Equipment lubricating oil intelligent regulation and control system based on big data

Abstract

The invention discloses an equipment lubricating oil intelligent regulation and control system based on big data, and the system comprises the steps: collecting and fusing multi-dimensional lubricating state original observation data and operation environment parameters in real time through a multi-modal sensing network, constructing a structure-standardized multi-source observation matrix, and completing the preliminary abnormality screening at an edge calculation node; multi-scale gradient response differential features are extracted, a state dynamic feature vector is generated, and a state evolution cluster set reflecting a lubrication state evolution path is constructed; constructing a nested hierarchical graph structure based on the evolution cluster, and generating a high-interpretability state inference map by using a graph attention mechanism; the deviation between the current state and the optimal lubrication window is evaluated in real time, and a lubrication strategy is output in a self-adaptive mode through a feedback enhancement type control algorithm; and finally, the execution module implements a control instruction and feeds back a result to form a closed-loop self-learning regulation and control system. Through self-learning quantitative characterization and a high-interpretability intelligent regulation and control mechanism, precise monitoring and dynamic management of the state of the lubricating oil are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of lubricating oil control, and in particular to an intelligent control system for equipment lubricating oil based on big data. Background Art

[0002] As industrial equipment evolves toward intelligent and highly reliable systems, lubricating oil, a critical operating medium, has a direct impact on its lifespan and operational safety. Existing online monitoring systems based on big data already collect key physical and chemical indicators of lubricating oil, such as viscosity, acid value, moisture content, particle contamination, and metal abrasive concentration, in real time. By analyzing this multidimensional data, they dynamically monitor lubricating oil degradation, contamination, and mechanical wear, providing timely warnings of potential lubrication failures and major accidents, ensuring stable equipment operation. These systems are widely used in industries such as wind power, mining, and metallurgy, significantly advancing the intelligentization of equipment maintenance. However, current lubricating oil condition monitoring models are mostly based on expert experience and preset fixed rules. Their model development relies on expert knowledge and manually set thresholds, lacking the deep learning capabilities to capture the nonlinear correlations and dynamic evolution of multiple variables under complex operating conditions. This makes the systems difficult to adapt to the diverse operating environments and complex operating conditions. Furthermore, fixed rule models often struggle to self-optimize and dynamically adjust in real time, limiting the accuracy of lubrication condition warnings and the intelligent implementation of control strategies.

[0003] Faced with the complex and ever-changing equipment operating environment and the actual working conditions where multiple factors intertwine the lubricating oil state, traditional monitoring methods and empirical rule models have difficulty in achieving accurate quantitative characterization of the lubricating oil state. The lack of a systematic self-learning mechanism leads to poor state recognition and may even result in false alarms or missed alarms. At the same time, the existing models have a strong "black box" characteristic and are unable to provide a clear and explainable state evolution path and decision-making basis, which limits the operation and maintenance personnel's understanding of the health status of the equipment and their confidence in intervention operations. Therefore, how to build a lubricating oil state quantitative characterization model with self-learning capabilities based on big data technology, which can automatically extract key features from complex multi-source time series data, capture the dynamic evolution of oil degradation and contamination, and further realize the model's explainable reasoning and intelligent regulation, is a core problem in improving the level of intelligent equipment lubrication management. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides an intelligent control system for equipment lubricating oil based on big data.

[0005] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0006] An intelligent control system for equipment lubricating oil based on big data, comprising:

[0007] Lubrication status multi-source data acquisition module: This module builds a multimodal sensor network to collect and integrate multi-dimensional lubrication status raw observation data and operating environment parameters in real time, constructs a standardized lubrication status multi-source observation matrix, and performs preliminary anomaly screening through edge computing nodes.

[0008] State representation generation module: used to extract multi-scale gradient response differential features from the observation matrix, generate state dynamic feature vectors with mutation response and trend modeling capabilities, and construct a state evolution cluster set reflecting the evolution path of the lubrication state based on this;

[0009] Self-learning state parsing module: This module is used to construct a nested hierarchical graph structure of cross-time window and cross-variable dependencies based on state evolution clusters. It uses a graph attention mechanism to mine the interaction paths between key variables and generate a highly interpretable inference graph of the lubrication state.

[0010] Lubrication strategy optimization module: This module calculates the deviation between the current lubrication state and the optimal lubrication window in real time, selects a dynamic lubrication intervention strategy through a feedback-enhanced deviation control algorithm, and achieves adaptive evolution of the strategy and optimal lubrication control output.

[0011] Control execution module: It is used to execute the control instructions output by the lubrication strategy generation module, complete the lubrication operation through the control valve group, and transmit the execution results and operation responses back to the observation matrix to form a closed-loop self-learning circuit to improve the control performance.

[0012] Further: the multimodal sensing network includes:

[0013] An oil state sensing subsystem, an equipment operation state sensing subsystem, and an environmental context sensing subsystem are constructed. By performing time synchronization and feature normalization on the data collected by the above three subsystems, a multi-source observation matrix of lubrication state in a unified format is constructed, and the standardized time series input matrix is ​​generated by stacking the sampling time series.

[0014] Further: the state representation generation module includes:

[0015] receiving a lubrication state multi-source observation matrix output by a lubrication state multi-source data acquisition module;

[0016] Construct a multi-scale differential response tensor, define multiple time scale windows, and extract the differential response set of each variable in the time domain of the multi-source observation matrix of the lubrication state;

[0017] Extract the local mutation response factor to measure the multi-scale local fluctuation intensity of the oil state at each moment;

[0018] Construct a global evolution trend code, use a sliding window and extract principal component vectors through PCA to model long-term trend evolution;

[0019] Integrate local mutations with global trends to construct a state evolution tensor;

[0020] The density-difference collaborative metric criterion and the dynamic density-driven clustering algorithm are used to cluster the state evolution tensor under unsupervised conditions and output the state evolution cluster label sequence.

[0021] Further: the self-learning state analysis module includes:

[0022] Receive the state evolution tensor and state cluster label sequence output by the state representation generation module;

[0023] Construct a nested hierarchical graph structure and construct a subgraph at each time step. The subgraph nodes represent the feature dimensions, and the edge weights are defined jointly by the Pearson correlation coefficient and the first-order difference gradient of the variable.

[0024] Construct a parent graph structure, with nodes corresponding to each subgraph. The edge weight is calculated by the trend similarity of the state tensors of adjacent time steps, and at the same time, it is determined whether they belong to the same state evolution cluster;

[0025] Design a nested hierarchical graph attention mechanism, the subgraph attention mechanism is used to extract the key dependency structure between local variables, and the cross-graph attention mechanism is used to model the time evolution path;

[0026] The nested graph structure is integrated with the attention weight to generate a state reasoning graph that includes the importance of subgraph structures and cross-time evolution pathways.

[0027] Outputs a structured oil state reasoning graph, the evolution embedding vectors at each time point, and the sub-graph level variable attention scores.

[0028] Further: the lubrication strategy optimization module includes:

[0029] Receive the state reasoning graph and the evolution embedding vector at each time point output by the state parsing module;

[0030] Construct a state reconstruction function to map the evolved embedding vector back to the state structure feature space and calculate the state deviation between the current state and the target lubrication window;

[0031] Identify deviation types based on state deviation values, including excessive degradation, sudden pollution increase, and abnormal wear;

[0032] Matching intervention strategy sets according to different deviation types;

[0033] Construct a state-deviation-intervention-feedback trajectory, introduce a strategy value function based on reinforcement learning ideas, and review and evaluate historical intervention effects;

[0034] Adopting feedback-enhanced deviation control algorithm, dynamically adjust intervention action priority, strengthen the weight of previous effective intervention path, and generate the current optimal lubrication control strategy;

[0035] Output the intervention action, parameter combination and strategy confidence of the current time step for the control execution module to call and implement.

[0036] Further: the control execution module includes:

[0037] Receive the lubrication intervention action, parameter combination and strategy confidence output by the lubrication strategy optimization module;

[0038] Call the multi-channel oil circuit control valve group integrated with the lubrication system to control the execution channel of the replenishment, replacement, dilution or regeneration operation according to the instructions;

[0039] The control execution module accurately controls the execution process of various lubrication operations based on the operating flow, duration and proportional coefficient in the parameter combination;

[0040] Real-time acquisition of key control node status during operation;

[0041] After the lubrication operation is completed, the short-cycle rapid sampling mechanism is activated to collect the latest round of original observation data on the oil status and equipment operation response indicators;

[0042] The state feature difference before and after the lubrication operation and the execution action are encapsulated into an execution-feedback event package and pushed to the lubrication state multi-source data acquisition module to complete the construction of the closed-loop data path.

[0043] Compared with the prior art, the present invention has the following technical advances:

[0044] First, by introducing a multi-source, multi-dimensional data acquisition mechanism, combined with a proprietary multi-scale gradient response differential feature extraction algorithm, this invention achieves dynamic quantitative characterization of lubricating oil conditions. This allows for automatic identification of subtle changes in oil degradation, contamination, and mechanical wear from massive amounts of time-series data, effectively avoiding the risk of misjudgment caused by artificial threshold setting. This effectively enhances the system's ability to sensitively capture changing lubrication conditions under complex operating conditions and significantly improves the accuracy of condition identification.

[0045] Secondly, the present invention adopts a nested hierarchical graph attention mechanism to construct a self-learning state parsing model, which has the ability to model dependencies across time windows and variables, can adaptively capture the inherent laws of lubrication state evolution, and achieve highly explainable reasoning of lubricating oil state changes, providing intuitive and reliable diagnostic basis for operation and maintenance personnel, solving the pain point that traditional "black box" models are difficult to explain the operating status, thereby improving the scientificity and transparency of equipment maintenance decisions.

[0046] Thirdly, combined with the feedback-enhanced deviation control algorithm, the present invention has the ability to optimize dynamic lubrication strategies. It can automatically generate the optimal lubrication intervention plan based on the deviation between the real-time monitoring status and the ideal lubrication window, support diversified operations such as replenishment, replacement, dilution and regeneration, effectively cope with complex and changeable operating environments, realize intelligent and adaptive lubrication control, and avoid the risk of equipment failure caused by strategy lag or inaccuracy.

[0047] Finally, the closed-loop intelligent control execution module of the present invention ensures a complete closed loop from strategy generation to execution and feedback, and uses real-time feedback data to continuously optimize model performance and control strategies, truly realizing self-learning and dynamic evolution of lubricating oil status, greatly improving the safety and reliability of equipment operation.

[0048] In summary, the present invention realizes accurate monitoring and dynamic management of lubricating oil status through big data-driven self-learning quantitative representation and highly interpretable intelligent control mechanism, significantly enhancing the system's adaptability and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0050] In the attached figure:

[0051] Figure 1 This is a system structure diagram of the present invention. DETAILED DESCRIPTION

[0052] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below with reference to the accompanying drawings.

[0053] like Figure 1 As shown, the present invention discloses an intelligent control system for equipment lubricating oil based on big data, comprising:

[0054] Lubrication status multi-source data acquisition module: This module builds a multimodal sensor network to collect and integrate multi-dimensional lubrication status raw observation data and operating environment parameters in real time, constructs a standardized lubrication status multi-source observation matrix, and performs preliminary anomaly screening through edge computing nodes.

[0055] State representation generation module: used to extract multi-scale gradient response differential features from the observation matrix, generate state dynamic feature vectors with mutation response and trend modeling capabilities, and construct a state evolution cluster set reflecting the evolution path of the lubrication state based on this;

[0056] Self-learning state parsing module: This module is used to construct a nested hierarchical graph structure of cross-time window and cross-variable dependencies based on state evolution clusters. It uses a graph attention mechanism to mine the interaction paths between key variables and generate a highly interpretable inference graph of the lubrication state.

[0057] Lubrication strategy optimization module: This module calculates the deviation between the current lubrication state and the optimal lubrication window in real time, selects a dynamic lubrication intervention strategy through a feedback-enhanced deviation control algorithm, and achieves adaptive evolution of the strategy and optimal lubrication control output.

[0058] Control execution module: It is used to execute the control instructions output by the lubrication strategy generation module, complete the lubrication operation through the control valve group, and transmit the execution results and operation responses back to the observation matrix to form a closed-loop self-learning circuit to improve the control performance.

[0059] Module 1 Lubrication status multi-source data acquisition module

[0060] This module builds a multimodal sensor network to collect multi-source observation data on lubrication status in real time, including oil physical and chemical characteristics, mechanical state characteristics, and environmental context variables. Through multi-level data fusion and edge computing, it constructs a well-structured multi-source observation matrix of lubrication status, providing a high-quality data foundation for subsequent dynamic characterization generation modules. Specifically, it includes:

[0061] 1. Construction and Deployment of Multimodal Sensing Networks

[0062] A distributed multimodal sensing network covering the core components of device operation is used. The multimodal sensing network includes three types of sensing subsystems:

[0063] 1. Oil state sensing subsystem: used to collect real-time physical and chemical characteristics of oil, including:

[0064] Viscosity μ t ,temperature Moisture content Acid value A t , particle pollution P t , metal abrasive concentration M t .

[0065] 2. Equipment operation status sensing subsystem: used to capture mechanical operation load information, including:

[0066] Vibration amplitude V t , load torque L t , current fluctuation ΔI t .

[0067] 3. Environmental context sensing subsystem: used to obtain operating environment parameters, including:

[0068] Ambient temperature Ambient humidity

[0069] 2. Construction method of multi-source observation matrix of lubrication state

[0070] By performing time synchronization and feature normalization on the data collected by the above three subsystems, a unified format multi-source observation matrix of lubrication state is constructed. Where m = 11, the variables are arranged as follows:

[0071]

[0072] This matrix is ​​generated by stacking the sampled time series:

[0073]

[0074] X 1:T A standardized time series input matrix consisting of oil characteristics, equipment status, and environmental variables serves as the data entry for the system and will run through the processing flow of all subsequent modules. All state representation, reasoning, and strategies are based on this structure.

[0075] 3. Edge Computing and Preprocessing Mechanism

[0076] To ensure data quality and data inflow efficiency, the lubrication status multi-source data acquisition module integrates the following three-layer preprocessing process:

[0077] 1. Signal synchronization layer:

[0078] All sensor channel signals are normalized by sampling frequency (in this embodiment, the sampling is unified at 10 Hz), and timestamp alignment is achieved through a clock synchronization chip. Dynamic window sliding mean correction is used for sudden change signals to suppress instantaneous interference fluctuations.

[0079] 2. Edge computing layer:

[0080] The primary algorithm module deployed on the device-side microprocessor performs abnormal data point elimination (such as 3σ outlier elimination) and missing value interpolation (such as based on forward filling or sliding mean), filters out short-term fluctuations, and improves data continuity and stability.

[0081] 3. Feature Normalization Layer:

[0082] All physical quantities are uniformly converted into standardized dimensions (such as normalized to the interval [0,1]) to maintain the weight balance between variables and provide clean input for subsequent feature extraction modules.

[0083] The output of this module is X 1:TAs the basis for system state observation, it will serve as the direct input of the state dynamic representation generation module in module 2 to extract the time series representation tensor with both sudden response and gradual evolution characteristics, and will run through the entire state reasoning and lubrication strategy formulation process in subsequent modules.

[0084] Module 2: State representation generation module driven by timing features

[0085] This module is used to receive the oil state multi-source observation matrix output by the oil state multi-source observation matrix construction module, and perform time-series structured modeling on the matrix through the multi-scale gradient response differential feature extraction algorithm to generate a state evolution tensor that can reflect the dynamic evolution characteristics of the oil operation state. And construct its corresponding state cluster label sequence y 1:T , as the basis for graph modeling in subsequent modules.

[0086] 1. Input and data structure

[0087] The input is the multi-source observation matrix of the oil state output by module 1, which is expressed as:

[0088]

[0089] Where T represents the time window length, Represents the D-dimensional feature vector collected at the tth time step, including but not limited to oil viscosity, oil temperature, moisture content, acid value, particle contamination, metal abrasive concentration, vibration, load, current fluctuation and environmental parameters.

[0090] 2. Multi-scale gradient response differential feature extraction algorithm

[0091] The algorithm extracts dynamic time series features that combine local mutation responses with global evolution trends through the following three steps:

[0092] 2.1 Multi-scale difference construction

[0093] With multiple time domain window scales k∈{k1,k2,...,k m}Build the gradient response tensor:

[0094]

[0095] Get the multi-scale differential response set:

[0096]

[0097] 2.2 Local mutation-sensitive coding

[0098] Define the local mutation response factor:

[0099]

[0100] This indicator is used to reflect the local fluctuation intensity of the oil state at multiple scales at the current moment.

[0101] 2.3 Global Evolution Trend Modeling

[0102] Use sliding windows to construct evolution trend encoding:

[0103]

[0104] Where ω is the sliding window length, and PCA extracts the first d principal components to form the trend vector.

[0105] 3. State Evolution Tensor Generation

[0106] Finally, the local mutation code and the global trend vector are integrated to form the state evolution representation at time point t:

[0107]

[0108] The state evolution tensor that constitutes the entire time series:

[0109]

[0110] 4. State cluster label sequence generation (y 1:T )

[0111] In order to further explore the evolution law of state representation, a weakly supervised adaptive clustering mechanism is used to map the state evolution tensor S into a semantically interpretable state evolution pattern:

[0112] The density-divergence criterion (DDC) is defined as follows:

[0113]

[0114] Where ρ(i) represents the local density estimate at time point i, which is implemented based on K nearest neighbors:

[0115]

[0116] A dynamic density-driven clustering discriminant algorithm is used to perform non-parametric clustering of the state space based on the above metric function, that is, the following operations are performed:

[0117] The initial center point of each cluster was determined by local density peak detection;

[0118] Density propagation is performed based on the DDC adjacency weight, and non-center points are classified into similar density areas;

[0119] For points with discontinuous density (such as mutation nodes), their isolated labels are retained to form abnormal state clusters.

[0120] After the above clustering process is completed, a corresponding cluster label is assigned to each time step, and the cluster label of each time point is finally output:

[0121]

[0122] 5. The final output includes:

[0123] State Evolution Tensor

[0124] State cluster label sequence

[0125] The above output will serve as the graph structure modeling input of Module 3: Self-learning State Parsing Module Based on Nested Hierarchical Graph Attention Mechanism, forming graph node embeddings and node type annotations respectively, driving the generation and reasoning of highly interpretable state graphs.

[0126] Module 3: Self-learning state parsing module based on nested hierarchical graph attention mechanism

[0127] This module is used to receive the state evolution tensor output by the state representation generation module driven by time series features. and its corresponding state cluster label sequence By constructing a nested hierarchical graph structure with the ability to model dependencies across time windows and feature dimensions, combined with a custom-designed Nested Hierarchical Graph Attention Mechanism (NH-GAM), the interaction structure and evolution path between variables are automatically learned, thereby generating a highly interpretable oil state inference graph, realizing adaptive inference and quantitative representation of lubrication status under complex operating conditions.

[0128] 1. Input structure

[0129] State evolution tensor: in

[0130] State cluster label sequence: y 1:T =[y1,y2,...,y T ].

[0131] 2. Graph structure modeling: nested hierarchical graph construction

[0132] This module introduces a two-layer nested graph structure:

[0133] 2.1 Sublayer (variable relationship diagram within the time window)

[0134] Construct a subgraph for each time slice t in:

[0135] Represents each feature dimension at time point t;

[0136] ε (t) : The edge weights generated by combining the Pearson correlation coefficient with the mutation response constitute the inter-variable dependency matrix:

[0137]

[0138] in represents the first-order differential gradient of the i-th dimension with respect to time, and α is the weight adjustment coefficient.

[0139] 2.2 Parent layer (cross-time state dependency graph)

[0140] Construct a parent graph in:

[0141] node A subgraph representing each time point;

[0142] Edge weights are modeled by trend similarity based on the state evolution tensor:

[0143]

[0144] in Indicates belonging to the same state evolution cluster, otherwise it is 0.

[0145] 3. Nested Hierarchical Graph Attention Mechanism (NH-GAM)

[0146] To achieve end-to-end training of graph structures, we design a nested hierarchical graph attention mechanism NHGAM, which consists of two parts:

[0147] 3.1 Intra-Graph Attention

[0148] For each subgraph Based on the learnable attention weight matrix A (t) To encode a variable node:

[0149]

[0150] 3.2 Inter-Graph Attention

[0151] Based on subgraph embedding vector Tectonic state evolution path:

[0152]

[0153]

[0154] 4. State Reasoning Graph Generation and Explanatory Construction

[0155] The final evolution path vector z t Combined into a dynamic graph:

[0156]

[0157] By visualizing the changing trends and attention scores of graph nodes, a graphical explanation path for key processes such as oil condition degradation, contamination mutation, and mechanical wear is provided, which serves as the core logical input for the generation of subsequent lubrication intervention strategies.

[0158] 5. The final output of this module:

[0159] High-dimensional interpretable state reasoning graph

[0160] The evolving embedding vector z at each time point t ;

[0161] Subgraph-level variable importance attention score

[0162] This result will serve as the core input of Module 4: Lubrication Strategy Optimization Module Based on Feedback Enhanced Deviation Control Algorithm, which will be used to dynamically identify the current state deviation and formulate corresponding intervention plans.

[0163] Module 4: Lubrication strategy optimization module based on feedback enhanced deviation control algorithm

[0164] This module is used to receive the state reasoning graph output by the self-learning state parsing module of the nested hierarchical graph attention mechanism Combined with the designed lubrication parameter range of equipment operation, based on the proposed Feedback-Enhanced Deviation Regulation Algorithm (FE-DRA), a multi-dimensional deviation space between the current lubrication state and the target optimal lubrication range is constructed in real time, the deviation type and severity are dynamically identified, and the optimal lubrication intervention operation mode (such as replenishment, replacement, dilution, regeneration, etc.) is decided accordingly. Finally, a standardized strategy control instruction stream is output for execution by the subsequent intelligent control execution module.

[0165] 1. Lubrication state deviation modeling

[0166] Assume that the optimal lubrication interval corresponding to the ideal operation of the equipment is the tensor Represents the upper and lower limits of each state dimension; the state representation vector at the current time point is It is deduced from module 3. First, we need to construct a mapping function to embed the latent space into z t Mapping back to the original state space:

[0167]

[0168] where f inv Reconstruct the inverse state function obtained by pseudo-inverse learning or residual regression training, and then define the bias vector:

[0169]

[0170] in,

[0171] 2. Bias Classification and Intervention Selection

[0172] According to Δ t The deviation type is identified based on the sign and magnitude of each dimension, forming three core intervention scenarios:

[0173] Deterioration beyond limit type (D-type): such as viscosity, acid value, and moisture exceeding the upper limit;

[0174] Sudden pollution increase type (C-type): such as sudden increase in particle contamination and metal abrasive concentration;

[0175] Abnormal wear type (W-type): Such as vibration, load fluctuation and severe deviation.

[0176] Construct the deviation label function:

[0177]

[0178] The set of intervention strategies corresponding to each type of deviation The definition is as follows:

[0179]

[0180]

[0181] 3. Strengthen feedback path construction

[0182] In order to improve the long-term effect and adaptability of the intervention strategy, FE-DRA constructs the state deviation intervention history trajectory matrix {(s τ ,Δ τ ,a τ ,r τ )} τ＜t , where a τ represents intervention operation, r τ Provide immediate rewards based on the degree of system improvement after intervention.

[0183] Estimate the reinforcement path through the policy value function:

[0184]

[0185] Adopting a soft-update reinforcement learning weight mechanism to adaptively optimize the strategy:

[0186]

[0187] 4. The module outputs a structured lubrication intervention strategy:

[0188] The optimal intervention action at the current time point:

[0189] Multi-channel control parameter package (including flow, time, ratio, etc.):

[0190] Intervention type identifier and confidence level: t ,Confidence t )

[0191] All outputs are encoded into a standard control instruction stream and fed into the intelligent control execution module for operational-level execution. Simultaneously, intervention results and oil response are fed back to the acquisition end of Module 1 via a closed-loop path, completing the data self-learning chain and forming a closed-loop self-evolution mechanism for the system.

[0192] Module 5 Control Execution Module

[0193] This module receives standardized control command streams from the lubrication strategy optimization module, which is based on a feedback-enhanced deviation control algorithm. It then executes selected lubrication interventions, including replenishment, replacement, dilution, and regeneration, through a multi-channel oil circuit control valve network deeply integrated with the equipment's lubrication system. Simultaneously, it collects raw observation data and operational response indicators of the oil state within a short post-lubrication period in real time and transmits them back to the multi-source lubrication state data acquisition module, closing the system's state control self-learning loop. This allows for dynamic verification, self-correction, and continuous evolution of the lubrication control strategy under real-world operating conditions.

[0194] 1. Control instruction flow structure and receiving mechanism

[0195] The control instruction stream output by module 4 is defined as a triple:

[0196]

[0197] in:

[0198] The optimal intervention operation type at the current time point (e.g., “regeneration,” “dilution,” or “filtration”);

[0199] Indicates a specific operation parameter package (such as flow rate, duration, scale factor, etc.);

[0200] Confidence t ∈[0,1] represents the confidence of the policy decision, which is used to perform fault tolerance threshold judgment.

[0201] The instruction flow is converted into a low-level control signal compatible with the on-site PLC or DCS through a communication protocol converter (such as CAN, Modbus or EtherCAT), and the corresponding operation is dispatched to each lubrication subsystem through the distributed execution unit.

[0202] 2. Multi-channel lubrication operation execution mechanism

[0203] This module integrates a multi-channel oil circuit control valve group network and supports the following typical execution paths:

[0204] Channel A: oil replenishment (precisely control injection volume and flow rate);

[0205] Channel B: Full oil replacement (controls the timing of old oil extraction and new oil injection);

[0206] Channel C: molecular dilution (adjusting the diluent injection concentration based on the proportional valve);

[0207] Channel D: Offline regeneration (combined with thermal cracking or magnetic adsorption module for oil reprocessing).

[0208] Before and after lubrication tasks are executed, all operations are verified by an embedded status confirmation submodule to collect physical signals from key nodes, such as valve opening and closing status, flow counting feedback, and pressure fluctuation response, to ensure the safety of the closed-loop operation.

[0209] 3. Control response feedback and data return path

[0210] After each lubrication operation is completed, this module will trigger a short-term data rapid sampling action to collect control response data groups including but not limited to the following:

[0211] Oil status original observation data (same format as module 1)

[0212] Fast reconstruction of state representation vector

[0213] State before operation The evolution difference of:

[0214]

[0215] Current execution action and response time τt .

[0216] Finally, a complete execution feedback event package is formed:

[0217]

[0218] This feedback data structure will be synchronously transmitted back to the edge cache area of ​​the lubrication status multi-source data acquisition module and marked as an "intervention sample" for use in the subsequent training of all state feature extraction, state analysis and strategy generation modules, closing the self-learning path and promoting continuous self-optimization of the strategy system in a real environment.

[0219] Also includes Module 6: Interpretable Knowledge Generation and Human-Computer Interaction

[0220] This module is used to convert the oil state inference graph output by the self-learning state parsing module based on the nested hierarchical graph attention mechanism into The strategy evolution path generated by the lubrication strategy optimization module based on the feedback-enhanced deviation control algorithm is converted into graphical, semantically interpretable knowledge content that can be understood by humans and used as a reference for operation and maintenance decision-making. At the same time, feedback from operation and maintenance personnel is collected through a multi-channel interactive interface to build a human-machine collaborative adaptation mechanism that supports iterative learning, thereby improving the system's response robustness and expert co-construction capabilities under unknown working conditions.

[0221] 1. Graph structure visualization and semantic interpretation generation

[0222] The system first transforms the structured reasoning graph Including local subgraphs (Variable interaction structure within the time window) and the global parent graph (cross-temporal evolution dependency path), perform visualization and expansion, use the graph neural layout compression algorithm LayoutRedue(·) to extract high semantic density node areas, and generate a structural diagram:

[0223]

[0224] In conjunction with the Natural Language Generation Engine (NLG Engine), this module performs event-driven semantic filling for each key node-edge structure in the graph, using a template-based interpretation generation formula:

[0225]

[0226] in:

[0227] z (i) : the temporal state representation vector corresponding to the current node;

[0228] δz (i) : The differential evolution trend before and after the state change;

[0229] The optimal control action type that triggers this state transition.

[0230] For example:

[0231] [Mutation Template]: "A sudden increase of [Δx] in the variable [metal wear particle concentration] was detected within [time window T], and the [filtration + replenishment] combined lubrication strategy has been activated."

[0232] [Trend template]: "Over the past [N hours], [oil acid value] has shown a slowly increasing trend and is predicted to exceed the warning threshold at [t+Δx]Δt].

[0233] Finally, a set of semantic interpretation sequences is formed:

[0234] ε={Explanation1,Explanation2,...Explanation n}

[0235] The semantic set will be output synchronously with the graph visualization results for operation and maintenance personnel to interactively view in the interface.

[0236] 2. Human-computer interaction and operation and maintenance feedback closed loop

[0237] This module is configured with a WebHMI-based interactive interface and API middle platform, which supports operation and maintenance personnel to provide three types of feedback after each round of lubrication strategy execution:

[0238] Evaluation feedback (Quality Rating): subjective rating of lubrication effect from 0 to 5;

[0239] Expert Notes: Provides suggestions for strategy replacement, label corrections, and environmental descriptions.

[0240] Manual Override: Detailed record of deviations between manual intervention and system instructions.

[0241] All feedback information will be encoded into a standard structure:

[0242] mathcal{R}_t=(text{Score}_t, text{Note}_t, text{Override}_t)

[0243]

[0244] It is then transmitted back to the lubrication state multi-source data acquisition module and the state representation generation module to participate in subsequent state cluster extraction and graph structure update in the form of "high-confidence labeled samples".

[0245] 3. Construction of interactive knowledge base and abnormal dictionary

[0246] This module continuously records confirmed strategy maps, explanation semantics and manual feedback into the interpretability knowledge base. The build includes:

[0247] Typical state evolution sequence → strategy matching path → execution effect score;

[0248] Abnormal mutation pattern → high response variable path → causal inference annotation;

[0249] Feedback instruction deviation set → strategy adjustment rules → self-supervisory re-injection path.

[0250] In addition, the module constructs a "semantic dictionary of abnormal working conditions" to enhance the ability to recognize low-confidence graphs in unknown scenarios and expand the self-explanatory capabilities under weak labels.

[0251] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. An intelligent control system for equipment lubricating oil based on big data, characterized in that: include: Lubrication status multi-source data acquisition module: This module builds a multimodal sensor network to collect and integrate multi-dimensional lubrication status raw observation data and operating environment parameters in real time, constructs a standardized lubrication status multi-source observation matrix, and performs preliminary anomaly screening through edge computing nodes. State representation generation module: used to extract multi-scale gradient response differential features from the observation matrix, generate state dynamic feature vectors with mutation response and trend modeling capabilities, and construct a state evolution cluster set reflecting the evolution path of the lubrication state based on this; Self-learning state parsing module: This module is used to construct a nested hierarchical graph structure of cross-time window and cross-variable dependencies based on state evolution clusters. It uses a graph attention mechanism to mine the interaction paths between key variables and generate a highly interpretable inference graph of the lubrication state. Lubrication strategy optimization module: This module calculates the deviation between the current lubrication state and the optimal lubrication window in real time, selects a dynamic lubrication intervention strategy through a feedback-enhanced deviation control algorithm, and achieves adaptive evolution of the strategy and optimal lubrication control output. Control execution module: It is used to execute the control instructions output by the lubrication strategy generation module, complete the lubrication operation through the control valve group, and transmit the execution results and operation responses back to the observation matrix to form a closed-loop self-learning circuit to improve the control performance.

2. The intelligent control system for equipment lubricating oil based on big data according to claim 1 is characterized in that: The multimodal sensing network comprises: An oil state sensing subsystem, an equipment operation state sensing subsystem, and an environmental context sensing subsystem are constructed. By performing time synchronization and feature normalization on the data collected by the above three subsystems, a multi-source observation matrix of lubrication state in a unified format is constructed, and the standardized time series input matrix is ​​generated by stacking the sampling time series.

3. The intelligent control system for equipment lubricating oil based on big data according to claim 2 is characterized in that: The state representation generation module includes: receiving a lubrication state multi-source observation matrix output by a lubrication state multi-source data acquisition module; Construct a multi-scale differential response tensor, define multiple time scale windows, and extract the differential response set of each variable in the time domain of the multi-source observation matrix of the lubrication state; Extract the local mutation response factor to measure the multi-scale local fluctuation intensity of the oil state at each moment; Construct a global evolution trend code, use a sliding window and extract principal component vectors through PCA to model long-term trend evolution; Integrate local mutations with global trends to construct a state evolution tensor; The density-difference collaborative metric criterion and the dynamic density-driven clustering algorithm are used to cluster the state evolution tensor under unsupervised conditions and output the state evolution cluster label sequence.

4. The intelligent control system for equipment lubricating oil based on big data according to claim 3 is characterized in that: The self-learning state analysis module includes: Receive the state evolution tensor and state cluster label sequence output by the state representation generation module; Construct a nested hierarchical graph structure and construct a subgraph at each time step. The subgraph nodes represent the feature dimensions, and the edge weights are defined jointly by the Pearson correlation coefficient and the first-order difference gradient of the variable. Construct a parent graph structure, with nodes corresponding to each subgraph. The edge weight is calculated by the trend similarity of the state tensors of adjacent time steps, and at the same time, it is determined whether they belong to the same state evolution cluster; Design a nested hierarchical graph attention mechanism, the subgraph attention mechanism is used to extract the key dependency structure between local variables, and the cross-graph attention mechanism is used to model the time evolution path; The nested graph structure is integrated with the attention weight to generate a state reasoning graph that includes the importance of subgraph structures and cross-time evolution pathways. Outputs a structured oil state reasoning graph, the evolution embedding vectors at each time point, and the sub-graph level variable attention scores.

5. The intelligent control system for equipment lubricating oil based on big data according to claim 4 is characterized in that: The lubrication strategy optimization module includes: Receive the state reasoning graph and the evolution embedding vector at each time point output by the state parsing module; Construct a state reconstruction function to map the evolved embedding vector back to the state structure feature space and calculate the state deviation between the current state and the target lubrication window; Identify deviation types based on state deviation values, including excessive degradation, sudden pollution increase, and abnormal wear; Matching intervention strategy sets according to different deviation types; Construct a state-deviation-intervention-feedback trajectory, introduce a strategy value function based on reinforcement learning ideas, and review and evaluate historical intervention effects; Adopting feedback-enhanced deviation control algorithm, dynamically adjust intervention action priority, strengthen the weight of previous effective intervention path, and generate the current optimal lubrication control strategy; Output the intervention action, parameter combination and strategy confidence of the current time step for the control execution module to call and implement.

6. The intelligent control system for equipment lubricating oil based on big data according to claim 5 is characterized in that: The control execution module includes: Receive the lubrication intervention action, parameter combination and strategy confidence output by the lubrication strategy optimization module; Call the multi-channel oil circuit control valve group integrated with the lubrication system to control the execution channel of the replenishment, replacement, dilution or regeneration operation according to the instructions; The control execution module accurately controls the execution process of various lubrication operations based on the operating flow, duration and proportional coefficient in the parameter combination; Real-time acquisition of key control node status during operation; After the lubrication operation is completed, the short-cycle rapid sampling mechanism is activated to collect the latest round of original observation data on the oil status and equipment operation response indicators; The state feature difference before and after the lubrication operation and the execution action are encapsulated into an execution-feedback event package and pushed to the lubrication state multi-source data acquisition module to complete the construction of the closed-loop data path.

Images