Method for predicting the remaining useful life of an engine oil of a vehicle

Abstract

The application relates to a method for predicting the remaining service life of engine oil of a vehicle, comprising: acquiring a current expected time point; acquiring first data continuously detected at multiple detection time points from an initial time t0 when the engine oil is not degraded or from a last expected time point when a last prediction is performed to the current expected time point as model input, each first data comprising multiple parameters related to engine operation at the same time; predicting the remaining service life of the engine oil at the current expected time point based on the acquired first data by using a trained model according to the current expected time point, the model being trained to be capable of determining the degradation degree of the engine oil at each time point based on the operation condition at the time point characterized by the first data, so that the remaining service life of the engine oil at the current expected time point can be obtained by time integration of the degradation degrees at the time points; and outputting the predicted remaining service life.

Description

Technical Field

[0001] This invention relates to a method for predicting the remaining service life of engine oil in a vehicle. Background Technology

[0002] The engine is the primary power source of a vehicle. This is especially true for construction machinery used in mining areas or on construction sites, where maintaining the engine in good working order is crucial. Engine oil plays a key role in the operation of an internal combustion engine. Its main functions include lubrication, cooling, cleaning, sealing, protection, noise reduction, and fuel efficiency. However, with increased operating time, the condition of engine oil gradually deteriorates. In extremely harsh operating environments or under high vehicle loads, engine oil may even degrade prematurely. Therefore, regularly checking and changing the engine oil at predetermined intervals is essential to ensuring the effective implementation of these functions and maintaining the overall performance and health of the engine.

[0003] However, in the field of construction machinery, it is often difficult to monitor engine oil health in real time. The common practice is to use scheduled oil sampling (SOS). But this method requires obtaining an oil sample from the engine and then sending it to an SOS laboratory for analysis. This is not only time-consuming (around two weeks) but also very expensive. Another option is to install an oil quality sensor in the vehicle. However, this requires not only adding the sensor but also numerous accessories such as brackets, wiring harnesses, and displays, making it very costly and generally undesirable for end customers.

[0004] Furthermore, even when engine oil health deterioration is detected, it may not be possible to immediately stop the engine and take corrective action. In such cases, knowing the remaining service life of the engine oil is crucial. Knowing how much longer the engine can continue to operate allows operators to rationally plan vehicle work and maintenance schedules, preventing the vehicle from breaking down during work or en route to a repair shop, or premature maintenance that wastes the remaining service life of the engine oil. However, because engine speed, torque, power, and other operating data are constantly changing, the amount of engine oil degradation per unit time also varies with operating conditions. Therefore, engine oil degradation is not a linear process, making it difficult to determine the amount of degradation and, consequently, the remaining service life of the engine oil. Summary of the Invention

[0005] Therefore, the object of the present invention is to provide a method for predicting the remaining service life of engine oil in a vehicle, which overcomes at least some of the aforementioned defects of the prior art.

[0006] This objective is achieved through a method having the following characteristics.

[0007] According to one aspect of the present invention, a method for predicting the remaining service life of engine oil in a vehicle is provided. The method includes: a current expected time point acquisition step: in the current expected time point acquisition step, acquiring the current expected time point for predicting the remaining service life of the engine oil; a first data acquisition step: in the first data acquisition step, acquiring first data continuously detected at multiple detection times from the initial time t0 when the engine oil has not degraded or from the previous expected time point where the last prediction was performed to the current expected time point as model input, each first data including multiple parameters related to engine operation at the same time; a model prediction step: in the model prediction step, based on the current expected time point and the acquired first data, using a trained model to predict the remaining service life of the engine oil at the current expected time point, the model being trained to determine the degree of degradation of the engine oil at each time point based on the operating conditions represented by the first data at each time point, thereby obtaining the remaining service life of the engine oil at the current expected time point by integrating the degree of degradation at each time point over time; and a prediction result output step: in the prediction result output step, outputting the predicted remaining service life. Preferably, the method further includes a post-processing step in which appropriate treatment measures are taken based on the predicted remaining useful life.

[0008] The method of this invention proposes to establish, calibrate, and standardize a supervised machine learning model based on actual sampled data from Caterpillar's SOS platform database and a database of historical engine oil replacement times. This model is driven to predict the remaining service life of the engine oil in real time. Besides the sensors already present in the vehicle, this method requires no additional physical sensors, thus resulting in lower commercialization costs.

[0009] Furthermore, the remaining service life prediction method of this invention can be used in conjunction with the monitoring process of engine oil health status, for example, after determining that the engine oil has a high aging level. By first classifying the contamination level of the engine oil and then making segmented predictions based on the classification results, the waste of computing power and resources is avoided, making it more practical. However, the remaining service life prediction method of this invention can also be used independently, that is, skipping engine oil health status monitoring and directly performing predictions, making it very flexible. By continuously training and optimizing the LSTM model during model use, and combining the reward and penalty algorithm with LSTM (Long Short-Term Memory) network, the model can be better guided to learn the pattern of engine oil remaining service life, and the accuracy and stability of the prediction can be improved.

[0010] According to one implementation, the number and type of multiple parameters in each first dataset are selected from the engine operating data during the model training process by performing correlation analysis on the engine operating data and the results of periodic oil sample analysis of engine oil degradation. The advantage here is that, by performing correlation analysis, the parameters most relevant to engine oil degradation can be selected from a massive amount of vehicle operating data as model input.

[0011] Preferably, the correlation analysis is performed using the Pearson correlation coefficient, which is calculated using the following formula:

[0012]

[0013] Where i is the number of observations, X i Y i Let be the i-th observation values ​​of two variables X and Y, respectively. These are the average values ​​of two variables, where variable X is one of the engine operating data and variable Y is the result of periodic oil sample analysis of engine oil degradation. An advantage of this approach is that correlation analysis can be easily performed.

[0014] Preferably, operational data whose absolute value of the Pearson correlation coefficient with the periodic oil sample analysis results of the engine oil degradation amount is greater than a preset threshold is selected as the first data. In particular, the preset threshold is 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, or 0.3. The advantage here is that operational data with high correlation can be determined to accurately predict the degree of engine oil degradation and thus its remaining service life.

[0015] According to another implementation, when selecting the number and type of multiple parameters in each set of first data during model training, the load factor is defined as the ratio of average engine power to maximum engine output power by examining the correlation between operating conditions and load. Furthermore, a correlation analysis is performed on the load factor and the results of periodic oil sample analysis of engine oil degradation. The advantage here is that by examining the correlation between operating conditions and load, multiple operating conditions can be characterized using a single parameter, simplifying the model structure.

[0016] According to another embodiment, in the first data acquisition step, each piece of first data includes engine speed, load factor, engine coolant temperature, and aftertreatment exhaust gas temperature, wherein the load factor is the ratio of the engine's average power to its maximum output power during the time interval between two adjacent detection moments. The advantage here is that it significantly reduces the amount of model input data, lowers the dimensionality of the model input, and simplifies the model structure.

[0017] According to another embodiment, the model input further includes second and third data corresponding to the first data. The second data is selected from the following: engine oil type and specifications, engine model, and vehicle model. The third data includes the theoretical service life (L) of the engine oil. s .

[0018] According to another embodiment, in the first data acquisition step, first data is detected by a corresponding sensor installed on the vehicle. The advantage here is that the first data can be easily acquired for predicting the remaining engine oil life.

[0019] According to another embodiment, in the first data acquisition step, the detection of the first data is performed uniformly, and / or the time interval between any two adjacent detection moments is the same, except for the current desired time point, said time interval being, for example, 0.5 seconds, 1 minute, 5 minutes, 8 minutes or 10 minutes.

[0020] According to another embodiment, in the step of obtaining the current desired time point, the current desired time point is obtained by receiving an instruction representing the user's needs. Preferably, the instruction representing the user's needs is input in real time by the user through a human-machine interface on the vehicle, such as the input interface of an in-vehicle display, or data containing one or more current desired time points is pre-stored in the vehicle or in the background. The advantage here is that it facilitates the user in predicting the remaining service life of the engine oil at the desired time.

[0021] According to another embodiment, the trained model used in the model prediction step is an LSTM model, and / or the remaining lifetime at each time step is calculated using the following formula in the trained model:

[0022] L(i)=L(i-1)-D(n) i ,F i ,Tc i ,Tg i )·Δt i ·L s i = 1, 2, 3, ..., m, ...

[0023] Where L(i) and L(i-1) are the remaining lifetimes at the i-th time step and the previous time step, respectively, and L(0) = 100%L s , D(n i ,F i ,Tc i ,Tg i ) represents the operating condition (n) at the i-th time step. i F i Tc i Tg i The degree of degradation under ) , Δt iLet L be the step size between the i-th time step and the previous time step. s This is the theoretical service life.

[0024] According to another embodiment, in the post-processing step, the processing measures include scheduling vehicle maintenance trips based on the predicted remaining service life.

[0025] According to another implementation, during the model training process, a reward and penalty algorithm is combined to train the model, and / or the model's overall loss function adopts the following formula:

[0026]

[0027] Where L^(i) is the actual remaining engine oil life at the i-th time step, L(i) is the remaining engine oil life predicted by the model at the i-th time step, R(i) is the reward signal at the i-th time step, and α and β are the weights of the prediction error term and the reward signal term in the comprehensive loss function, respectively.

[0028] According to another embodiment, the trained model is further updated and optimized using the remaining useful life value predicted in the model prediction step and the actual useful life value. The advantage here is that the model can be continuously trained and optimized during use after it has been trained, thereby improving the accuracy of the model's prediction results over time.

[0029] According to another aspect of the invention, a computer program product is provided, comprising instructions that are implemented to perform the method of the invention.

[0030] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the scope of protection of the present invention. Attached Figure Description

[0031] The invention will now be described in detail with reference to the accompanying drawings. In the drawings,

[0032] Figure 1 A flowchart illustrating a method for training an engine oil remaining service life prediction model according to the present invention is shown.

[0033] Figure 2 A flowchart of a method for predicting the remaining service life of engine oil according to the present invention is shown. Detailed Implementation

[0034] Engine oil gradually degrades during engine operation, and its degradation rate (D), i.e., the amount of degradation per unit time, is closely related to the operating conditions. If the engine operates under standard conditions throughout its entire lifespan, the engine oil will also degrade at a constant rate; this constant rate of degradation is called the standard degradation rate (D).s The total time it takes for engine oil to go from being completely undamaged (i.e., not degraded) to being 100% degraded under standard operating conditions is called the theoretical service life (L) of the engine oil. s Therefore, D s ·L s =100%. However, in practice, engines do not always operate under standard conditions. In fact, because mining machinery needs to perform complex maneuvers, such as forward and backward movement, driving on flat ground, going uphill and downhill, digging, lifting, loading, and unloading, the engine operating data defining the operating conditions, such as speed, torque, and power, are constantly changing. Therefore, the degradation degree D, which is closely related to the operating conditions, is also a constantly changing quantity, D = D(t). Furthermore, engine oil undergoes numerous operating conditions throughout its lifespan. Under some operating conditions, the degradation degree D of the engine oil is higher than that under standard conditions. s The faster the oil degrades, the greater the amount of engine oil degrades within the same working time compared to standard operating conditions. This means more of the engine oil's lifespan is consumed, resulting in a shorter final lifespan (L). r Compared to theoretical service life L s Shorter; however, under certain operating conditions, the degradation degree D of the engine oil is shorter than that under standard operating conditions. s The process is slower, resulting in less degradation of engine oil over the same working time compared to standard operating conditions. This means less oil is consumed, leading to a longer actual service life (L). r Compared to theoretical service life L s Therefore, to facilitate a clear understanding of whether the actual service life under a given operating condition is longer or shorter than the theoretical service life, this invention uses the standard degradation degree D under standard operating conditions. s and theoretical service life L s As a benchmark, the degradation degree D of engine oil is defined as follows: for a given operating condition, the degradation degree D is defined as the percentage of the engine oil's service life L consumed per unit time. s The percentage, that is:

[0035]

[0036] Among them, L D L represents the lifetime consumed within the time period Δt. s Let Δt be the theoretical service life. When Δt is very small and approaches 0 (i.e., Δt→0), then:

[0037]

[0038] Where d() / dt represents the time derivative of the variable within the parentheses.

[0039] The consumed lifespan L can be derived from the above formula (2). D And the formulas for calculating the remaining useful life L are as follows:

[0040] L D =∫D(t)dt·L s ,

[0041] L = L s - L D =(1-∫D(t)dt)·L s (3)

[0042] Therefore, for a life cycle of engine oil, at a certain moment t m To determine the remaining service life L(t) of engine oil m When calculating the time from the initial time t0 when the engine oil is intact (not degraded) to time t... m The service life (in L) of engine oil consumed under various operating conditions up to this point. D (t m And subtract the consumed service life from the theoretical service life, that is

[0043]

[0044] Where, Δt i Let t be the duration of the i-th operating condition. i t i-1 Let D(t) represent the i-th and (i-1)-th time points, respectively. i ) represents the degree of degradation of the engine oil at the i-th time point.

[0045] Considering D s ·L s =1, and can also be expressed as theoretical service life L. s and standard degradation degree D s For each time period Δt i and the corresponding degradation degree D(t) i ) Perform dimensionless transformation:

[0046] p i = D(t) i ) / D s ,Δτ i = Δt i / L s (5)

[0047] Therefore, formula (3) can also be transformed into the following form:

[0048]

[0049] To make it easier to understand, here's a simple example. Assume we start from the initial moment t0 = 0, before the engine oil degrades, and at the first moment t1 = 0.2L... s The corresponding degradation degree under the first operating condition is D(t1) = 1.2D. s That is, the degree of degradation D compared to standard operating conditions. s The value is large, and in the first time interval Δt1 = t1 – t0 = 0.2L. s Substituting into formula (4), the remaining service life L(t1) is:

[0050] L(t1)=(1-D(t1)·Δt1)·L s =(1–1.2D) s 0.2L s )·L s =76% L s .

[0051] Theoretically, the current remaining service life should be 80% L. s (=L s -t1). Therefore, the actual remaining useful life is shorter than the theoretical remaining useful life.

[0052] Next, assume that at the second time t2 = 0.5L s The corresponding degradation degree under the second operating condition is D(t2) = 0.75D. s That is, the degree of degradation D compared to standard operating conditions. s The value is small, so in the second time interval Δt2 = t2 – t1 = 0.5L. s -0.2L s =0.3L s Substituting into formula (4), the remaining service life L(t2) is:

[0053] L(t2)=(1-D(t1)·Δt1–D(t2)·Δt2)·L s

[0054] =(1–1.2D) s 0.2L s -0.75D s 0.3L s )·L s =53.5% L s .

[0055] Theoretically, the current remaining service life should be 50%L. s (=L s Therefore, after experiencing these two operating conditions, the current actual remaining service life is longer than the theoretical remaining service life.

[0056] The example above is purely illustrative. In practice, it is almost impossible to maintain a constant degree of degradation over such a long period of time. As mentioned above, the degree of degradation changes almost every moment. However, in principle, if the degree of degradation of the engine oil at each moment is known, the remaining service life of the engine oil at each moment can be calculated by integrating the degree of degradation over time using the above formula (4) or (6).

[0057] The problem now is that the degree of degradation is usually unknown. Therefore, this invention proposes a method to determine the degree of degradation based on operating conditions using a machine learning model; that is, to model and find the functional relationship or mapping relationship between each operating condition and the corresponding degree of engine oil degradation.

[0058] Figure 1 A flowchart of a method for training a remaining useful life prediction model according to the present invention is shown. In step S101, considering the correlation between operating conditions and load, parameters included in the first operating data, which is the main part of the model input, are determined based on the correlation analysis between engine operating data and the SOS result of engine oil degradation. That is, the number of parameters included in the first operating data and the specific parameters are determined. It is well known that operating conditions can be defined by engine operating data. However, there are more than 20 types of engine operating data, such as vehicle model, engine model, engine oil type and specification, total vehicle operating hours, idling time, engine speed, vehicle speed, fuel pressure, intake pressure or vacuum, throttle opening, engine coolant temperature, actual engine torque, engine friction torque, aftertreatment exhaust gas temperature, inlet and outlet temperatures of the SCR catalytic converter or exhaust aftertreatment device, maximum engine torque, and maximum engine output power, etc. If all these operating data are considered when modeling the functional relationship or mapping relationship between operating conditions and degradation, the model will be very complex. Furthermore, some engine operating data are not closely related to the degree of engine oil degradation. Considering this operating data would actually increase data noise, thus negatively affecting the accuracy of the mapping relationship. Therefore, before building the model, a correlation analysis was first performed on the engine operating data and the amount of engine oil degradation to select suitable operating data (i.e., the first operating data) as the model input.

[0059] Engine operating data can be obtained from, for example, a Caterpillar measurement database stored on a server external to the vehicle, such as in the background. This measurement database stores time series of all engine operating data collected periodically from multiple vehicles. Engine operating data is collected by the Engine Control Module (ECM) and transmitted to the background via, for example, a telematics unit. Engine oil degradation can be analyzed from the Caterpillar SOS (Scheduled Oil Sampling) database, which is also stored in the background. By periodically extracting oil samples from the engine and sending them to the SOS laboratory for analysis, a time series of SOS test data characterizing engine oil quality can be obtained. Each SOS test data point includes indicators such as engine oil viscosity, acid value, base value, contamination level, wear particle impurities, oxidation degree, and degradation status. From these indicators, the degree of engine oil degradation can be analyzed, i.e., the remaining degradation amount of the engine oil, and thus the SOS result of the degradation amount can be obtained (the sum of the remaining degradation amount and the degradation amount is 100%). For example, empirical data can be combined with analysis of various indicators in SOS test data to obtain the degradation status of engine oil under corresponding indicator conditions, i.e., its remaining degradation amount. Therefore, by finding the values ​​and remaining degradation amounts of various operating data detected at the same time (e.g., the same total vehicle operating hours) or within the same time interval under the same combination of engine oil type and specification and / or engine model and / or vehicle model in the measurement database and SOS database, interrelated time series can be formed. Based on these interrelated time series, correlation analysis can be used to filter operating data related to engine degradation (i.e., the first operating data) from the collected operating data as model input. Here, for example, the Pearson correlation coefficient is used. The Pearson correlation coefficient between two variables X and Y is calculated by the following formula:

[0060]

[0061] Where i is the number of observations, X i Y i Let be the i-th observation values ​​of two variables X and Y, respectively. These are the average values ​​of the two variables. Here, variable X is one of the engine operating data, and variable Y is the amount of residual engine oil degradation obtained from the SOS test results. The Pearson correlation coefficient is compared with a preset threshold. If the absolute value of the Pearson correlation coefficient between the two variables X and Y is greater than or equal to the preset threshold, then the two variables X and Y are considered correlated; otherwise, if the absolute value of the Pearson correlation coefficient between the two variables X and Y is less than the preset threshold, then the two variables X and Y are considered uncorrelated. By selecting different preset thresholds, different quantities and types of first operating data can be filtered out, that is, different dimensions of model input can be selected or determined for the model. Preset thresholds can be, for example, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, or 0.3, etc.

[0062] Furthermore, to simplify the model, when determining the first operating data, efforts are made to characterize multiple operating conditions using a single parameter. Here, by examining the correlation between operating conditions and load, this invention defines a load factor, F, which is defined as the ratio of the engine's average power over a period of time to the engine's maximum output power over that period, i.e., load factor = average power / maximum output power, where power is proportional to the product of speed and torque. In other words, instead of directly using raw operating data such as engine speed, torque, friction torque, and output power as model input, a combination of their mathematical operations is used. A correlation analysis was also performed on the load factor and the residual degradation (SOS) result, for example, using the Pearson correlation coefficient described above, to verify whether the load factor is suitable for use as the first operating data.

[0063] In a preferred embodiment, the final selected first operating data may include four parameters: engine speed, load factor, engine coolant temperature, and aftertreatment exhaust gas temperature. Therefore, in this invention, the degree of engine oil degradation D is designed as a function of the parameters engine speed n, load factor F, coolant temperature Tc, and aftertreatment exhaust gas temperature Tg, and these four parameters n, F, Tc, and Tg themselves change with time t, i.e.

[0064] D=D(n,F,Tc,Tg)=D(n(t),F(t),Tc(t),Tg(t)).

[0065] After determining the initial operating data, a model can be built and trained, i.e., the model parameters can be adjusted. In addition to the initial operating data, the model input also includes second operating data, which is a combination of engine oil type and specification and / or engine model and / or vehicle model, corresponding to the first operating data. Here, the second operating data acts as a model switch, allowing different combinations of model parameters to be trained for different second operating data. Alternatively, it is possible to train a single combination of model parameters applicable to multiple different second operating data. The model input also includes the theoretical service life of the engine oil under the determined combination of engine oil type and specification and / or engine model and / or vehicle model.

[0066] In step S102, training data is acquired, which includes training input data and training output data. The training input data includes: values ​​of the first operating data determined in step S101, detected at various times, i.e., at various moments or time intervals. Each first operating data value includes engine speed n, load factor F, engine coolant temperature Tc, and aftertreatment exhaust gas temperature Tg; second operating data corresponding to the first operating data, i.e., engine oil type and specification and / or engine model and / or vehicle model; and the theoretical service life L of the engine oil corresponding to the second operating data. s The training output data represents the actual service life (L) of the engine oil. r Here, the theoretical service life (L) of engine oil is given. s This information is available from Caterpillar's recommended engine oil change time database. This database describes how often the engine oil should be changed for a given combination of engine oil type and specification and / or engine and / or vehicle model. First, second, and training output data can be obtained, for example, from a historical actual engine oil change time database stored in Caterpillar's backend. This historical actual change time database stores data collected over decades from various types and specifications of engine oil used in various engine and / or vehicle models, showing the engine oil's lifecycle from a healthy, undamaged state to a completely degraded state (i.e., from the last change to the current change) as it was applied. For this purpose, within each lifecycle, the historical actual change time database contains detailed data at very small time intervals Δt. i (For example, operational data detected every 0.5 seconds, 1 minute, 5 minutes, 8 minutes, or 10 minutes). In other words, for each lifecycle, this historical actual replacement time database records the total time the engine oil has been in operation. i Δt i (This total duration is the actual service life of the engine oil in L) r), and how many operating conditions (n) it experienced during its lifecycle. i F i Tc i Tg i ) and the duration Δt of each operating condition i Therefore, there is

[0067] Σ i D(n i ,F i ,Tc i ,Tg i )·(t i -t i-1 ) = 100%, and Σ i Δt i =L r ;

[0068] or

[0069] Σ i p(n i , F i , Tc i , Tg i ) ·(τ i - τ i-1 ) = 100%, and Σ i Δτ i = L r / L s (7)

[0070] Where, n i F i Tc i Tg i They represent time t respectively i Engine speed, load factor, coolant temperature, and aftertreatment exhaust gas temperature n(t) i ), F(t) i ), Tc(t i ), Tg(t i In other words, at the end of each life cycle, after experiencing its actual useful life L... r After that, the remaining service life of the engine oil is zero.

[0071] After collecting training data, in step S103, the remaining useful life prediction model of the present invention is trained based on the acquired training data. The remaining useful life prediction model of the present invention is constructed as an LSTM model suitable for processing time series data, which includes an input layer, an LSTM layer, and an output layer. As a special type of recurrent neural network (RNN) model, the LSTM model has memory cells containing three gating mechanisms (i.e., forget gate, input gate, and output gate) in addition to the conventional hidden layers in the LSTM layer, in order to effectively maintain and process long-term dependent information. The memory cells are used to store long-term information and control the inflow and outflow of information through gating mechanisms. The forget gate uses a sigmoid layer to determine which information to discard from the cell state. The input gate is responsible for updating the cell state, which consists of two parts: the sigmoid layer determines which values ​​will be updated, and the tanh layer creates a new candidate value vector that will be added to the cell state. The output gate controls which information should be output to the current hidden state, which contains information about the previous input and is used for prediction. Here, the output of the output gate is calculated using a sigmoid layer. This output is then multiplied by the tanh operation result of the current cell state to calculate the current hidden state. This element-wise multiplication operation ensures that information can flow only when the gate is "open." Here, sigmoid is a widely used activation function in neural networks, mapping any value to between 0 and 1. tanh is another activation function, mapping any value to between -1 and 1. The recurrent mechanism of the RNN model allows the results generated by the hidden layer in the previous time step to influence the output of the current time step as part of the input. That is, the input of the current time step includes not only the normal input but also the hidden layer output from the previous time step. Therefore, the LSTM model can help capture temporal dependencies in the data, thereby better predicting the remaining service life of engine oil. In this embodiment, the model input of the LSTM model is the theoretical service life L of the engine oil. s The model output includes the first and second running data for each time step. Each first running data includes engine speed n, load factor F, engine coolant temperature Tc, and aftertreatment exhaust gas temperature Tg. The second running data includes a combination of vehicle model and / or engine model and / or engine oil type and specification. The model output is the remaining service life of the engine oil. Therefore, formula (4) is rewritten as:

[0072] L(i) = L(i-1) - D(n) i , F i , Tc i , Tg i ) · Δt i · L s, i = 1, 2, 3, …, m,… (8)

[0073] Where L(i) and L(i-1) are the remaining lifetimes at the i-th time step and the previous time step, respectively, and L(0) = 100%L s , D(n i ,F i ,Tc i ,Tg i ) represents the operating condition (n) at the i-th time step. i F i Tc i Tg i The degree of degradation under ) , Δt i Let L be the step size between the i-th time step and the previous time step. s This represents the theoretical service life. Here, the remaining service life L(i-1) from the previous time step is also input into the current time step as a hidden layer output. Through continuous iteration, the remaining service life for each time step, starting from the initial moment t0 when the engine oil has not degraded or from the time point of the last prediction, can be calculated. Therefore, by training the model based on the training data obtained in step S102, the mapping relationship D(n) between each operating condition and the corresponding degree of degradation can be found. i ,F i ,Tc i ,Tg i Thus, the remaining service life after experiencing the corresponding operating conditions at each time can be calculated using formula (4) or (6) or formula (8).

[0074] During model training, the model is also provided with reward or penalty feedback according to reward and penalty rules. The reward and penalty rules can be designed such that a reward is given when the model's prediction result is close to the actual remaining service life (e.g., accuracy ≥ 80%), and a penalty is given when the prediction deviates from the actual value (e.g., accuracy < 80%). The reward and penalty algorithm of this invention uses a Q-Learning method based on deep neural networks, namely the DQN (Deep Q Network) algorithm. When combining the reward and penalty algorithm with an LSTM model to predict the remaining service life of engine oil, a comprehensive loss function can be defined, combining the prediction error (usually using mean squared error) with the reward signal. The comprehensive loss function can be expressed as follows:

[0075]

[0076] Where L^(i) is the actual remaining engine oil life at time step i, L(i) is the remaining engine oil life predicted by the model at time step i, R(i) is the reward signal at time step i, and α and β are the weights of the prediction error term and the reward signal term in the comprehensive loss function, respectively, used to balance the impact of prediction accuracy and reward penalty. (L(i)-L^(i)) 2 • α is the mean squared error loss term, used to measure the model's prediction error. β·R(i) is the reward signal term, used to guide the model to adjust based on the reward signal. By adjusting the values ​​of α and β, the emphasis placed on prediction error and reward signal by the model during training can be controlled. During training, the optimization objective is to minimize this comprehensive loss function so that the model considers both prediction accuracy and reward / penalty effects when predicting the remaining engine oil life. By iteratively adjusting the model and reward / penalty rules, the model's prediction accuracy and stability are improved.

[0077] The DQN algorithm helps models learn complex relationships between states and behaviors to better predict the remaining lifespan of engine oil. By combining a reward and penalty algorithm with training data to train an LSTM model, the model parameters can be optimized by maximizing the reward or minimizing the penalty.

[0078] Advantageously, after training, the trained model can be deployed on multiple test vehicles for evaluation, checking its predictive performance and accuracy. Here, the model's effectiveness is evaluated by comparing actual engine oil life with the model's predictions.

[0079] Of course, in addition to the LSTM model, other suitable machine learning models are also possible, as long as the model can determine the corresponding degree of engine oil degradation from the operating data or operating conditions and calculate the remaining service life of the engine oil by time integration of the degree of degradation using formula (4), formula (6), or formula (8).

[0080] Figure 2A flowchart of a method for predicting the remaining service life of engine oil according to the present invention is shown. In step S201, the current expected time point for predicting the remaining service life of the engine oil is obtained. For example, the current expected time point can be obtained by receiving an instruction representing a user's needs. The instruction representing the user's needs can be input by the user at any time through a human-machine interface on the vehicle, such as the input interface of an in-vehicle display, for example by clicking the corresponding button on the input interface. That is, each current expected time point is selected by the user in real time at any moment. Alternatively, data containing one or more current expected time points can be pre-stored in the vehicle or in the background and retrieved and called by the model. For example, the time interval between every two adjacent expected time points can be preset, for example, once a week. Thus, when the remaining service life prediction algorithm is deployed in the vehicle's display or controller, the predicted remaining service life can be displayed on the vehicle display in real time at preset time intervals. Alternatively, when the remaining service life prediction algorithm is deployed in the background, the predicted remaining service life can be periodically pushed to the user, for example, through a platform used by dealers and end customers (e.g., Caterpillar's Fleet Vision and Doer platform) or a mobile APP. The frequency of the push, i.e., the time interval, can be customized according to user needs.

[0081] In step S202, first data is continuously detected at multiple detection times, starting from the initial time t0 when the engine oil has not degraded, or from the previous expected time point to the current expected time point, and is used as part of the model input. Here, the detection can be performed uniformly except for the current expected time point. That is, the time interval between any two adjacent detection times is the same in all detection times except the current expected time point. This time interval can be set according to user needs, for example, it can be 1 minute, 5 minutes, 8 minutes, or 10 minutes. For high accuracy requirements, the time interval can even be 0.5 seconds, i.e., acquiring the first data at a frequency of 2Hz. If the duration between the current expected time point and the initial time t0 or the previous expected time point is not an integer multiple of the time interval, i.e., the current expected time point does not coincide with a detection time, then an additional first data point is detected at the current expected time point. In other words, the current expected time point is also a detection time point. Each set of first data may include values ​​for parameters such as engine speed, load factor, engine coolant temperature, and aftertreatment exhaust gas temperature at the same time. The load factor is the ratio of the engine's average power to its maximum output power during the time interval between two adjacent detection moments (i.e., within each time interval and between the current desired time point and the nearest previous detection moment). The number and type of parameters in each set of first data (i.e., how many parameters each set contains and what these parameters specifically are) are determined by considering the correlation between operating conditions and load, and by filtering from the operating data based on the correlation analysis between engine operating data and the SOS (Self-Operating Operating System) result of engine oil degradation, obtained from the SOS test results of the engine oil. In addition to the first data, the model input may include second and third data. The second data can be selected from the following: engine oil type and specification, engine model, and vehicle model. In particular, the second data can act as a model switch; different second data trigger different models or different combinations of model parameters. Multiple sets of second data can also trigger the same model or the same combination of model parameters. The third data includes the theoretical service life L of the engine oil, corresponding to the second data. s The model inputs, especially the first data, can be obtained from corresponding sensors installed on the vehicle. Alternatively, the second and third data can also be fixed in the model, especially when the remaining useful life prediction algorithm is deployed on a specific vehicle model.

[0082] In step S203, based on the current expected time point, the remaining service life of the engine oil at the current expected time point is predicted using a trained model based on the acquired first data. The trained model has been prepared according to the above reference... Figure 1The described method has been trained. That is, the trained model has been trained to determine the degree of degradation at each time point based on the first data at each time point, so that the degree of degradation at each time point can be integrated over time using formula (4) or (6) or formula (8) to obtain the remaining degradation amount of engine oil and thus the remaining service life at the current desired time point. The trained model is preferably an LSTM model.

[0083] In step S204, the predicted remaining useful life is output. The predicted remaining useful life can be output to a display on the vehicle or to a remote operator, for example, in a back office.

[0084] Then, in the optional step S205, appropriate measures are taken based on the predicted remaining service life, such as reducing the engine load to extend the remaining service life of the engine oil, or reasonably arranging the vehicle's working time and maintenance schedule based on the remaining service life to avoid the vehicle breaking down halfway before reaching the repair shop or changing the engine oil prematurely before the remaining service life of the engine oil is used up.

[0085] It should be noted that the remaining service life prediction method of the present invention can be used in conjunction with the monitoring process of engine oil health status. For example, the remaining service life prediction can be performed only after the aging level of the engine oil is determined to be high. The aging level of the engine oil can be determined either through model prediction or through offline laboratory testing using existing technology. By first classifying the contamination level of the engine oil and then performing segmented predictions based on the classification results, the waste of computing power and resources is avoided, making it more practical. Of course, this remaining service life prediction method can also be used independently, that is, skipping engine oil health status monitoring and directly performing predictions. In this case, the health status of the engine oil can be inferred from the predicted remaining service life. Therefore, the remaining service life prediction method of the present invention is very flexible in its use.

[0086] It should also be noted that the model of this invention can be continuously updated and optimized during use. That is, after the model is trained, the predicted remaining service life value and the actual service life value are used together as training data, and the model is retrained using a reward and penalty algorithm, such as the DQN algorithm, thereby continuously improving the accuracy of the model's predictions. Therefore, the remaining service life of the engine oil is always predicted for the user using the latest model.

[0087] The present invention also relates to a computer program product comprising instructions which are implemented to perform the methods described above.

[0088] Industrial applicability

[0089] The method for training an engine oil remaining service life prediction model and the method for predicting engine oil remaining service life using a trained model based on detected engine operating data of the present invention can be applied to various vehicles, especially construction machinery in mining areas and construction sites, but also to motor vehicles on roads, such as trucks and passenger cars. For a given desired time point where the user wants to know the remaining service life of the engine oil, the trained model is input with first data continuously detected at multiple detection times at predetermined time intervals, starting from the initial time t0 when the engine oil has not degraded or the previous desired time point where the prediction was last performed, up to the current desired time point. If necessary, a first data point is also detected at the current desired time point. Each first data point includes engine speed, load factor, engine coolant temperature, and aftertreatment exhaust gas temperature at the same time. The load factor is the ratio of the engine's average power to its maximum output power during the time interval between two adjacent detection times or between the current desired time point and the previous detection time point. The trained model then predicts the remaining service life of the engine oil at the current desired time point based on the acquired first data and outputs the predicted remaining service life to the vehicle's display or to a remote operator, for example, in the background. The vehicle or background operator can then take appropriate action based on the model's prediction.

[0090] The foregoing description is merely an exemplary embodiment relating to the spirit and principles of the present invention. Those skilled in the art will understand that various changes can be made to the described examples without departing from the spirit and principles, and such changes and their various equivalents are contemplated by the inventors and fall within the scope defined by the claims of the present invention.

Claims

1. A method for predicting the remaining service life of engine oil in a vehicle, characterized in that, The method includes: Current expected time point acquisition step: In this current expected time point acquisition step, the current expected time point for the expected remaining service life of the engine oil is obtained; First data acquisition step: In this first data acquisition step, first data continuously detected at multiple detection times, starting from the initial time t0 when the engine oil has not degraded or from the previous expected time point when the last prediction was implemented to the current expected time point, is acquired as model input. Each first data includes multiple parameters related to engine operation at the same time. Model prediction step: In this model prediction step, based on the current expected time point and the acquired first data, a trained model is used to predict the remaining service life of the engine oil at the current expected time point. The model has been trained to determine the degree of engine oil degradation at each time point based on the operating conditions represented by the first data, thereby calculating the remaining service life of the engine oil at the current expected time point by integrating the degradation degree at each time point over time. Prediction result output step: In this prediction result output step, the predicted remaining useful life is output. Preferably, the method further includes a post-processing step in which appropriate treatment measures are taken based on the predicted remaining useful life.

2. The method according to claim 1, characterized in that, In the first data acquisition step, each piece of first data includes engine speed, load factor, engine coolant temperature, and aftertreatment exhaust gas temperature. The load factor is the ratio of the engine's average power to its maximum output power during the time interval between two adjacent detection moments. In particular, the model input also includes second and third data corresponding to the first data. The second data is selected from the following: engine oil type and specifications, engine model, and vehicle model. The third data includes the theoretical service life (L) of the engine oil. s .

3. The method according to claim 1 or 2, characterized in that, In the first data acquisition step, the first data is detected by the corresponding sensor installed on the vehicle.

4. The method according to any one of claims 1 to 3, characterized in that, In the first data acquisition step, the detection of the first data is performed uniformly, and / or the time interval between any two adjacent detection moments is the same, except for the current desired time point, such time intervals are, for example, 0.5 seconds, 1 minute, 5 minutes, 8 minutes or 10 minutes.

5. The method according to any one of claims 1 to 4, characterized in that, In the step of obtaining the current desired time point, the current desired time point is obtained by receiving an instruction that represents the user's needs. Preferably, the instruction representing the user's needs is input by the user in real time through the human-machine interface on the vehicle, such as the input interface of the in-vehicle display, or data containing one or more current desired time points is pre-stored in the vehicle or in the background.

6. The method according to any one of claims 1 to 5, characterized in that, The trained model used in the model prediction step is an LSTM model. and / or The remaining lifetime at each time step is calculated using the following formula in the trained model: L(i)=L(i-1)-D(n i ,F i ,Tc i ,Tg i )·Δt i ·L s ,i=1,2,3,…,m,… Where L(i) and L(i-1) are the remaining lifetimes at the i-th time step and the previous time step, respectively, and L(0) = 100%L s , D(n i ,F i ,Tc i ,Tg i ) represents the operating condition (n) at the i-th time step. i F i 、Tc i Tg i The degree of degradation under ) , Δt i Let L be the step size between the i-th time step and the previous time step. s This is the theoretical service life.

7. The method according to any one of claims 1 to 6, characterized in that, In the post-processing steps, the processing measures include scheduling vehicle maintenance trips based on the predicted remaining service life.

8. The method according to any one of claims 1 to 7, characterized in that, The number and type of multiple parameters in each primary dataset were selected from the engine operating data during the model training process by performing correlation analysis on the engine operating data and the results of periodic oil sample analysis of engine oil degradation.

9. The method according to claim 8, characterized in that, The correlation analysis was performed using the Pearson correlation coefficient, which is calculated using the following formula: Where i is the number of observations, X i Y i Let be the i-th observation values ​​of two variables X and Y, respectively. These are the averages of the two variables. Wherein, variable X is one of the engine operating data, and variable Y is the result of periodic oil sample analysis of the engine oil degradation amount.

10. The method according to claim 9, characterized in that, The first data is selected from the operational data whose absolute value of the Pearson correlation coefficient with the periodic oil sample analysis results of the engine oil degradation amount is greater than a preset threshold. In particular, the preset threshold is 0.9, 0.8, 0.7, 0.6, 0.5, 0.4 or 0.

3.

11. The method according to any one of claims 8 to 10, characterized in that, When selecting the number and type of multiple parameters in each first data point during the model training process, the correlation between operating conditions and load is also examined. The ratio of the engine's average power to its maximum output power is defined as the load factor, and a correlation analysis is performed on the load factor and the results of periodic oil sample analysis of engine oil degradation.

12. The method according to any one of claims 1 to 11, characterized in that, During model training, a reward and penalty algorithm is used to train the model, and / or the overall loss function of the model is expressed by the following formula: Where L^(i) is the actual remaining engine oil life at the i-th time step, L(i) is the remaining engine oil life predicted by the model at the i-th time step, R(i) is the reward signal at the i-th time step, and α and β are the weights of the prediction error term and the reward signal term in the comprehensive loss function, respectively.

13. The method according to any one of claims 1 to 12, characterized in that, The trained model is then updated and optimized using the remaining useful life values ​​predicted in the model prediction step and the actual useful life values.

14. A computer program product comprising instructions that are implemented to perform the method according to any one of claims 1 to 13.

Images