A lubricating oil abrasive particle and gum pollution decoupling detection method and device based on transmission-scattering multi-source signal fusion and a storage medium

Abstract

The application belongs to the technical field of photoelectric detection, and particularly relates to a lubricating oil abrasive particle and gum pollution decoupling detection method and device based on transmission-scattering multi-source signal fusion and a storage medium. The method synchronously acquires transmission time sequence signals and non-forward scattering time sequence signals of the same oil detection area, and selectively fuses inductive abrasive particle signals with space-time correlation. After compensation preprocessing, the optical signals are decomposed into slow baseline components and transient event components. The transmission and scattering transient events are paired to identify particle type pollution events to estimate abrasive particle pollution contribution. After deducting the particle contribution from the baseline, the aging gum pollution contribution is estimated, and the two types of pollution contribution and confidence are corrected through physical constraint inversion or model fusion. The application can realize decoupling detection of abrasive particles and aging gum pollution, support degraded operation when the inductive channel fails, and is suitable for state monitoring of various mechanical equipment using lubricating and hydraulic working fluid.

Description

Technical Field

[0001] This invention belongs to the field of photoelectric detection technology, and specifically relates to a method, equipment and storage medium for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion. Background Technology

[0002] The total turbidity of lubricating oil is formed by a combination of factors, including absorption by base oil and additives, absorption and scattering by aging products, extinction of metallic and non-metallic particles, water droplets, bubbles, and window adhesion. A single transmittance value cannot definitively determine the source of contamination. Even when two average light intensities are measured simultaneously, the retrieval may still be ill-conditioned when unknown variables include particle size, concentration, morphology, temperature, and oil type.

[0003] Discrete particles typically produce local transient responses when passing through a finite beam of light, while uniformly or submicron-dispersed aged colloids are more likely to form a continuous background and drift slowly. The transmission and scattering channels exhibit different but correlated responses to the same event. Combining temporal scale differences, dual-channel correlation, and physical constraints in the inversion process can improve the identifiability of the contributions from both types of pollution.

[0004] Existing machine learning-based oil identification schemes often directly input average signals into classification models, easily learning random correlations between oil type, temperature, or equipment operating conditions, but lacking interpretation of particulate events, baselines, and failure boundaries. Therefore, a decoupling method is needed that is tailored to specific optical observation processes, has traceability features, and outputs uncertainty. Summary of the Invention

[0005] To address the aforementioned issues, this invention aims to provide a contamination decoupling method that combines physical scale separation, dual-channel event pairing, interference identification, and constraint inversion. Without relying on a single black-box model, it obtains the contributions of abrasive contamination and aging colloid contamination separately, and identifies the reliability and applicability of the results.

[0006] The present invention is implemented according to the following scheme: A method for decoupling detection of abrasive particles and gum contamination in lubricating oil based on transmission-scattering multi-source signal fusion includes the following steps: Acquire transmission time-series signals and non-forward scattering time-series signals generated from the same oil detection area and collected based on a common time reference; Preprocessing is performed on the transmission timing signal and the scattering timing signal; The preprocessed signal is decomposed into a baseline component characterizing a continuous or slowly changing response and a transient event component characterizing a contaminant crossing the detection area. Identifying particulate pollution events based on the temporal correspondence and / or correlation between transmission transient events and scattering transient events; Candidate contributions of abrasive contamination are determined based on the statistical characteristics of particulate contamination events; the impact of particulate contamination events is compensated or subtracted from the baseline components, and the contribution of aged colloid contamination is determined based on the compensated transmission baseline characteristics and scattering baseline characteristics. The transmitted signal, the scattered signal, and optional auxiliary sensing signals are fused at the feature level, the event level, or the decision level to correct the confidence of the abrasive contamination candidate contribution, the aging gum contamination contribution, or both; and output at least two of the abrasive contamination index and the aging gum contamination index, or output the relative contribution of the two.

[0007] Preferably, the preprocessing includes at least one of dark level subtraction, reference light normalization, ambient light suppression, synchronization demodulation, noise reduction, gain correction, temperature compensation, flow compensation, pressure compensation, viscosity compensation, window contamination compensation, and new oil baseline normalization.

[0008] Preferably, the signal decomposition employs low-pass and high-pass filtering, band-pass filtering, moving average, local polynomial fitting, wavelet decomposition, empirical mode decomposition, variational mode decomposition, Kalman filtering, robust trend estimation, or a combination thereof.

[0009] Preferably, the identification of particulate pollution events includes: detecting candidate pulses in the transmission transient event component and the scattering transient event component respectively; pairing candidate pulses within a preset time window; and determining that when the cross-correlation coefficient, amplitude relationship, polarity relationship, duration or waveform similarity of the paired pulses meet preset conditions, they are events generated by the same pollutant through a common detection volume. The statistical characteristics of particulate pollution events include at least two of the following: transmission drop amplitude, scattering response amplitude, pulse width, rise time, fall time, pulse integral area, event frequency, event interval, kurtosis, skewness, quantile, transmission-scattering amplitude ratio, transmission-scattering area ratio, and two-channel time delay. The events are divided into multiple equivalent particle size or severity intervals based on the amplitude or area distribution of particulate pollution events, and the contribution of abrasive pollution is formed based on the number of events, cumulative area or weighted sum of each interval.

[0010] This method uses at least one of the following: rule model, statistical classifier, support vector machine, decision tree, random forest, gradient boosting tree, neural network, probabilistic graphical model or clustering model to distinguish abrasive events, bubble events, water droplet events, soot aggregation events and other abnormal events.

[0011] Preferably, the contribution of aging gum contamination is determined by at least two of the following: transmission baseline, scattering baseline, transmission-scattering baseline ratio, baseline slope, baseline variability, change relative to new oil, and change relative to clean window state after event interval removal, interpolation, or model reconstruction.

[0012] A forward response model is established between the observed characteristics and the contributions of abrasive contamination and aging gum contamination. The contributions are then solved by inversion with non-negativity constraints, monotonicity constraints, mass conservation constraints, energy conservation constraints, sparsity constraints, smoothness constraints, or prior distribution constraints.

[0013] The forward response model should include at least the oil-based absorption term, particle extinction term, particle scattering term, colloidal absorption or scattering term, and window and environmental influence term, and allow each term to be a linear, piecewise linear or nonlinear function.

[0014] A calibration dataset was established using pure abrasive samples, pure aged adhesive samples, mixed samples of abrasive and adhesive in a known ratio, and interference samples containing water, air bubbles, dust, or soot.

[0015] The calibration dataset covers at least two particle sizes, at least two contamination concentrations, at least two temperature points, and at least one base oil or formulation oil. The model's generalization ability across oil types or across equipment is evaluated using leave-one-out-of-oil validation, leave-one-out-of-equipment validation, or domain adaptation methods.

[0016] Save the corresponding new oil baseline, model parameters and alarm thresholds for different oil types, equipment or operation stages, and update the model parameters according to oil change events, laboratory re-inspection results or equipment maintenance results.

[0017] Output abrasive contamination index, aging gum contamination index, and their equivalent extinction contribution rate or equivalent scattering contribution rate, and simultaneously output confidence interval, uncertainty, data validity flag or out-of-calibration range flag.

[0018] Preferably, the auxiliary sensing signal includes at least one of inductive abrasive signal, magnetic field modulation response, temperature, flow rate, pressure, viscosity, dielectric constant, moisture or vibration signal; when the auxiliary sensing signal includes inductive abrasive signal, at least one of the amplitude, polarity, phase, pulse width or resonance parameter of inductive pulse is used to confirm or probability-correct whether the optical particle event belongs to metal abrasive.

[0019] The detection area of ​​the inductive abrasive sensor is located inside the optical detection area, at least partially overlaps with the optical detection area, or is located upstream or downstream of the optical detection area. When the two detection areas are separated from each other, the theoretical passage delay is determined based on the distance between the detection areas and the oil flow rate, and the optical particle event is paired with the inductive event within the matching window corresponding to the theoretical passage delay.

[0020] Theoretically, the time delay is determined by dividing the equivalent distance between detection areas along the flow direction by the oil flow velocity. When the flow velocity is unknown or fluctuates, the time delay is obtained by the cross-correlation peak of the continuous event sequence, dynamic time warping, probability data correlation, or adaptive state estimation.

[0021] Events exhibiting both transmission-scattering pairing responses and inductive pairing responses are identified as metal abrasive events; events exhibiting only transmission-scattering pairing responses are identified as candidates for non-metallic particles, bubbles, metal abrasive particles below the inductive detection limit, or unknown particles; and events exhibiting only inductive responses are identified as candidates for optical missed detection, electromagnetic interference, or window contamination.

[0022] Preferably, ferromagnetic metal abrasive particles and non-ferromagnetic metal abrasive particles are distinguished based on the polarity, phase, frequency response, or impedance change direction of the inductive pulse, and the material category probability and transmission-scattering event characteristics are used together to calculate the contribution of abrasive particle contamination.

[0023] Preferably, when the inductive abrasive signal is missing, the sensor fails, the signal-to-noise ratio is below the threshold, or the abrasive size is below the inductive detection limit, the transmission-scattering signal is used to perform degradation decoupling; when the inductive abrasive signal is restored to a valid state, the optical abrasive severity model or event classification model is corrected online using the matched metal abrasive events.

[0024] The present invention also provides a lubrication equipment condition diagnosis method, which performs the above-mentioned decoupled detection method for lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion, and outputs mechanical wear inspection suggestions when the contribution of abrasive particle contamination exceeds a first threshold, outputs oil change, purification or oil temperature inspection suggestions when the contribution of aged gum contamination exceeds a second threshold, and outputs composite abnormality suggestions when both exceed the corresponding thresholds.

[0025] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a program, and when the program is executed by the processor, it implements the above-described method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion.

[0026] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion.

[0027] The working principle of this invention is: First, the transmission T(t) and scattering S(t) are obtained based on a common time reference; then, the standardized signals Tn(t) and Sn(t) are obtained through dark level subtraction, reference light normalization, and temperature compensation.

[0028] Then, a multi-scale decomposition is performed: Tn = BT + PT + NT, Sn = BS + PS + NS, where BT and BS are slowly varying baselines, PT and PS are transient components, and NT and NS are noise or unmodeled terms. The decomposition method can be fixed filtering, adaptive wavelet, state-space estimation, or robust trend decomposition.

[0029] Candidate pulses are detected in PT and PS; candidate pulses with a time difference not exceeding Δt and satisfying cross-correlation, polarity and amplitude relationships are regarded as the same particle crossing event; an event feature vector x=[AT,AS,WT,WS,QT,QS,AS / AT,Δτ,…] is constructed; abnormally wide pulses, saturated pulses or pulses synchronized with flow changes are bubbled and anomaly removed.

[0030] The contribution of abrasive contamination, Cw, can be obtained by weighting the number of events and event characteristics, or by calibration regression. The contribution of colloidal contamination, Cv, is estimated by BT, BS, their ratios, and rates of change after removing events. For high-concentration multiple scattering scenarios, a nonlinear forward model y=F(Cw,Cv,z)+ε can be established, where z is a covariate such as temperature, oil type, and window state, and the inverse problem with nonnegativity and monotonic constraints can be solved.

[0031] Machine learning models are used to correct for complex nonlinearities and interference boundaries, rather than replacing physical decomposition. Training is performed using pure components, mixtures with known proportions, and interference samples; testing is conducted to perform cross-oil type or cross-device validation. If the input is outside the calibration range, a low-confidence output or a rejection of the quantitative result is given.

[0032] In a high-reliability implementation, the auxiliary sensing signal is an inductive abrasive grain signal. The inductive detection area can overlap with the optical co-detection area, or it can be located upstream or downstream of it. When the detection areas separate along the flow direction, event delay compensation is performed based on the distance between them and the oil flow rate, thereby establishing a correspondence between optical events and inductive events generated by the same abrasive grain.

[0033] The inductive channel does not directly handle the detection of adhesive particles, but rather serves as an anchor point for metallic properties: when transmission-scattering pairing events and inductive events occur simultaneously, the probability that the event belongs to metallic abrasive particles is increased; when only optical events occur, the amplitude, width, and interference model are used to determine whether the event is a small-sized metallic abrasive particle, non-metallic particle, bubble, or unknown particle; when only inductive events occur, the event is determined to be a candidate for optical missed detection, electromagnetic interference, or window anomaly.

[0034] Inductive detection can employ differential coils, three-coil systems, planar coils, microfluidic coils, resonant coils, impedance detectors, or other structures capable of sensing changes in permeability or conductivity caused by metal particles. Ferromagnetic and non-ferromagnetic abrasive particles can be distinguished based on the polarity, phase, or direction of impedance change of the inductive pulse.

[0035] This invention retains the degradation operation mechanism. When there is no inductor channel or the inductor signal is invalid, dual-source decoupling continues to be implemented based on transmission-scattering events and baseline; when the inductor signal is valid, three-source fusion is performed and the optical event classifier and severity model are corrected online using confirmed metal events.

[0036] The beneficial effects of this invention are: (1) This invention utilizes the difference between the time-domain transient and the slowly varying baseline to reduce the aliasing of the two types of pollution.

[0037] (2) The present invention improves the authenticity of events by pairing the same particles in the transmission and scattering channels.

[0038] (3) The present invention reduces the bias of abrasive particles in aging judgment by estimating the colloid after event subtraction.

[0039] (4) This invention combines physical constraints with machine learning, and has good interpretability and stability across working conditions.

[0040] (5) This invention avoids misreporting out-of-model samples as definitive results by outputting uncertainty and superdomain flag; (6) The inductive channel of the present invention provides independent evidence of metallic properties, reducing the probability of misjudging bubbles, colloidal agglomerates or non-metallic particles as metallic abrasive particles.

[0041] (7) This invention supports flow rate compensation and event registration when the optical and inductive detection areas are separated.

[0042] (8) The present invention can still be downgraded to dual optical decoupling when the inductor channel is invalid; (9) The method provided by the present invention is not limited to a specific type of equipment and can be used in all mechanical systems with lubricating or hydraulic working fluids. Attached Figure Description

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0044] Figure 1 This is a flowchart of the method of the present invention.

[0045] Figure 2 This is a flowchart of the dual-channel preprocessing and multi-scale decomposition method of the present invention.

[0046] Figure 3 This is a flowchart of the dual-channel event pairing and interference removal process of the present invention.

[0047] Figure 4 This is a block diagram of the physical constraint decoupling and uncertainty output of the method of the present invention.

[0048] Figure 5 This is a flowchart illustrating the calibration, training, and cross-oil type verification process of the method of this invention.

[0049] Figure 6 This is a flowchart of the spatiotemporal registration and fusion of transmission-scattering-inductive events in the method of the present invention. Detailed Implementation

[0050] The present invention will now be clearly described with reference to specific embodiments. These descriptions are merely illustrative and are not intended to limit the scope of the invention. Any modifications, equivalent substitutions, or improvements made by those skilled in the art based on the embodiments of the present invention without inventive effort to obtain all other embodiments should be included within the scope of protection of the present invention.

[0051] Example 1 like Figure 1-6 As shown, this embodiment provides a method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion, including the following steps: S101 Synchronous Acquisition: Acquires transmission time-series signals and non-forward scattering time-series signals generated from the same oil detection area and acquired based on a common time reference; S102 Preprocessing: Preprocessing the transmission timing signal and the scattering timing signal; S103 scale decomposition: decompose at least one preprocessed signal into a baseline component characterizing a continuous or slowly changing response and a transient event component characterizing a pollutant crossing the detection area. S104 Event Identification: Identifying particulate pollution events based on the temporal correspondence and / or correlation between transmission transient events and scattering transient events; S105 Contribution Inversion: Candidate contributions of abrasive contamination are determined based on the statistical characteristics of particulate contamination events; the impact of particulate contamination events is compensated or subtracted from the baseline components, and the contribution of aged colloid contamination is determined based on the compensated transmission baseline characteristics and scattering baseline characteristics. S106 Output Confidence: The transmitted signal, the scattered signal, and the optional auxiliary sensing signal are fused at the feature level, event level, or decision level to correct the confidence of the candidate contribution of abrasive contamination, the contribution of aging gum contamination, or both; output at least two of the abrasive contamination index and the aging gum contamination index, or output the relative contribution of the two.

[0052] The above preprocessing includes at least one of the following: dark level subtraction, reference light normalization, ambient light suppression, synchronization demodulation, noise reduction, gain correction, temperature compensation, flow compensation, pressure compensation, viscosity compensation, window contamination compensation, and new oil baseline normalization.

[0053] Signal decomposition employs low-pass and high-pass filtering, band-pass filtering, moving average, local polynomial fitting, wavelet decomposition, empirical mode decomposition, variational mode decomposition, Kalman filtering, robust trend estimation, or combinations thereof.

[0054] The identification of particulate pollution events includes: detecting candidate pulses in the transmission transient event component and the scattering transient event component respectively; pairing candidate pulses within a preset time window; and identifying the paired pulses as events generated by the same pollutant through a common detection volume when the cross-correlation coefficient, amplitude relationship, polarity relationship, duration or waveform similarity of the paired pulses meet preset conditions. The statistical characteristics of particulate pollution events include at least two of the following: transmission drop amplitude, scattering response amplitude, pulse width, rise time, fall time, pulse integral area, event frequency, event interval, kurtosis, skewness, quantile, transmission-scattering amplitude ratio, transmission-scattering area ratio, and two-channel time delay. The events are divided into multiple equivalent particle size or severity intervals based on the amplitude or area distribution of particulate pollution events, and the abrasive pollution contribution is formed based on the number of events, cumulative area or weighted sum of each interval.

[0055] This method uses at least one of the following: rule model, statistical classifier, support vector machine, decision tree, random forest, gradient boosting tree, neural network, probabilistic graphical model or clustering model to distinguish abrasive events, bubble events, water droplet events, soot aggregation events and other abnormal events.

[0056] The contribution of aging gum contamination is determined by at least two of the following: transmission baseline, scattering baseline, transmission-scattering baseline ratio, baseline slope, baseline variability, change relative to new oil, and change relative to clean window status after event interval removal, interpolation, or model reconstruction.

[0057] A forward response model is established between the observed characteristics and the contributions of abrasive contamination and aging gum contamination. The contributions are then solved by inversion with non-negativity constraints, monotonicity constraints, mass conservation constraints, energy conservation constraints, sparsity constraints, smoothness constraints, or prior distribution constraints.

[0058] The forward response model should include at least the oil-based absorption term, particle extinction term, particle scattering term, colloidal absorption or scattering term, and window and environmental influence term, and allow each term to be a linear, piecewise linear or nonlinear function.

[0059] A calibration dataset was established using pure abrasive samples, pure aged adhesive samples, mixed samples of abrasive and adhesive in a known ratio, and interference samples containing water, air bubbles, dust, or soot.

[0060] The calibration dataset covers at least two particle sizes, at least two contamination concentrations, at least two temperature points, and at least one base oil or formulation oil. The model's generalization ability across oil types or across equipment is evaluated using leave-one-out-of-oil validation, leave-one-out-of-equipment validation, or domain adaptation methods.

[0061] Save the corresponding new oil baseline, model parameters and alarm thresholds for different oil types, equipment or operation stages, and update the model parameters according to oil change events, laboratory re-inspection results or equipment maintenance results.

[0062] Output abrasive contamination index, aging gum contamination index, and their equivalent extinction contribution rate or equivalent scattering contribution rate, and simultaneously output confidence interval, uncertainty, data validity flag or out-of-calibration range flag.

[0063] The auxiliary sensing signal includes at least one of inductive abrasive signal, magnetic field modulation response, temperature, flow rate, pressure, viscosity, dielectric constant, moisture or vibration signal; when the auxiliary sensing signal includes inductive abrasive signal, at least one of the amplitude, polarity, phase, pulse width or resonance parameter of inductive pulse is used to confirm or probability-correct whether the optical particle event belongs to metal abrasive.

[0064] The detection area of ​​the inductive abrasive sensor is located inside the optical detection area, at least partially overlaps with the optical detection area, or is located upstream or downstream of the optical detection area. When the two detection areas are separated from each other, the theoretical passage delay is determined based on the distance between the detection areas and the oil flow rate, and the optical particle event is paired with the inductive event within the matching window corresponding to the theoretical passage delay.

[0065] Theoretically, the time delay is determined by dividing the equivalent distance between detection areas along the flow direction by the oil flow velocity. When the flow velocity is unknown or fluctuates, the time delay is obtained by the cross-correlation peak of the continuous event sequence, dynamic time warping, probability data correlation, or adaptive state estimation.

[0066] Events exhibiting both transmission-scattering pairing responses and inductive pairing responses are identified as metal abrasive events; events exhibiting only transmission-scattering pairing responses are identified as candidates for non-metallic particles, bubbles, metal abrasive particles below the inductive detection limit, or unknown particles; and events exhibiting only inductive responses are identified as candidates for optical missed detection, electromagnetic interference, or window contamination.

[0067] Ferromagnetic and nonferromagnetic abrasive particles are distinguished based on the polarity, phase, frequency response, or impedance change direction of the inductive pulse, and the material category probability and transmission-scattering event characteristics are used together to calculate the contribution of abrasive particle contamination.

[0068] When the inductive abrasive signal is missing, the sensor fails, the signal-to-noise ratio is below the threshold, or the abrasive size is below the inductive detection limit, the transmission-scattering signal is used to perform degradation decoupling; when the inductive abrasive signal is restored to a valid state, the optical abrasive severity model or event classification model is corrected online using the matched metal abrasive events.

[0069] Example 2 Signal standardization. Let the original transmission, scattering, and reference channel voltages be VT, VS, and VR, respectively, and the dark levels be DT, DS, and DR, respectively. Then, Tn = (VT - DT) / (VR - DR) and Sn = (VS - DS) / (VR - DR). Based on the temperature θ, establish correction functions gT(θ) and gS(θ) to obtain Tc = Tn / gT(θ) and Sc = Sn / gS(θ). For devices without a reference channel, the light source drive current, periodic reference measurements, or a long-term drift model can be used as substitutes.

[0070] Example 3 Scale separation. Robust low-pass filtering with a cutoff frequency fc is used to obtain BT and BS, where PT = Tc - BT and PS = Sc - BS. fc is determined based on the flow velocity v, the effective beam width d, and the expected event duration τ ≈ d / v. To avoid mismatch between different flow velocities with a fixed fc, it can be adjusted in real time according to the flow sensor, or automatically selected using wavelet scale energy.

[0071] Example 4 Event detection. Signal-to-noise ratio (SNR) thresholds are set for the negative peak of PT and the positive or bipolar peak of PS, respectively. For candidate event i, the time center ti, amplitude Ai, full width at half maximum (FWHM) Wi, and integral area Qi are calculated. Among candidate pairs where |tT-tS| ≤ Δt, the pair with the highest cross-correlation is selected. Events satisfying a correlation coefficient r ≥ r0 and a width ratio within a predetermined interval are marked as granular events.

[0072] Example 5 Abrasive particle contribution calculation. Events are mapped to K severity intervals based on dual-channel features, Cw = Σk wk Nk / (Q·Δtobs), where Nk is the number of events in the k-th interval, wk is the calibration weight, Q is the flow rate, and Δtobs is the observation time. When no flow rate is measured, the event index per unit time can be output instead of the mass concentration. For cases with dense particles and overlapping pulses, sparse deconvolution or statistical fluctuation models can be used.

[0073] Example 6 Colloidal contribution calculation. The intervals corresponding to particulate events are removed from Tc and Sc and interpolated to obtain Tg and Sg. The extinction feature ET = -ln(Tg / Tnew) / L and the scattering feature ES = (Sg - Snew) / L are defined, where L is the optical path length, and Tnew and Snew are baseline values ​​for similar new oil. The colloidal index is obtained using Cv = h(ET,ES,θ,oil). h can be a calibration curve, constrained regression, or a small machine learning model.

[0074] Example 7 Joint inversion. The observation vector y includes ET, ES, event density, pulse amplitude quantile, and transmission-scattering ratio. The formula y = F(Cw,Cv,z) + ε is established, which is solved by minimizing ||W[yF(Cw,Cv,z)]||² + λR(Cw,Cv), where Cw ≥ 0, Cv ≥ 0, and R is a smoothing or prior regularization term. The uncertainty U is obtained through bootstrapping, Bayesian posterior, or model ensemble.

[0075] Example 8 Calibration samples. Abrasive samples can be standard iron powder, steel powder, copper powder, or abrasive particles collected from actual equipment, and graded according to particle size; colloidal samples can be controlled oxidation oil, on-site aged oil, samples reconstituted after membrane filtration, or representative aging products. Samples can be prepared as pure abrasive particles, pure colloidal particles, mixtures of both, and samples containing water and air bubbles. Laboratory reference values ​​can be provided by methods such as particle counting, ferrography, ICP, FTIR, acid value, or membrane colorimetry.

[0076] Example 9 Diagnostic output. When Cw is high and Cv is low, the output is "Mainly mechanical wear"; when Cw is low and Cv is high, the output is "Mainly oil aging"; when both are high, the output is "Combined anomaly"; when the uncertainty exceeds the threshold or the detection window is contaminated, the output is "Cleaning or laboratory retesting required". The above textual conclusions are not necessary limitations of this invention; the system may also output only the numerical value and confidence level.

[0077] Example 10 Scope of Application. This method can be applied to vehicles, wind power, mining, construction machinery, ships, rail transportation, aviation, compressors, pumps, turbines, industrial gearboxes, and hydraulic equipment. Regardless of whether the object being tested is called engine oil, gear oil, hydraulic oil, or transmission fluid, as long as the transmission and non-forward scattering of the same detection area are observed simultaneously and the contaminant contribution is separated according to this method, it falls within the scope of application of this invention.

[0078] Example 11 The overall flow of the core decoupling algorithm is as follows: The algorithm sequentially performs synchronous demodulation and normalization, robust baseline separation, dual-channel candidate event detection, transmission-scattering event pairing, bubble and other interference removal, abrasive particle contribution estimation, particle baseline contribution subtraction, colloid contribution estimation, joint physical constraint correction, and confidence calculation. The input includes at least the synchronously acquired raw transmission sequence VT[n] and raw scattering sequence VS[n]; optional inputs include the reference light sequence VR[n], temperature θ[n], flow rate q[n], and oil type identifier.

[0079] Example 12 Synchronous demodulation and normalization. When the detection light uses on-off modulation, the difference between the average bright period value and the average dark period value of each channel is calculated to obtain the transmission demodulation amount XT[n], the scattering demodulation amount XS[n], and the reference demodulation amount XR[n]. The normalized signal is calculated according to equations (1) and (2), respectively:

[0080]

[0081] Where DT, DS, and DR represent the dark levels of the corresponding channels, and ε is a positive number to prevent the denominator from being zero. When there is no reference optical channel, the light source drive current, periodic reference measurement, or light source drift model can be used to replace XR[n]. The temperature and flow correction functions are calibrated using the same type of new oil under different temperature and flow conditions, and the corrected signal is calculated according to equation (3):

[0082] Example 13 Robust baseline separation. The corrected transmitted and scattered signals are decomposed into slowly varying baseline, transient component, and noise, respectively, as expressed by equation (4):

[0083] Iterative weighted smoothing is preferred to obtain BT[n] and BS[n]. In the k-th iteration, the objective function shown in equation (5) is solved; the weights of the negative pulse points of the transmission channel and the positive pulse points of the scattering channel are reduced so that short-time particle events are not absorbed into the slowly varying baseline:

[0084] In the formula, y[n] is the signal to be decomposed, b[n] is the baseline to be determined, wn is the robust weight, λ is the smoothing parameter, and Δ² represents the second-order difference. λ is determined based on the flow velocity, the effective width of the detection beam, and the typical particle passage time; equivalent separation can also be achieved by wavelet decomposition, empirical mode decomposition, Kalman filtering, or dual time constant filtering.

[0085] Example 14 Candidate event detection and pairing. Transient noise is estimated using median absolute deviation, and σT and σS are calculated according to equation (6); transmission candidate events satisfy PT[n]<-kTσT, and scattering candidate events satisfy |PS[n]|>kSσS, where kT and kS are determined by the noise level or calibration samples:

[0086] Adjacent threshold sampling points are merged into candidate events, and the peak time, amplitude, half-width at half-maximum (FWHM), integral area, rise time, and fall time are extracted. For each transmission candidate event, scattering candidate events are searched within a preset time window, and the pairwise score is calculated according to equation (7):

[0087] Where corr represents local cross-correlation, Δt represents the peak time difference between the two channels, WT and WS represent pulse widths, R(A,Q) represents the consistency of amplitude ratio and area ratio with the calibration range, and w1 to w4 are non-negative weights. The candidate pair with the highest score that exceeds the threshold M0 is selected as the particulate event generated by the same pollutant through a common detection volume.

[0088] Example 15 Interference removal. A feature vector xi is constructed from paired events, including at least the transmission amplitude AT, scattering amplitude AS, pulse widths WT and WS, integral areas QT and QS, amplitude ratio, area ratio, two-channel time difference, correlation coefficient, local baseline, temperature, and flow rate. Thresholding rules, Mahalanobis distance, random forests, gradient boosting trees, support vector machines, or neural networks can be used to obtain the probability pi that an event belongs to abrasive particles. Events with abnormally wide pulses, saturated pulses, amplitude ratios exceeding the abrasive particle calibration range, or events synchronized with sudden changes in flow rate are marked as bubbles, droplets, or unknown interference candidates.

[0089] Example 16 Abrasive contribution estimation. For events with abrasive probability not lower than a set threshold, calculate the non-negative severity h(xi), and form the abrasive contribution per unit sampling volume according to equation (8):

[0090] Where Vobs is the volume of oil flowing through the detection area during the observation period; when there is no flow rate information, the observation time is used instead of Vobs and the event index per unit time is output. The severity function h can be a non-negative linear model, monotonic regression, or a machine learning model with monotonic constraints. One possible implementation is shown in equation (9):

[0091] Among them, β0 to β5 are non-negative coefficients obtained from abrasive samples with known particle size or known contamination level. In addition to cumulative contribution, the event rate, maximum event severity, 95th percentile severity, and outbreak index can also be output.

[0092] Example 17 Particle baseline contribution subtraction. For baseline shifts caused by high concentration of particles or event overlap, an event statistics vector zw is first constructed, which includes at least the event rate, cumulative QT, cumulative QS, amplitude quantile, and pulse overlap rate. A particle baseline model fw is established using pure abrasive particles or calibration samples dominated by abrasive particles, and the candidate baseline for colloidal particles is obtained according to equation (10):

[0093] The preferred particle baseline model is a non-negative monotonic spline, non-negative regression, or gradient boosting model with monotonic constraints, so that the predicted particle extinction contribution and scattering contribution do not decrease when the number of events or the cumulative pulse area increases.

[0094] Example 18 Estimation of gum contribution. The equivalent gum extinction characteristic Eg and equivalent gum scattering characteristic Gg are calculated relative to similar new oil baselines Tnew and Snew according to Equation (11):

[0095] Where L represents the detection optical path. Eg, Gg, Eg / Gg, baseline change rate, temperature, and oil type identifier are input into a non-negative monotonic model fv to obtain the candidate contribution Cv,base of the gum. The model can be a non-negative linear regression, monotonic splines, a gradient boosting tree with monotonic constraints, or a monotonic neural network. When a reliable mass concentration reference is not available, the output is an equivalent gum aging index instead of a mass percentage.

[0096] Example 19 Joint physical constraint correction. To reduce the staged estimation error, an observation vector y and a forward response model F(Cw,Cv,z) are constructed, where z includes temperature, flow rate, oil type, and window state. The final abrasive contribution is obtained by solving equation (12). w and collagen contribution v:

[0097] Where W is the feature weight matrix, Cw,event is the abrasive candidate contribution obtained from the event model, Cv,base is the gum candidate contribution obtained from the baseline model, and λw and λv are regularization coefficients. The forward response model includes at least a base oil absorption term, an abrasive extinction / scattering term, a gum absorption / scattering term, and a window environment term, and can be superimposed with a nonlinear residual model obtained by machine learning.

[0098] The final output equivalent optical contribution rate is calculated according to equation (13):

[0099] The contribution rate represents the equivalent optical pollution contribution defined by the model used, and is not directly equivalent to the mass percentage of the two types of pollutants.

[0100] Example 20 Confidence and superdomain judgment. The data quality index (QI) comprehensively references optical stability, channel saturation rate, effective event pairing rate, bubble ratio, window contamination score, and flow stability. The distance of the sample relative to the calibration data distribution is denoted as Ddomain, and the model integration prediction variance or inversion posterior variance is denoted as Umodel. The confidence level can be calculated according to equation (14):

[0101] When the confidence level is lower than the set threshold, the channel becomes saturated, the window contamination score exceeds the threshold, or the input exceeds the calibration range, the system outputs a trend value, a low confidence level indicator, or a laboratory retest prompt, instead of outputting a definitive quantitative conclusion.

[0102] Example 21 Before specific equipment experimental data is available, the above thresholds, weights, and model parameters can be determined by subsequent calibration. The instruction manual does not pre-set specific accuracy rates or detection errors. The preferred combination for initial implementation is: robust smoothing baseline, adaptive threshold based on median absolute deviation, dual-channel maximum score pairing, interference identification combining rules and tree models, non-negative abrasive regression, monotonic colloid regression, and joint correction of two-dimensional non-negative constraints.

[0103] Example 22 Inductive detection structure. The inductive abrasive sensor employs two anti-parallel series excitation / detection coils and one compensation coil, or a three-coil differential structure; the oil flow path sequentially passes through the optical detection area and the inductive detection area. The inductive detection circuit can output pulse amplitude AI, pulse width WI, phase φI, polarity signI, and resonant frequency shift ΔfI. Ferromagnetic particles and non-ferromagnetic conductive particles exhibit different responses in the direction of polarity, phase, or impedance change; specific classification thresholds are calibrated using standard iron, steel, copper, or aluminum particles.

[0104] Example 23 Spatiotemporal registration. When the equivalent distance along the flow direction between the center of the optical detection region and the center of the inductive detection region is d, and the local average flow velocity is v, the theoretical value of the same particle is determined by time delay according to equation (15):

[0105] An allowable window δ is set near the theoretical time delay. If the optical event center time tO and the inductor event center time tI satisfy equation (16), then the two are listed as candidate pairs:

[0106] When the flow rate cannot be directly measured, the cross-correlation between the optical event sequence and the inductive event sequence can be calculated in a continuous event window, and d / v can be estimated by the time delay corresponding to the correlation peak. Alternatively, dynamic time warping, probabilistic data correlation, or Kalman filtering can be used to track the time delay that varies with the flow rate.

[0107] Example 24 Three-source event fusion. For the optical event feature xO and the inductive event feature xI, the probability pm of metal abrasive particles is calculated according to equation (17):

[0108] fmetal can be derived from logistic regression, Naive Bayes, random forest, gradient boosting tree, neural network, or rule-based probability mapping. xO includes at least transmission amplitude, scattering amplitude, pulse width, area, and correlation coefficient; xI includes at least inductance amplitude, polarity, phase, and pulse width; rOI represents the matching score after time delay compensation.

[0109] The abrasive contribution after the fusion of three sources is calculated according to formula (18):

[0110] The hm value provides the severity of an event based on both optical and inductive characteristics. Events with strong inductive confirmation are given higher weights; events with only optical responses but whose size may be below the inductive detection limit are given non-zero weights to avoid systematic undercounting caused by the lower limit of inductive detection.

[0111] Example 25 Online correction and degradation. The system uses inductor confirmation events with high paired scores as weak labels to periodically update the optical event classifier or severity model. If the inductor channel experiences saturation, disconnection, excessive noise, or no reasonable events for a predetermined period, the online update is frozen, the inductor channel is marked as invalid, and it degrades to a transmission-scattering dual-source decoupling. After recovery, the time delay and amplitude normalization parameters are re-estimated through a stable observation window.

[0112] Example 26 Multi-source decision output. The system outputs the metal abrasive index, non-metallic particle candidate index, adhesive aging index, ferromagnetic / non-ferromagnetic class probability, and fusion confidence score separately. When optical and inductive conclusions conflict, a definitive class is not forced; instead, a low confidence score or re-examination prompt is output based on the window state, detection limit, saturation state, and model domain distance.

[0113] Example 27 This embodiment also provides a lubrication equipment condition diagnosis method, which executes the lubricating oil abrasive and gum contamination decoupling detection method based on transmission-scattering multi-source signal fusion in Embodiment 1 above, and outputs mechanical wear inspection suggestions when the contribution of abrasive contamination exceeds a first threshold, outputs oil change, purification or oil temperature inspection suggestions when the contribution of aged gum contamination exceeds a second threshold, and outputs composite abnormality suggestions when both exceed the corresponding thresholds.

[0114] Example 28 This embodiment also provides an electronic device, including a processor and a memory. The memory stores a program, and when the program is executed by the processor, it implements the decoupling detection method for lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion in Embodiment 1 above.

[0115] Example 29 This embodiment also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the decoupling detection method for lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion in Embodiment 1 above.

[0116] The thresholds, models, filters, and learners mentioned above can all be adjusted by those skilled in the art according to the equipment and oil type. Achieving the same event-baseline separation and dual-contamination contribution output using equivalent time-frequency analysis, probabilistic inference, or physical inversion methods does not depart from the concept of this invention.

[0117] The above embodiments are merely illustrative and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion, characterized in that: Includes the following steps: Acquire transmission time-series signals and non-forward scattering time-series signals generated from the same oil detection area and collected based on a common time reference; Preprocessing is performed on the transmission timing signal and the scattering timing signal; The preprocessed signal is decomposed into a baseline component characterizing a continuous or slowly changing response and a transient event component characterizing a contaminant crossing the detection area. Identifying particulate pollution events based on the temporal correspondence and / or correlation between transmission transient events and scattering transient events; Candidate contributions for abrasive contamination were determined based on the statistical characteristics of particulate contamination events. The impact of particulate contamination events is compensated or subtracted from the baseline components, and the contribution of aged colloid contamination is determined based on the compensated transmission baseline characteristics and scattering baseline characteristics. The transmitted signal, the scattered signal, and optional auxiliary sensing signals are fused at the feature level, the event level, or the decision level to correct the confidence of the abrasive contamination candidate contribution, the aging gum contamination contribution, or both; and output at least two of the abrasive contamination index and the aging gum contamination index, or output the relative contribution of the two.

2. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 1, characterized in that: Preprocessing includes at least one of the following: dark level subtraction, reference light normalization, ambient light suppression, synchronization demodulation, noise reduction, gain correction, temperature compensation, flow compensation, pressure compensation, viscosity compensation, window contamination compensation, and new oil baseline normalization.

3. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 1, characterized in that: Signal decomposition employs low-pass and high-pass filtering, band-pass filtering, moving average, local polynomial fitting, wavelet decomposition, empirical mode decomposition, variational mode decomposition, Kalman filtering, robust trend estimation, or combinations thereof.

4. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 1, characterized in that: The identification of particulate pollution events includes: detecting candidate pulses in the transmission transient event component and the scattering transient event component respectively; pairing candidate pulses within a preset time window; and identifying the paired pulses as events generated by the same pollutant through a common detection volume when the cross-correlation coefficient, amplitude relationship, polarity relationship, duration or waveform similarity of the paired pulses meet preset conditions. The statistical characteristics of particulate pollution events include at least two of the following: transmission drop amplitude, scattering response amplitude, pulse width, rise time, fall time, pulse integral area, event frequency, event interval, kurtosis, skewness, quantile, transmission-scattering amplitude ratio, transmission-scattering area ratio, and two-channel time delay. The events are divided into multiple equivalent particle size or severity intervals based on the amplitude or area distribution of particulate pollution events, and the abrasive pollution contribution is formed based on the number of events, cumulative area or weighted sum of each interval.

5. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 1, characterized in that: The contribution of aging gum contamination is determined by at least two of the following: transmission baseline, scattering baseline, transmission-scattering baseline ratio, baseline slope, baseline variability, change relative to new oil, and change relative to clean window status after event interval removal, interpolation, or model reconstruction.

6. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 1, characterized in that: The auxiliary sensing signal includes at least one of inductive abrasive signal, magnetic field modulation response, temperature, flow rate, pressure, viscosity, dielectric constant, moisture or vibration signal; when the auxiliary sensing signal includes inductive abrasive signal, at least one of the amplitude, polarity, phase, pulse width or resonance parameter of inductive pulse is used to confirm or probability-correct whether the optical particle event belongs to metal abrasive.

7. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 6, characterized in that: Ferromagnetic and nonferromagnetic abrasive particles are distinguished based on the polarity, phase, frequency response, or impedance change direction of the inductive pulse, and the material category probability and transmission-scattering event characteristics are used together to calculate the contribution of abrasive particle contamination.

8. The method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion according to claim 6, characterized in that: When the inductive abrasive signal is missing, the sensor fails, the signal-to-noise ratio is below the threshold, or the abrasive size is below the inductive detection limit, the transmission-scattering signal is used to perform degradation decoupling; when the inductive abrasive signal is restored to a valid state, the optical abrasive severity model or event classification model is corrected online using the matched metal abrasive events.

9. An electronic device comprising a processor and a memory, characterized in that: The memory stores a program, which, when executed by the processor, implements the method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the method for decoupling detection of lubricating oil abrasive particles and gum contamination based on transmission-scattering multi-source signal fusion as described in any one of claims 1-8.

Images