A method and system for detecting the degree of deterioration of lubricating oil for a gear

Abstract

The application discloses a kind of degradation degree detection method and system for lubricating fluid of gear, it is related to lubricating fluid detection technical field, the system includes: data acquisition module, data processing module, collaborative fusion module, analog prediction module and decision feedback module.The application, two-way collaborative attention interaction mechanism is designed, and the three modal characteristic vectors of chemistry, physics and abrasive particle are calculated in two-way attention between each other, form complete cross-modal collaborative feature, can effectively mine the implicit association and complementary information between different modal data, actively model the asymmetric relationship between modal, so that the feature after fusion not only contains independent information of each modal, also carries cross-modal collaborative knowledge, to provide more comprehensive, more discriminative comprehensive state representation for subsequent prediction.

Description

Technical Field

[0001] This invention relates to the field of lubricant testing technology, and in particular to a method and system for detecting the deterioration of lubricant used in gears. Background Technology

[0002] The gearbox is the core equipment of the unit, and lubricating oil is the "blood" of the gearbox, playing a vital role in maintaining its normal operation. The performance and quality of lubricating oil have become key factors affecting the safe and reliable operation of the unit and extending its service life. At the same time, as a carrier, lubricating oil contains rich information on the tribological state of the moving parts of the equipment, which is an important information source for gearbox fault diagnosis and health management. In recent years, online oil monitoring has become an important means of monitoring wind turbine gearbox faults.

[0003] Currently, in existing technologies, multimodal data of lubricants are often analyzed in isolation or simply spliced ​​together during the testing process, ignoring the inherent coupling and co-evolutionary relationship between the modes and lacking a hierarchical and interactive fusion mechanism. This results in the incomplete utilization of cross-modal complementary information, making it difficult to comprehensively characterize the true deterioration state of the lubricant.

[0004] Therefore, a method and system for detecting the deterioration of lubricating fluid for gears is proposed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for detecting the deterioration of lubricating fluid for gears, in order to solve the problems mentioned in the background above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a lubricating fluid deterioration detection system for gears, the system comprising: a data acquisition module, a data processing module, a collaborative fusion module, a simulation prediction module, and a decision feedback module;

[0007] The data acquisition module is used to acquire multimodal data of the lubricating fluid through a multimodal sensor;

[0008] The data processing module is used to preprocess and extract features from the multimodal raw data to obtain the chemical degradation index, abrasive degradation index and physical state vector;

[0009] The collaborative fusion module is used to perform hierarchical collaborative tensor fusion of chemical degradation index, abrasive degradation index and physical state vector. The hierarchical collaborative tensor fusion includes: extracting the internal temporal dependencies of each mode, performing bidirectional collaborative attention interaction, and adaptively controlling the fusion contribution ratio through learnable gating vectors to output a comprehensive state vector.

[0010] The simulation prediction module is used to input the comprehensive state vector into the digital twin model for dynamic simulation, and output the remaining effective life and comprehensive degradation index through hybrid intelligent prediction.

[0011] The decision feedback module is used to generate and output maintenance recommendations based on the remaining effective life and the comprehensive degradation index.

[0012] Preferably, the data acquisition module includes a physical acquisition unit, a chemical acquisition unit, and an abrasive particle acquisition unit;

[0013] The physical acquisition unit is used to acquire the physical parameters of the lubricating fluid in real time through a MEMS sensor array, including viscosity, dielectric constant, moisture content, and temperature.

[0014] The chemical acquisition unit is used to acquire the full-band Raman spectrum of lubricating oil using a Raman spectrometer;

[0015] The abrasive particle acquisition unit is used to emit micro-focus X-rays to penetrate the lubricating fluid flow. Based on the difference in the attenuation of X-rays during the penetration process, it statistically analyzes the total macroscopic concentration and large particle distribution data of abrasive particles in the lubricating fluid, and directly captures microscopic morphology images of abrasive particles in the lubricating fluid through a high-resolution CCD optical system. The microscopic morphology images include data on the size, number, shape, and transmittance of the abrasive particles.

[0016] Preferably, the data processing module includes a preprocessing unit, a chemical degradation unit, an abrasive degradation unit, and a physical state unit;

[0017] The preprocessing unit is used to perform noise reduction, temporal alignment, normalization, and image enhancement on the data acquired in the data acquisition module.

[0018] Preferably, the chemical degradation unit calculates the skeleton isomerization parameters based on the peak intensity ratio of the CC stretching vibration peak and the CH2 torsion peak in the Raman spectrum, and calculates the fluorescence and conjugation indicator parameters based on the fluorescence background intensity and the integral intensity of the conjugated double bond characteristic peak in the Raman spectrum. The two parameters are used to construct a two-dimensional coordinate point, which is compared with a pre-established degradation trajectory database to calculate the Mahalanobis distance between the current point and the standard new oil point and the failure point, thus obtaining the chemical degradation index.

[0019] The abrasive degradation unit classifies abrasive morphology based on the pre-processed abrasive image and inputs it into a pre-trained convolutional neural network, outputting the statistical quantity of each type of abrasive. The classification includes normal sliding abrasive, cutting abrasive, fatigue spalling block, and oxide cluster. The grayscale image acquired by the micro-focal X-ray imaging unit is thresholded, and the total pixel area of ​​the abrasive is counted. This is then converted into abrasive concentration by combining the calibration curve, and the proportion of large particles is calculated. The normalized abrasive concentration, the proportion of large particles, the proportion of cutting abrasive, and the proportion of fatigue spalling block are input into a weighted fusion formula to obtain the abrasive degradation index.

[0020] The physical state unit directly combines the pre-processed and time-aligned viscosity value, dielectric constant value, moisture content, and temperature value into a four-dimensional physical state vector. Each dimension is normalized to its maximum and minimum values, and the normalization range is mapped to [0,1] or [-1,1] to obtain the physical state vector.

[0021] Preferably, the collaborative fusion module includes a first fusion unit, a second fusion unit, and a third fusion unit;

[0022] The first fusion unit inputs the time series of chemical degradation index, the time series of abrasive degradation index and the time series of physical state vector into three independent modal self-attention encoders. Each encoder adopts a multi-head self-attention mechanism, with 4 to 8 attention heads and 2 to 4 encoder layers.

[0023] Each encoder first generates a query matrix, a key matrix, and a value matrix through a linear transformation, then calculates the self-attention weights and sums them by weight, then processes them through residual connections and layer normalization, and finally obtains the intra-modal feature encoding vectors of the three modalities through a feedforward neural network.

[0024] The historical window length of the time series is 10 to 50 consecutive sampling points, and the time interval between each sampling point is 1 to 10 seconds.

[0025] Furthermore, each modal self-attention encoder incorporates a learnable position embedding vector to preserve temporal sequence information.

[0026] Preferably, the second fusion unit is used for multiple intramodal feature encoding vectors to perform bidirectional collaborative attention interactions between pairs;

[0027] For chemical and physical modes, the feature vectors of the chemical mode are used as queries and the feature vectors of the physical mode are used as keys and values ​​to calculate the chemical-to-physical collaborative attention feature vectors.

[0028] Simultaneously, using the feature vector of the physical mode as the query and the feature vector of the chemical mode as the key and value, the collaborative attention feature vector from physics to chemistry is calculated;

[0029] By concatenating the collaborative attention feature vectors from these two directions, we obtain the chemical-physical cross-modal collaborative features.

[0030] Using the exact same method, the cross-modal synergistic characteristics of chemical and abrasive particles and the cross-modal synergistic characteristics of physical and abrasive particles were calculated separately.

[0031] Scaled dot product attention is used in all attention calculations, and a dropout rate of 0.1 to 0.3 is set to prevent overfitting.

[0032] Preferably, the third fusion unit performs adaptive fusion based on three intramodal feature encoding vectors and three cross-modal collaborative feature vectors;

[0033] The three intramodal feature encoding vectors and the three cross-modal collaborative feature vectors are concatenated and input into a small gating network;

[0034] The small gated network is a fully connected neural network with one hidden layer. The number of neurons in the hidden layer is 16 to 64, the output activation function is the sigmoid function, and the output dimension is equal to the total number of channels of the feature vector to be fused.

[0035] The gating network outputs the fusion weights for each feature channel in real time, with the weights ranging from 0 to 1.

[0036] Finally, all feature vectors are weighted and summed according to their channel dimensions to output a high-dimensional collaborative tensor as the comprehensive state vector.

[0037] Preferably, the simulation prediction module includes a model building unit and an intelligent early warning unit;

[0038] The model building unit constructs a lubricant state evolution model based on the oxidation kinetics equation and the wear particle generation model, using the comprehensive state vector as the model input;

[0039] The consumption rate of additives in the lubricating fluid, the growth rate of the total acid value, and the accumulation rate of abrasive particles are described by differential equations.

[0040] Using real-time acquired multimodal data, Kalman filtering or recursive least squares method is used to correct the model parameters in the differential equation online;

[0041] When a sudden change in cross-modal cooperative features is detected in the integrated state vector, the frequency is automatically corrected and the state prediction step size is adjusted.

[0042] Output the physical evolution trend sequence, including the total acid value, viscosity change and predicted abrasive concentration within a predetermined time period;

[0043] The intelligent early warning unit inputs the historical comprehensive state vector sequence into the temporal prediction model of the long short-term memory network, learns the temporal dependencies between the comprehensive state vectors, and outputs data-driven trend prediction results.

[0044] The physical evolution trend sequence and the trend prediction results are fused using Bayesian methods to obtain the fused state distribution.

[0045] Based on the preset failure threshold, Monte Carlo sampling is performed on the fused state distribution to calculate the probability distribution of the remaining effective lifetime. The median of the distribution is taken as the estimated value of the remaining effective lifetime, and a confidence interval with a confidence level of 90% or 95% is given.

[0046] The Euclidean distance between the current comprehensive state vector and the preset failure threshold vector is then normalized to obtain the basic degradation degree. The trend correction coefficient is calculated based on the ratio of the remaining effective life to the expected total life. The basic degradation degree and the trend correction coefficient are then weighted and fused to obtain the comprehensive degradation degree index.

[0047] Preferably, the decision feedback unit includes a hierarchical maintenance unit and a visualization unit;

[0048] The graded maintenance unit classifies the current lubricant based on its remaining effective life and comprehensive deterioration index.

[0049] Set a first threshold, a second threshold, and a third threshold;

[0050] When the overall degradation index is less than the first threshold and the remaining effective lifespan is greater than the preset first time threshold, it is judged as a low-risk level, output continues to run, and it is recommended to perform a routine check once a week.

[0051] When the comprehensive deterioration index is greater than or equal to the first threshold and less than the second threshold, or the remaining effective life is less than or equal to the first time threshold and greater than the second time threshold, it is judged as a medium risk level, and a signal is output to perform lubricant sampling and testing at the next shutdown and prepare for oil change plan.

[0052] When the comprehensive deterioration index is greater than or equal to the second threshold and less than the third threshold, or when the remaining effective life is less than or equal to the second time threshold and greater than the third time threshold, it is judged as a high-risk level; output a signal to arrange an oil change as soon as possible and to check the wear condition of the gearbox.

[0053] When the comprehensive deterioration index is greater than or equal to the third threshold, or the remaining effective life is less than or equal to the third time threshold, it is judged as a serious risk level; output a signal to immediately stop the machine, change the oil, and check the wear of gears and bearings.

[0054] The visualization unit displays the remaining effective lifespan, overall degradation index, risk level, and maintenance recommendations on the remote monitoring terminal.

[0055] A method for detecting the deterioration degree of lubricating fluid for gears includes the following steps:

[0056] S1: Collect multimodal data of lubricating fluid through a multimodal sensor, and perform preprocessing and feature extraction to obtain intramodal feature encoding vectors for the three modes;

[0057] S2: The historical time series of the three modalities are sequentially processed by self-attention to extract temporal dependencies, bidirectional collaborative attention to perform cross-modal interaction, and then adaptively fused using a gating network to output a comprehensive state vector;

[0058] S3: Based on the prediction of the digital twin model and the time series prediction model respectively, the remaining effective lifetime and the comprehensive degradation index are obtained by Bayesian fusion.

[0059] S4: Based on the remaining lifespan and degradation index compared with the threshold, four risk levels are classified and corresponding maintenance suggestions are given, which are displayed on the remote terminal.

[0060] The present invention has the following beneficial effects:

[0061] 1. In this invention, the time series of chemical degradation index, abrasive degradation index and physical state vector are respectively input into an independent modal self-attention encoder through the first fusion unit. This can effectively extract the long-term temporal dependencies within each mode, fully preserve the evolution rhythm and trend information of different physicochemical phenomena during lubricant degradation, and enable subsequent cross-modal interactions to be based on more realistic temporal characteristics, significantly improving the system's comprehensive identification ability for slow degradation and sudden anomalies.

[0062] 2. In this invention, a bidirectional collaborative attention interaction mechanism is designed to perform bidirectional attention calculations on the feature vectors of the chemical, physical, and abrasive modalities in pairs, forming complete cross-modal collaborative features. This mechanism can effectively mine the implicit correlations and complementary information between different modal data, actively model the asymmetric relationships between modalities, and ensure that the fused features not only contain the independent information of each modality, but also carry cross-modal collaborative knowledge, thereby providing a more comprehensive and discriminative integrated state representation for subsequent predictions.

[0063] 3. In this invention, all intramodal feature encoding vectors are concatenated with three cross-modal collaborative feature vectors and then input into a fully connected neural network with a single hidden layer. The fusion weights of each feature channel are output in real time through the Sigmoid activation function. This allows for dynamic adjustment of the contribution ratio of each feature channel based on the current operating conditions and data quality, thereby suppressing the introduction of interference information and greatly improving the robustness and generalization ability of the system under complex operating conditions. This establishes a more stable and reliable feature foundation for remaining effective lifetime prediction and degradation assessment. Attached Figure Description

[0064] Figure 1 This is a flowchart of a method for detecting the deterioration of lubricating fluid for gears according to the present invention. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] Please see Figure 1 The present invention provides a technical solution: a lubricant deterioration detection system for gears, the system comprising: a data acquisition module, a data processing module, a collaborative fusion module, a simulation prediction module and a decision feedback module;

[0067] The data acquisition module is used to acquire multimodal data of the lubricating fluid through a multimodal sensor;

[0068] The data processing module is used to preprocess and extract features from the multimodal raw data to obtain the chemical degradation index, abrasive degradation index, and physical state vector.

[0069] The collaborative fusion module is used to perform hierarchical collaborative tensor fusion of chemical degradation index, abrasive degradation index and physical state vector. The hierarchical collaborative tensor fusion includes: extracting the internal temporal dependencies of each mode, performing bidirectional collaborative attention interaction, and adaptively controlling the fusion contribution ratio through learnable gating vectors to output a comprehensive state vector.

[0070] The simulation prediction module is used to input the comprehensive state vector into the digital twin model for dynamic simulation, and output the remaining effective life and comprehensive degradation index through hybrid intelligent prediction.

[0071] The decision feedback module is used to generate and output maintenance recommendations based on the remaining effective life and the comprehensive degradation index.

[0072] The data acquisition module includes a physical acquisition unit, a chemical acquisition unit, and an abrasive particle acquisition unit;

[0073] The physical acquisition unit is used to acquire the physical parameters of the lubricating fluid in real time through a MEMS sensor array, including viscosity, dielectric constant, moisture content, and temperature.

[0074] The chemical acquisition unit is used to acquire the full-band Raman spectrum of lubricating oil using a Raman spectrometer;

[0075] The abrasive particle acquisition unit is used to emit micro-focus X-rays to penetrate the lubricating fluid flow. Based on the difference in the attenuation of X-rays during the penetration process, it statistically analyzes the total macroscopic concentration and large particle distribution data of abrasive particles in the lubricating fluid. It also directly captures microscopic morphology images of abrasive particles in the lubricating fluid through a high-resolution CCD optical system. The microscopic morphology images include data on the size, number, shape, and transmittance of the abrasive particles.

[0076] The data processing module includes a preprocessing unit, a chemical degradation unit, an abrasive degradation unit, and a physical state unit;

[0077] The preprocessing unit is used to perform noise reduction, time alignment, normalization, and image enhancement on the data acquired in the data acquisition module.

[0078] The chemical degradation unit calculates the skeleton isomerization parameters based on the peak intensity ratio of the CC stretching vibration peak and the CH2 torsion peak in the Raman spectrum, and calculates the fluorescence and conjugation indicator parameters based on the fluorescence background intensity and the integral intensity of the conjugated double bond characteristic peak in the Raman spectrum. The two parameters are used to construct two-dimensional coordinate points, which are compared with the pre-established degradation trajectory database to calculate the Mahalanobis distance between the current point and the standard new oil point and the failure point, thus obtaining the chemical degradation index.

[0079] Among them, two-dimensional Raman spectral feature points of different deterioration stages were obtained by accelerated aging test, the standard deterioration trajectory curve from new oil point to failure point was fitted, and the covariance matrix was calculated to establish a deterioration trajectory database.

[0080] Specifically, the formula for calculating the skeleton isomerization parameter is as follows:

[0081] ;

[0082] In the formula, Represents the skeleton isomerization parameter. Represents the intensity of the C-S stretching vibration peak in the Raman spectrum. The integral intensity represents the CH2 torsion peak in the Raman spectrum;

[0083] The formulas for calculating fluorescence and conjugate indicator parameters are as follows:

[0084] ;

[0085] In the formula, Represents fluorescence and conjugation indicator parameters. The fluorescence background intensity represents the Raman spectrum. The integral intensity of the characteristic peak of the conjugated double bond;

[0086] Two-dimensional coordinate points are represented as: ;

[0087] The formula for calculating Mahalanobis distance is:

[0088] ;

[0089] In the formula, The mean vector of standard new oil in two-dimensional coordinates. Represents the covariance matrix in the deterioration trajectory database. Represents Mahalanobis distance;

[0090] The chemical degradation index is obtained by normalization:

[0091] ;

[0092] In the formula, The chemical degradation index is represented by the value of [0,1][0,1]. When the oil sample is new, the numerator is 0 and the index is 0. When the oil sample deteriorates to the point of failure, the numerator equals the denominator and the index is 1. Values ​​between the two reflect the relative position of the current oil sample on the degradation trajectory. The larger the value, the more severe the chemical degradation. Represents the coordinates of the oil failure point. Represents the coordinates of the new oil;

[0093] The abrasive degradation unit classifies abrasive morphology based on the pre-processed abrasive image and inputs it into a pre-trained convolutional neural network. It outputs the statistical quantity of each type of abrasive, including normal sliding abrasive, cutting abrasive, fatigue spalling block, and oxide cluster. The grayscale image acquired by the micro-focal X-ray imaging unit is thresholded, and the total pixel area of ​​the abrasive is counted. It is then converted into abrasive concentration by combining the calibration curve and the proportion of large particles is calculated. The normalized abrasive concentration, the proportion of large particles, the proportion of cutting abrasive, and the proportion of fatigue spalling block are input into a weighted fusion formula to obtain the abrasive degradation index.

[0094] The training set for the convolutional neural network includes: oil samples collected during laboratory gearbox bench tests, oil samples from gearboxes in operation in the field, and publicly available abrasive particle pattern databases;

[0095] Specifically, abrasive particle concentration normalization The calculation formula is:

[0096] ;

[0097] In the formula, Represents the current total concentration of abrasive particles. and These represent the maximum and minimum concentration values ​​corresponding to the calibration curve, respectively.

[0098] Large particle ratio The calculation formula is:

[0099] ;

[0100] In the formula, This represents the total area of ​​abrasive pixels with a diameter greater than 10 micrometers. Represents the total area of ​​all abrasive pixels;

[0101] The formulas for calculating the proportion of cutting abrasive particles and the proportion of fatigue spalling are as follows:

[0102] ;

[0103] ;

[0104] In the formula, This represents the number of abrasive grains in the cutting process. This represents the number of fatigue spalling fragments. It represents the sum of normal sliding abrasive grains, cutting abrasive grains, fatigue spalling blocks, and oxide clusters;

[0105] Each weight in the weighted fusion formula is determined by expert experience or multiple linear regression.

[0106] The physical state unit directly combines the pre-processed and time-aligned viscosity value, dielectric constant value, moisture content, and temperature value into a four-dimensional physical state vector. Each dimension is normalized to its maximum and minimum values, and the normalization range is mapped to [0,1] or [-1,1] to obtain the physical state vector.

[0107] The collaborative fusion module includes a first fusion unit, a second fusion unit, and a third fusion unit;

[0108] The first fusion unit inputs the time series of chemical degradation index, abrasive degradation index and physical state vector into three independent modal self-attention encoders respectively. Each encoder adopts a multi-head self-attention mechanism, with 4 to 8 attention heads and 2 to 4 encoder layers.

[0109] Each encoder first generates a query matrix, a key matrix, and a value matrix through a linear transformation, then calculates the self-attention weights and sums them by weight, then processes them through residual connections and layer normalization, and finally obtains the intra-modal feature encoding vectors of the three modalities through a feedforward neural network.

[0110] The historical window length for the time series is 10 to 50 consecutive sampling points, and the time interval between each sampling point is 1 to 10 seconds.

[0111] Furthermore, each modal self-attention encoder incorporates a learnable position embedding vector to preserve temporal sequence information.

[0112] Specifically, positional embedding is represented as:

[0113] ;

[0114] In the formula, Represents the learnable position embedding matrix. Represents the original input sequence matrix. This represents the output matrix after adding location information;

[0115] The linear transformation that generates the query matrix, key matrix, and value matrix is ​​specifically represented as follows:

[0116] ;

[0117] In the formula, , and These represent learnable linear transformation weight matrices, used to generate queries, keys, and values, respectively.

[0118] The self-attention weights are represented as follows:

[0119] ;

[0120] In the formula, This represents the query matrix, key matrix, and value matrix obtained after linear transformation. The dimension representing the query / key, i.e., the dimension of each attention head. This represents the product of the query matrix and the transpose of the key matrix. This represents performing softmax normalization on each row of the matrix, so that the sum of the attention weights of each position to all positions is 1. Represents the output of the attention mechanism;

[0121] Residual connectivity and layer normalization are represented as follows:

[0122] ;

[0123] In the formula, Represents a multi-head attention function. This represents the intermediate result after residual connection and first layer normalization;

[0124] The feedforward network is represented as:

[0125] ;

[0126] In the formula, This represents the intra-modal feature encoding vector of the final output. The representative layer normalization operation normalizes the feature dimensions of each sample. Represents a feedforward neural network;

[0127] The second fusion unit is used for bidirectional collaborative attention interaction between multiple intramodal feature encoding vectors.

[0128] For chemical and physical modes, the feature vectors of the chemical mode are used as queries and the feature vectors of the physical mode are used as keys and values ​​to calculate the chemical-to-physical collaborative attention feature vectors.

[0129] Simultaneously, using the feature vector of the physical mode as the query and the feature vector of the chemical mode as the key and value, the collaborative attention feature vector from physics to chemistry is calculated;

[0130] By concatenating the collaborative attention feature vectors from these two directions, we obtain the chemical-physical cross-modal collaborative features.

[0131] Using the exact same method, the cross-modal synergistic characteristics of chemical and abrasive particles and the cross-modal synergistic characteristics of physical and abrasive particles were calculated separately.

[0132] Scaled dot product attention is used in all attention calculations, and a dropout rate of 0.1 to 0.3 is set to prevent overfitting.

[0133] Specifically, taking chemistry → physics as an example, the cross-modal cooperative feature is represented as follows:

[0134] in, ;

[0135] ;

[0136] Physics → Chemistry

[0137] ;

[0138] After splicing, chemical-physical cross-modal cooperative features are obtained. ;

[0139] in, This represents a matrix that uses chemical modal features as queries. The intra-modal feature encoding vector representing the chemical mode. This represents the query transformation weight matrix from chemistry to physics. This represents a matrix with physical modal characteristics as bonds. The intra-modal feature encoding vector representing the physical mode. The weight matrix represents the bond transformation from chemical to physical. This represents a matrix whose values ​​are physical modal characteristics. This represents a matrix whose values ​​are physical modal characteristics. This represents the collaborative attention feature vector pointing from the chemical mode to the physical mode. This represents the collaborative attention feature vector pointing from the physical mode to the chemical mode;

[0140] The third fusion unit performs adaptive fusion based on three intramodal feature encoding vectors and three cross-modal collaborative feature vectors;

[0141] The three intramodal feature encoding vectors and three cross-modal collaborative feature vectors are concatenated and then input into a small gated network;

[0142] Small gated networks are fully connected neural networks with one hidden layer, the number of neurons in the hidden layer is 16 to 64, the output activation function is the sigmoid function, and the output dimension is equal to the total number of channels of the feature vector to be fused.

[0143] The gating network outputs the fusion weights for each feature channel in real time, with the weights ranging from 0 to 1.

[0144] Finally, all feature vectors are weighted and summed according to their channel dimensions to output a high-dimensional collaborative tensor as the comprehensive state vector.

[0145] Specifically, the gating vector production representation is as follows:

[0146] ;

[0147] In the formula, Represents the fusion weights for the corresponding feature channels. This represents the weight matrix of the second fully connected layer. This represents the weight matrix of the first fully connected layer. This represents the concatenated total feature vector. and Represents the bias vector

[0148] The combined state vector is represented as:

[0149] ;

[0150] In the formula, The final output is represented by the combined state vector.

[0151] The simulation prediction module includes a model building unit and an intelligent early warning unit;

[0152] The model building unit constructs a lubricant state evolution model based on the oxidation kinetics equation and the wear particle generation model, using the comprehensive state vector as the model input;

[0153] The consumption rate of additives in the lubricating fluid, the growth rate of the total acid value, and the accumulation rate of abrasive particles are described by differential equations.

[0154] Using real-time acquired multimodal data, Kalman filtering or recursive least squares method is used to correct the model parameters in the differential equation online;

[0155] When a sudden change in cross-modal cooperative features is detected in the integrated state vector, the frequency is automatically corrected and the state prediction step size is adjusted.

[0156] Output the physical evolution trend sequence, including the total acid value, viscosity change and predicted abrasive concentration within a predetermined time period;

[0157] Specifically, in this scheme, the consumption rate of additives in the lubricating fluid is obtained through oxidation kinetic equations (Arrhenius-type oxidation kinetics):

[0158] ;

[0159] In the formula, This represents the remaining concentration of the additive. Represents temperature. Represents the rate constant. Represents activation energy. Represents the gas constant. Represents dissolved oxygen concentration. Represents the reaction order;

[0160] The rate of increase in total acid value was obtained by the linear relationship between the increase in total acid value and the consumption of antioxidants;

[0161] ;

[0162] In the formula, Represents the total acid value. Represents the conversion factor. Represents the base oil's background acid production rate;

[0163] The cumulative rate of abrasive particles was obtained through a model of abrasive particle generation, and its mathematical form was used without any substantial modification.

[0164] ;

[0165] In the formula, Represents gear load. Represents rotational speed. Represents the wear generation coefficient;

[0166] The oxidation kinetics equation and the wear particle generation model are both well-known existing technologies in the field of lubricant monitoring, and will not be elaborated on here.

[0167] Meanwhile, this scheme uses extended Kalman filtering for online correction without modifying the core filtering algorithm. It is only used to correct the parameters of this model. The existing technologies known to us will not be elaborated on here.

[0168] In its specific implementation, the state vector is defined as follows: The observation vector is taken from the total acid value and abrasive particle concentration in the real-time multimodal data. The state transition is based on the discretization of the above differential equation. The prediction-update step of EKF is performed according to the standard formula, which will not be elaborated here.

[0169] Calculate the rate of change within the sliding window:

[0170] ;

[0171] In the formula, Representative moment The rate of change of the overall state vector, The rate of change of the overall state vector at the current moment. Represents the time span of the sliding window;

[0172] when A change is defined as a change in the value of more than three standard deviations from the historical mean.

[0173] The corrected model is then integrated forward to output a sequence of physical evolution trends for a predetermined future time period, represented as:

[0174] ;

[0175] In the formula, Represents a sequence of physical evolution trends. This represents the predicted total acid value obtained through integration. This represents the change in viscosity relative to the current moment. This represents the predicted abrasive particle concentration obtained through integration. Represents the number of predicted time points. This represents the j-th predicted time in the future;

[0176] The intelligent early warning unit inputs the historical comprehensive state vector sequence into the time-series prediction model of the long short-term memory network, learns the temporal dependencies between the comprehensive state vectors, and outputs data-driven trend prediction results.

[0177] Among them, the temporal prediction model of the Long Short-Term Memory Network is a well-known existing technology. No modifications were made to the core gating structure or training algorithm of the temporal prediction model of the Long Short-Term Memory Network. It is only used as a temporal prediction tool to learn the long-term dependencies of the integrated state vector. Its backpropagation or training details will not be elaborated here.

[0178] Example of parameters for a time-series prediction model of a Long Short-Term Memory network

[0179] Historical window length = 30 (corresponding to the past 30 sampling points, with a sampling interval of 5 seconds, i.e., 150 seconds of historical data).

[0180] The overall state vector dimension is 64 (output by the collaborative fusion module).

[0181] LSTM hidden layer dimension = 128;

[0182] Number of LSTM layers: 2 (stacked LSTM);

[0183] Output prediction step size: Simultaneously predict the trend forecast results for the next 1 minute, 5 minutes, 10 minutes, 30 minutes, and 1 hour.

[0184] Training method: Using historical data from the past 7 days, with mean squared error (MSE) as the loss function, Adam optimizer, learning rate 0.001, and batch size 32;

[0185] The physical evolution trend sequence and the trend prediction results are fused using Bayesian methods to obtain the fused state distribution.

[0186] The fused state distribution is represented as follows:

[0187] ;

[0188] In the formula, This represents the combined state vector after fusion. Represents a multivariate Gaussian distribution. The vector representing the average of the merged distributions. The covariance matrix represents the distribution after fusion;

[0189] Based on the preset failure threshold, Monte Carlo sampling is performed on the fused state distribution to calculate the probability distribution of the remaining effective lifetime. The median of the distribution is taken as the estimated value of the remaining effective lifetime, and a confidence interval with a confidence level of 90% or 95% is given.

[0190] Monte Carlo sampling, median estimation, and quantile confidence intervals are all well-known methods in statistics and reliability engineering. Example parameter settings are provided below.

[0191] Monte Carlo sample size = 10,000;

[0192] Simulation time step = 1 minute;

[0193] Maximum simulation time = 1000 hours;

[0194] Failure determination is based on the Euclidean distance between the comprehensive state vector and the comprehensive state vector corresponding to the new oil. When this distance is greater than or equal to the Euclidean distance between the comprehensive state vector corresponding to the failure threshold and the comprehensive state vector corresponding to the new oil, the lubricant is determined to have failed.

[0195] Output: Median remaining effective lifetime and 90% confidence interval;

[0196] The Euclidean distance between the current comprehensive state vector and the preset failure threshold vector is then normalized to obtain the basic degradation degree. The trend correction coefficient is calculated based on the ratio of the remaining effective life to the expected total life. The basic degradation degree and the trend correction coefficient are then weighted and fused to obtain the comprehensive degradation degree index.

[0197] The decision feedback unit includes a hierarchical maintenance unit and a visualization unit;

[0198] The graded maintenance unit classifies the current lubricant based on its remaining effective life and overall deterioration index;

[0199] Set a first threshold, a second threshold, and a third threshold;

[0200] When the overall degradation index is less than the first threshold and the remaining effective lifespan is greater than the preset first time threshold, it is judged as a low-risk level, output continues to run, and it is recommended to perform a routine check once a week.

[0201] When the comprehensive deterioration index is greater than or equal to the first threshold and less than the second threshold, or the remaining effective life is less than or equal to the first time threshold and greater than the second time threshold, it is judged as a medium risk level, and a signal is output to perform lubricant sampling and testing at the next shutdown and prepare for oil change plan.

[0202] When the comprehensive deterioration index is greater than or equal to the second threshold and less than the third threshold, or when the remaining effective life is less than or equal to the second time threshold and greater than the third time threshold, it is judged as a high-risk level; output a signal to arrange an oil change as soon as possible and to check the wear condition of the gearbox.

[0203] When the comprehensive deterioration index is greater than or equal to the third threshold, or the remaining effective life is less than or equal to the third time threshold, it is judged as a serious risk level; output a signal to immediately stop the machine, change the oil, and check the wear of gears and bearings.

[0204] The visualization unit displays the remaining effective lifespan, overall degradation index, risk level, and maintenance recommendations on the remote monitoring terminal.

[0205] A method for detecting the deterioration degree of lubricating fluid for gears includes the following steps:

[0206] S1: Collect multimodal data of lubricating fluid through a multimodal sensor, and perform preprocessing and feature extraction to obtain intramodal feature encoding vectors for the three modes;

[0207] S2: The historical time series of the three modalities are sequentially processed by self-attention to extract temporal dependencies, bidirectional collaborative attention to perform cross-modal interaction, and then adaptively fused using a gating network to output a comprehensive state vector;

[0208] S3: Based on the prediction of the digital twin model and the time series prediction model respectively, the remaining effective lifetime and the comprehensive degradation index are obtained by Bayesian fusion.

[0209] S4: Based on the remaining lifespan and degradation index compared with the threshold, four risk levels are classified and corresponding maintenance suggestions are given, which are displayed on the remote terminal.

[0210] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0211] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for detecting the deterioration of lubricating fluid for gears, characterized in that, The system includes: a data acquisition module, a data processing module, a collaborative fusion module, a simulation and prediction module, and a decision feedback module; The data acquisition module is used to acquire multimodal data of the lubricating fluid through a multimodal sensor; The data processing module is used to preprocess and extract features from the multimodal raw data to obtain the chemical degradation index, abrasive degradation index and physical state vector; The collaborative fusion module is used to perform hierarchical collaborative tensor fusion of chemical degradation index, abrasive degradation index and physical state vector. The hierarchical collaborative tensor fusion includes: extracting the internal temporal dependencies of each mode, performing bidirectional collaborative attention interaction, and adaptively controlling the fusion contribution ratio through learnable gating vectors to output a comprehensive state vector. The simulation prediction module is used to input the comprehensive state vector into the digital twin model for dynamic simulation, and output the remaining effective life and comprehensive degradation index through hybrid intelligent prediction. The decision feedback module is used to generate and output maintenance recommendations based on the remaining effective life and the comprehensive degradation index.

2. The deterioration detection system for lubricating fluid in gears according to claim 1, characterized in that, The data acquisition module includes a physical acquisition unit, a chemical acquisition unit, and an abrasive particle acquisition unit; The physical acquisition unit is used to acquire the physical parameters of the lubricating fluid in real time through a MEMS sensor array, including viscosity, dielectric constant, moisture content, and temperature. The chemical acquisition unit is used to acquire the full-band Raman spectrum of lubricating oil using a Raman spectrometer; The abrasive particle acquisition unit is used to emit micro-focus X-rays to penetrate the lubricating fluid flow. Based on the difference in the attenuation of X-rays during the penetration process, it statistically analyzes the total macroscopic concentration and large particle distribution data of abrasive particles in the lubricating fluid, and directly captures microscopic morphology images of abrasive particles in the lubricating fluid through a high-resolution CCD optical system. The microscopic morphology images include data on the size, number, shape, and transmittance of the abrasive particles.

3. The deterioration detection system for lubricating fluid used in gears according to claim 1, characterized in that, The data processing module includes a preprocessing unit, a chemical degradation unit, an abrasive degradation unit, and a physical state unit; The preprocessing unit is used to perform noise reduction, temporal alignment, normalization, and image enhancement on the data acquired in the data acquisition module.

4. The deterioration detection system for lubricating fluid used in gears according to claim 3, characterized in that, The chemical degradation unit calculates the skeleton isomerization parameters based on the peak intensity ratio of the CC stretching vibration peak and the CH2 torsion peak in the Raman spectrum, and calculates the fluorescence and conjugation indicator parameters based on the fluorescence background intensity and the integral intensity of the conjugated double bond characteristic peak in the Raman spectrum. The two parameters are used to construct a two-dimensional coordinate point, which is compared with the pre-established degradation trajectory database to calculate the Mahalanobis distance between the current point and the standard new oil point and the failure point, thus obtaining the chemical degradation index. The abrasive degradation unit classifies abrasive morphology based on the pre-processed abrasive image and inputs it into a pre-trained convolutional neural network, and outputs the statistical quantity of each type of abrasive. The classification includes normal sliding abrasive, cutting abrasive, fatigue spalling block and oxide cluster. Threshold segmentation is performed on the grayscale image acquired by the micro-focal X-ray imaging unit, the total pixel area of ​​abrasive particles is counted, and the abrasive particle concentration is converted by combining the calibration curve. The proportion of large particles is calculated, and the normalized abrasive particle concentration, the proportion of large particles, the proportion of cutting abrasive particles, and the proportion of fatigue spalling blocks are input into the weighted fusion formula to obtain the abrasive particle deterioration index. The physical state unit directly combines the pre-processed and time-aligned viscosity value, dielectric constant value, moisture content, and temperature value into a four-dimensional physical state vector. Each dimension is normalized to its maximum and minimum values, and the normalization range is mapped to [0,1] or [-1,1] to obtain the physical state vector.

5. The deterioration detection system for lubricating fluid in gears according to claim 1, characterized in that, The collaborative fusion module includes a first fusion unit, a second fusion unit, and a third fusion unit; The first fusion unit inputs the time series of chemical degradation index, the time series of abrasive degradation index and the time series of physical state vector into three independent modal self-attention encoders. Each encoder adopts a multi-head self-attention mechanism, with 4 to 8 attention heads and 2 to 4 encoder layers. Each encoder first generates a query matrix, a key matrix, and a value matrix through a linear transformation, then calculates the self-attention weights and sums them by weight, then processes them through residual connections and layer normalization, and finally obtains the intra-modal feature encoding vectors of the three modalities through a feedforward neural network. The historical window length of the time series is 10 to 50 consecutive sampling points, and the time interval between each sampling point is 1 to 10 seconds. Furthermore, each modal self-attention encoder incorporates a learnable position embedding vector to preserve temporal sequence information.

6. The deterioration detection system for lubricating fluid used in gears according to claim 5, characterized in that, The second fusion unit is used for bidirectional collaborative attention interaction between multiple intramodal feature encoding vectors; For chemical and physical modes, the feature vectors of the chemical mode are used as queries and the feature vectors of the physical mode are used as keys and values ​​to calculate the chemical-to-physical collaborative attention feature vectors. Simultaneously, using the feature vector of the physical mode as the query and the feature vector of the chemical mode as the key and value, the collaborative attention feature vector from physics to chemistry is calculated; By concatenating the collaborative attention feature vectors from these two directions, we obtain the chemical-physical cross-modal collaborative features. Using the exact same method, the cross-modal synergistic characteristics of chemical and abrasive particles and the cross-modal synergistic characteristics of physical and abrasive particles were calculated separately. Scaled dot product attention is used in all attention calculations, and a dropout rate of 0.1 to 0.3 is set to prevent overfitting.

7. The deterioration detection system for lubricating fluid in gears according to claim 5, characterized in that, The third fusion unit performs adaptive fusion based on three intramodal feature encoding vectors and three cross-modal collaborative feature vectors; The three intramodal feature encoding vectors and the three cross-modal collaborative feature vectors are concatenated and input into a small gating network; The small gated network is a fully connected neural network with one hidden layer. The number of neurons in the hidden layer is 16 to 64, the output activation function is the sigmoid function, and the output dimension is equal to the total number of channels of the feature vector to be fused. The gating network outputs the fusion weights for each feature channel in real time, with the weights ranging from 0 to 1. Finally, all feature vectors are weighted and summed according to their channel dimensions to output a high-dimensional collaborative tensor as the comprehensive state vector.

8. The deterioration detection system for lubricating fluid used in gears according to claim 1, characterized in that, The simulation prediction module includes a model building unit and an intelligent early warning unit; The model building unit constructs a lubricant state evolution model based on the oxidation kinetics equation and the wear particle generation model, using the comprehensive state vector as the model input; The consumption rate of additives in the lubricating fluid, the growth rate of the total acid value, and the accumulation rate of abrasive particles are described by differential equations. Using real-time acquired multimodal data, Kalman filtering or recursive least squares method is used to correct the model parameters in the differential equation online; When a sudden change in cross-modal cooperative features is detected in the integrated state vector, the frequency is automatically corrected and the state prediction step size is adjusted. Output the physical evolution trend sequence, including the total acid value, viscosity change and predicted abrasive concentration within a predetermined time period; The intelligent early warning unit inputs the historical comprehensive state vector sequence into the temporal prediction model of the long short-term memory network, learns the temporal dependencies between the comprehensive state vectors, and outputs data-driven trend prediction results. The physical evolution trend sequence and the trend prediction results are fused using Bayesian methods to obtain the fused state distribution. Based on the preset failure threshold, Monte Carlo sampling is performed on the fused state distribution to calculate the probability distribution of the remaining effective lifetime. The median of the distribution is taken as the estimated value of the remaining effective lifetime, and a confidence interval with a confidence level of 90% or 95% is given. The Euclidean distance between the current integrated state vector and the preset failure threshold vector is then normalized to obtain the basic degradation degree. The trend correction coefficient is calculated based on the ratio of remaining effective life to expected total life. The basic deterioration degree and the trend correction coefficient are weighted and combined to obtain the comprehensive deterioration degree index.

9. A deterioration detection system for lubricating fluid in gears according to claim 8, characterized in that, The decision feedback unit includes a hierarchical maintenance unit and a visualization unit; The graded maintenance unit classifies the current lubricant based on its remaining effective life and comprehensive deterioration index. Set a first threshold, a second threshold, and a third threshold; When the overall degradation index is less than the first threshold and the remaining effective lifespan is greater than the preset first time threshold, it is judged as a low-risk level, output continues to run, and it is recommended to perform a routine check once a week. When the comprehensive deterioration index is greater than or equal to the first threshold and less than the second threshold, or the remaining effective life is less than or equal to the first time threshold and greater than the second time threshold, it is judged as a medium risk level, and a signal is output to perform lubricant sampling and testing at the next shutdown and prepare for oil change plan. When the comprehensive deterioration index is greater than or equal to the second threshold and less than the third threshold, or when the remaining effective life is less than or equal to the second time threshold and greater than the third time threshold, it is judged as a high-risk level; output a signal to arrange an oil change as soon as possible and to check the wear condition of the gearbox. When the comprehensive deterioration index is greater than or equal to the third threshold, or the remaining effective life is less than or equal to the third time threshold, it is judged as a severe risk level; Output a signal to immediately stop the machine, change the oil, and check the wear of gears and bearings. The visualization unit displays the remaining effective lifespan, overall degradation index, risk level, and maintenance recommendations on the remote monitoring terminal.

10. A method for detecting the deterioration degree of a lubricating fluid for gears, comprising a lubricating fluid deterioration detection system for gears as described in any one of claims 1-9, characterized in that, Includes the following steps: S1: Collect multimodal data of lubricating fluid through a multimodal sensor, and perform preprocessing and feature extraction to obtain intramodal feature encoding vectors for the three modes; S2: The historical time series of the three modalities are sequentially processed by self-attention to extract temporal dependencies, bidirectional collaborative attention to perform cross-modal interaction, and then adaptively fused using a gating network to output a comprehensive state vector; S3: Based on the prediction of the digital twin model and the time series prediction model respectively, the remaining effective lifetime and the comprehensive degradation index are obtained by Bayesian fusion. S4: Based on the remaining lifespan and degradation index compared with the threshold, four risk levels are classified and corresponding maintenance suggestions are given, which are displayed on the remote terminal.

Images