Lubricating oil filtering system based on machine learning

Abstract

The invention belongs to the technical field of lubricating oil filtering, and discloses a lubricating oil filtering system based on machine learning. Comprising a data acquisition and processing module used for acquiring lubricating oil sample data and working condition environment data corresponding to the lubricating oil sample data, and preprocessing the lubricating oil sample data to obtain sample feature data; the component identification module is used for inputting the sample feature data into a pre-constructed improved deep belief network model to obtain a dependent acyclic graph; the evaluation module is used for acquiring a preliminary degradation trend sequence in the t period of time in the future based on the working condition environment data and the dependency acyclic graph; and the decision module is used for outputting a filtering decision by utilizing an attention fusion algorithm based on the preliminary degradation trend sequence, and sending the filtering decision to a filtering control terminal, so that the utilization rate of equipment is greatly improved, the maintenance cost is reduced, and the stability and the safety of production are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of lubricating oil filtration, and more specifically, to a lubricating oil filtration system based on machine learning. Background Art

[0002] Traditional lubricating oil filtration systems mainly rely on fixed filtration cycles or simple threshold judgments, and are unable to dynamically make accurate filtration decisions based on actual operating conditions and lubricating oil status. This method has the problem of inaccurate filtration time, leading to premature or delayed filtration. Secondly, existing lubricating oil status monitoring methods mainly rely on manual detection, which cannot comprehensively analyze the changes in the physical and chemical properties of the lubricating oil, and it is difficult to accurately predict the degradation trend of the lubricating oil. Existing filtration decisions are mainly based on a single indicator or simple rules, and cannot fully integrate multiple operating conditions and lubricating oil performance data, making it difficult to make the best filtration decision.

[0003] In view of this, the present invention proposes a lubricating oil filtration system based on machine learning to solve the above problems. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention provides the following technical solutions: A lubricating oil filtration system based on machine learning, comprising: a data acquisition and processing module, used to collect lubricating oil sample data and working condition environment data corresponding to the lubricating oil sample data, pre-process the lubricating oil sample data, and obtain sample feature data;

[0005] The component identification module is used to input the sample feature data into the pre-built improved deep belief network model to obtain a dependency acyclic graph;

[0006] The evaluation module obtains the preliminary degradation trend sequence in the future t period of time based on the working environment data and the dependent acyclic graph;

[0007] The decision module outputs a filtering decision based on the preliminary degradation trend sequence using an attention fusion algorithm, and sends the filtering decision to a filtering control terminal. Each module is connected to another by wire and / or wireless means.

[0008] Furthermore, the lubricating oil sample data includes lubricating oil physical property data and lubricating oil chemical property data; the working environment data includes equipment operation data and use environment data;

[0009] The physical property data of lubricating oil include viscosity data, density data, acid value data and water content data; the chemical property data of lubricating oil include base oil composition, additive composition and pollutant content;

[0010] Equipment operation data includes equipment speed, equipment load, equipment temperature and equipment pressure; operating environment data includes ambient temperature, ambient humidity and ambient vibration.

[0011] Furthermore, the method of preprocessing the lubricating oil sample data includes:

[0012] Build an adaptive encoder; the basic structure of the adaptive encoder includes an encoder and a decoder; define the loss function of the adaptive encoder

[0013] Among them, N(μ,σ 2 ) represents a normal distribution, KL() represents the Kullback-Leibler divergence, and β is the weight coefficient;

[0014] By collecting lubricating oil sample data within a fixed historical time period as training data for the adaptive encoder, in the process of training the adaptive encoder, the parameters of the encoder and decoder are optimized by minimizing the loss function, thereby obtaining a constructed adaptive encoder;

[0015] The lubricating oil sample data is divided according to data types, and the data corresponding to each data type is sent as a sample to the encoder in the adaptive encoder to obtain a low-dimensional potential feature vector

[0016] The low-dimensional potential feature vector is clustered to obtain a clustering result, which is N1 clusters. A clustering sample threshold is preset, and clusters with a sample number less than or equal to the clustering sample threshold in the N1 clusters are regarded as abnormal clusters; for each sample, the distance between it and each abnormal cluster is calculated as its abnormal score;

[0017] The low-dimensional potential feature vector is used as a clustering sample, and the clustering samples of the abnormal score are statistically analyzed to fit their probability density function P(s), where s is the abnormal score; according to the probability density function of the abnormal score, an adaptive abnormal detection threshold τ is set; and the abnormal detection threshold is adaptively updated; samples with an abnormal score greater than or equal to τ are detected as abnormal samples;

[0018] τ=inf{s:P(S≤s)≥α}; where P(S≤s) represents the cumulative distribution function of the set S of anomaly scores, and α is the confidence level;

[0019] The abnormal samples are deleted from the lubricating oil sample data to obtain a complete lubricating oil sample data set, and the data in the complete lubricating oil sample data set are normalized to obtain sample feature data.

[0020] Furthermore, the encoder maps the input lubricating oil sample data x to potential features that obey Gaussian distribution Where μ is the mean of the latent feature, σ is the standard deviation of the latent feature; ε is the random noise of the standard normal distribution, and ε~N(0,I), where N(0,I) represents a standard normal distribution with a mean of 0, a variance of 1, and each variable is independent of each other, I is the unit matrix, represents element-by-element product;

[0021] The decoder reconstructs the potential feature z back to the lubricating oil sample data input by the encoder, and the reconstruction formula is: in, is the lubricating oil sample data reconstructed from the latent feature z by the decoder; f dec is the mapping function of the decoder; θ dec Parameters for the decoder.

[0022] Furthermore, the method of clustering low-dimensional potential features includes:

[0023] Define the neighborhood radius r and the minimum number of samples MinPts for clustering; take the low-dimensional potential feature vector as the sample point, and for each sample point, calculate its Euclidean distance from all other sample points. Then join in Neighborhood in, represents the jth sample point among all sample points, represents the i-th sample point among all sample points, is the Euclidean distance between the i-th sample point and the j-th sample point;

[0024] For each sample point, calculate its density weight. The calculation formula of density weight is:

[0025] Among them, ρ i is the density weight of the i-th sample point;

[0026] Sort the density weights of all sample points in ascending order to obtain a sorted sequence and define a density threshold ρ th , where ρ th =ρ (n-MinPts+1) ; Where n is the total number of sample points;

[0027] For the i-th sample point, if ρ i ≥ρ th , then Marked as a core object, if ρ i Less than ρ th , then calculate The highest density weight ρ in its neighborhood max , ifmax ≥ρ th , then Marked as a boundary object, if ρ max Still less than ρ th , then Mark as noise object;

[0028] If there is a arrive The sequence of but Depend on Density can be reached if there is a sample point Make and All by The density can be reached, and Density connected;

[0029] All density-connected core objects and boundary objects form a cluster, and finally N1 clusters are obtained.

[0030] Furthermore, the construction method of the improved deep belief network model includes:

[0031] The basic architecture of the improved deep belief network model is defined to include N2 restricted Boltzmann machines, each of which includes a visible layer and a hidden layer. The visible layer and the hidden layer are connected through a weight matrix, and the hidden layer of the I-th layer restricted Boltzmann machine is used as the visible layer of the I+1-th layer restricted Boltzmann machine; finally, an output layer of supervised training is added to form a complete basic architecture of the improved deep belief network model;

[0032] Define the energy function E(v (I) ,h (I) ); collect historical data, pre-process the historical data to obtain model input data, use the model input data to train the improved deep belief network model, and optimize the parameters of the improved deep belief network model by minimizing the KL divergence between the distribution of historical data and the distribution of the improved deep belief network model; obtain a preliminary improved deep belief network model;

[0033] Label the historical data with the target DAG value to obtain a sample data set containing labeled samples; use the sample data set to fine-tune the preliminary improved deep belief network model, and define the goal of fine-tuning to minimize the error between the DAG output by the preliminary improved deep belief network model and the target DAG value. The error is calculated using the square error loss function or the cross entropy loss function;

[0034] The back propagation algorithm is used to update the parameters of the entire improved deep belief network model, including each restricted Boltzmann machine and the added output layer of the improved deep belief network model, to obtain the completed improved deep belief network model.

[0035] Furthermore, the energy function E(v (I) ,h (I) ) is expressed as:

[0036]

[0037] Among them, GE(v (I) ,h (I) ) is the connector function; FE(v (I) ,h (I) ) is a cross-level sub-function;

[0038]

[0039] ; where v( I ) is the unit vector of the visible layer of the I-th layer restricted Boltzmann machine, h( I ) is the unit vector of the hidden layer of the I-th layer restricted Boltzmann machine, is the weight between the Jth visible layer and the Kth hidden layer in the Ith layer restricted Boltzmann machine, is the bias of the Jth visible layer in the Ith layer restricted Boltzmann machine, is the Jth visible unit in the Ith layer restricted Boltzmann machine; is the state or value of the Kth hidden unit in the Ith layer restricted Boltzmann machine; is the bias of the Kth hidden layer in the Ith layer restricted Boltzmann machine, n v is the number of units in the visible layer, n h is the number of units in the hidden layer; is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I+1th restricted Boltzmann machine, is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I-1th restricted Boltzmann machine, is the lth visible unit in the I+1th layer restricted Boltzmann machine, is the state or value of the lth hidden unit in the I-1th layer restricted Boltzmann machine;

[0040] Furthermore, the calculation method of the preliminary degradation trend sequence includes:

[0041] Clean the working environment data to obtain pure working environment data, and extract environmental features based on the pure working environment data. The environmental features include time features, statistical features, and cross features.

[0042] Extract sample features from the dependent acyclic graph. Sample features include the shortest undirected path length feature between nodes and node centrality. Node centrality includes degree centrality, betweenness centrality and closeness centrality.

[0043] The method of extracting the sample features includes:

[0044] Convert the directed acyclic graph to an undirected graph. For each pair of nodes in the undirected graph, use breadth-first search or Dijkstra algorithm to calculate the shortest path length between them, and express it with an adjacency matrix. Finally, a matrix D is obtained, which is the shortest undirected path length feature between nodes.

[0045] For node u in an undirected graph, calculate the number of neighbor nodes ku in the undirected graph;

[0046] Degree Centrality Where N3 is the total number of nodes; for node u in an undirected graph, calculate the number of shortest path lengths of all other pairs of nodes njk. Among the njk shortest paths, the number of paths passing through node u is njk(u); then the betweenness centrality is

[0047] For a node u in an undirected graph, calculate the shortest path length dau from it to all other nodes a, then the proximity centrality

[0048] The sample features and environmental features are input into the constructed and trained random forest model to obtain the degradation degree change value of each node in the next t time steps. The degradation degree change values ​​are connected in series to obtain the preliminary degradation trend sequence of the node in the next t time steps.

[0049] Furthermore, the random forest model is constructed in the following manner:

[0050] Collect historical comprehensive data, which includes lubricating oil sample data within a fixed historical period and working condition environment data corresponding to the lubricating oil sample data;

[0051] The historical comprehensive data is engineered to obtain the corresponding environmental characteristics and sample characteristics; the environmental characteristics and sample characteristics are combined into a data set, and the data set is annotated with node degradation degree labels;

[0052] The labeled data set is divided into a training set and a test set in chronological order. For the training set, a decision tree is selected as the base learner, and N5 decision trees are constructed. For each decision tree, some features are randomly extracted from the sample features and the environmental features with replacement as the split features of the decision tree nodes. At each decision tree node, an optimal feature is randomly selected from the split features for splitting. The maximum height is defined, and each decision tree stops growing when it reaches the maximum height to obtain the initial random forest.

[0053] Input the training set into the initial random forest. After the training is completed, each decision tree corresponds to a prediction result; and the initial random forest model is obtained;

[0054] The loss function is defined as the root mean square error function, and the function value of the loss function of the preliminary random forest model is calculated on the test set. When the function value of the loss function does not decrease for p consecutive times on the test set, a grid search or random search method is used in this process to adjust the hyperparameters of the random forest model to obtain a constructed and trained random forest model.

[0055] Furthermore, the output method of the filtering decision includes:

[0056] The preliminary degradation trend sequence contains the degradation degree change values ​​of each node in the future t time steps, and these degradation degree change values ​​are used as the input of the attention fusion algorithm;

[0057] Constructing a filtering decision output model, which includes an attention fusion layer, a calculation layer, and a decision layer;

[0058] The attention fusion layer consists of a query vector, a key vector, and a value vector. The query vector is the initial degradation trend sequence of the node making the decision, and the key vector and value vector are the initial degradation trend sequences of all other nodes.

[0059] The calculation layer calculates the similarity between any query vector q' and all key vectors to obtain the corresponding attention score e q' =q' T ·K' w ; where q' T represents the transpose of the query vector q', K' w is the w-th key vector; the attention score is scaled by the Softmax function to obtain the corresponding final attention score;

[0060] Perform weighted summation of the final attention score and the corresponding value vector to obtain the attention value vector;

[0061] The fully connected layer is used as the decision layer, the attention value vector is used as the input of the decision layer, and the preliminary filtering decision probability distribution o = σ(W'·A'+b') is output; where W' is the weight matrix of the fully connected layer, b' is the bias vector, σ is the Softmax activation function, and the category with the largest probability in the preliminary filtering decision probability distribution is taken as the decision result, that is, the filtering decision; the categories include early filtering, normal filtering and delayed filtering.

[0062] Technical effects and advantages of a lubricating oil filtration system based on machine learning of the present invention:

[0063] The present invention can comprehensively analyze the performance changes of lubricating oil, accurately predict future degradation trends, fully mine data, and establish an in-depth understanding of the laws of lubricating oil state changes, so as to make more accurate predictions and provide a reliable basis for subsequent decision-making. Secondly, it intelligently integrates various operating environment data and lubricating oil performance data to make dynamic and optimized filtration decisions. It can not only fully consider the actual operating conditions of the equipment, but also adjust the filtration strategy in time according to changes in different operating conditions, greatly improving the pertinence and efficiency of filtration. In addition, it also has the ability of self-learning and self-adaptation, can automatically detect abnormal situations, and continuously optimize its own analysis models and decision-making strategies, ensuring a high degree of flexibility and robustness in complex and changeable working conditions, greatly improving equipment utilization, reducing maintenance costs, and ensuring production stability and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of a lubricating oil filtration system based on machine learning of the present invention;

[0065] Figure 2 Schematic diagram of a lubricating oil filtration method based on machine learning of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] Example 1

[0068] See also Figure 1 As shown, the lubricating oil filtration system based on machine learning described in this embodiment includes:

[0069] A data acquisition and processing module is used to collect lubricating oil sample data and working condition environment data corresponding to the lubricating oil sample data, and pre-process the lubricating oil sample data to obtain sample characteristic data;

[0070] The component identification module is used to input the sample feature data into the pre-built improved deep belief network model to obtain a dependency acyclic graph;

[0071] The evaluation module obtains the preliminary degradation trend sequence in the future t period of time based on the working environment data and the dependent acyclic graph;

[0072] The decision module outputs the filtering decision based on the preliminary degradation trend sequence using the attention fusion algorithm and sends the filtering decision to the filtering control terminal. The modules are connected by wire and / or wireless means to realize data transmission between modules.

[0073] Furthermore, the lubricant sample data includes lubricant physical property data and lubricant chemical property data; the operating environment data includes equipment operation data and use environment data;

[0074] The physical property data of lubricating oil include viscosity data, density data, acid value data and water content data; the chemical property data of lubricating oil include base oil composition, additive composition and pollutant content;

[0075] Equipment operation data includes equipment speed, equipment load, equipment temperature and equipment pressure; operating environment data includes ambient temperature, ambient humidity and ambient vibration.

[0076] It should be noted that the viscosity data is measured using a viscometer (such as a falling ball viscometer, a rotational viscometer, etc.); the density data is measured using a densitometer (such as a liquid densitometer, a digital densitometer, etc.); the acid value data is measured using an acid value meter (such as a potentiometric titration, a titration method, etc.); and the water content data is measured using a Karl Fischer moisture meter or other moisture meter.

[0077] The base oil component is obtained by analyzing the base oil content of the lubricant using infrared spectroscopy; the additive component is obtained by analyzing the additive content of the lubricant using infrared spectroscopy, and the pollutant content is obtained by analyzing the pollutant content in the lubricant using atomic absorption spectroscopy.

[0078] The equipment speed is obtained through a speed sensor or encoder installed on the equipment; the equipment load is obtained through a force sensor or torque sensor installed on the equipment; the equipment temperature is obtained through a temperature sensor (such as thermocouple, thermistor, etc.) installed on the equipment; the equipment pressure is obtained through a pressure sensor installed on the equipment.

[0079] A temperature sensor (such as a thermocouple, a thermistor, etc.) is used to measure the temperature of the environment in which the device is located as the ambient temperature. A humidity sensor is used to measure the relative humidity of the environment in which the device is located as the ambient humidity. A vibration sensor (such as an accelerometer, a velocity sensor, etc.) is used to obtain the ambient vibration.

[0080] Furthermore, the method of preprocessing the lubricating oil sample data includes:

[0081] Build an adaptive encoder; the basic structure of the adaptive encoder includes an encoder and a decoder;

[0082] The encoder maps the input lubricant sample data x to latent features that follow a Gaussian distribution Where μ is the mean of the latent feature, which is output by the encoder and represents the center of the latent variable distribution. σ is the standard deviation of the latent feature, which is also output by the encoder and represents the scale of the latent variable distribution. ε is the random noise of the standard normal distribution, ε~N(0,I), where N(0,I) represents a standard normal distribution with a mean of 0, a variance of 1, and each variable is independent of each other. I is the unit matrix, the diagonal elements of the unit matrix are 1, and the off-diagonal elements are 0. It represents element-by-element product, that is, the corresponding elements of two vectors are multiplied.

[0083] The decoder reconstructs the latent feature z back to the lubricant sample data input by the encoder. The reconstruction formula is: in, is the lubricating oil sample data reconstructed from the latent feature z by the decoder; f dec is the mapping function of the decoder, which transforms the latent feature z back to f dec The decoder parameters θ dec Determine; dec is the decoder parameter, θ dec Contains all the functions used to map the latent feature z back to The weights and biases.

[0084] The mapping function of the decoder is composed of a multi-layer fully connected neural network, and all weights and bias parameters in the mapping function are obtained through training and learning to obtain the mapping function of the decoder.

[0085] Define the loss function for the adaptive encoder

[0086] Among them, N(μ,σ 2) represents a normal distribution, KL() represents the Kullback-Leibler divergence, and β is a weight coefficient used to balance the reconstruction error and the Kullback-Leibler divergence. By adjusting β, the generation ability and reconstruction ability of the adaptive encoder can be controlled;

[0087] By collecting lubricating oil sample data within a fixed historical time period as training data for the adaptive encoder, during the training process, the parameters of the encoder and decoder are optimized by minimizing the loss function to obtain the constructed adaptive encoder.

[0088] The lubricating oil sample data is divided according to data types, and the data corresponding to each data type is sent as a sample to the encoder in the adaptive encoder to obtain a low-dimensional potential feature vector The low-dimensional latent feature vector can reflect the intrinsic structure and characteristics of the data. Its dimension is much lower than that of the original data. The low-dimensional latent feature vector is clustered to obtain the clustering result, which is N1 clusters. The clustering sample threshold is preset, and the clusters with the number of samples less than or equal to the clustering sample threshold in the N1 clusters are regarded as abnormal clusters; for each sample, its distance to each abnormal cluster is calculated as its abnormal score.

[0089] The low-dimensional latent feature vector is used as the clustering sample, and the clustering samples of the anomaly score are statistically analyzed to fit their probability density function P(s), where s is the anomaly score; according to the statistical distribution of the anomaly score, an adaptive anomaly detection threshold τ is set; and the anomaly detection threshold is adaptively updated; samples with anomaly scores greater than or equal to τ are detected as anomaly samples; τ=inf{s:P(S≤s)≥α}; where P(S≤s) represents the cumulative distribution function of the set S of anomaly scores, and α is the confidence level.

[0090] The confidence α is taken as a parameter that needs to be learned, and the precision and recall based on anomaly detection are defined. By optimizing the precision and recall, an optimal confidence value is obtained.

[0091] The abnormal samples are deleted from the lubricating oil sample data to obtain a perfect lubricating oil sample data set, and the data in the perfect lubricating oil sample data set are normalized to obtain a sample feature data set.

[0092] The method of performing adaptive updating includes:

[0093] Set the sliding window size N2, collect the initial N2 clustering samples of anomaly scores, and use these N2 samples to fit the probability density function P0(s) of the initial anomaly score.

[0094] According to the preset confidence level α, the initial anomaly detection threshold τ0 is determined; when a new cluster sample with anomaly score arrives, it is added to the sliding window, and the earliest cluster sample is removed to form a new sliding window. The N2 cluster samples in the new sliding window are used to re-estimate the probability density function P1(s) of the anomaly score. According to P1(s), a new anomaly detection threshold τ1 is calculated. τ1 is used as the current anomaly detection threshold to determine whether the newly arrived cluster sample is abnormal. For each new sample that arrives subsequently, the new anomaly detection threshold is repeatedly calculated, the sliding window is moved, and the earliest cluster sample is removed each time, and a newly arrived cluster sample is added. The cluster samples in the new sliding window are used to re-estimate the probability density function of the anomaly score. According to the new density function, a new anomaly detection threshold is calculated to determine the newly arrived cluster sample.

[0095] The method of clustering low-dimensional potential features includes:

[0096] Define the neighborhood radius r and the minimum number of samples MinPts of the cluster. The neighborhood radius r is used to define the neighborhood of the sample point, and the minimum number of samples MinPts is used to define the minimum number of sample points required for a dense area. Usually, visualization tools or heuristic methods are used to estimate the appropriate values ​​of the neighborhood radius r and the minimum number of samples MinPts.

[0097] The low-dimensional potential feature vector is used as a sample point. For each sample point, the Euclidean distance between it and all other sample points is calculated. If Then join in Neighborhood in, represents the jth sample point among all sample points, represents the i-th sample point among all sample points, is the Euclidean distance between the i-th sample point and the j-th sample point.

[0098] For each sample point, calculate its density weight. The calculation formula of density weight is:

[0099] Among them, ρ i is the density weight of the i-th sample point;

[0100] Sort the density weights of all sample points in ascending order to obtain a sorted sequence and define a density threshold ρ th , where ρ th =ρ (n-MinPts+1) ; Where n is the total number of sample points;

[0101] For the i-th sample point, if ρ i ≥ρ th, then Marked as a core object, if ρ i Less than ρ th , then calculate The highest density weight ρ in its neighborhood max , if max ≥ρ th , then Marked as a boundary object, if ρ max Still less than ρ th , then Marked as a noise object.

[0102] If there is a arrive The sequence of but Depend on Density can be reached if there is a sample point Make and All by The density can be reached, and Density connected; all density-connected core objects and boundary objects form a cluster, and finally N1 clusters are obtained.

[0103] Furthermore, the construction method of the improved deep belief network model includes:

[0104] The infrastructure of the improved deep belief network model is defined to include N2 restricted Boltzmann machines. Each restricted Boltzmann machine includes a visible layer and a hidden layer. The visible layer and the hidden layer are connected through a weight matrix. The hidden layer of the I-th restricted Boltzmann machine is used as the visible layer of the I+1-th restricted Boltzmann machine. Finally, an output layer for supervised training is added to form a complete infrastructure of the improved deep belief network model.

[0105] Define the energy function E(v (I) ,h (I) );

[0106]

[0107] ; Among them, GE (v (I) ,h (I) ) is the connector function; FE(v (I) ,h (I) ) is a cross-level sub-function;

[0108]

[0109] ; where v (I) is the unit vector of the visible layer of the I-th layer restricted Boltzmann machine, h(I) is the unit vector of the hidden layer of the I-th layer restricted Boltzmann machine, is the weight between the Jth visible layer and the Kth hidden layer in the Ith layer restricted Boltzmann machine, is the bias of the Jth visible layer in the Ith layer restricted Boltzmann machine, is the Jth visible unit in the Ith layer of the restricted Boltzmann machine. The visible unit is the node when the input data is transmitted through the network. The input data is stored and processed in the input layer. is the state or value of the Kth hidden unit in the Ith layer of the restricted Boltzmann machine. The hidden unit is a node used to learn and represent implicit patterns or features in the data, and interacts with the visible units through weight connections in the network; is the bias of the Kth hidden layer in the Ith layer restricted Boltzmann machine, n v is the number of units in the visible layer, n h is the number of units in the hidden layer; is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I+1th restricted Boltzmann machine, is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I-1th restricted Boltzmann machine, is the lth visible unit in the I+1th layer restricted Boltzmann machine, is the state or value of the lth hidden unit in the I-1th layer restricted Boltzmann machine.

[0110] Collect historical data, where the historical data is lubricating oil sample data within a fixed historical time period; preprocess the historical data to obtain model input data, use the model input data to train an improved deep belief network model, and optimize the parameters of the improved deep belief network model by minimizing the KL divergence between the distribution of historical data and the distribution of the improved deep belief network model; obtain a preliminary improved deep belief network model.

[0111] Label the historical data with the target DAG value to obtain a sample data set containing labeled samples; use the sample data set to fine-tune the preliminary improved deep belief network model, and define the goal of fine-tuning as minimizing the error between the DAG output by the preliminary improved deep belief network model and the target DAG value. The error is calculated using the square error loss function or the cross entropy loss function.

[0112] It should be noted that by analyzing each sample in the historical data, the expected DAG structure is manually constructed as the target DAG value, and it is used as the label of the sample.

[0113] The back propagation algorithm is used to update the parameters of the entire improved deep belief network model, including each restricted Boltzmann machine and the added output layer of the improved deep belief network model, to obtain the completed improved deep belief network model.

[0114] Through the fine-tuning process, the improved deep belief network model further learns the mapping relationship from the input features to the target DAG, thereby improving the prediction performance of the model; by minimizing the KL divergence, the model can be effectively optimized in unsupervised pre-training, thereby improving the overall performance of the model.

[0115] Furthermore, the method for obtaining the dependency acyclic graph includes:

[0116] The sample feature data is input into the constructed improved deep belief network model. As the input sample feature data is passed layer by layer in the improved deep belief network model, each hidden layer will generate a set of activation values. The activation values ​​reflect the representation of the input in the hidden layer. According to the relationship between the activation values ​​of each hidden layer, a dependent acyclic graph is constructed.

[0117] The nodes in the dependency acyclic graph represent hidden layer neurons, and the directed edges represent the dependencies between nodes; the weights of the edges are determined based on the correlation or causality of the activation values ​​of the nodes.

[0118] This dependency acyclic graph represents the complex dependencies and causal relationships between sample feature data.

[0119] Furthermore, the calculation method of the preliminary degradation trend sequence includes:

[0120] Clean the working environment data to obtain pure working environment data, and extract environmental features based on the pure working environment data. The environmental features include time features, statistical features, and cross features.

[0121] Data cleaning methods include processing missing values ​​and removing outliers; processing missing values ​​means filling missing values ​​of continuous values ​​with mean / median / nearest neighbor methods, and filling missing values ​​of categorical values ​​with the category with the highest frequency; removing outliers means identifying and removing outliers based on statistical methods (such as the 3σ principle).

[0122] The time feature is to add the date / time component (year, month, day, hour, etc.) to the original data to capture periodic patterns. The statistical feature is to calculate the statistics (mean, variance, maximum / minimum value) of the pure working environment data in different time windows (such as hours / days / weeks). The cross feature is to construct the cross terms of equipment operation data and usage environment data, such as temperature x speed, humidity x load, etc.

[0123] For environmental features with heteroscedasticity, standardization (such as Z-Score standardization) is performed, and for environmental features with homoscedasticity, normalization (such as Min-Max normalization) is performed.

[0124] Extract sample features from dependent acyclic graphs. Sample features include the shortest undirected path length features between nodes and node centrality. Node centrality includes degree centrality, betweenness centrality and closeness centrality.

[0125] The method of extracting the sample features includes:

[0126] Convert the directed acyclic graph to an undirected graph. For each pair of nodes in the undirected graph, use breadth-first search (BFS) or Di jkstra algorithm to calculate the shortest path length between them, and express it with the adjacency matrix. Finally, a matrix D is obtained, which is the shortest undirected path length feature between the nodes.

[0127] For node u in an undirected graph, calculate the number of neighbor nodes ku in the undirected graph;

[0128] Degree Centrality Among them, N3 is the total number of nodes; the higher the degree centrality, the more core the node is;

[0129] For node u in an undirected graph, calculate the number njk of the shortest path lengths of all other pairs of nodes. Among the njk shortest paths, the number of paths passing through node u is njk(u);

[0130] Betweenness Centrality A high betweenness centrality indicates that node u plays a large role as a “bridge” connecting other pairs of nodes;

[0131] For a node u in an undirected graph, calculate the shortest path length dau from it to all other nodes a.

[0132] Close to centrality A high proximity centrality indicates that the average distance from node u to other nodes is short and it is in a central position.

[0133] The sample features and environmental features are input into the constructed and trained random forest model to obtain the degradation degree change value of each node in the next t time steps. The degradation degree change values ​​are connected in series to obtain the preliminary degradation trend sequence of the node in the next t time steps.

[0134] Furthermore, the random forest model is constructed in the following manner:

[0135] Collect historical comprehensive data, which includes lubricating oil sample data within a fixed historical period and operating environment data corresponding to the lubricating oil sample data.

[0136] The historical comprehensive data is engineered to obtain the corresponding environmental characteristics and sample characteristics; the environmental characteristics and sample characteristics are merged into a data set, and the data set is labeled with node degradation degree labels; according to the historical operation data, maintenance records, etc. of the nodes, the degradation degree of the nodes is scored or graded as the node degradation degree label; for example, the node degradation degree is divided into 5 levels, level 0 is intact, level 1 is slightly degraded, level 2 is moderately degraded, level 3 is severely degraded, and level 4 is seriously failed.

[0137] The labeled data set is divided into a training set and a test set in chronological order, for example, in a ratio of 7:3. For the training set, a decision tree is selected as the base learner, and N5 decision trees are constructed. For each decision tree, some features are randomly extracted from the sample features and the environmental features with replacement as the split features of the decision tree nodes. At each decision tree node, an optimal feature is randomly selected from the split features for splitting. The maximum height is defined, and each decision tree stops growing when it reaches the maximum height; the initial random forest is obtained.

[0138] The training set contains sample features, environmental features and corresponding node degradation degree labels. The training set is input into the initial random forest. After the training is completed, each decision tree corresponds to a prediction result; a preliminary random forest model is obtained.

[0139] The loss function is defined as the root mean square error function, and the function value of the loss function of the preliminary random forest model is calculated on the test set. When the function value of the loss function does not decrease for p consecutive times on the test set, a grid search or random search method is used in this process to adjust the hyperparameters of the random forest model to obtain a constructed and trained random forest model.

[0140] Furthermore, the output method of the filtering decision includes:

[0141] The preliminary degradation trend sequence contains the degradation degree change values ​​of each node in the future t time steps, and these degradation degree change values ​​are used as the input of the attention fusion algorithm;

[0142] Constructing a filtering decision output model, which includes an attention fusion layer, a calculation layer, and a decision layer;

[0143] The attention fusion layer consists of a query vector, a key vector, and a value vector. The query vector is the preliminary degradation trend sequence of the node making the decision, and the key vector and value vector are the preliminary degradation trend sequences of all other nodes.

[0144] The calculation layer calculates the similarity between any query vector q' and all key vectors to obtain the corresponding attention score e q' =q' T ·K'w ; where q' T represents the transpose of the query vector q', K' w is the w-th key vector; the attention score is scaled by the Softmax function to obtain the corresponding final attention score;

[0145] The final attention score is weightedly summed with the corresponding value vector to obtain the attention value vector, which integrates the degradation trend information of all nodes.

[0146] The fully connected layer is used as the decision layer, all neurons in the previous layer are connected to each neuron in the current layer, the attention value vector is used as the input of the decision layer, and the preliminary filtering decision o = σ(W'·A'+b') is output; where W' is the weight matrix of the fully connected layer, b' is the bias vector, and σ is the Softmax activation function; the preliminary filtering decision is converted into a probability distribution through the Softmax activation function, and the category with the largest probability in the probability distribution is taken as the decision result, that is, the filtering decision; the categories include early filtering, normal filtering and delayed filtering.

[0147] For example, the output of the decision layer may be [0.2, 0.6, 0.2], indicating that the probability of normal filtering is the highest, so the final decision is normal filtering;

[0148] Different attention weights are adaptively assigned to each node, and the degradation trend information of all nodes is integrated to generate the optimal filtering decision.

[0149] This embodiment can comprehensively analyze the performance changes of lubricating oil, accurately predict future degradation trends, fully mine data, and establish an in-depth understanding of the laws of lubricating oil state changes, so as to make more accurate predictions and provide a reliable basis for subsequent decision-making. Secondly, it intelligently integrates various operating environment data and lubricating oil performance data to make dynamic and optimized filtration decisions. It can not only fully consider the actual operating conditions of the equipment, but also adjust the filtration strategy in time according to changes in different operating conditions, greatly improving the pertinence and efficiency of filtration. In addition, it also has the ability of self-learning and self-adaptation, can automatically detect abnormal situations, and continuously optimize its own analysis models and decision-making strategies, ensuring that it maintains a high degree of flexibility and robustness in complex and changeable working conditions, greatly improving equipment utilization, reducing maintenance costs, and ensuring production stability and safety.

[0150] Example 2

[0151] See also Figure 2As shown, the part not described in detail in this embodiment is described in Example 1, and a lubricating oil filtration method based on machine learning is provided, comprising: S1, collecting lubricating oil sample data and working condition environment data corresponding to the lubricating oil sample data, preprocessing the lubricating oil sample data, and obtaining sample feature data;

[0152] S2, inputting the sample feature data into the pre-built improved deep belief network model to obtain a dependent acyclic graph;

[0153] S3, based on the working environment data and the dependent acyclic graph, obtain the preliminary degradation trend sequence in the future t period of time;

[0154] S4. Based on the preliminary degradation trend sequence, the attention fusion algorithm is used to output the filtration decision, and the filtration decision is sent to the filtration control terminal.

[0155] Example 3

[0156] This embodiment discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the operation mode of the lubricating oil filtration method based on machine learning provided above is implemented.

[0157] Since the electronic device introduced in this embodiment is an electronic device used to implement a lubricating oil filtration method based on machine learning in the embodiment of this application, based on the lubricating oil filtration method based on machine learning introduced in the embodiment of this application, the technical personnel of this field can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application is not described in detail here. As long as the technical personnel of this field implement the electronic device used in the lubricating oil filtration method based on machine learning in the embodiment of this application, it belongs to the scope of protection of this application.

[0158] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0159] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A lubricating oil filtration system based on machine learning, It is characterized in that include: A data acquisition and processing module is used to collect lubricating oil sample data and working condition environment data corresponding to the lubricating oil sample data, and pre-process the lubricating oil sample data to obtain sample characteristic data; The component identification module is used to input the sample feature data into the pre-built improved deep belief network model to obtain a dependency acyclic graph; The evaluation module obtains the preliminary degradation trend sequence in the future t period of time based on the working environment data and the dependent acyclic graph; The decision module outputs a filtering decision based on the preliminary degradation trend sequence using an attention fusion algorithm, and sends the filtering decision to a filtering control terminal. Each module is connected to another by wire and / or wireless means.

2. A lubricating oil filtration system based on machine learning according to claim 1, It is characterized in that The lubricating oil sample data includes lubricating oil physical property data and lubricating oil chemical property data; the working environment data includes equipment operation data and use environment data; The physical property data of lubricating oil include viscosity data, density data, acid value data and water content data; the chemical property data of lubricating oil include base oil composition, additive composition and pollutant content; Equipment operation data includes equipment speed, equipment load, equipment temperature and equipment pressure; operating environment data includes ambient temperature, ambient humidity and ambient vibration.

3. The lubricating oil filtration system based on machine learning according to claim 2, It is characterized in that The method of preprocessing the lubricating oil sample data includes: Build an adaptive encoder; the basic structure of the adaptive encoder includes an encoder and a decoder; Define the loss function for the adaptive encoder Among them, N(μ,σ 2 ) represents a normal distribution, KL() represents the Kullback-Leibler divergence, and β is the weight coefficient; By collecting lubricating oil sample data within a fixed historical time period as training data for the adaptive encoder, in the process of training the adaptive encoder, the parameters of the encoder and decoder are optimized by minimizing the loss function, thereby obtaining a constructed adaptive encoder; The lubricating oil sample data is divided according to data types, and the data corresponding to each data type is sent as a sample to the encoder in the adaptive encoder to obtain a low-dimensional potential feature vector The low-dimensional potential feature vector is clustered to obtain a clustering result, which is N1 clusters. A clustering sample threshold is preset, and clusters with a sample number less than or equal to the clustering sample threshold in the N1 clusters are regarded as abnormal clusters; for each sample, the distance between it and each abnormal cluster is calculated as its abnormal score; The low-dimensional potential feature vector is used as a clustering sample, and the clustering samples of the abnormal score are statistically analyzed to fit their probability density function P(s), where s is the abnormal score; according to the probability density function of the abnormal score, an adaptive abnormal detection threshold τ is set; and the abnormal detection threshold is adaptively updated; samples with an abnormal score greater than or equal to τ are detected as abnormal samples; τ=inf{s:P(S≤s)≥α}; where P(S≤s) represents the cumulative distribution function of the set S of anomaly scores, and α is the confidence level; The abnormal samples are deleted from the lubricating oil sample data to obtain a complete lubricating oil sample data set, and the data in the complete lubricating oil sample data set are normalized to obtain sample feature data.

4. The lubricating oil filtration system based on machine learning according to claim 3, It is characterized in that The encoder maps the input lubricant sample data x to latent features that obey a Gaussian distribution Where μ is the mean of the latent feature, σ is the standard deviation of the latent feature; ε is the random noise of standard normal distribution, and ε~N(0,I), where N(0,I) represents a standard normal distribution with mean 0, variance 1, and each variable is independent of each other, I is the unit matrix, represents element-by-element product; The decoder reconstructs the potential feature z back to the lubricating oil sample data input by the encoder, and the reconstruction formula is: in, is the lubricating oil sample data reconstructed from the latent feature z by the decoder; f dec is the mapping function of the decoder; θ dec Parameters for the decoder.

5. The lubricating oil filtration system based on machine learning according to claim 4, It is characterized in that The method of clustering low-dimensional potential features includes: Define the neighborhood radius r and the minimum number of samples MinPts for clustering; take the low-dimensional potential feature vector as the sample point, and for each sample point, calculate its Euclidean distance from all other sample points. Then join in Neighborhood in, represents the jth sample point among all sample points, represents the i-th sample point among all sample points, is the Euclidean distance between the i-th sample point and the j-th sample point; For each sample point, calculate its density weight. The calculation formula of density weight is: Among them, ρ i is the density weight of the i-th sample point; Sort the density weights of all sample points in ascending order to obtain a sorted sequence and define a density threshold ρ th , where ρ th =ρ (n-MinPts+1) ; Where n is the total number of sample points; For the i-th sample point, if ρ i ≥ρ th , then Marked as a core object, if ρ i Less than ρ th , then calculate The highest density weight ρ in its neighborhood max , if max ≥ρ th , then Marked as a boundary object, if ρ max Still less than ρ th , then Mark as noise object; If there is a arrive The sequence of but Depend on Density can be reached if there is a sample point Make and All by The density can be reached, and Density connected; All density-connected core objects and boundary objects form a cluster, and finally N1 clusters are obtained.

6. A lubricating oil filtration system based on machine learning according to claim 5, It is characterized in that The improved deep belief network model is constructed in the following manner: The basic architecture of the improved deep belief network model is defined to include N2 restricted Boltzmann machines, each of which includes a visible layer and a hidden layer. The visible layer and the hidden layer are connected through a weight matrix, and the hidden layer of the I-th layer restricted Boltzmann machine is used as the visible layer of the I+1-th layer restricted Boltzmann machine; finally, an output layer of supervised training is added to form a complete basic architecture of the improved deep belief network model; Define the energy function E(v (I) ,h (I) ); collect historical data, pre-process the historical data to obtain model input data, use the model input data to train the improved deep belief network model, and optimize the parameters of the improved deep belief network model by minimizing the KL divergence between the distribution of historical data and the distribution of the improved deep belief network model; obtain a preliminary improved deep belief network model; Label the historical data with the target DAG value to obtain a sample data set containing labeled samples; use the sample data set to fine-tune the preliminary improved deep belief network model, and define the goal of fine-tuning to minimize the error between the DAG output by the preliminary improved deep belief network model and the target DAG value. The error is calculated using the square error loss function or the cross entropy loss function; The back propagation algorithm is used to update the parameters of the entire improved deep belief network model, including each restricted Boltzmann machine and the added output layer of the improved deep belief network model, to obtain the completed improved deep belief network model.

7. The lubricating oil filtration system based on machine learning according to claim 6, It is characterized in that The energy function E(v (I) ,h (I) ) is expressed as: Among them, GE(v (I) ,h (I) ) is the connector function; FE(v (I) ,h (I) ) is a cross-level sub-function; Among them, v( I ) is the unit vector of the visible layer of the I-th layer restricted Boltzmann machine, h( I ) is the unit vector of the hidden layer of the I-th layer restricted Boltzmann machine, is the weight between the Jth visible layer and the Kth hidden layer in the Ith layer restricted Boltzmann machine, is the bias of the Jth visible layer in the Ith layer restricted Boltzmann machine, is the Jth visible unit in the Ith layer restricted Boltzmann machine; is the state or value of the Kth hidden unit in the Ith layer restricted Boltzmann machine; is the bias of the Kth hidden layer in the Ith layer restricted Boltzmann machine, n v is the number of units in the visible layer, n h is the number of units in the hidden layer; is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I+1th restricted Boltzmann machine, is the cross-layer connection weight between the Jth visible layer and the lth hidden layer in the Ith to I-1th restricted Boltzmann machine, is the lth visible unit in the I+1th layer restricted Boltzmann machine, is the state or value of the lth hidden unit in the I-1th layer restricted Boltzmann machine.

8. The lubricating oil filtration system based on machine learning according to claim 7, It is characterized in that The calculation method of the preliminary degradation trend series includes: Clean the working environment data to obtain pure working environment data, and extract environmental features based on the pure working environment data. The environmental features include time features, statistical features, and cross features. Extract sample features from the dependent acyclic graph. Sample features include the shortest undirected path length feature between nodes and node centrality. Node centrality includes degree centrality, betweenness centrality and closeness centrality. The method of extracting the sample features includes: Convert the directed acyclic graph to an undirected graph. For each pair of nodes in the undirected graph, use breadth-first search or Dijkstra algorithm to calculate the shortest path length between them, and express it with an adjacency matrix. Finally, a matrix D is obtained, which is the shortest undirected path length feature between nodes. For node u in an undirected graph, calculate the number of neighbor nodes ku in the undirected graph; Degree Centrality Among them, N 3 is the total number of nodes; for node u in an undirected graph, calculate the number of shortest path lengths of all other pairs of nodes njk. Among the njk shortest paths, the number of paths passing through node u is njk(u); then the betweenness centrality For a node u in an undirected graph, calculate the shortest path length dau from it to all other nodes a. Close to centrality The sample features and environmental features are input into the constructed and trained random forest model to obtain the degradation degree change value of each node in the next t time steps. The degradation degree change values ​​are connected in series to obtain the preliminary degradation trend sequence of the node in the next t time steps.

9. The lubricating oil filtration system based on machine learning according to claim 8, It is characterized in that The random forest model is constructed in the following ways: Collect historical comprehensive data, which includes lubricating oil sample data within a fixed historical period and working condition environment data corresponding to the lubricating oil sample data; The historical comprehensive data is engineered to obtain the corresponding environmental characteristics and sample characteristics; the environmental characteristics and sample characteristics are combined into a data set, and the data set is annotated with node degradation degree labels; The labeled data set is divided into a training set and a test set in chronological order. For the training set, a decision tree is selected as the base learner, and N5 decision trees are constructed. For each decision tree, some features are randomly extracted from the sample features and the environmental features with replacement as the split features of the decision tree nodes. At each decision tree node, an optimal feature is randomly selected from the split features for splitting. The maximum height is defined, and each decision tree stops growing when it reaches the maximum height to obtain the initial random forest. Input the training set into the initial random forest. After the training is completed, each decision tree corresponds to a prediction result; and the initial random forest model is obtained; The loss function is defined as the root mean square error function, and the function value of the loss function of the preliminary random forest model is calculated on the test set. When the function value of the loss function does not decrease for p consecutive times on the test set, a grid search or random search method is used in this process to adjust the hyperparameters of the random forest model to obtain a constructed and trained random forest model.

10. The lubricating oil filtration system based on machine learning according to claim 9, It is characterized in that The output of the filtering decision includes: The preliminary degradation trend sequence contains the degradation degree change values ​​of each node in the future t time steps, and these degradation degree change values ​​are used as the input of the attention fusion algorithm; Constructing a filtering decision output model, which includes an attention fusion layer, a calculation layer, and a decision layer; The attention fusion layer consists of a query vector, a key vector, and a value vector. The query vector is the initial degradation trend sequence of the node making the decision, and the key vector and value vector are the initial degradation trend sequences of all other nodes. The calculation layer calculates the similarity between any query vector q' and all key vectors to obtain the corresponding attention score e q' =q' T ·K' w ; where q' T represents the transpose of the query vector q', K' w is the w-th key vector; the attention score is scaled by the Softmax function to obtain the corresponding final attention score; Perform weighted summation of the final attention score and the corresponding value vector to obtain the attention value vector; The fully connected layer is used as the decision layer, the attention value vector is used as the input of the decision layer, and the preliminary filtering decision probability distribution o = σ(W'·A'+b') is output; where W' is the weight matrix of the fully connected layer, b' is the bias vector, σ is the Softmax activation function, and the category with the largest probability in the preliminary filtering decision probability distribution is taken as the decision result, that is, the filtering decision; the categories include early filtering, normal filtering and delayed filtering.

Images