A Machine Learning-Based Method and System for Optimizing Corrosion-Resistant Marine Gear Oil Formulations

Abstract

This invention belongs to the field of gear oil formulation optimization technology, specifically involving a machine learning-based method and system for optimizing anti-corrosion marine gear oil formulations. The method includes: generating several candidate formulations based on the formulation components of the anti-corrosion marine gear oil; simulating the intrusion process of moisture and salt spray on the anti-corrosion marine gear oil under marine conditions, and measuring the lubricating film thickness to obtain a lubricating film thickness decay sequence; performing feature identification on the lubricating film thickness decay sequence, constructing a lubricating film thickness decay feature vector, and establishing a mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion using a gradient boosting regression algorithm; and, based on the mapping relationship, aggregating the lubricating film thickness decay feature vector into a comprehensive decay index, using the minimum comprehensive decay index as the optimization objective, to obtain the optimized formulation of the anti-corrosion marine gear oil. This invention effectively improves the anti-corrosion reliability of anti-corrosion marine gear oil formulations under complex marine conditions.

Description

Technical Field

[0001] This invention belongs to the field of gear oil formulation optimization technology, specifically relating to a machine learning-based method and system for optimizing the formulation of corrosion-resistant marine gear oils. Background Technology

[0002] Corrosion-resistant marine gear oils are key lubricating materials that ensure the long-term reliable operation of gear transmission systems in marine engineering equipment such as ships and offshore platforms under harsh environments such as salt spray and high humidity. Their core function is to form a lubricating film on the gear surface, isolating direct metal-to-metal contact and resisting corrosion and wear caused by seawater and salt.

[0003] However, during actual operation at sea, external moisture and salt spray continuously seep into the gear oil through the gearbox breather or weak seals, accumulating over time. As moisture and salt spray continuously penetrate, the gear oil's protective film-forming ability is gradually compromised. Once the oil film fails, gear corrosion and abnormal wear accelerate. Unlike lubrication under stable operating conditions, this intrusion process is continuously dynamic. The changes in the lubricating film of anti-corrosion marine gear oils with different formulations vary significantly during this intrusion process. Relying solely on test results under stable operating conditions is insufficient to effectively characterize the correspondence between formulation composition and actual film-forming protective ability, making it difficult for optimized formulations to meet the anti-corrosion requirements under complex marine conditions. Summary of the Invention

[0004] (1) Technical problems to be solved

[0005] The purpose of this invention is to provide a machine learning-based method and system for optimizing the formulation of anti-corrosion marine gear oils, in order to solve the problem that in practical application scenarios, the relationship between the composition of anti-corrosion marine gear oil formulations and the actual lubricating film thickness is difficult to be effectively characterized through testing under stable operating conditions, resulting in the film-forming protection effect of the optimized formulation failing to meet the actual needs of complex marine operating conditions.

[0006] (2) Technical solution

[0007] To achieve the above objectives, on the one hand, the present invention provides a method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning, the method comprising:

[0008] Several candidate formulations were generated based on the formulation components of the anti-corrosion marine gear oil formulation.

[0009] The invasion process of water and salt spray on anti-corrosion marine gear oil under simulated marine working conditions was investigated for each candidate formulation. The thickness of the lubricating film was measured during the invasion process to obtain the lubricating film thickness decay sequence.

[0010] Feature identification was performed on the lubricating film thickness decay sequence to construct a lubricating film thickness decay feature vector. Using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, a gradient boosting regression algorithm was used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion.

[0011] Based on the mapping relationship, the components of the lubricating film thickness decay feature vector are aggregated into a comprehensive decay index. The optimization objective is to minimize the comprehensive decay index to optimize the formulation components and obtain the optimized formulation of anti-corrosion marine gear oil.

[0012] Furthermore, the method for generating several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation includes:

[0013] Obtain the mass fraction range of each component in the anti-corrosion marine gear oil formulation; generate an initial candidate formulation with the constraint that the mass fraction of each component is within its corresponding mass fraction range and the sum of the mass fractions of all components equals the preset total mass fraction.

[0014] The corrosion resistance values ​​of the initial candidate formulations were measured, and the initial candidate formulations with corrosion resistance values ​​not lower than the preset corrosion resistance threshold were retained, resulting in several groups of candidate formulations.

[0015] Furthermore, the method for simulating the intrusion process of moisture and salt spray on anti-corrosion marine gear oil under marine operating conditions for each candidate formulation, and measuring the lubricating film thickness during the intrusion process to obtain the lubricating film thickness decay sequence includes:

[0016] Collect moisture and salt spray data of the target sea area; prepare simulated seawater based on the moisture and salt spray data of the target sea area.

[0017] Simulated seawater was added to the anti-corrosion marine gear oil samples of each candidate formulation in several batches; the thickness of the lubricating film was measured after each addition of simulated seawater; and a simulated seawater addition-lubricating film thickness curve was constructed based on the cumulative amount of simulated seawater added and the corresponding lubricating film thickness at the end of the addition process.

[0018] Based on the simulated seawater addition-lubricating film thickness curve, the cumulative amount of simulated seawater added when the lubricating film thickness decreases at the maximum rate is taken as the intrusion characteristic quantity.

[0019] Using zero as the lower bound and the intrusion characteristic quantity as the upper bound, the range of simulated cumulative seawater addition is divided into N equal parts; the lubricating film thickness value corresponding to each division point is extracted sequentially according to the simulated seawater addition amount-lubricating film thickness curve to obtain the lubricating film thickness decay sequence of each candidate formulation.

[0020] Furthermore, the method of adding simulated seawater in stages to the anti-corrosion marine gear oil samples of each candidate formulation includes:

[0021] Simulated seawater was added to the pre-samples of anti-corrosion marine gear oil for each candidate formulation and mixed uniformly. After uniform mixing, the pre-samples of anti-corrosion marine gear oil were centrifuged and the termination time was when the volume of the free water phase no longer increased. The time interval from the completion of uniform mixing to the termination time was taken as the demulsification time.

[0022] The amount of simulated seawater added in a single step is obtained by multiplying the preset salt spray deposition rate, the oil surface area of ​​each candidate formulation of the anti-corrosion marine gear oil sample, and the corresponding demulsification time. Simulated seawater of the same amount is then added to the anti-corrosion marine gear oil sample in multiple steps.

[0023] Furthermore, the method of using the cumulative amount of simulated seawater added at the point where the lubricating film thickness decreases at the maximum rate, based on the simulated seawater addition amount-lubricating film thickness curve, as the intrusion characteristic quantity includes:

[0024] Cubic spline fitting was performed on the simulated seawater addition amount-lubricating film thickness curves of each candidate formulation to obtain a cubic spline function of the lubricating film thickness with respect to the simulated cumulative seawater addition amount.

[0025] Within the horizontal axis range of the simulated seawater addition amount-lubricating film thickness curve, all local maxima of the lubricating film thickness decrease rate corresponding to the cubic spline function are extracted; the cumulative amount of simulated seawater addition corresponding to the local maxima of the lubricating film thickness decrease rate is taken as the intrusion characteristic quantity of each candidate formulation.

[0026] If, within the abscissa range of the simulated seawater addition amount-lubricating film thickness curve, the lubricating film thickness decrease rate corresponding to the cubic spline function has no local maximum point, then the cumulative simulated seawater addition amount corresponding to the right endpoint of the abscissa is taken as the intrusion characteristic quantity.

[0027] Furthermore, the method for feature identification of the lubricating film thickness decay sequence and construction of a lubricating film thickness decay feature vector includes:

[0028] A lubricating film thickness decay difference sequence is constructed based on the decrease in lubricating film thickness between adjacent equal division points in the lubricating film thickness decay sequence; the lubricating film thickness decay difference sequence is smoothed and then normalized based on the initial lubricating film thickness to obtain the normalized lubricating film thickness decay sequence; the initial lubricating film thickness is the lubricating film thickness value corresponding to the simulated cumulative addition of seawater being zero.

[0029] Feature recognition is performed on the normalized lubricating film thickness decay sequence to obtain the lubricating film thickness decay feature vector.

[0030] Furthermore, the method for obtaining a lubricating film thickness decay feature vector by feature recognition of the normalized lubricating film thickness decay sequence includes:

[0031] For the normalized lubricating film thickness decay sequence of each candidate formulation, enumerate all pairs of dividing points that can divide the entire equally divided interval into three continuous non-empty subsets; the equally divided interval is the interval obtained by dividing the range of simulated cumulative seawater addition into N equal parts; for each pair of dividing points, calculate the sum of the squares of the differences between the normalized lubricating film thickness reduction and the mean of the subset in each equally divided part of the three subsets; determine the optimal pair of dividing points for each candidate formulation based on minimizing the sum of the sums in the three subsets.

[0032] Based on the optimal segmentation point, the normalized lubricating film thickness decay sequence is divided into a slow decay interval, a gradual decay interval, and a rapid decay interval; the sum of the normalized lubricating film thickness reduction between each equal partition in the slow decay interval, the gradual decay interval, and the rapid decay interval is calculated respectively, and a lubricating film thickness decay feature vector is constructed.

[0033] Furthermore, the method for dividing the normalized lubricating film thickness decay sequence into a slow decay interval, a gradual decay interval, and a rapid decay interval based on the optimal segmentation point includes:

[0034] Based on the optimal split point, the normalized lubricating film thickness decay sequence is divided into the first candidate interval, the second candidate interval, and the third candidate interval. The least squares linear fitting is performed on each candidate interval with the sequence number between equal partitions as the independent variable and the reduction in normalized lubricating film thickness as the dependent variable to obtain the slope of the decay trend.

[0035] The candidate interval with the steepest slope of the decline trend is marked as the acute decline interval, the candidate interval with the smallest slope is marked as the mild decline interval, and the candidate interval with a moderate slope is marked as the gradual decline interval.

[0036] Furthermore, the method for aggregating the components of the lubricating film thickness degradation feature vector into a comprehensive degradation index based on the mapping relationship, and optimizing the formulation components with the minimum comprehensive degradation index as the optimization objective, to obtain the optimized formulation of the anti-corrosion marine gear oil includes:

[0037] The range of the components corresponding to each decay interval of the lubricating film thickness decay feature vector is obtained by calculating the difference between the maximum and minimum values ​​of the components in all candidate formulations.

[0038] The constraints are that the mass fraction of each formulation component is not lower than the preset minimum effective mass fraction of each formulation component, not higher than the preset mass fraction of each formulation component, and the sum of the mass fractions of each formulation component equals the preset total mass fraction. A genetic algorithm is used to iteratively generate candidate formulation components under the constraints. The candidate formulation components are input into the mapping relationship to obtain the predicted value of the lubricating film thickness decay feature vector. The components corresponding to the slow decay interval, gradual decay interval, and rapid decay interval in the predicted value of the lubricating film thickness decay feature vector are divided by the range of the corresponding components in the same decay interval, and then multiplied by the corresponding decay intensity index. The products obtained from the three decay intervals are added together to obtain the comprehensive decay index. The decay intensity index is: 1 for the slow decay interval, 2 for the gradual decay interval, and 3 for the rapid decay interval. The candidate formulation component with the smallest comprehensive decay index is selected as the optimized formulation of the anti-corrosion marine gear oil.

[0039] Based on the same inventive concept, this invention also provides a machine learning-based system for optimizing the formulation of corrosion-resistant marine gear oils, the system comprising:

[0040] The formulation generation module is used to generate several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation.

[0041] The lubricating film measurement module is used to simulate the intrusion process of water and salt spray on anti-corrosion marine gear oil under marine working conditions for each candidate formulation. During the intrusion process, the thickness of the lubricating film is measured to obtain the lubricating film thickness decay sequence.

[0042] The lubricating film feature learning module is used to identify features of the lubricating film thickness decay sequence and construct a lubricating film thickness decay feature vector. Using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, the gradient boosting regression algorithm is used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion.

[0043] The formulation optimization module is used to aggregate the components of the lubricating film thickness decay feature vector into a comprehensive decay index according to the mapping relationship, and optimize the formulation components with the minimum comprehensive decay index as the optimization objective to obtain the optimized formulation of anti-corrosion marine gear oil.

[0044] (3) Beneficial effects

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. By simulating the intrusion process of water and salt spray under marine conditions, the decay sequence of lubricating film thickness was obtained. The gradient boosting regression algorithm was used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion. The formulation components were optimized with the minimum comprehensive decay index as the optimization objective. This effectively solved the problem that the formulation composition was difficult to correlate with the actual film-forming protection ability during the actual intrusion process, so that the optimized formulation could truly meet the anti-corrosion needs under complex marine conditions.

[0047] 2. By constructing a characteristic vector of lubricating film thickness decay and using the comprehensive decay index as the optimization target, the optimized formula maintains a strong film-forming protection capability in all stages of the invasion cycle, further improving the reliability of corrosion prevention. Attached Figure Description

[0048] Figure 1 This is a flowchart of the machine learning-based method for optimizing the formulation of anti-corrosion marine gear oil according to Embodiment 1 of the present invention.

[0049] Figure 2 This is a schematic diagram of the module composition of the machine learning-based anti-corrosion marine gear oil formulation optimization system according to Embodiment 2 of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0051] Before providing examples, it is necessary to describe the application scenarios of this invention. This invention is applicable to marine transmission equipment that is exposed to salt spray and high humidity environments for extended periods.

[0052] Example 1: As Figure 1 As shown in this embodiment, a machine learning-based method for optimizing the formulation of corrosion-resistant marine gear oil is provided. The method includes:

[0053] S1. Generate several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation.

[0054] S2. Simulate the invasion process of water and salt spray on anti-corrosion marine gear oil under marine working conditions for each candidate formulation, and measure the thickness of the lubricating film during the invasion process to obtain the lubricating film thickness decay sequence.

[0055] S3. Perform feature identification on the lubricating film thickness decay sequence and construct a lubricating film thickness decay feature vector; use the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, and use the gradient boosting regression algorithm to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion.

[0056] For example, using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, a gradient boosting regression algorithm is used to train the mapping relationship between the formulation components and the lubricating film thickness decay feature vector. Gradient boosting regression constructs multiple decision trees iteratively, with each tree fitting the residual from the previous step. This is suitable for scenarios where there is a non-linear mapping relationship between the formulation components and the lubricating film thickness decay distribution, and where the training sample size is limited. Each component of the lubricating film thickness decay feature vector is trained using an independent regressor. 80% of the 200 candidate formulations are randomly divided as the training set, and 20% as the validation set. A grid search is performed within the range of 100-500 trees, a learning rate of 0.01-0.1, and a maximum tree depth of 3-6. The minimum root mean square mean of the prediction error of each component of the lubricating film thickness decay feature vector on the validation set is used as the criterion to determine the number of trees as 200, the learning rate as 0.05, and the maximum tree depth as 4.

[0057] S4. Based on the mapping relationship, the components of the lubricating film thickness decay feature vector are aggregated into a comprehensive decay index. The minimum comprehensive decay index is used as the optimization objective to optimize the formulation components and obtain the optimized formulation of the anti-corrosion marine gear oil.

[0058] The method for generating several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation includes:

[0059] Obtain the mass fraction range of each component in the anti-corrosion marine gear oil formulation; generate an initial candidate formulation with the constraint that the mass fraction of each component is within its corresponding mass fraction range and the sum of the mass fractions of all components equals the preset total mass fraction.

[0060] The corrosion resistance values ​​of the initial candidate formulations were measured, and the initial candidate formulations with corrosion resistance values ​​not lower than the preset corrosion resistance threshold were retained, resulting in several groups of candidate formulations.

[0061] For example, consider a corrosion-resistant marine gear oil for a gear transmission system on an offshore platform in a specific sea area. The formulation consists of base oil and six types of functional additives: rust inhibitor, extreme pressure anti-wear agent, demulsifier, metal passivator, defoamer, and antioxidant, totaling seven components. Based on the recommended dosages from various additive suppliers and industry experience, and considering the requirements for corrosion protection and demulsification performance under the high-humidity salt spray conditions of the target sea area, the mass fraction range of each component is determined as follows: base oil 88%–95%, rust inhibitor 1.5%–4%, extreme pressure anti-wear agent 1%–3%, demulsifier 0.8%–2%, metal passivator 0.1%–0.5%, defoamer 0.05%–0.2%, and antioxidant 0.5%–2%, with a preset total mass fraction of 100%.

[0062] The input consists of pre-determined formulation components and their corresponding mass fraction ranges for the target usage scenario. Latin hypercube sampling is used to generate sampling points within the mass fraction range of each formulation component. Each sampling result is scaled proportionally to a total mass fraction of 100%. Sampling results where the mass fraction of each component exceeds the corresponding range after scaling are removed, generating an initial candidate formulation that meets the formulation constraints.

[0063] The single-batch testing cycle for invasive simulation experiments is relatively long. Including formulations with significantly substandard basic corrosion resistance would generate a large number of invalid training samples and introduce training noise. Therefore, static corrosion resistance performance tests were conducted on each initial candidate formulation according to the ASTM D665B method (using synthetic seawater as the medium), with the percentage of non-corroded area on the metal specimen surface used as the corrosion resistance performance value. The preset corrosion resistance threshold was determined based on the operation and maintenance standards of equipment in the target marine area; in this embodiment, it was set to 95%. Initial candidate formulations with corrosion resistance performance values ​​of not less than 95% were retained, resulting in several candidate formulation groups. In this embodiment, Latin hypercube sampling was used to generate 400 initial candidate formulation groups. After screening using the ASTM D665B static corrosion resistance performance test, formulations with corrosion resistance performance values ​​of not less than 95% were retained, resulting in 200 candidate formulation groups.

[0064] The method for simulating the invasion process of moisture and salt spray on anti-corrosion marine gear oil under marine operating conditions for each candidate formulation, and measuring the lubricating film thickness during the invasion process to obtain the lubricating film thickness decay sequence includes:

[0065] Collect moisture and salt spray data of the target sea area; prepare simulated seawater based on the moisture and salt spray data of the target sea area.

[0066] Simulated seawater was added to the anti-corrosion marine gear oil samples of each candidate formulation in several batches; the thickness of the lubricating film was measured after each addition of simulated seawater; and a simulated seawater addition-lubricating film thickness curve was constructed based on the cumulative amount of simulated seawater added and the corresponding lubricating film thickness at the end of the addition process.

[0067] Based on the simulated seawater addition-lubricating film thickness curve, the cumulative amount of simulated seawater added when the lubricating film thickness decreases at the maximum rate is taken as the intrusion characteristic quantity.

[0068] Using zero as the lower bound and the intrusion characteristic quantity as the upper bound, the range of simulated cumulative seawater addition is divided into N equal parts; the lubricating film thickness value corresponding to each division point is extracted sequentially according to the simulated seawater addition amount-lubricating film thickness curve to obtain the lubricating film thickness decay sequence of each candidate formulation.

[0069] For example, seawater samples were collected from the target sea area. Chloride ion content was determined according to GB / T 7476, total dissolved solids according to GB / T 13200, and the mass concentrations of sodium, magnesium, sulfate, calcium, and potassium ions were determined using conventional ion chromatography. The pH of the seawater was measured using a precision pH meter. Based on the ASTM D1141 standard artificial seawater formulation, the dosage of each ionic component was adjusted proportionally according to the ratio of the measured ion mass concentration in the target sea area to the ASTM D1141 standard value. The volume was brought to the target volume with deionized water, and the pH was adjusted to match the measured pH of the target sea area using dilute hydrochloric acid or sodium hydroxide solution. After preparation, the overall salinity was verified using a conductivity meter.

[0070] Taking a specific target sea area as an example, the measured main parameters were: sodium chloride 21.2 g / L, magnesium chloride 9.5 g / L, sodium sulfate 4.0 g / L, calcium chloride 1.1 g / L, potassium chloride 0.6 g / L, pH 8.1, and total salinity approximately 36.4 g / L. Based on these parameters, the amounts of each salt were adjusted according to ASTM D1141 standards. Simulated seawater was then prepared with deionized water, the pH was adjusted to 8.1, and the conductivity was verified to be consistent with the seawater sampled on-site before being used in subsequent simulation experiments.

[0071] Simulated seawater was added to the anti-corrosion marine gear oil samples of each candidate formulation in stages. After each addition, the thickness of the lubricating film was measured using optical interferometry. The measuring device was a ball-and-disc thin film thickness gauge, with a contact pair consisting of a ceramic ball and an optical glass disk. After each addition of simulated seawater to the anti-corrosion marine gear oil sample, it was allowed to stand for (5±0.5) min. An appropriate amount of oil was taken from the sample and placed on the glass disk. The measurement program was started, and the average thickness of the central film during the steady-state phase was recorded as the corresponding lubricating film thickness. A simulated seawater addition amount-lubricating film thickness curve was constructed for each candidate formulation, with the cumulative amount of simulated seawater added as the x-axis and the corresponding lubricating film thickness as the y-axis. The simulated seawater addition amount-lubricating film thickness curve was analyzed, and the cumulative amount of simulated seawater added corresponding to the maximum rate of decrease in lubricating film thickness was extracted as the intrusion characteristic quantity.

[0072] When simulated seawater is added in stages, the difference in lubricating film thickness before and after each addition is used as the reduction in lubricating film thickness for that stage. The maximum single reduction in lubricating film thickness and its corresponding stage are tracked in real time. When, after the stage corresponding to the maximum recorded single reduction in lubricating film thickness, the reduction in lubricating film thickness after at least 5 consecutive additions is lower than that maximum value, the peak value of the lubricating film thickness decrease rate is determined to have stably fallen within the current measurement range, and the addition is terminated. Setting the number of consecutive additions to at least 5 is to prevent a brief drop in lubricating film thickness reduction near the peak value from being mistakenly interpreted as the peak value having passed, and to ensure that the peak point does not fall at the end of the measurement range but has sufficient support for subsequent measurement points.

[0073] The upper bound of the sequence is set using the intrusion characteristic amount rather than a fixed addition amount because different candidate formulations have different resistance to water intrusion. If a uniform fixed addition amount is used as the cutoff point, formulations with strong resistance will not have entered the critical degradation stage at the cutoff point, while formulations with weak resistance may have already failed completely. The degradation stages experienced by formulations with different resistance at the same fixed cutoff amount are inherently different, and the resulting curve segments are not comparable. Using the simulated cumulative addition amount of seawater at the point where the lubricating film thickness decreases at its maximum as the upper bound allows the degradation description of each candidate formulation to be uniformly locked at the critical stage where the intrusion impact is most concentrated.

[0074] It should be noted that the actual measurement points of different candidate formulations are not evenly spaced or consistent in number on the horizontal axis, making direct cross-formulation comparisons impossible. The purpose of equal division is to establish a unified positional correspondence among the candidate formulations, ensuring that the lubricating film thickness decay sequences of different formulations have the same dimension and comparable positions. The lubricating film thickness values ​​at each division point are not derived from actual measurement points, but rather extracted by interpolation from the simulated seawater addition-lubricating film thickness curve. Therefore, the value of N is determined based on the principle that each division interval contains at least one actual measurement point. The number of actual measurement points for each candidate formulation within the range from zero to the intrusion characteristic amount is counted, and the minimum value among all candidate formulations is selected as N. Subsequent feature identification steps divide the normalized lubricating film thickness decay sequence into three continuous subsets. To ensure that each subset contains at least three equal division intervals to support statistical analysis within each subset, N must not be less than 9. If the minimum number of actual measurement points obtained is less than 9, the single simulated seawater addition amount for the relevant candidate formulation should be appropriately reduced and additional measurements should be taken until the number of actual measurement points for each candidate formulation within the range from zero to the intrusion characteristic amount is not less than 9, at which point the minimum value of N is recalculated. In this embodiment, the minimum number of actual measurement points for each candidate formulation in the range from zero to the invasive characteristic amount is 20, which satisfies the requirement that N is not less than 9. Therefore, N=20 is taken. The lubricating film thickness values ​​corresponding to 21 equally divided points are extracted sequentially from the simulated seawater addition amount-lubricating film thickness curve to obtain the lubricating film thickness decay sequence of each candidate formulation.

[0075] The method of adding simulated seawater in stages to the anti-corrosion marine gear oil samples of each candidate formulation includes:

[0076] Simulated seawater was added to the pre-samples of anti-corrosion marine gear oil for each candidate formulation and mixed uniformly. After uniform mixing, the pre-samples of anti-corrosion marine gear oil were centrifuged and the termination time was when the volume of the free water phase no longer increased. The time interval from the completion of uniform mixing to the termination time was taken as the demulsification time.

[0077] The amount of simulated seawater added in a single step is obtained by multiplying the preset salt spray deposition rate, the oil surface area of ​​each candidate formulation of the anti-corrosion marine gear oil sample, and the corresponding demulsification time. Simulated seawater of the same amount is then added to the anti-corrosion marine gear oil sample in multiple steps.

[0078] For example, the demulsification time is determined on a preparatory sample of the corrosion-resistant marine gear oil, rather than directly on the sample used for invasive simulation testing. This is because the demulsification test requires the addition of simulated seawater and forced mixing, which causes irreversible disturbance to the oil's state. If the sample were directly used, it would be contaminated before the formal invasive simulation begins, failing to reflect the initial state of the formulation. It should be noted that the demulsification time in this invention is determined by centrifugal separation, unlike standard methods such as ASTM D1401 which evaluate oil-water separation by natural stratification under static conditions. The purpose of this invention is to quantify the time required for the formulation to spontaneously recover separation after forced disturbance, thereby characterizing the emulsification stability of the formulation under actual invasive conditions, rather than evaluating its static demulsification performance.

[0079] An equal volume of simulated seawater was added to a preparatory sample of anti-corrosion marine gear oil and mixed thoroughly. Immediately after mixing, the mixture was transferred to a standard 50mL graduated centrifuge tube and centrifuged at 3000 rpm. The volume of the free aqueous phase was recorded every 5 minutes. When the difference between two consecutive readings did not exceed 0.05mL, the volume of the free aqueous phase was considered to have stopped increasing, and this moment was recorded as the termination point. The time interval from the completion of uniform mixing to the termination point was taken as the demulsification time. For example, the demulsification time for a candidate formulation was measured to be 45 minutes. The centrifugation speed was set to 3000 rpm, which allowed the free aqueous phase to settle sufficiently within a few minutes without forcibly breaking down the still-emulsified droplets, ensuring that the measured demulsification time reflected the spontaneous separation capability of the formulation. The centrifugation speed, reading interval, and termination criteria were consistent for all candidate formulations in the same batch.

[0080] The preset salt spray deposition rate is characterized by the equivalent liquid water volume deposition rate, which is the equivalent liquid water volume received per unit area of ​​oil surface per unit time. Standard salt spray collectors are deployed at representative locations in the target sea area using a wet sedimentation method. The collected liquid volume is measured and divided by the collector orifice area and collection time to obtain a daily average value; in this embodiment, it is taken as 2.0 mL / (m²·h). The oil surface area of ​​the corrosion-resistant marine gear oil sample is taken as the cross-sectional area of ​​the sample container's inner cavity. In this embodiment, a flat-bottomed cylindrical container with an inner diameter of approximately 28 cm is selected. After measuring the container's inner diameter with calipers, the oil surface area is calculated to be approximately 616 cm².

[0081] In actual marine conditions, salt spray continuously deposits on the oil surface at an approximately constant rate. The deposited water separates again after one demulsification cycle. Therefore, the amount of water deposited on the oil surface during one demulsification cycle is the natural equivalent water volume added each time in the simulated actual intrusion process. Substituting the preset salt spray deposition rate of 2.0 mL / (m²·h), the oil surface area of ​​616 cm², and the demulsification time of 45 min, the amount of simulated seawater added in a single batch is 2.0 × (616 / 10000) × (45 / 60), which is approximately equal to 0.09 mL. Simulated seawater is added in portions of 0.09 mL to the anti-corrosion marine gear oil sample.

[0082] The method of using the cumulative amount of simulated seawater added at the point where the lubricating film thickness decreases at the maximum rate as an intrusion characteristic quantity, based on the simulated seawater addition amount-lubricating film thickness curve, includes:

[0083] Cubic spline fitting was performed on the simulated seawater addition amount-lubricating film thickness curves of each candidate formulation to obtain a cubic spline function of the lubricating film thickness with respect to the simulated cumulative seawater addition amount.

[0084] Within the horizontal axis range of the simulated seawater addition amount-lubricating film thickness curve, all local maxima of the lubricating film thickness decrease rate corresponding to the cubic spline function are extracted; the cumulative amount of simulated seawater addition corresponding to the local maxima of the lubricating film thickness decrease rate is taken as the intrusion characteristic quantity of each candidate formulation.

[0085] If, within the abscissa range of the simulated seawater addition amount-lubricating film thickness curve, the lubricating film thickness decrease rate corresponding to the cubic spline function has no local maximum point, then the cumulative simulated seawater addition amount corresponding to the right endpoint of the abscissa is taken as the intrusion characteristic quantity.

[0086] For example, in the experiment, both the simulated cumulative seawater addition and the lubricating film thickness were measured at discrete points. Directly calculating the quotient of adjacent differences as the rate of descent for discrete points resulted in significant noise interference, making it difficult to accurately locate the rate extrema. Cubic spline fitting was performed on the simulated seawater addition-lubricating film thickness curves for each candidate formulation, using natural boundary conditions (i.e., setting the second derivative at both endpoints to zero). This yielded a continuously differentiable cubic spline function of the lubricating film thickness with respect to the simulated cumulative seawater addition. The rate of change was obtained by differentiating the cubic spline function. The value is negative throughout the entire intrusion range (the lubricating film thickness decreases with increasing water content); take its opposite value. , defined as the rate of decrease in lubricating film thickness at each location, can be used to stably calculate the rate of decrease in lubricating film thickness at any location within the range of the horizontal axis.

[0087] Instead of directly taking the maximum value after differentiating the cubic spline function, the local maxima of the lubricating film thickness decay rate are extracted because different functional additives in anti-corrosion marine gear oils exhibit different failure sequences during water intrusion, and the lubricating film thickness often shows a multi-stage decay characteristic, with multiple local maxima in the lubricating film thickness decay rate. Taking a candidate formulation as an example, local maxima of the lubricating film thickness decay rate appear at simulated cumulative seawater addition amounts of 0.63 mL and 1.8 mL, corresponding to lubricating film thickness decay rates of 3.2 nm / mL and 8.7 nm / mL, respectively. The decay rate is largest at 1.8 mL, and 1.8 mL is taken as the intrusion characteristic quantity of this candidate formulation. The extraction process is strictly limited to the abscissa range of the simulated seawater addition amount-lubricating film thickness curve to avoid spurious extrema generated by the cubic spline function in the extrapolated region outside the measurement range.

[0088] For candidate formulations where the lubricating film thickness decay rate corresponding to the cubic spline function within the aforementioned abscissa range has no local maxima, this typically means that the formulation's film-forming protective ability continuously declines after the simulated seawater intrusion begins. The lubricating film thickness decay rate decreases monotonically throughout, and the intrusion impact peak occurs at the initiation stage of intrusion rather than within the measurement range, lacking identifiable internal extreme points. In such cases, using the cumulative amount of simulated seawater added at the measurement termination time (i.e., the right endpoint of the abscissa) as the intrusion characteristic quantity indicates that the formulation is in a state of continuous high-speed decay throughout the entire measurement interval. Subsequent lubricating film thickness decay sequences will cover the entire measurement range, ensuring that the decay behavior is fully reflected in cross-formulation comparisons.

[0089] The method for identifying features of the lubricating film thickness decay sequence and constructing a lubricating film thickness decay feature vector includes:

[0090] A lubricating film thickness decay difference sequence is constructed based on the decrease in lubricating film thickness between adjacent equal division points in the lubricating film thickness decay sequence; the lubricating film thickness decay difference sequence is smoothed and then normalized based on the initial lubricating film thickness to obtain the normalized lubricating film thickness decay sequence; the initial lubricating film thickness is the lubricating film thickness value corresponding to the simulated cumulative addition of seawater being zero.

[0091] Feature recognition is performed on the normalized lubricating film thickness decay sequence to obtain the lubricating film thickness decay feature vector.

[0092] For example, the lubricating film thickness decay sequence records the absolute lubricating film thickness value at each division point. Since the initial lubricating film thickness varies among different candidate formulations, directly using the absolute lubricating film thickness value as a learning feature would introduce an initial baseline difference unrelated to the invasive decay behavior. A lubricating film thickness decay difference sequence is constructed by selecting the reduction in lubricating film thickness between adjacent division points, allowing the sequence to reflect the local decay magnitude within each division interval.

[0093] Measurement noise exists in the lubricating film thickness decay difference sequence. If directly normalized, the small fluctuations caused by the noise will be proportionally amplified after normalization, interfering with subsequent feature recognition. Therefore, a Savitzky-Golay filter is used to smooth the lubricating film thickness decay difference sequence, with a window length of 5 and a polynomial order of 2. Using the initial lubricating film thickness as a reference, the smoothed lubricating film thickness decay difference sequence is normalized to obtain the normalized lubricating film thickness decay sequence.

[0094] Feature identification was performed on the normalized lubricating film thickness decay sequence to obtain the lubricating film thickness decay feature vector for each candidate formulation.

[0095] The method for obtaining a lubricating film thickness decay feature vector by feature recognition of the normalized lubricating film thickness decay sequence includes:

[0096] For the normalized lubricating film thickness decay sequence of each candidate formulation, enumerate all pairs of dividing points that can divide the entire equally divided interval into three continuous non-empty subsets; the equally divided interval is the interval obtained by dividing the range of simulated cumulative seawater addition into N equal parts; for each pair of dividing points, calculate the sum of the squares of the differences between the normalized lubricating film thickness reduction and the mean of the subset in each equally divided part of the three subsets; determine the optimal pair of dividing points for each candidate formulation based on minimizing the sum of the sums in the three subsets.

[0097] Based on the optimal segmentation point, the normalized lubricating film thickness decay sequence is divided into a slow decay interval, a gradual decay interval, and a rapid decay interval; the sum of the normalized lubricating film thickness reduction between each equal partition in the slow decay interval, the gradual decay interval, and the rapid decay interval is calculated respectively, and a lubricating film thickness decay feature vector is constructed.

[0098] For example, dividing the normalized lubricating film thickness decay sequence into three segments represents a trade-off between descriptive accuracy and feature dimensionality: if divided into only two segments, it becomes impossible to distinguish between the two dynamic features of continuously accelerating decay rate and rapid oil film collapse, causing the optimization objective to lose awareness of the most dangerous rapid failure stage; if divided into four or more segments, the feature vector dimensionality increases, leading to a greater risk of overfitting under limited training sample size, and the differences in decay behavior between adjacent intervals tend to be small, resulting in decreased discriminative power. The three-segment division can fully capture the three typical decay dynamics during the intrusion process—from gradual to accelerated to rapid—at a lower feature dimensionality, while maintaining the generalization ability of the mapping model.

[0099] The lubricating film thickness decay during the invasion process is not uniformly distributed among different candidate formulations. Some formulations show concentrated decay in the early stages of invasion, while others show significant decay abruptly in the middle or late stages. If the three intervals are divided at fixed locations, incorrect segmentation will occur for formulations with decay concentrated near the boundaries. Enumerating the optimal segmentation point pairs based on minimizing the sum of variances within each group can adaptively determine the segmentation boundaries according to the distribution characteristics of the normalized lubricating film thickness decay sequence of each candidate formulation, making the decay behavior within each interval as consistent as possible.

[0100] Since N=20 in this embodiment, the normalized lubricating film thickness decay sequence is divided into 20 equal intervals. Two dividing points are each located at a non-overlapping position on the boundary of intervals 1 to 19, resulting in a total of C(19,2)=171 dividing point pairs. For each dividing point pair, the sum of the squares of the differences between the normalized lubricating film thickness reduction and the mean of the corresponding subset within each of the three subsets is calculated. The dividing point pair with the smallest sum is the optimal dividing point pair. Based on the optimal dividing point pair, the normalized lubricating film thickness decay sequence is divided into slow decay intervals, gradual decay intervals, and rapid decay intervals. The sum of the normalized lubricating film thickness reduction within each decay interval is calculated, and a lubricating film thickness decay feature vector is constructed. Taking a candidate formulation as an example, the sums of the normalized lubricating film thickness reductions in the three intervals are 0.18, 0.54, and 0.28, respectively, and the lubricating film thickness decay feature vector is [0.18, 0.54, 0.28].

[0101] The method for dividing the normalized lubricating film thickness decay sequence into slow decay intervals, gradual decay intervals, and rapid decay intervals based on the optimal segmentation point includes:

[0102] Based on the optimal split point, the normalized lubricating film thickness decay sequence is divided into the first candidate interval, the second candidate interval, and the third candidate interval. The least squares linear fitting is performed on each candidate interval with the sequence number between equal partitions as the independent variable and the reduction in normalized lubricating film thickness as the dependent variable to obtain the slope of the decay trend.

[0103] The candidate interval with the steepest slope of the decline trend is marked as the acute decline interval, the candidate interval with the smallest slope is marked as the mild decline interval, and the candidate interval with a moderate slope is marked as the gradual decline interval.

[0104] For example, the optimal segmentation point pairs are divided based on the uniformity of the distribution of the reduction in normalized lubricating film thickness in each equally divided interval. Different candidate formulations may exhibit different trends in the reduction in normalized lubricating film thickness within the same candidate interval. Simply assigning the candidate intervals directly to the slow decay interval, gradual decay interval, and rapid decay interval based on their positional order cannot guarantee that the labeling results reflect the actual dynamic characteristics of the decay behavior within each interval. Determining the labeling assignment based on the slope of the decay trend within each candidate interval ensures that the slow decay interval, gradual decay interval, and rapid decay interval consistently correspond to the same dynamic decay meaning across different candidate formulations.

[0105] Taking a candidate formulation as an example, the optimal segmentation point is located at the boundary between the 6th and 14th equal division intervals. The first candidate interval is equal division interval 1-6, the second candidate interval is equal division interval 7-14, and the third candidate interval is equal division interval 15-20. Using the sequence number of the equal division intervals within each candidate interval as the independent variable and the normalized reduction in lubricating film thickness as the dependent variable, least squares linear fitting is performed on each interval. The slopes of the decay trends for the three candidate intervals are 0.0046, 0.0086, and 0.0067, respectively. The second candidate interval (equal division interval 7-14, slope 0.0086) with the largest decay trend slope is marked as the rapid decay interval; the first candidate interval (equal division interval 1-6, slope 0.0046) with the smallest slope is marked as the slow decay interval; and the third candidate interval (equal division interval 15-20, slope 0.0067) with a moderate slope is marked as the gradual decay interval.

[0106] It should be noted that the markings for the slow decay interval, gradual decay interval, and rapid decay interval are based on the dynamic intensity of decay within each candidate interval, rather than their temporal position in the invasion process. In this embodiment, rapid decay is concentrated in the middle of the invasion, while the decay trend in the later stage of invasion is relatively slow, which is marked as the gradual decay interval. This reflects the actual decay behavior of the additives failing in the middle stage of the formulation and the oil film entering a relatively stable residual protective state in the later stage.

[0107] The method for optimizing the formulation of anti-corrosion marine gear oil by aggregating the components of the lubricating film thickness degradation feature vector into a comprehensive degradation index based on the mapping relationship, and using the minimum comprehensive degradation index as the optimization objective, includes:

[0108] The range of the components corresponding to each decay interval of the lubricating film thickness decay feature vector is obtained by calculating the difference between the maximum and minimum values ​​of the components in all candidate formulations.

[0109] With the constraints that the mass fraction of each formulation component is not lower than the preset minimum effective mass fraction of each formulation component, not higher than the preset mass fraction of each formulation component, and the sum of the mass fractions of each formulation component equals the preset total mass fraction, a genetic algorithm is used to iteratively generate candidate formulation components under the constraints. The candidate formulation components are input into the mapping relationship to obtain the predicted value of the lubricating film thickness decay feature vector. The components corresponding to the slow decay interval, gradual decay interval, and rapid decay interval in the predicted value of the lubricating film thickness decay feature vector are divided by the range of the corresponding components in the same decay interval, and then multiplied by the corresponding decay intensity index. The products obtained from the three decay intervals are added together to obtain the comprehensive decay index. The decay intensity index is: 1 for the slow decay interval, 2 for the gradual decay interval, and 3 for the rapid decay interval. The candidate formulation component with the smallest comprehensive decay index is selected as the optimized formulation of the anti-corrosion marine gear oil.

[0110] For example, the lubricating film thickness decay feature vector is a three-dimensional vector, with three components corresponding to the sum of normalized lubricating film thickness reductions in the slow decay, gradual decay, and rapid decay intervals, respectively. The absolute numerical ranges of the three components may differ significantly. If the weighted sum of the three components is directly used as the optimization objective, the component with the larger numerical range will dominate the optimization direction, obscuring the influence of other components. Normalization using the range of the components corresponding to each decay interval can eliminate the differences in numerical ranges between components. Based on the lubricating film thickness decay feature vectors of all 200 candidate formulations, the differences between the maximum and minimum values ​​of the components corresponding to the slow decay, gradual decay, and rapid decay intervals in the 200 candidate formulations are calculated. In this embodiment, the ranges of the components corresponding to the three decay intervals are 0.31, 0.42, and 0.38, respectively.

[0111] The preset minimum effective mass fraction of each formulation component is determined based on the minimum dosage required for each component to exert its functional effect. Below this minimum dosage, the functional contribution of the component is negligible. The preset minimum effective mass fraction can be referenced from the recommended minimum dosage provided by each additive supplier, based on the dosage at which the anti-corrosion performance begins to decline significantly. The upper limit of the preset mass fraction range for each formulation component is consistent with the upper limit of the mass fraction range for each formulation component in the formulation generation step. In this embodiment, the upper limit of the preset mass fraction range for each component is: base oil 95%, rust inhibitor 4%, extreme pressure anti-wear agent 3%, demulsifier 2%, metal passivator 0.5%, defoamer 0.2%, and antioxidant 2%. This ensures that the search space of the genetic algorithm is consistent with the coverage of the training candidate formulations, guaranteeing the reliability of the mapping relationship in predicting the candidate formulation components generated during the optimization process.

[0112] The genetic algorithm iteratively generates candidate formulation components under constraints: a population size of 200, a maximum number of generations of 500, a crossover probability of 0.85, and a mutation probability of 0.05. The specific genetic operations are as follows:

[0113] The selection operation uses a tournament selection process with a tournament size of 2. Each time, two individuals are randomly selected from the current population, and the one with the smaller overall decline index is selected as the parent to participate in the subsequent crossover operation. The crossover operation uses overall arithmetic crossover. For the two selected parent individuals, a random weight λ is uniformly sampled on [0,1]. The quality fraction of each ingredient in offspring 1 is the product of λ and the corresponding ingredient quality fraction of parent 1 plus the product of (1-λ) and the corresponding ingredient quality fraction of parent 2. Offspring 2 takes complementary weights (1-λ) and λ. The mutation operation uses Gaussian mutation. For the selected individuals to be mutated, a Gaussian random perturbation with a mean of 0 and a standard deviation of 5% of the range of the ingredient quality fraction is independently superimposed on each of its ingredient quality fractions. After mutation, the individuals are normalized.

[0114] Equality constraints are enforced by proportionally normalizing the components of each candidate formulation group to a total mass fraction of 100% after population initialization and each mutation. Inequality constraints are achieved by rejecting any formulation component whose mass fraction exceeds the upper limit of the corresponding preset mass fraction range or falls below the preset minimum effective mass fraction. The candidate formulation components are input into a mapping relationship to obtain the predicted value of the lubricating film thickness decay feature vector.

[0115] The components corresponding to the slow, gradual, and rapid decline intervals in the predicted value of the lubricating film thickness decline feature vector are divided by the range of the corresponding components within the same decline interval, and then multiplied by the corresponding decline intensity index. The products obtained for each of the three decline intervals are summed to obtain the comprehensive decline index. The decline intensity index uses an arithmetic progression of 1, 2, and 3: the more severe the decline interval, the stronger the damage to lubrication protection capability; assigning a higher decline intensity index allows the comprehensive decline index to impose a greater penalty on the rapid decline stage. The arithmetic progression design is not intended to precisely quantify the absolute hazard ratio of the three types of decline, but rather to ensure the priority of optimization directions with monotonically increasing weights, based on the fact that range normalization has eliminated the differences in the numerical ranges of each component: among candidate formulations with similar total film-forming performance decline, the formulation with the smallest rapid decline component is prioritized, thereby avoiding anti-corrosion failure caused by localized rapid decline. The candidate formulation component with the smallest comprehensive decline index is used as the optimized formulation for the anti-corrosion marine gear oil.

[0116] In this embodiment, after 500 generations of iteration and convergence, the optimal formula with the minimum comprehensive degradation index has the following component mass fractions: base oil 92.3%, rust inhibitor 3.2%, extreme pressure anti-wear agent 1.8%, demulsifier 1.5%, metal passivator 0.3%, defoamer 0.1%, and antioxidant 0.8%, totaling 100%. The mass fractions of each component are all within the preset range and not lower than the corresponding preset minimum effective mass fraction. The formula components of this optimized formula are input into the trained mapping relationship to obtain the predicted value of the lubricating film thickness degradation feature vector as [0.11, 0.28, 0.15]. Substituting the ranges of the corresponding components in the three degradation intervals (slow degradation 0.31, gradual degradation 0.42, rapid degradation 0.38) and the degradation intensity index, the comprehensive degradation index is (0.11 / 0.31)×1+(0.28 / 0.42)×2+(0.15 / 0.38)×3≈2.87. As a control, the formula with the lowest overall degradation index among the 200 training candidate formulas corresponds to a lubricating film thickness degradation feature vector of [0.14, 0.33, 0.19], and an overall degradation index of (0.14 / 0.31)×1+(0.33 / 0.42)×2+(0.19 / 0.38)×3≈3.52. The optimized formula's overall degradation index is about 18.5% lower than the optimal value among the training candidate formulas.

[0117] Example 2: Based on the same inventive concept, such as Figure 2 As shown, this embodiment also provides a machine learning-based system for optimizing the formulation of corrosion-resistant marine gear oils, including:

[0118] The formulation generation module is used to generate several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation.

[0119] The lubricating film measurement module is used to simulate the intrusion process of water and salt spray on anti-corrosion marine gear oil under marine working conditions for each candidate formulation. During the intrusion process, the thickness of the lubricating film is measured to obtain the lubricating film thickness decay sequence.

[0120] The lubricating film feature learning module is used to identify features of the lubricating film thickness decay sequence and construct a lubricating film thickness decay feature vector. Using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, the gradient boosting regression algorithm is used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion.

[0121] The formulation optimization module is used to aggregate the components of the lubricating film thickness decay feature vector into a comprehensive decay index according to the mapping relationship, and optimize the formulation components with the minimum comprehensive decay index as the optimization objective to obtain the optimized formulation of anti-corrosion marine gear oil.

[0122] It should be noted that the specific ways in which each module operates in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0123] Finally, it should be noted that although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for optimizing the formulation of corrosion-resistant marine gear oils, characterized in that, The method includes: Several candidate formulations were generated based on the formulation components of the anti-corrosion marine gear oil formulation; The invasion process of water and salt spray on anti-corrosion marine gear oil under simulated marine working conditions was investigated for each candidate formulation. The thickness of the lubricating film was measured during the invasion process to obtain the lubricating film thickness decay sequence. Feature identification was performed on the lubricating film thickness decay sequence to construct a lubricating film thickness decay feature vector; using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, a gradient boosting regression algorithm was used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion. Based on the mapping relationship, the components of the lubricating film thickness decay feature vector are aggregated into a comprehensive decay index. The optimization objective is to minimize the comprehensive decay index to optimize the formulation components and obtain the optimized formulation of anti-corrosion marine gear oil.

2. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 1, characterized in that, The method for generating several candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation includes: Obtain the mass fraction range of each component in the anti-corrosion marine gear oil formulation; generate an initial candidate formulation with the constraint that the mass fraction of each component is within its corresponding mass fraction range and the sum of the mass fractions of all components equals the preset total mass fraction. The corrosion resistance values ​​of the initial candidate formulations were measured, and the initial candidate formulations with corrosion resistance values ​​not lower than the preset corrosion resistance threshold were retained, resulting in several groups of candidate formulations.

3. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 2, characterized in that, The method for simulating the invasion process of moisture and salt spray on anti-corrosion marine gear oil under marine operating conditions for each candidate formulation, and measuring the lubricating film thickness during the invasion process to obtain the lubricating film thickness decay sequence includes: Collect moisture and salt spray data of the target sea area; prepare simulated seawater based on the moisture and salt spray data of the target sea area; Simulated seawater was added to the anti-corrosion marine gear oil samples of each candidate formulation in several batches; the thickness of the lubricating film was measured after each addition of simulated seawater; and a simulated seawater addition-lubricating film thickness curve was constructed based on the cumulative amount of simulated seawater added and the corresponding lubricating film thickness when the addition of simulated seawater was completed. Based on the simulated seawater addition amount-lubricating film thickness curve, the cumulative amount of simulated seawater added when the lubricating film thickness decreases at the maximum rate is taken as the intrusion characteristic quantity; Using zero as the lower bound and the intrusion characteristic quantity as the upper bound, the range of simulated cumulative seawater addition is divided into N equal parts; the lubricating film thickness value corresponding to each division point is extracted sequentially according to the simulated seawater addition amount-lubricating film thickness curve to obtain the lubricating film thickness decay sequence of each candidate formulation.

4. The method for optimizing the formulation of anti-corrosion marine gear oil based on machine learning according to claim 3, characterized in that, The method of adding simulated seawater in stages to the anti-corrosion marine gear oil samples of each candidate formulation includes: Simulated seawater was added to the pre-samples of anti-corrosion marine gear oil for each candidate formulation and mixed evenly. After the mixture was evenly mixed, the pre-samples of anti-corrosion marine gear oil were centrifuged and the process was terminated when the volume of the free water phase no longer increased. The time interval from the completion of the even mixing to the termination time was taken as the demulsification time. The amount of simulated seawater added in a single step is obtained by multiplying the preset salt spray deposition rate, the oil surface area of ​​each candidate formulation of the anti-corrosion marine gear oil sample, and the corresponding demulsification time. Simulated seawater of the same amount is then added to the anti-corrosion marine gear oil sample in multiple steps.

5. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 4, characterized in that, The method of using the cumulative amount of simulated seawater added at the point where the lubricating film thickness decreases at the maximum rate as an intrusion characteristic quantity, based on the simulated seawater addition amount-lubricating film thickness curve, includes: Cubic spline fitting was performed on the simulated seawater addition amount-lubricating film thickness curves of each candidate formulation to obtain a cubic spline function of the lubricating film thickness with respect to the simulated cumulative seawater addition amount; Within the horizontal axis range of the simulated seawater addition amount-lubricating film thickness curve, all local maxima of the lubricating film thickness decrease rate corresponding to the cubic spline function are extracted; the cumulative amount of simulated seawater addition corresponding to the local maxima of the lubricating film thickness decrease rate is used as the intrusion characteristic quantity of each candidate formulation. If, within the abscissa range of the simulated seawater addition amount-lubricating film thickness curve, the lubricating film thickness decrease rate corresponding to the cubic spline function has no local maximum point, then the cumulative simulated seawater addition amount corresponding to the right endpoint of the abscissa is taken as the intrusion characteristic quantity.

6. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 1, characterized in that, The method for identifying features of the lubricating film thickness decay sequence and constructing a lubricating film thickness decay feature vector includes: A lubricating film thickness decay difference sequence is constructed based on the decrease in lubricating film thickness between adjacent equal division points in the lubricating film thickness decay sequence; the lubricating film thickness decay difference sequence is smoothed and then normalized based on the initial lubricating film thickness to obtain the normalized lubricating film thickness decay sequence; the initial lubricating film thickness is the lubricating film thickness value corresponding to the simulated cumulative addition of seawater being zero. Feature recognition is performed on the normalized lubricating film thickness decay sequence to obtain the lubricating film thickness decay feature vector.

7. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 6, characterized in that, The method for obtaining a lubricating film thickness decay feature vector by feature recognition of the normalized lubricating film thickness decay sequence includes: For the normalized lubricating film thickness decay sequence of each candidate formulation, enumerate all pairs of dividing points that can divide the entire equally divided interval into three continuous non-empty subsets; the equally divided interval is the interval obtained by dividing the range of simulated cumulative seawater addition into N equal parts; for each pair of dividing points, calculate the sum of the squares of the differences between the normalized lubricating film thickness reduction and the mean of the subset in each equally divided part of the three subsets; determine the optimal pair of dividing points for each candidate formulation based on minimizing the sum of the sums in the three subsets. Based on the optimal segmentation point, the normalized lubricating film thickness decay sequence is divided into a slow decay interval, a gradual decay interval, and a rapid decay interval; the sum of the normalized lubricating film thickness reduction between each equal partition in the slow decay interval, the gradual decay interval, and the rapid decay interval is calculated respectively, and a lubricating film thickness decay feature vector is constructed.

8. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 7, characterized in that, The method for dividing the normalized lubricating film thickness decay sequence into slow decay intervals, gradual decay intervals, and rapid decay intervals based on the optimal segmentation point includes: Based on the optimal split point, the normalized lubricating film thickness decay sequence is divided into the first candidate interval, the second candidate interval, and the third candidate interval. The least squares linear fitting is performed with the sequence number between equal partitions in each candidate interval as the independent variable and the reduction in normalized lubricating film thickness as the dependent variable to obtain the decay trend slope. The candidate interval with the steepest slope of the decline trend is marked as the acute decline interval, the candidate interval with the smallest slope is marked as the mild decline interval, and the candidate interval with a moderate slope is marked as the gradual decline interval.

9. The method for optimizing the formulation of corrosion-resistant marine gear oil based on machine learning according to claim 8, characterized in that, The method for optimizing the formulation of anti-corrosion marine gear oil by aggregating the components of the lubricating film thickness degradation feature vector into a comprehensive degradation index based on the mapping relationship, and using the minimum comprehensive degradation index as the optimization objective, includes: The difference between the maximum and minimum values ​​of the components corresponding to each decay interval of the lubricating film thickness decay feature vector in all candidate formulations is calculated to obtain the range of the components corresponding to the decay interval. The constraints are that the mass fraction of each formulation component is not lower than the preset minimum effective mass fraction of each formulation component, not higher than the preset mass fraction of each formulation component, and the sum of the mass fractions of each formulation component equals the preset total mass fraction. A genetic algorithm is used to iteratively generate candidate formulation components under the constraints. The candidate formulation components are input into the mapping relationship to obtain the predicted value of the lubricating film thickness decay feature vector. The components corresponding to the slow decay interval, gradual decay interval, and rapid decay interval in the predicted value of the lubricating film thickness decay feature vector are divided by the range of the corresponding components in the same decay interval, and then multiplied by the corresponding decay intensity index. The products obtained from the three decay intervals are added together to obtain the comprehensive decay index. The decay intensity index is: 1 for the slow decay interval, 2 for the gradual decay interval, and 3 for the rapid decay interval. The candidate formulation component with the smallest comprehensive decay index is selected as the optimized formulation of the anti-corrosion marine gear oil.

10. A machine learning-based system for optimizing the formulation of corrosion-resistant marine gear oils, used to perform the method according to any one of claims 1 to 9, characterized in that, The system includes: The formulation generation module is used to generate several sets of candidate formulations based on the formulation components of the anti-corrosion marine gear oil formulation. The lubricating film measurement module is used to simulate the intrusion process of water and salt spray on anti-corrosion marine gear oil under marine working conditions for each candidate formulation. During the intrusion process, the thickness of the lubricating film is measured to obtain the lubricating film thickness decay sequence. The lubricating film feature learning module is used to identify features of the lubricating film thickness decay sequence and construct a lubricating film thickness decay feature vector. Using the formulation components of each candidate formulation as training input and the corresponding lubricating film thickness decay feature vector as training label, the gradient boosting regression algorithm is used to establish the mapping relationship between the formulation components and the change law of lubricating film thickness with the degree of intrusion. The formulation optimization module is used to aggregate the components of the lubricating film thickness decay feature vector into a comprehensive decay index according to the mapping relationship, and optimize the formulation components with the minimum comprehensive decay index as the optimization objective to obtain the optimized formulation of anti-corrosion marine gear oil.

Images