory methods, such as the open-cup flash point method (GB / T3536-2008) and the closed-cup flash point method (GB / T261-2021). While these methods offer high measurement accuracy, they suffer from drawbacks such as complex sampling, long testing cycles, high levels of human intervention, and the inability to achieve continuous online monitoring. For critical equipment requiring real-time risk warnings, this offline testing approach is no longer sufficient to meet the demands of intelligent operation and maintenance. Therefore, how to utilize online monitoring methods to dynamically predict flash point or related characteristics (such as fuel dilution) has become a research hotspot in the field of oil monitoring.

[0003] With the development of the Industrial Internet of Things (IIoT) and intelligent sensing technology, an increasing number of studies are exploring the use of online sensors to measure various physical quantities of oil in real time, including temperature, dielectric constant, density, and viscosity, in order to build oil quality prediction models. Multi-parameter fusion sensing can achieve real-time monitoring of characteristics such as lubricating oil viscosity and dielectric constant; however, most research remains at the level of health assessment or lifespan prediction, with models largely based on static empirical regression and lacking dynamic flash point prediction mechanisms across time scales. Furthermore, while some device-type patents or industrial solutions possess the ability to simultaneously measure temperature, viscosity, and density online, their algorithms mainly rely on empirical formulas or rule-based judgments, and a complete time-correction and prediction coupling framework has not yet been formed.

[0004] While the aforementioned parameters provide a data foundation for flash point prediction, they are all highly sensitive to temperature. Without temperature correction, measurement results under different operating conditions will exhibit systematic deviations, leading to unstable model outputs or even model failure. Besides temperature, the aging degree of lubricating oil (i.e., oil age) is also a crucial factor affecting flash point. In this invention, oil age refers to the cumulative effect of multiple processes since the oil was put into use, including high-temperature oxidation, fuel dilution, mechanical shearing, and additive dissipation. Experimental studies show that as oil age increases, the volatilization of light components, accumulation of oxidation products, and increased fuel dilution lead to a significant downward trend in flash point. This process exhibits a significant time dependence, with substantial differences in the decay rate under different operating conditions. Therefore, establishing a static model based solely on instantaneous physical quantities, while ignoring the dynamic influence of oil age, makes it difficult to accurately reflect the true degree of fuel dilution and safety risks of the oil.

[0005] In existing research, some scholars have attempted to predict flash point using single-parameter or simplified multivariate linear models, but these models often fail to incorporate a coupled correction mechanism for temperature and oil age. Current research on flash point prediction models shows that most models are based on static physical properties (such as boiling point, vapor pressure, and molecular structure characteristics). While existing pure-substance flash point prediction models possess high accuracy under static conditions, they also fail to consider the effects of cyclical or aging processes. In summary, existing flash point prediction techniques are mostly static or semi-empirical models, generally lacking a joint correction mechanism that simultaneously considers temperature drift and oil age decay.

[0006] With the development of remote intelligent operation and maintenance and condition monitoring technologies for mechanical equipment, higher requirements are placed on online oil prediction models: they must not only achieve real-time performance and high accuracy, but also possess adaptive capabilities to changes in ambient temperature and oil usage cycles. To address these issues, this invention proposes an open-face flash point prediction framework based on multi-physical quantity fusion. This framework simultaneously introduces temperature correction and oil age correction mechanisms at the model input, standardizing instantaneous physical quantities to a unified reference temperature. Oil age is identified through the cumulative time integration of oil temperature, achieving dynamic compensation for flash point decay over usage time. This method maintains consistency and interpretability of prediction results under different operating conditions and stages, providing a new technical approach for early warning of lubricant quality degradation and risk control of fuel dilution in industrial settings. Summary of the Invention

[0007] This invention aims to overcome the shortcomings of existing technologies and proposes an online prediction method and system for oil flash point based on multiple physical quantities, temperature, and usage time correction. Through multi-source sensor data fusion and dual correction algorithms, it can predict the flash point and water content of lubricating oil, and is used to determine in real time whether the quality of lubricating oil has declined and whether there is a risk of diesel fuel mixing.

[0008] This invention provides an online prediction method for oil flash point based on multiple physical quantities, temperature, and usage time correction, characterized by the following steps:

[0009] S1: Data Acquisition and Preprocessing. Online signals are acquired through multi-physical quantity sensors (temperature, dynamic viscosity, density, dielectric constant) of the oil, and timestamp synchronization, interpolation alignment, outlier removal, and sample matrix construction are performed. The data acquisition module can adopt a split-type or multi-parameter integrated structure to achieve multi-point synchronous sampling.

[0010] S2: Temperature correction processing. For temperature-sensitive variables (dynamic viscosity, density, dielectric constant), a standard conversion model is used to standardize the temperature, converting all physical quantities to the reference temperature to ensure the comparability of data under different operating conditions and the stability of model input.

[0011] S3: Lubricating oil usage time identification and correction. The system monitors oil temperature in real time. When the temperature exceeds a set threshold, the accumulated high-temperature running time is used as an indicator of oil age. Based on empirical formulas and exponential decay models, the flash point is corrected over time to compensate for the flash point decrease caused by aging, thereby reflecting the true deterioration state of the oil.

[0012] S4: Multi-physical quantity fusion prediction. Using temperature-corrected and oil age-corrected multi-physical quantity vectors as input, a multi-model fusion prediction framework is constructed, including linear regression, partial least squares regression, support vector regression, and small-sample adaptive models. Flash point prediction values ​​are generated through weighted fusion or stacking strategies, achieving real-time estimation.

[0013] S5: Prediction and evidence chain log output compares the predicted flash point with the baseline value. When the decline exceeds a threshold, a tiered warning is issued. Simultaneously, a structured evidence chain log is generated, recording input parameters, model version, prediction results, and threshold settings to ensure the traceability of warning results.

[0014] In one embodiment of the present invention, step S1 of the data acquisition and preprocessing method includes the following steps:

[0015] S101: The object of online monitoring and the measurement parameters are the circulating oil of a running diesel engine or industrial lubrication system. The monitoring parameters include the following physical quantities that can be measured online.

[0016] Oil temperature T(t): A platinum resistance temperature sensor (Pt100 / NTC type) or an integrated thermistor is used, installed in the main oil circuit or return oil branch, to reflect the thermal state of the oil in real time.

[0017] Dynamic viscosity μ(t): obtained by a vibratory microfluidic viscometer, a resonant fork sensor, or an integrated multi-parameter probe;

[0018] Density ρ(t): Measured in real time using a density sensor based on the principles of vibrating tube, ultrasonic, or piezoelectric resonance;

[0019] Dielectric constant ε(t): obtained through capacitive or microwave resonant dielectric sensors to reflect changes in the polar components and contamination level of the oil.

[0020] Water-active a w (t): Measured in real time by a water activity sensor to characterize the thermodynamic activity state of dissolved water in the oil;

[0021] Moisture content w(t): Measured by a capacitive or infrared absorption moisture sensor, used to characterize changes in the moisture content of oil.

[0022] S102: Sensor structure and sampling method:

[0023] Split-type sensor deployment: Each sensor is independently installed at a different location in the oil circuit and collects its own signals through a standard interface, which is suitable for multi-point deployment;

[0024] Multi-parameter integrated sensor: It adopts an integrated module to simultaneously measure multiple physical quantities such as temperature, density, dielectric constant, and viscosity. It has the advantages of easy installation and high data synchronization accuracy, and is suitable for on-site online monitoring.

[0025] Sensors are typically located in the flow stabilization bypass section or the main oil circuit sampling branch to ensure that the fluid temperature is consistent with the main system and to avoid bubble interference. The sampling frequency can be set from 1 to 10 Hz to meet dynamic response requirements. All channels are equipped with signal conditioning circuits and A / D conversion modules to output digital signals in a uniform format.

[0026] S103: Timestamp Synchronization and Data Alignment. Due to differences in sampling periods and communication delays among different sensors, the system adds a timestamp to each data entry using a unified clock reference to achieve synchronization of multi-channel data. The main steps include: performing linear or spline interpolation on channels with inconsistent sampling periods to align all physical quantities on the same time axis; and employing a sliding window mean or last valid value hold strategy for short-term missing data points to ensure the continuity of time-series data.

[0027] S104: Outlier identification, using statistical methods within a sliding window (such as the 3σ criterion or IQR detection) to eliminate transient changes and bubble interference.

[0028] S105: Data matrix construction and cache management; synchronized multi-channel data is arranged into a sample matrix according to time series.

[0029] X(t i )=[T(t i ),μ(t i ),ρ(t i ),ε(t i ),a w (t i ),w(t i )]

[0030] Where t i This represents the i-th sampling time. The sample matrix has dimensions n×m, where n is the number of sampling points and m is the dimension of the physical quantity. The matrix can be cached in blocks within the edge computing unit and updated in real time to support online model inference.

[0031] In one embodiment of the present invention, step S2 of the temperature correction processing method includes the following steps:

[0032] S201: Determination of Temperature-Sensitive Variables. Among the multiple physical quantities monitored in online oil flow, dynamic viscosity, density, and dielectric constant are the three most significantly affected by temperature. Dynamic viscosity decreases significantly with increasing temperature, exhibiting an approximately exponential relationship; density decreases slightly with increasing temperature, which can be considered a linear relationship; dielectric constant decreases slowly with increasing temperature, showing an approximately linear trend. Temperature-based corrections are needed for these three variables to eliminate measurement biases caused by instantaneous temperature changes.

[0033] S202: Reference Temperature Setting and Conversion Principles: The system sets the reference temperature T based on equipment characteristics and oil type. ref Typically, 25℃ (laboratory conditions) or 40℃ (industrial standard conditions) is used. The current temperature is defined as T. All online temperature-sensitive quantities must be converted to T. ref The equivalent value is determined to ensure data consistency under different times and operating conditions.

[0034] S203: Temperature Correction Model and Calculation Formulas. The conversion formulas for each temperature-sensitive variable are shown below:

[0035] Dynamic viscosity temperature correction: The Andrade model is used to describe the exponential relationship between dynamic viscosity and temperature.

[0036]

[0037] Where T K =T+273.15, T ref,K =T ref +273.15, constants A and B can be determined by the two-point calibration method of standard oil samples.

[0038] Density-temperature correction: using linear thermal expansion relationship.

[0039]

[0040] Where α is the volume expansion coefficient, and its value ranges from 6 × 10⁻⁶. -4 ~9×10 -4 ℃ -1 .

[0041] Temperature correction for dielectric constant: using an empirical linear model.

[0042]

[0043] Where λ is the dielectric temperature sensitivity coefficient, typically ranging from (2 to 5) × 10⁻⁶. -3 ℃ -1 .

[0044] S204: Correction parameter calibration experiment calibration: Through constant temperature bath experiment, standard oil sample data were collected at different temperatures, and the parameters A, B, α, λ were obtained by least squares fitting.

[0045] S205: Collect the current oil temperature T(t) and the original values ​​of each physical quantity μ(T), ρ(T), ε(T), and set the reference temperature T. ref The calibrated parameters are then corrected to obtain the temperature-corrected μ(T). ref ),ρ(T ref ),ε(T ref );

[0046] In one embodiment of the present invention, step S3 of the temperature correction processing method includes the following steps:

[0047] S301: Identify usage time. The system uses the oil temperature signal T(t) for real-time monitoring. When the oil temperature exceeds the set high temperature threshold T... hot When the temperature reaches 60℃ (typically taken as 60℃), it is considered to be in a high-temperature operating state, and the time accumulation begins. The hot-state operating condition is identified using the judgment function δ(·), and the formula for calculating the accumulated high-temperature operating time H(t) is as follows:

[0048]

[0049] Where: H(t) is the cumulative high-temperature operating time of the lubricating oil, in hours; T(τ) is the oil temperature at time τ; T hot δ(x) is the high temperature threshold; δ(x) is the step function, which takes the value of 1 when x≥0 and 0 otherwise.

[0050] In practical discrete sampling systems, this integral form can be discretized into a summation calculation:

[0051] H k =H k-1 +t s δ(T k -T hot )

[0052] Where t s T represents the sampling interval; k H represents the oil temperature at the k-th sampling time. k-1 The cumulative high-temperature running time at the previous moment; δ(x) is the step function, which takes a value of 1 when x≥0, and 0 otherwise; the calculated H k This represents the cumulative high-temperature operating time of the lubricating oil (in hours). The system can update this value in real time and record it in the edge computing module with each sampling cycle for subsequent oil age correction and flash point decay analysis. The following text will not specifically distinguish between H(t) and its discrete form Ht. k , which is denoted as H(t).

[0053] S302: With increasing usage time, the flash point of the fuel continuously decreases due to phenomena such as the gradual volatilization of light hydrocarbon components, accumulation of oxidation products, fuel dilution, and shear thinning. Experiments have shown that the flash point decreases with fuel age in a quasi-linear or saturated manner. To compensate for this decrease in flash point with usage time, a linear or saturated time correction term is introduced into the flash point prediction model. The correction formula is as follows:

[0054]

[0055] in: This is the predicted flash point value after oil age correction; k is the flash point value predicted by the model after temperature correction. t The decay coefficient of flash point with oil age (unit: °C / h) can be obtained through experimental calibration or online learning; H max This is the duration saturation value, used to prevent over-correction.

[0056] When H(t) < H max When H(t) ≥ H, the flash point decreases linearly with usage time; when H(t) ≥ H max At this point, the degradation tends to saturate. This model can effectively describe the flash point change characteristics of lubricating oil at different aging stages.

[0057] S303: Calibration of correction parameters, obtained by detecting the flash point of oil samples with different usage times and using linear regression or exponential decay fitting to obtain k. t and H max ;

[0058] S304: After system maintenance and oil change, the timer is automatically reset to start a new oil age cycle.

[0059] S4: Multi-physical quantity fusion prediction, using the corrected variable vector Using the input as input, construct a predictive model to output the flash point.

[0060] In one embodiment of the present invention, step S4 of the temperature correction processing method includes the following steps:

[0061] S401: Input variable composition; the oil state parameters corrected in steps S2 and S3 constitute the model input feature vector.

[0062]

[0063] Where: X corr (t) The input feature vector; T ref To set a reference temperature; This is the temperature-corrected dynamic viscosity; ρ Tref Density after temperature correction; H(t) is the temperature-corrected dielectric constant; H(t) is the cumulative high-temperature service time of the lubricating oil (oil age).

[0064] S402: Model selection and small sample matching principle. Considering the limited number of samples actually available, this invention prioritizes model architectures with small sample adaptability and high interpretability. This mainly includes:

[0065] Linear Regression (LR) Model: This method is suitable for scenarios where variables have an approximately linear relationship. It does not require complex data preprocessing and model parameter tuning, and can quickly build a stable baseline prediction model, providing an intuitive reference benchmark for subsequent optimization iterations.

[0066] Partial Least Squares Regression (PLS): Suitable for small sample scenarios with significant multivariate collinearity, and has good stability when the data dimension is high;

[0067] Support Vector Regression (SVR): Uses kernel functions (RBF, Polynomial, etc.) to capture non-linear relationships, suitable for medium-sized datasets;

[0068] Lightweight Random Forest (RF_lite) or Extreme Gradient Boosting (XGBoost_small): In the case of small samples, overfitting can be suppressed by subsampling and shallow tree structure;

[0069] Model fusion mechanism (Ensemble): Integrates the prediction results of multiple models through weighted averaging or stacking to improve robustness.

[0070] S403: Multi-model fusion prediction mechanism. When the system has multiple candidate models, the fusion prediction value is obtained by weighted combination.

[0071]

[0072] in, Let ω be the predicted output of the i-th model. i Its dynamic weights satisfy ∑ i ω i =1. Weights can be determined based on the performance of each model on the historical validation set, such as:

[0073]

[0074] This mechanism enables the dynamic binding of model weights with prediction reliability, achieving complementary integration of different algorithms.

[0075] S404: Model Training and Validation. Data Preparation: Construct a training set using experimental samples; Feature Preprocessing: Standardize or Z-score normalize each input variable to avoid dimensional differences; Cross-validation Strategy: Use leave-one-out cross-validation (LOO-CV) or five-fold cross-validation to evaluate small-sample generalization performance; Evaluation Metrics: Root Mean Square Error (RMSE) and Coefficient of Determination (R²) are used. 2 As a key performance indicator; Model update: As new monitoring data accumulates, the model can be retrained and the weight parameters replaced periodically (e.g., quarterly).

[0076] In one embodiment of the present invention, step S5 of the temperature correction processing method includes the following steps:

[0077] S501: Anomaly detection rules and threshold settings; the system uses new oil or initial operating data as the baseline flash point value F. base Real-time comparison of the current predicted flash point The relative deviation from the baseline. When A value significantly lower than the baseline indicates that the lubricating oil may be affected by fuel dilution or other factors. The judgment criteria are as follows:

[0078]

[0079] If satisfied

[0080]

[0081] This triggers an anomaly warning. Where η F This represents the percentage of the flash point drop threshold, typically set within the range of 5%-10%.

[0082] S502: Tiered early warning strategy. To avoid false alarms and ignoring potential risks, the system is designed with a three-tiered early warning mechanism:

[0083] Level I (Minor Risk): η F =5%, the flash point is slightly below the baseline, indicating an early sign of flash point decline;

[0084] Level II (Medium Risk): η F =8%, flash point decreased significantly, it is recommended to retest the oil sample;

[0085] Level III (Severe Risk): η F If the flash point drops rapidly by ≥10%, it is considered a high-risk condition, and the system will immediately issue an alarm and recommend stopping the machine for inspection.

[0086] The warning thresholds at each level are customized in the configuration file according to different equipment types, operating conditions and safety requirements.

[0087] S503: Evidence Chain Log Structure and Traceability Mechanism. To ensure the transparency and verifiability of the model output, this invention establishes a standardized "evidence chain log" structure. Each flashpoint prediction and warning event generates an independent record, the main contents of which include:

[0088] Time information: Alarm time

[0089] Parameter information: timestamp, corrected variable vector X corr (t), oil age H(t), temperature T ref ;

[0090] Model information: Model name;

[0091] Prediction results: Outputs of each model Fusion results Prediction confidence level;

[0092] Warning status: anomaly level, trigger threshold;

[0093] Another objective of this invention is to provide a system for implementing the above-mentioned online prediction method of oil open flash point based on multiple physical quantities, temperature, and oil age correction. The system includes a sensor module, a data acquisition and preprocessing module, a prediction and analysis module, and an early warning output module.

[0094] Sensor module: Used for real-time monitoring of various physical parameters of lubricating oil, including temperature, dynamic viscosity, density, and dielectric constant. Sensors can be installed in a separate unit or as a multi-parameter integrated sensor to meet the installation and maintenance needs of different application scenarios. All sensors have temperature compensation capabilities and connect to the acquisition module via a digital signal output interface (such as RS485 or CAN bus).

[0095] Data Acquisition and Preprocessing Module: This module is responsible for acquiring signals from various sensor channels and performing time synchronization, interpolation alignment, outlier detection, and temperature benchmark correction. It incorporates a real-time clock (RTC) for timestamp synchronization and possesses edge computing capabilities, enabling real-time calculations for temperature correction and oil age identification. The processed data is stored as a time-series matrix, providing standardized input for subsequent predictive models.

[0096] Predictive Analysis Module: Based on the modified multi-physical quantity input vector, it calls the embedded prediction model (including linear regression, partial least squares, support vector regression or multi-model fusion structure) to calculate the predicted flash point value of lubricating oil.

[0097] Early warning output module: Monitors predicted flash point results in real time and compares them with baseline values. When the flash point drop exceeds a set threshold, the system automatically determines and outputs a tiered early warning. Warning information is displayed visually on the local monitoring interface and can automatically generate structured evidence chain logs.

[0098] The system features a clear modular design, real-time performance, scalability, and traceability. It can dynamically predict the flash point of lubricating oil, detect anomalies, and provide intelligent early warnings in industrial settings, offering strong technical support for the safe operation of machinery and equipment.

[0099] Compared with the prior art, the present invention has the following beneficial effects: (1) By introducing a multi-physical quantity fusion and a dual correction mechanism of temperature and oil age, the online dynamic prediction of the flash point of lubricating oil is realized, overcoming the shortcomings of traditional methods that rely only on a single parameter or static model and cannot reflect the performance changes of oil under different temperatures and usage stages; (2) By adopting a multi-model fusion structure (including linear regression, partial least squares, and support vector regression) adapted to small sample conditions, the prediction stability and generalization ability of the model under limited data conditions are significantly improved, ensuring that high-precision output is maintained under different working conditions and operating stages; (3) The system has the features of hierarchical early warning and traceable evidence chain log, which can automatically identify risks, output multi-level alarms and generate evidence chain logs when the flash point drops beyond the set threshold, realizing closed-loop management of the entire process from data collection, intelligent prediction to risk early warning, and providing high-reliability technical guarantee for the safe operation of diesel engines and key equipment. Attached Figure Description

[0100] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0101] Figure 1 : Flowchart of the method of this invention;

[0102] Figure 2 System structure diagram;

[0103] Figure 3 Temperature correction diagram;

[0104] Figure 4 : Duration correction diagram;

[0105] Figure 5 Flash point fitting model diagram;

[0106] Figure 6 : Comparison chart of experimental and predicted results. Detailed Implementation

[0107] To provide a clearer understanding of the technical solution of this invention, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. These embodiments are illustrative and do not limit the scope of protection of this invention.

[0108] This embodiment focuses on a diesel engine lubricating oil circulation system. Utilizing multi-physical quantity monitoring data, and after dual correction based on temperature and oil age, the data is input into a flash point prediction model to achieve dynamic estimation and risk assessment of the flash point. The overall process is as follows: Figure 1 As shown, it includes the following steps:

[0109] S1: Data Acquisition and Preprocessing. Online signals are acquired through multi-physical quantity sensors (temperature, dynamic viscosity, density, dielectric constant) of the oil, and timestamp synchronization, interpolation alignment, outlier removal, and sample matrix construction are performed. The data acquisition module can adopt a split-type or multi-parameter integrated structure to achieve multi-point synchronous acquisition.

[0110] S2: Temperature correction processing. For temperature-sensitive variables (dynamic viscosity, density, dielectric constant), a standard conversion model is used to standardize the temperature, converting all physical quantities to the reference temperature to ensure the comparability of data under different operating conditions and the stability of model input.

[0111] S3: Lubricating oil usage time identification and correction

[0112] The system monitors oil temperature in real time, and the accumulated high-temperature running time when the temperature exceeds a set threshold is used as an indicator of oil age. Flash point is corrected for time based on empirical formulas and an exponential decay model to compensate for the decrease in flash point caused by aging, thus reflecting the true deterioration state of the oil.

[0113] S4: Multi-physics fusion prediction

[0114] Using temperature-corrected and oil age-corrected multi-physical quantity vectors as inputs, a multi-model fusion prediction framework is constructed, including linear regression, partial least squares regression, and support vector regression small-sample adaptive models. Flash point prediction values ​​are generated through weighted fusion or stacking strategies to achieve real-time estimation.

[0115] S5: Warning and Evidence Chain Log Output

[0116] The predicted flash point is compared with the baseline value, and a tiered warning is issued when the decline exceeds a threshold. Simultaneously, a structured evidence chain log is generated, recording input parameters, model version, prediction results, and threshold settings to ensure the traceability of the warning results.

[0117] In one embodiment of the present invention, in step S1, the data acquisition and preprocessing method includes the following steps:

[0118] S101: Online monitoring objects and measurement parameters. The object of data collection is the circulating oil of a running diesel engine or industrial lubrication system. The monitoring parameters include the following physical quantities that can be measured online:

[0119] Oil temperature T(t): A platinum resistance temperature sensor (Pt100 / NTC type) or an integrated thermistor is used, installed in the main oil circuit or return oil branch, to reflect the thermal state of the oil in real time.

[0120] Dynamic viscosity μ(t): obtained by a vibratory microfluidic viscometer, a resonant fork sensor, or an integrated multi-parameter probe;

[0121] Density ρ(t): Measured in real time using a density sensor based on the principles of vibrating tube, ultrasonic, or piezoelectric resonance;

[0122] Dielectric constant ε(t): obtained through capacitive or microwave resonant dielectric sensors to reflect changes in the polar components and contamination level of the oil.

[0123] Water-active a w (t): Measured in real time by a water activity sensor to characterize the thermodynamic activity state of dissolved water in the oil;

[0124] Moisture content w(t): Measured by a capacitive or infrared absorption moisture sensor, used to characterize changes in the moisture content of oil.

[0125] S102: Sensor structure and sampling method:

[0126] Split-type sensor deployment: Each sensor is independently installed at a different location in the oil circuit and collects its own signals through a standard interface, which is suitable for multi-point deployment;

[0127] Multi-parameter integrated sensor: It adopts an integrated module to simultaneously measure multiple physical quantities such as temperature, density, dielectric constant, and viscosity. It has the advantages of easy installation and high data synchronization accuracy, and is suitable for on-site online monitoring.

[0128] Sensors are typically located in the flow stabilization bypass section or the main oil circuit sampling branch to ensure that the fluid temperature is consistent with the main system and to avoid bubble interference. The sampling frequency can be set from 1 to 10 Hz to meet dynamic response requirements. All channels are equipped with signal conditioning circuits and A / D conversion modules to output digital signals in a uniform format.

[0129] S103: Timestamp Synchronization and Data Alignment. Due to differences in sampling periods and communication delays among different sensors, the system adds a timestamp to each data entry using a unified clock reference to achieve synchronization of multi-channel data. The main steps include: performing linear or spline interpolation on channels with inconsistent sampling periods to align all physical quantities on the same time axis; and employing a sliding window mean or last valid value hold strategy for short-term missing data points to ensure the continuity of time-series data.

[0130] S104: Outlier identification, using statistical methods within a sliding window (such as the 3σ criterion or IQR detection) to eliminate transient changes and bubble interference.

[0131] S105: Data matrix construction and cache management; synchronized multi-channel data is arranged into a sample matrix according to time series.

[0132] X(t i )=[T(t i ),μ(t i ),ρ(t i ),ε(t i ),a w (t i ),w(t i )]

[0133] Where t i This represents the i-th sampling time. The sample matrix has dimensions n×m, where n is the number of sampling points and m is the dimension of the physical quantity. The matrix can be cached in blocks within the edge computing unit and updated in real time to support online model inference.

[0134] In one embodiment of the present invention, step S2 of the temperature correction processing method includes the following steps:

[0135] S201: Determination of Temperature-Sensitive Variables. Among the multiple physical quantities monitored in online oil flow, dynamic viscosity, density, and dielectric constant are the three most significantly affected by temperature. Dynamic viscosity decreases significantly with increasing temperature, exhibiting an approximately exponential relationship; density decreases slightly with increasing temperature, which can be considered a linear relationship; dielectric constant decreases slowly with increasing temperature, showing an approximately linear trend. Temperature-based corrections are needed for these three variables to eliminate measurement biases caused by instantaneous temperature changes.

[0136] S202: Reference Temperature Setting and Conversion Principles: The system sets the reference temperature T based on equipment characteristics and oil type. ref Typically, 25℃ (laboratory conditions) or 40℃ (industrial standard conditions) is used. The current temperature is defined as T. All online temperature-sensitive quantities must be converted to T. ref The equivalent value is determined to ensure data consistency under different times and operating conditions.

[0137] S203: Temperature Correction Model and Calculation Formulas. The conversion formulas for each temperature-sensitive variable are shown below:

[0138] Dynamic viscosity temperature correction: The Andrade model is used to describe the exponential relationship between dynamic viscosity and temperature.

[0139]

[0140] Where T K =T+273.15, T ref,K =T ref +273.15, constants A and B can be determined by the two-point calibration method of standard oil samples.

[0141] Density-temperature correction: using linear thermal expansion relationship.

[0142]

[0143] Where α is the volume expansion coefficient, and its value ranges from 6 × 10⁻⁶. -4 ~9×10 -4 ℃ -1 .

[0144] Temperature correction for dielectric constant: using an empirical linear model.

[0145]

[0146] Where λ is the dielectric temperature sensitivity coefficient, typically ranging from (2 to 5) × 10⁻⁶. -3 ℃ -1 .

[0147] S204: Correction parameter calibration experiment calibration: Through constant temperature bath experiment, standard oil sample data were collected at different temperatures, and the parameters A, B, α, λ were obtained by least squares fitting.

[0148] S205: Collect the current oil temperature T(t) and the original values ​​of each physical quantity μ(T), ρ(T), ε(T), and set the reference temperature T. ref The calibrated parameters are then corrected to obtain the temperature-corrected μ(T). ref ),ρ(T ref ),ε(T ref );

[0149] In one embodiment of the present invention, step S3 of the temperature correction processing method includes the following steps:

[0150] S301: Identify usage time. The system uses the oil temperature signal T(t) for real-time monitoring. When the oil temperature exceeds the set high temperature threshold T... hotWhen the temperature is typically 60℃, it is considered to be in a high-temperature operating state, and the time accumulation begins. The hot-state operating condition is identified using the judgment function δ(·), and the formula for calculating the accumulated hot-state running time H(t) is as follows:

[0151]

[0152] Where: H(t): cumulative high-temperature operating time of lubricating oil, in hours; T(τ): oil temperature at time τ; T hot : High temperature threshold; δ(x): Step function, which takes the value of 1 when x≥0, and 0 otherwise.

[0153] In practical discrete sampling systems, this integral form can be discretized into a summation calculation:

[0154] H k =H k-1 +t s ·δ(T k -T hot )

[0155] Where t s T represents the sampling interval; k H represents the oil temperature at the k-th sampling time. k-1 The cumulative high-temperature running time at the previous moment; δ(x) is the step function, which takes a value of 1 when x≥0, and 0 otherwise; the calculated H k This represents the cumulative high-temperature operating time of the lubricating oil (in hours). The system can update this value in real time and record it in the edge computing module with each sampling cycle for subsequent oil age correction and flash point decay analysis. The following text will not specifically distinguish between H(t) and its discrete form Ht. k , which is denoted as H(t).

[0156] S302: With increasing usage time, the flash point of the oil continuously decreases due to phenomena such as the gradual volatilization of light hydrocarbon components, accumulation of oxidation products, fuel dilution, and shear thinning. Experiments have shown that the flash point decreases with oil age in a quasi-linear or saturated manner. To compensate for this decreasing trend in flash point with usage time, this invention introduces a linear or saturated time correction term into the flash point prediction model. The correction formula is as follows:

[0157]

[0158] in: This is the predicted flash point value after oil age correction; k is the flash point value predicted by the model after temperature correction. t The decay coefficient of flash point with oil age (unit: °C / h) can be obtained through experimental calibration or online learning; H maxThis is the duration saturation value, used to prevent over-correction.

[0159] When H(t) < H max When H(t) ≥ H, the flash point decreases linearly with usage time; when H(t) ≥ H max At this point, the degradation tends to saturate. This model can effectively describe the flash point change characteristics of lubricating oil at different aging stages.

[0160] S303: Calibration of correction parameters, obtained by detecting the flash point of oil samples with different usage times and using linear regression or exponential decay fitting to obtain k. t and H max ;

[0161] S304: After system maintenance and oil change, the timer is automatically reset to start a new oil age cycle.

[0162] S4: Multi-physical quantity fusion prediction, using the corrected variable vector Using the input as input, construct a predictive model to output the flash point.

[0163] In one embodiment of the present invention, step S4 of the temperature correction processing method includes the following steps:

[0164] S401: Input Variable Composition: The model input feature vector is composed of oil state parameters corrected by steps S2 and S3.

[0165]

[0166] Where: X corr (t) The input feature vector; T ref To set a reference temperature; This is the temperature-corrected dynamic viscosity; Density after temperature correction; H(t) is the temperature-corrected dielectric constant; H(t) is the cumulative high-temperature service time of the lubricating oil (oil age).

[0167] S402: Model selection and small sample matching principle. Considering the limited number of samples actually available, this invention prioritizes model architectures with small sample adaptability and high interpretability. This mainly includes:

[0168] Linear regression (LR) model: suitable for scenarios with approximately linear relationships between variables, and can quickly establish baseline predictions;

[0169] Partial Least Squares Regression (PLS): Suitable for small sample scenarios with significant multivariate collinearity, and has good stability when the data dimension is high;

[0170] Support Vector Regression (SVR): Uses kernel functions (RBF, Polynomial, etc.) to capture non-linear relationships, suitable for medium-sized datasets;

[0171] Lightweight Random Forest (RF_lite) or Extreme Gradient Boosting (XGBoost_small): In the case of small samples, overfitting can be suppressed by subsampling and shallow tree structure;

[0172] Model fusion mechanism (Ensemble): Integrates the prediction results of multiple models through weighted averaging or stacking to improve robustness.

[0173] S403: Multi-model fusion prediction mechanism. When the system has multiple candidate models, the fusion prediction value is obtained by weighted combination.

[0174]

[0175] in, Let ω be the predicted output of the i-th model. i Its dynamic weights satisfy ∑ i ω i =1. Weights can be determined based on the performance of each model on the historical validation set, such as:

[0176]

[0177] This mechanism enables the dynamic binding of model weights with prediction reliability, achieving complementary integration of different algorithms.

[0178] S404: Model Training and Validation. Data Preparation: Construct a training set using experimental samples; Feature Preprocessing: Standardize or Z-score normalize each input variable to avoid dimensional differences; Cross-validation Strategy: Use leave-one-out cross-validation (LOO-CV) or five-fold cross-validation to evaluate small-sample generalization performance; Evaluation Metrics: Root Mean Square Error (RMSE) and Coefficient of Determination (R²) are used. 2 As a key performance indicator; Model update: As new monitoring data accumulates, the model can be retrained and the weight parameters can be replaced as needed.

[0179] In one embodiment of the present invention, step S5 of the temperature correction processing method includes the following steps:

[0180] S501: Anomaly detection rules and threshold settings; the system uses new oil or initial operating data as the baseline flash point value F. base Real-time comparison of the current predicted flash point The relative deviation from the baseline. When A value significantly lower than the baseline indicates that the lubricating oil may be affected by fuel dilution or other factors. The judgment criteria are as follows:

[0181]

[0182] If satisfied

[0183]

[0184] This triggers an anomaly warning. Where η F This represents the percentage of the flash point drop threshold, typically set within the range of 5%-10%.

[0185] S502: Tiered early warning strategy. To avoid false alarms and ignoring potential risks, the system is designed with a three-tiered early warning mechanism:

[0186] Level I (Minor Risk): η F =5%, the flash point is slightly below baseline, indicating an early sign of lightning decline;

[0187] Level II (Medium Risk): η F =8%, flash point decreased significantly, it is recommended to retest the oil sample;

[0188] Level III (Severe Risk): η F If the flash point drops rapidly by ≥10%, it is considered a high-risk condition, and the system will immediately issue an alarm and recommend stopping the machine for inspection.

[0189] The warning thresholds at each level can be customized in the configuration file according to different equipment types, operating conditions and safety requirements.

[0190] S503: Evidence Chain Log Structure and Traceability Mechanism. To ensure the transparency and verifiability of the model output, this invention establishes a standardized "evidence chain log" structure. Each flashpoint prediction and warning event generates an independent record, the main contents of which include:

[0191] Time information: Alarm time

[0192] Parameter information: timestamp, corrected variable vector X corr (t), oil age H(t), reference temperature T ref ;

[0193] Model information: Model name;

[0194] Prediction results: Outputs of each model Fusion results Prediction confidence level;

[0195] Warning status: anomaly level, trigger threshold;

[0196] This invention also provides a system for implementing the above-mentioned online prediction method for the open-circuit flash point of oil based on multiple physical quantities, temperature, and oil age correction, and its system structure diagram is shown below. Figure 2As shown, it includes: a sensor module, a data acquisition and preprocessing module, a predictive analysis module, and an early warning output module.

[0197] Sensor module: Used for real-time monitoring of various physical parameters of lubricating oil, including temperature, dynamic viscosity, density, dielectric constant, and water activity. Sensors can be installed in a separate unit or as a multi-parameter integrated sensor to meet the installation and maintenance needs of different application scenarios. All sensors have temperature compensation capabilities and connect to the acquisition module via a digital signal output interface (such as RS485 or CAN bus).

[0198] Data Acquisition and Preprocessing Module: This module is responsible for acquiring signals from various sensor channels and performing time synchronization, interpolation alignment, outlier detection, and temperature benchmark correction. It incorporates a real-time clock (RTC) for timestamp synchronization and possesses edge computing capabilities, enabling real-time calculations for temperature correction and oil age identification. The processed data is stored as a time-series matrix, providing standardized input for subsequent predictive models.

[0199] Predictive Analysis Module: Based on the modified multi-physical quantity input vector, it calls the embedded prediction model (including linear regression, partial least squares, support vector regression or multi-model fusion structure) to calculate the predicted value of the flash point of lubricating oil.

[0200] Early warning output module: Monitors predicted flash point results in real time and compares them with baseline values. When the flash point drop exceeds a set threshold, the system automatically identifies it as a risk event and outputs a tiered early warning. Warning information is displayed visually on the local monitoring interface and can automatically generate structured evidence chain logs.

[0201] The system features a clear modular design, real-time performance, scalability, and traceability. It can dynamically predict the flash point of lubricating oil, detect anomalies, and provide intelligent early warnings in industrial settings, offering strong technical support for the safe operation of diesel engines and critical mechanical equipment.

[0202] Compared with the prior art, the present invention has the following beneficial effects: (1) By introducing a multi-physical quantity fusion and a dual correction mechanism of temperature and oil age, the online dynamic prediction of the flash point of lubricating oil is realized, overcoming the shortcomings of traditional methods that rely only on a single parameter or static model and cannot reflect the performance changes of oil under different temperatures and usage stages; (2) By adopting a multi-model fusion structure (including linear regression, partial least squares, and support vector regression) adapted to small sample conditions, the prediction stability and generalization ability of the model under limited data conditions are significantly improved, ensuring that high-precision output is maintained under different working conditions and operating stages; (3) The system has the features of hierarchical early warning and traceable evidence chain log, which can automatically identify risks, output multi-level alarms and generate evidence chain logs when the flash point drops beyond the set threshold, realizing closed-loop management of the entire process from data collection, intelligent prediction to risk early warning, and providing high-reliability technical guarantee for the safe operation of diesel engines and key equipment.

[0203] The specific implementation method is as follows:

[0204] Data Sources and Variable Descriptions: This embodiment uses 20 sets of sample data from the attached "Predicted Lightning Test Data.xlsx" file as the training and validation set. Considering that this patent no longer involves the moisture content parameter, the following online measurable physical quantities are selected as independent variables of the model: temperature T (°C), dynamic viscosity μ, density ρ, and dielectric constant ε. The target variable is the open-cup flash point F. The sample names indicate the mixing ratio of lubricating oil ("lubricant") and diesel oil ("diesel") and the state of the oil ("new / old" respectively, indicating new oil / oil used for 70 hours).

[0205] Table 1. Oil parameters and test data (approximately 25°C operating conditions)

[0206]

[0207] Table 2. Oil parameters and test data (approximately 40℃ operating condition)

[0208]

[0209] Data cleaning and preprocessing: Perform the following preprocessing on the original table to obtain {T,μ,ρ,ε,F};

[0210] Temperature Correction: The unified conversion formula for temperature-sensitive variables is as follows (the temperature difference in this dataset is small, so the temperature correction has a minor impact on the results and does not affect the implementation of the methodology). In this embodiment, each physical quantity is converted to the reference temperature. The correction formula is as follows:

[0211]

[0212] Wherein: T K =T+273.15, T ref,K =Tref +273.15, Kelvin value of temperature; B: Andrade constant; α: coefficient of volume expansion; λ: dielectric temperature sensitivity coefficient. Correction parameters were obtained through experimental calibration: B = 1000, α = 8 × 10⁻⁶. -4 λ = 3 × 10 -3 . Figure 3 This is a temperature correction graph.

[0213] Oil age correction: The system monitors oil temperature in real time, and when the temperature exceeds the threshold T... hot At 60℃, the accumulated high-temperature operating time H(t) begins. A step function δ(x) is defined to identify the thermal operating condition.

[0214]

[0215] In a discrete sampling system, this equation can be discretized as:

[0216] H k =H k-1 +t s ·δ(T k -T hot )

[0217] Where: t s T is the sampling interval. k The oil temperature is the temperature at the kth sampling point.

[0218] To compensate for the decrease in flash point with increasing oil age, this invention introduces an oil age correction model:

[0219]

[0220] in: The flash point after positive; k t : is the point attenuation coefficient (℃ / h), experimentally calibrated k t =0.0457; H max : Duration saturation value (70h). Figure 4 The oil age correction chart shows the trend of flash point values ​​of various oil samples as the service life changes.

[0221] To adapt to a small sample size of n=20, this embodiment constructs a lightweight "multi-model library":

[0222] LR (Linear Regression): Using standardized features as input, it fits a linear relationship as a baseline model.

[0223] PLS2 (Partial Least Squares, 2-component): Reduces dimensionality for small sample sizes to alleviate multicollinearity.

[0224] SVR (Support Vector Regression): It uses the RBF kernel and a grid search to obtain the optimal configuration of hyperparameters such as C and σ.

[0225] Mean fusion: To improve stability, the outputs of the three models are simply averaged, as shown in the following formula:

[0226]

[0227] All models were trained in a normalized feature space and their generalization performance was evaluated using LOO-CV.

[0228] An interpretable equation for linear regression (standardized feature space), with the following standardized variables:

[0229]

[0230] The predictive form of linear regression is:

[0231]

[0232] The coefficients of the fitting formula obtained on this dataset are as follows (based on the full fitting results trained with the same LOO-CV configuration):

[0233] b0 = 170.35, β T =17.4256,β μ =43.9702,β ρ =1.6956,β ε =1.3375

[0234] Explanation: The importance of coefficients can be directly compared in the standardized space; in this sample set, the dynamic viscosity (standardized coefficient β) μ The positive correlation with flash point is most significant; the temperature independent variable should be considered in the explicitly modified engineering implementation. Replace the original measurement.

[0235] The cross-validation results and model optimization using LOO-CV metrics are as follows (R 2 The higher the better, the lower the RMSE the better; (unit: °C)

[0236] Table 3 Model Validation Results

[0237]

[0238] Conclusion: On this dataset, the LR baseline achieved the best generalization performance, followed by PLS2; SVR did not show any advantage with extremely small samples. For engineering deployment, a dual-model redundancy strategy of "LR master model + PLS2 backup" can be adopted; as field data accumulates, more complex nonlinear models can be gradually introduced and re-compared. The results of each model are as follows: Figure 5 As shown.

[0239] Sample-level prediction versus actual results (LOO-CV, LR), experimental results compared to... Figure 6 As shown in the table below, the true values, predicted values, and residuals (predicted – true) of the 10 samples in Table 1 under LR cross-validation are given (unit: °C):

[0240] Table 4 Comparison of predicted and measured flash point values

[0241]

[0242] Observations: The overall residual distribution is in the middle and there is no obvious systematic bias; for pure lubricating oil samples (100% lubricant), the model has a slight underestimation or overestimation. In engineering deployment, the bias drift can be further suppressed by introducing "baseline flash point" and "time correction term".

[0243] Warning threshold and evidence chain log demonstration: Baseline flash point F of pure lubricating oil sample (e.g., "New-Run 100% Diesel 0%)" base =251℃. When the predicted value at a certain moment on site is... At that time, the relative decrease was:

[0244]

[0245] A Level III critical alert has been triggered. The system simultaneously writes the evidence chain log: timestamp, device number, {T,μ,ρ,ε} (or its T). ref The system includes correction values, model version and parameters (LR / PLS2), outputs of each sub-model and fusion weights (if fusion is enabled), as well as trigger thresholds and handling suggestions, to achieve auditability and traceability.

[0246] For initial deployment and retraining, it is recommended to use LR as the main model and PLS2 as a redundancy check. When the cumulative data volume is ≥100 sets and covers multiple operating conditions, the model library should be re-evaluated quarterly and nonlinear models such as SVR / GBDT should be considered. At the same time, H should be continuously recorded at the edge and an oil age correction term should be added to further reduce the risk of long-term drift.

[0247] The above description is merely a preferred embodiment of this application and an explanation of the technical principles used. Those skilled in the art should understand that the scope involved in this application is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept. For example, technical solutions formed by replacing the above-mentioned features with technical features with similar functions disclosed in this application (but not limited to) each other.

[0248] Apart from the technical features described in the specification, the other technical features are known to those skilled in the art. To highlight the innovative features of this invention, the other technical features will not be described in detail here.

Claims

1. A method for online prediction of oil flash point based on multiple physical quantities, temperature, and usage time correction, characterized in that, Includes the following steps: S1: Data acquisition and preprocessing. Online signals are acquired through multiple physical quantity sensors of oil, namely temperature, dynamic viscosity, density, and dielectric constant. Timestamp synchronization, interpolation alignment, outlier removal, and sample matrix construction are performed. The data acquisition module can adopt a split or multi-parameter integrated structure to achieve multi-point synchronous sampling. S2: Temperature correction processing. For temperature-sensitive variables such as dynamic viscosity, density, and dielectric constant, a standard conversion model is used to standardize the temperature, converting each physical quantity to the reference temperature to ensure the comparability of data under different working conditions and the stability of model input. S3: Lubricating oil usage time identification and correction. The system monitors the oil temperature in real time. When the temperature exceeds the set threshold, the accumulated high-temperature running time is used as an indicator of oil age. The flash point is corrected for time based on empirical formulas and exponential decay models to compensate for the flash point drop caused by aging, thereby reflecting the true deterioration state of the oil. S4: Multi-physical quantity fusion prediction. Using the multi-physical quantity vectors after temperature correction and oil age correction as input, a multi-model fusion prediction framework is constructed, including linear regression, partial least squares regression, support vector regression, and small sample adaptive models. Flash point prediction values ​​are generated through weighted fusion or stacking strategies to achieve real-time estimation. S5: Prediction and evidence chain log output compares the predicted flash point with the baseline value. When the decline exceeds the threshold, it is automatically judged as a risk and a graded warning is executed. At the same time, a structured evidence chain log is generated to record input parameters, model version, prediction results and threshold settings to ensure that the warning results are traceable.

2. The method according to claim 1, characterized in that: In step S1, the data acquisition and preprocessing method includes the following steps: S101: The object of online monitoring and the object of measurement are the circulating oil of a running diesel engine or industrial lubrication system. The monitoring parameters include the following physical quantities that can be measured online. Oil temperature T(t): A platinum resistance temperature sensor or an integrated thermistor is used, installed in the main oil flow or return branch, to reflect the thermal state of the oil in real time. Dynamic viscosity μ(t): obtained by a vibratory microfluidic viscometer, a resonant fork sensor, or an integrated multi-parameter probe; Density ρ(t): Measured in real time using a density sensor based on the principles of vibrating tube, ultrasonic, or piezoelectric resonance; Dielectric constant ε(t): obtained through capacitive or microwave resonant dielectric sensors to reflect changes in the polar components and contamination level of the oil. Water-active a w (t): Measured in real time by a water activity sensor to characterize the thermodynamic activity state of dissolved water in the oil; Moisture content w(t): Measured by a capacitive or infrared absorption moisture sensor, used to characterize changes in moisture content in oil. S102: Sensor structure and sampling method: Split-type sensor deployment: Each sensor is independently installed at a different location in the oil circuit and collects its own signals through a standard interface, which is suitable for multi-point deployment; Multi-parameter integrated sensor: It adopts an integrated module to simultaneously measure multiple physical quantities such as temperature, density, dielectric constant, and viscosity. It has the advantages of easy installation and high data synchronization accuracy, and is suitable for on-site online monitoring. The sensor is placed in the flow stabilization bypass section or the sampling branch of the main oil circuit; the sampling frequency can be set to 1–10Hz; all channels are equipped with signal conditioning circuits and A / D conversion modules to output digital signals in a uniform format. S103: Timestamp synchronization and data alignment. Due to differences in sampling periods and communication delays among different sensors, the system adds timestamps to each data entry using a unified clock reference to achieve synchronization of multi-channel data. The steps include: performing linear or spline interpolation on channels with inconsistent sampling periods to align all physical quantities on the same time axis. For data points that are missing for a short period of time, a sliding window mean or last valid value preservation strategy is adopted to ensure the continuity of time series data; S104: Outlier identification, using statistical methods within a sliding window (such as the 3σ criterion or IQR detection) to eliminate transient changes and bubble interference; S105: Data matrix construction and cache management; synchronized multi-channel data is arranged into a sample matrix according to time series. X(t i )=[T(t i ),μ(t i ),ρ(t i ),ε(t i ),a w (t i ),w(t i )] Where t i This represents the i-th sampling time; the sample matrix has an dimension of n×m, where n is the number of sampling points and m is the dimension of the physical quantity.

3. The method according to claim 1, characterized in that: In step S2, the temperature correction processing method includes the following steps: S201: Determination of temperature-sensitive variables. Among the multiple physical quantities monitored in online oil flow, dynamic viscosity, density, and dielectric constant are the three indicators most significantly affected by temperature. Dynamic viscosity: decreases significantly with increasing temperature, showing an approximately exponential relationship. Density: decreases slightly with increasing temperature, which can be considered a linear relationship. Dielectric constant: decreases slowly with increasing temperature, showing an approximately linear trend; temperature-based corrections are applied to the above three variables to eliminate measurement bias caused by instantaneous temperature changes; S202: Reference Temperature Setting and Conversion Principles: The system sets a unified reference temperature T based on equipment characteristics and oil type. ref Use 25℃ or 40℃; take the current temperature as T; all online temperature-sensitive quantities need to be converted to the value at T. ref The equivalent value below; S203: Temperature Correction Model and Calculation Formulas. The conversion formulas for each temperature-sensitive variable are shown below: Dynamic viscosity temperature correction: The Andrade model is used to describe the exponential relationship between dynamic viscosity and temperature. Where T K =T+273.15, T ref,K =T ref +273.15, constants A and B can be determined by the two-point calibration method of standard oil samples; Density-temperature correction: using linear thermal expansion relationship. Where α is the volume expansion coefficient, and its value ranges from 6 × 10⁻⁶. -4 ~9×10 -4 ℃ -1 ; Temperature correction for dielectric constant: using an empirical linear model. Where λ is the dielectric temperature sensitivity coefficient, typically ranging from (2 to 5) × 10⁻⁶. -3 ℃ -1 ; S204: Correction parameter calibration experiment calibration: Through constant temperature bath experiment, standard oil sample data were collected at different temperatures, and the parameters A, B, α, λ were obtained by least squares fitting. S205: Collect the current oil temperature T(t) and the original values ​​of each physical quantity μ(T), ρ(T), ε(T), and set the reference temperature T. ref The calibrated parameters are then corrected to obtain the temperature-corrected μ(T). ref ),ρ(T ref ),ε(T ref ).

4. The method according to claim 1, characterized in that: In step S3, the temperature correction processing method includes the following steps: S301: Identify usage time. The system uses the oil temperature signal T(t) for real-time monitoring. When the oil temperature exceeds the set high temperature threshold T... hot (Usually taken as 60℃), is considered to be in a high-temperature working state, and the time accumulation begins; the hot working condition is identified by the judgment function δ(·), and the formula for calculating the accumulated hot running time H(t) is as follows: Where: H(t): cumulative high-temperature operating time of lubricating oil, in hours; T(τ): oil temperature at time τ; T hot : High temperature threshold; δ(x): Step function, which takes the value of 1 when x≥0, and 0 otherwise; In practical discrete sampling systems, this integral form can be discretized into a summation calculation: H k =H k-1 +t s ·δ(T k -T hot ) Where t s T represents the sampling interval; k H represents the oil temperature at the k-th sampling time. k-1 The cumulative high-temperature running time at the previous moment; δ(x) is the step function, which takes a value of 1 when x≥0, and 0 otherwise; the calculated H k This refers to the cumulative high-temperature operating time of the lubricating oil (in hours). The system can update this value in real time and record it in the edge computing module with each sampling cycle for subsequent oil age correction and flash point decay analysis. The following text will not specifically distinguish between H(t) and its discrete form Ht. k Let H(t) be the common denominator. S302: As usage time increases, phenomena such as the gradual volatilization of light hydrocarbon components in the oil, accumulation of oxidation products, fuel dilution, and shear thinning lead to a continuous decrease in flash point. Experiments have shown that the flash point decreases with oil age in a quasi-linear or saturated manner. To compensate for the decreasing trend of flash point with increasing usage time, this invention introduces a linear or saturated time correction term into the flash point prediction model. The correction formula is as follows: in: This is the predicted flash point value after oil age correction; k is the flash point value predicted by the model after temperature correction. t The decay coefficient of flash point with oil age (unit: °C / h) can be obtained through experimental calibration or online learning; H max This is the duration saturation value, used to prevent over-correction; When H(t) < H max When H(t) ≥ H, the flash point decreases linearly with usage time; when H(t) ≥ H max When the flash point decays to saturation, the model can effectively describe the flash point change characteristics of lubricating oil at different aging stages. S303: Calibration of correction parameters, obtained by detecting the flash point of oil samples with different usage times and using linear regression or exponential decay fitting to obtain k. t and H max ; S304: After system maintenance and oil change, the timer is automatically reset to start a new oil age cycle; S4: Multi-physical quantity fusion prediction, using the corrected variable vector Using the input as input, construct a predictive model to output the flash point.

5. The method according to claim 1, characterized in that: In step S4, the temperature correction processing method includes the following steps: S401: Input variable composition; the oil state parameters corrected in steps S2 and S3 constitute the model input feature vector. Among them, X corr (t) The input feature vector; T ref To standardize reference temperatures; This is the temperature-corrected dynamic viscosity; Density after temperature correction; H(t) is the temperature-corrected dielectric constant; H(t) is the cumulative high-temperature service time of the lubricating oil (oil age). S402: Model selection and small sample matching principles. Considering the limited amount of samples actually available, this invention prioritizes model architectures with small sample adaptability and high interpretability; mainly including: Linear regression (LR) model: suitable for scenarios with approximately linear relationships between variables, and can quickly establish baseline predictions; Partial Least Squares Regression (PLS): Suitable for small sample scenarios with significant multivariate collinearity, and has good stability when the data dimension is high; Support Vector Regression (SVR): Uses kernel functions (RBF, Polynomial, etc.) to capture non-linear relationships, suitable for medium-sized datasets; Lightweight Random Forest (RF_lite) or Extreme Gradient Boosting (XGBoost_small): In the case of small samples, overfitting can be suppressed through subsampling and shallow tree structure; Model fusion mechanism (Ensemble): Integrates the prediction results of multiple models through weighted averaging or stacking to improve robustness; S403: Multi-model fusion prediction mechanism. When the system has multiple candidate models, the fusion prediction value is obtained by weighted combination. in, Let ω be the predicted output of the i-th model. i Its dynamic weights satisfy Σ i ω i =1; the weights can be determined based on the performance of each model on the historical validation set, such as: This mechanism enables the dynamic binding of model weights with prediction reliability, achieving complementary integration of different algorithms; S404: Model Training and Validation; Data Preparation: Constructing a training set using experimental samples; Feature Preprocessing: Standardizing or Z-score normalizing each input variable to avoid dimensional differences; Cross-validation Strategy: Using leave-one-out cross-validation (LOO-CV) or five-fold cross-validation to evaluate small-sample generalization performance; Evaluation Metrics: Root Mean Square Error (RMSE) and Coefficient of Determination R0 2 As a key performance indicator; Model update: As new monitoring data accumulates, the model can be retrained and the weight parameters replaced periodically.

6. The method according to claim 1, characterized in that: In step S5, the temperature correction processing method includes the following steps: S501: Anomaly detection rules and threshold settings, the system uses new oil or initial operating data as the baseline flash point value F. base Real-time comparison of the current predicted flash point The relative deviation from the baseline; when A value significantly lower than the baseline indicates that the lubricating oil may be affected by fuel dilution or other factors; the judgment criteria are as follows: If satisfied This triggers an anomaly warning; where η F This represents the percentage of the flash point drop threshold, typically set within a range of 5%-10%. S502: Tiered Early Warning Strategy: To avoid false alarms and ignoring potential risks, the system is designed with a three-tiered early warning mechanism: Level I (Minor Risk): η F =5%, flash point slightly below baseline, indicating early signs of fuel dilution; Level II (Medium Risk): η F =8%, flash point decreased significantly, it is recommended to retest the oil sample; Level III (Severe Risk): η F If the flash point decreases rapidly by ≥10%, it is considered a high-risk condition, and the system will immediately alarm and recommend stopping the machine for inspection. The warning thresholds at each level can be customized in the configuration file according to different equipment types, operating conditions and safety requirements; S503: Evidence Chain Log Structure and Traceability Mechanism: To ensure the transparency and verifiability of the model output, this invention establishes a standardized "evidence chain log" structure; each flashpoint prediction and warning event generates an independent record, the main contents of which include: Time information: Alarm time Parameter information: timestamp, corrected variable vector X corr (t), oil age H(t), temperature T ref ; Model information: Model name; Prediction results: Outputs of each model Fusion results Prediction confidence level; Warning status: Abnormal level, trigger threshold.

7. A system for online prediction of the open-circuit flash point of oil based on multiple physical quantities and oil age correction, characterized in that, include: Sensor module: Used to monitor various physical parameters of lubricating oil in real time, including temperature, dynamic viscosity, density, dielectric constant, etc. The sensors can be installed in a split configuration or as a multi-parameter integrated sensor to meet the installation and maintenance needs of different application scenarios; all sensors have temperature compensation function and are connected to the acquisition module through a digital signal output interface; Data acquisition and preprocessing module: responsible for acquiring signals from various sensor channels and performing time synchronization, interpolation alignment, outlier detection, and temperature benchmark correction; this module has a built-in real-time clock (RTC) for timestamp synchronization and also has edge computing capabilities, which can complete the calculation of temperature correction and oil age identification in real time; the processed data is stored in the form of a time series matrix to provide standardized input for subsequent prediction models; Predictive Analysis Module: Based on the modified multi-physical quantity input vector, it calls embedded prediction models, including linear regression, partial least squares, support vector regression, or multi-model fusion structure, to calculate the predicted value of the flash point of lubricating oil. Early warning output module: Real-time monitoring of predicted flash point results, comparison with baseline values, and automatic determination of risk events when the flash point drop exceeds a set threshold, and output of graded early warnings; early warning information is displayed in a visual manner on the local monitoring interface, and can automatically generate structured evidence chain logs.

Images