Oil liquid quality evaluation system based on big data

Abstract

The invention discloses an oil liquid quality evaluation system based on big data, and relates to the technical field of multi-source sensing and Internet of Things, and the system comprises a data collection module which collects full-dimensional data through a heterogeneous sensing array of an oil-containing liquid multi-parameter integrated sensor and a high-frequency vibration sensor in combination with a timestamp synchronous triggering mechanism; the space-time calibration module adopts a dynamic time warping-interpolation collaborative algorithm to align the vibration signal and the oil data, and the information loss is controlled to be lower than a preset threshold value; the feature fusion module generates enhanced fusion features through a pre-trained embedded model and a double-stage attention mechanism; the intelligent evaluation model outputs oil quality grade and fault early warning based on a deep neural network, and the federated learning optimization unit guarantees data privacy; and the visual output module visually displays the result. The system improves the evaluation precision and the data utilization rate, adapts to industrial predictive maintenance, and reduces the shutdown loss of equipment.

Description

Technical Field

[0001] This invention relates to the fields of multi-source sensing and Internet of Things (IoT) technology, specifically to an oil quality assessment system based on big data. Background Technology

[0002] Driven by the Industry 4.0 wave, global manufacturing is accelerating its transformation towards intelligent manufacturing, and the innovation of equipment maintenance models has become a key breakthrough for improving production efficiency. Traditional "periodic maintenance" relies on fixed-cycle inspections, often resulting in over-maintenance or missed fault diagnosis. In contrast, "predictive maintenance," which monitors equipment status in real time and provides early warnings of faults, has become a core industry requirement. As the "lubricating blood" of industrial equipment, the physical and chemical properties and contaminant content of oil directly reflect potential hidden dangers such as internal wear and seal failure. According to industry statistics, more than 80% of mechanical failures are directly related to abnormal oil conditions.

[0003] The limitations of traditional oil quality assessment methods are becoming increasingly apparent: offline laboratory testing cycles can take more than 24 hours, making it difficult to capture nonlinear degradation processes such as exponential oxidation after the oil's acid value exceeds the critical point, often leading to delayed fault warnings; monitoring of a single parameter is prone to misjudgment, such as failing to identify the risk of oil emulsification caused by water contamination, which can result in hidden damage to equipment; regular oil change strategies not only waste 30% to 50% of lubricating oil, but also make passive maintenance costs 4.7 times that of predictive maintenance, significantly increasing the operational burden on enterprises.

[0004] In recent years, breakthroughs in key technologies have provided strong support for online monitoring of oil quality. MEMS micro-sensing technology has enabled the large-scale application of six-in-one integrated sensors, which are small in size, have lower costs, and can adapt to harsh industrial environments such as high temperature and high pressure. The collaborative architecture of edge computing and cloud computing has broken through the bottleneck of data processing. The innovative application of federated learning frameworks has enabled cross-enterprise and cross-device joint training of models while ensuring data privacy and security, thereby improving the accuracy of oil fault early warning.

[0005] Therefore, there is an urgent need for an oil quality assessment system based on big data. This system faces challenges in the fusion of multi-source data, including spatiotemporal asynchrony, where the difference in timestamps between vibration signals and oil composition data leads to a 20% loss of information; weak feature correlation, where the semantic gap between heterogeneous data is difficult to bridge using traditional methods and requires complex algorithms such as attention mechanisms. Summary of the Invention

[0006] This invention aims to provide an oil quality assessment system based on big data, which solves the problems of asynchronous data in time and space, semantic gap and low assessment accuracy by heterogeneous sensor synchronous acquisition, spatiotemporal calibration, two-stage attention fusion and federated learning assessment.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A big data-based oil quality assessment system includes: The data acquisition module is equipped with a heterogeneous sensor array for synchronously acquiring oil physicochemical data, equipment vibration signals, environmental parameters and equipment operating status data. The heterogeneous sensor array includes an integrated oil multi-parameter sensor and a high-frequency vibration sensor, and has a preset timestamp synchronous triggering mechanism. The spatiotemporal calibration module has a built-in dynamic time warping-interpolation collaborative algorithm for time alignment processing of vibration signals and oil composition data. It generates the optimal time matching path by constructing a time distance matrix and using a dynamic programming algorithm, and supplements missing data by combining an adaptive interpolation strategy, keeping the information loss caused by timestamp differences below a preset threshold. The feature fusion module includes a feature unified mapping unit and a two-stage attention mechanism unit. The feature unified mapping unit transforms heterogeneous data into the same high-dimensional feature space through a pre-trained embedding model. The two-stage attention mechanism unit first calculates the internal correlation weights of similar features through self-attention, and then mines the semantic correlations between different types of features through cross-attention to generate enhanced fusion features. The intelligent evaluation model is built on a deep neural network, which includes a feature encoding layer, a multi-task prediction layer, and a federated learning optimization unit. It is used to receive enhanced fusion features and output oil quality grade evaluation results and equipment fault warning information. The visualization output module is used to display evaluation results, fault warning information, and feature correlation heatmaps in real time.

[0008] Preferably, the data acquisition module specifically includes: The integrated multi-parameter oil sensor is used to collect data on oil viscosity, water content, acid value, dielectric constant, particle size, and metal element concentration; the high-frequency vibration sensor is used to collect vibration acceleration, velocity, and displacement signals of the equipment; the timestamp synchronization triggering mechanism is based on hardware synchronization pulses to control the timestamp error of each sensor in the heterogeneous sensor array within ±1ms.

[0009] Preferably, the spatiotemporal calibration module specifically includes: The construction of the time distance matrix adopts a weighted Euclidean distance metric, in which the weight of vibration signal features is 1.5-3 times the weight of oil composition data features; the dynamic programming algorithm is set with a maximum time offset constraint, the maximum allowable time offset is 1 / 3 of the oil composition data acquisition cycle, in order to limit the excessive stretching and compression of nonlinear time alignment.

[0010] Preferably, the spatiotemporal calibration module specifically includes: The adaptive interpolation strategy includes using cubic spline interpolation for missing vibration signal data and weighted average interpolation based on historical similar working conditions for missing oil composition data; when the length of a single continuous missing data segment exceeds twice the oil composition data acquisition cycle, an anomaly marker is activated and a sensor status self-check is triggered; the preset threshold is 5%, and the amount of information loss is calculated by the change in feature entropy value before and after alignment.

[0011] Preferably, the feature fusion module specifically includes: The pre-trained embedding model adopts a Transformer-based bidirectional encoder structure, and the high-dimensional feature space is set to 256-1024 dimensions. The feature unification mapping unit directly performs standardized mapping on numerical oil physicochemical data, and extracts time-frequency domain features from time-series vibration signals before embedding and mapping, ensuring the consistency of the distribution of different types of data in the high-dimensional space.

[0012] Preferably, the feature fusion module specifically includes: The self-attention calculation in the two-stage attention mechanism unit adopts the scaled dot product attention algorithm, which calculates the similarity matrix between similar feature vectors and obtains the normalized internal association weights by combining the softmax function; the self-attention calculation process also introduces position encoding information to preserve the temporal sequence characteristics of time-series data.

[0013] Preferably, the feature fusion module specifically includes: The cross-attention mechanism enables the semantic association mining of different types of features by constructing a heterogeneous feature interaction matrix. The element values ​​of the interaction matrix are obtained by weighted calculation of cosine similarity and mutual information values. The generation of the enhanced fusion features adopts a combination of weighted summation and feature concatenation, and the overfitting risk is reduced by dropout regularization during the fusion process.

[0014] Preferably, the intelligent evaluation model specifically includes: The feature encoding layer adopts a hybrid structure of stacked convolutional neural networks (CNN) and gated recurrent units (GRU). The CNN extracts spatial correlation features of enhanced fusion features, and the GRU captures temporal dependencies. The multi-task prediction layer includes an oil quality grade assessment sublayer and a fault warning sublayer, which adopt fully connected neural networks and lightweight Transformer architectures, respectively. The oil quality grade is divided into four levels: excellent, good, medium and poor. The fault warning sublayer can output the probability of fault occurrence and the expected time window.

[0015] Preferably, the intelligent evaluation model specifically includes: The federated learning optimization unit adopts an improved federated averaging algorithm, which adjusts the aggregation ratio of parameters of each device by introducing a device condition weight coefficient. The condition weight coefficient is dynamically calculated based on the device's operating load, working time, and historical failure frequency. At the same time, a differential privacy protection module is configured to add Gaussian noise before the parameters are uploaded to ensure local data privacy and security, and the noise intensity is adaptively adjusted according to the parameter sensitivity.

[0016] Preferably, the visualization output module specifically includes: It supports multi-dimensional display modes, where oil quality assessment results are presented as dynamic trend curves combined with numerical dashboards, and fault warning information adopts a hierarchical reminder mechanism, with different color indicators and audible and visual alarms corresponding to level one to level three warnings; the feature association heat map can be drilled down to specific feature dimensions according to user operation, and it also supports time interval query and comparative analysis of historical assessment data.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The heterogeneous sensor array integrates multi-parameter oil sensors and high-frequency vibration sensors, covering all dimensions of oil physicochemical properties, equipment vibration, environment, and operating status data. Combined with a preset timestamp synchronization triggering mechanism, it ensures the spatiotemporal consistency of data from the source, avoiding subsequent analysis deviations caused by asynchronous acquisition, and laying a high-quality data foundation for multi-source data fusion.

[0018] The built-in dynamic time warping-interpolation collaborative algorithm generates the optimal matching path through the time distance matrix and dynamic programming. Combined with adaptive interpolation to supplement missing data, it can control the information loss caused by timestamp differences to below a preset threshold, completely solving the problem of 20% information loss caused by the asynchrony between vibration signal and oil composition data, and significantly improving data utilization and reliability.

[0019] The Feature Unified Mapping Unit transforms heterogeneous data into the same high-dimensional space through a pre-trained embedding model, and then generates enhanced fusion features through a two-stage attention mechanism. This breaks through the limitation of traditional methods that are difficult to associate heterogeneous data, provides high-value feature support for subsequent evaluation, and directly improves evaluation accuracy.

[0020] The intelligent evaluation model is based on deep neural networks to achieve multi-task output. The federated learning optimization unit improves the model's generalization ability while protecting data privacy, and is adapted to multi-device and cross-scenario applications. The visualization output module displays the results intuitively in the form of dynamic curves, hierarchical early warning, heat maps and other forms. It also supports data export, which not only facilitates quick decision-making by staff, but also lowers the threshold for using the system.

[0021] The overall system can significantly reduce unplanned equipment downtime losses, lower maintenance costs, and promote the efficient implementation of industrial equipment maintenance from "periodic maintenance" to "predictive maintenance". Attached Figure Description

[0022] Figure 1 This is an internal framework diagram of an oil quality assessment system based on big data. Detailed Implementation

[0023] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0024] Reference Figure 1 As shown, the oil quality assessment system based on big data includes: The data acquisition module is equipped with a heterogeneous sensor array for synchronously acquiring oil physicochemical data, equipment vibration signals, environmental parameters and equipment operating status data. The heterogeneous sensor array includes an integrated oil multi-parameter sensor and a high-frequency vibration sensor, and has a preset timestamp synchronous triggering mechanism. The spatiotemporal calibration module has a built-in dynamic time warping-interpolation collaborative algorithm for time alignment processing of vibration signals and oil composition data. It generates the optimal time matching path by constructing a time distance matrix and using a dynamic programming algorithm, and supplements missing data by combining an adaptive interpolation strategy, keeping the information loss caused by timestamp differences below a preset threshold. The feature fusion module includes a feature unified mapping unit and a two-stage attention mechanism unit. The feature unified mapping unit transforms heterogeneous data into the same high-dimensional feature space through a pre-trained embedding model. The two-stage attention mechanism unit first calculates the internal correlation weights of similar features through self-attention, and then mines the semantic correlations between different types of features through cross-attention to generate enhanced fusion features. The intelligent evaluation model is built on a deep neural network, which includes a feature encoding layer, a multi-task prediction layer, and a federated learning optimization unit. It is used to receive enhanced fusion features and output oil quality grade evaluation results and equipment fault warning information. The visualization output module is used to display evaluation results, fault warning information, and feature correlation heatmaps in real time.

[0025] It should be noted that the heterogeneous sensor array integrates multi-dimensional data such as oil physicochemical parameters (viscosity / acid value / moisture), high-frequency vibration signals (20kHz sampling), ambient temperature and humidity, and equipment load rate; The timestamp synchronization triggering mechanism ensures that the vibration peak and the oil parameter mutation point are accurately correlated in the time domain through hardware-level clock synchronization (PTP protocol) and software-level trigger compensation (μs-level error), laying the foundation for subsequent fusion; Spatiotemporal calibration and feature fusion: The dynamic time warping-interpolation collaborative algorithm not only solves the timestamp difference, but also constructs a physical-data dual-driven compensation model; the time distance matrix introduces the equipment rotation speed parameter as a weighting factor, so that the alignment path of the vibration signal and the oil data conforms to the laws of mechanical dynamics; the adaptive interpolation strategy injects tribological prior knowledge into the missing data, avoiding physical distortion caused by pure mathematical interpolation.

[0026] Feature fusion module: The self-attention layer analyzes the internal correlations of similar data, such as the differential characterization of gear wear by 1kHz and 5kHz components in the vibration spectrum, and generates physical consistency features; the cross-attention layer mines cross-domain semantic correlations, such as the causal chain between the abrupt change in ferrographic abrasive particle concentration and the sideband energy of the vibration envelope spectrum, and constructs a fault evolution knowledge graph.

[0027] Federated learning optimization unit: Under the premise of ensuring data privacy, it aggregates the degradation patterns of equipment in multiple factories through a gradient encryption sharing mechanism, thereby reducing the prediction error of the model in scenarios with few samples. Multi-task prediction layer: Oil quality assessment (regression task) and fault early warning (classification task) share the feature encoding layer, and achieve accurate resource allocation through dynamic adjustment of loss function weights.

[0028] Visualization output: Feature association heatmap, generated using a spectral clustering-entropy fusion algorithm, visually reveals: The correlation coefficient between key fault-sensitive factors, such as moisture content and vibration acceleration RMS value, reaches 0.89. Hidden risk transmission path: ambient humidity → oil water content ↑ → additive degradation rate ↑ → viscosity ↓ → bearing wear ↑; maintenance personnel can shorten the time to locate the root cause of the failure based on this.

[0029] Early warning information: Fault early warning information is associated with spatiotemporal segments of raw data. For example, "gearbox wear early warning" automatically marks the vibration EMD decomposition diagram and ferrography image of the previous 10 minutes, supporting retrospective root cause analysis; the output interface integrates standard threshold curves such as ISO4406 / SAEAS4059 to achieve automatic compliance benchmarking.

[0030] The data acquisition module specifically includes: The integrated multi-parameter oil sensor is used to collect data on oil viscosity, water content, acid value, dielectric constant, particle size, and metal element concentration; the high-frequency vibration sensor is used to collect vibration acceleration, velocity, and displacement signals of the equipment; the timestamp synchronization triggering mechanism is based on hardware synchronization pulses to control the timestamp error of each sensor in the heterogeneous sensor array within ±1ms.

[0031] It should be noted that the multi-parameter integrated sensor for oil uses a MEMS resonant viscometer, which calculates the viscosity value by measuring the resonant frequency offset (Δf) of the silicon cantilever beam in the oil, with a resolution of ±0.2 cP. Laser-induced breakdown spectroscopy (LIBS) for metal element detection: pulsed laser focusing on oil microchannels (Φ50μm) and excited plasma spectroscopy analysis of elements such as Fe / Cu / Pb, with the detection limit broken down to 1ppm. Dielectric constant-moisture joint sensing unit: The interdigital electrode measures the dielectric constant (ε), and the moisture content is calculated by combining the temperature compensation model, eliminating the misjudgment caused by the ionization of additives in the traditional capacitance method; Anti-interference enhancement, multi-parameter cross-validation mechanism, when the data from the moisture sensor and the dielectric constant sensor violate the Maxwell relation ( When the moisture content is 2.1 + 0.08W (where W is the moisture percentage), the self-cleaning process is automatically triggered. The oil matrix compensation algorithm dynamically adjusts the metal concentration detection calibration curve for different lubricant types (mineral oil / PAO / ester).

[0032] High-frequency vibration sensor: Triaxial synchronous acquisition, independent processing of axial / radial / tangential vibration signals. Axial vibration detection of thrust bearing failure (characteristic frequency band: 0.5-1 × rotational frequency); Radial vibration diagnosis of gear meshing faults (sideband modulation: meshing frequency ± fault characteristics); Wideband coverage design: Low frequency band (5-200Hz): MEMS accelerometers are used to monitor shaft misalignment; High frequency band (2-20kHz): Piezoelectric ceramic sensors capture shock waves from bearing defects; Enhanced tolerance to extreme working conditions, high-temperature oil immersion packaging, sensor housing made of Hastelloy C276, resistant to 150℃ gearbox hot oil environment; The electromagnetic interference (EMI) suppression design features a three-layer shielding structure (copper plating + iron-nickel alloy + conductive rubber) to suppress inverter radiation interference, and a common-mode choke to eliminate high-frequency noise from the motor windings.

[0033] Timestamp synchronization trigger mechanism: hardware-level synchronization, master-slave clock architecture, the master controller sends 1000 synchronization pulses per second (1kHz), and the slave sensor locks the trigger edge through a PLL phase-locked loop (clock jitter <10ns). Dynamic compensation for transmission delay:

[0034] In the formula, The compensated transmission delay time is given, where L is the cable length, v is the signal propagation speed (0.7c for coaxial cable), and C is the interface capacitance. The cutoff frequency; Software-level time correction strategy: Two-way delay measurement (TDDM) protocol. The controller sends a Sync message and records the transmission time T1. After receiving the message, the sensor records T2 and replies with Delay_Req. The controller calculates the two-way transmission delay.

[0035] In the formula, Δt is the bidirectional transmission delay, which is the total time required for the signal to be sent from the controller to the sensor and back to the controller; T1 is the timestamp of the controller sending the synchronization message; T2 is the timestamp recorded by the sensor after receiving the synchronization message; T3 is the timestamp of the sensor sending the reply message; and T4 is the timestamp of the controller receiving the reply message from the sensor.

[0036] The spatiotemporal calibration module specifically includes: The construction of the time distance matrix adopts a weighted Euclidean distance metric, in which the weight of vibration signal features is 1.5-3 times the weight of oil composition data features; the dynamic programming algorithm is set with a maximum time offset constraint, the maximum allowable time offset is 1 / 3 of the oil composition data acquisition cycle, in order to limit the excessive stretching and compression of nonlinear time alignment.

[0037] The adaptive interpolation strategy includes using cubic spline interpolation for missing vibration signal data and weighted average interpolation based on historical similar working conditions for missing oil composition data; when the length of a single continuous missing data segment exceeds twice the oil composition data acquisition cycle, an anomaly marker is activated and a sensor status self-check is triggered; the preset threshold is 5%, and the amount of information loss is calculated by the change in feature entropy value before and after alignment.

[0038] It should be noted that the time-distance matrix is ​​a weighted average of vibration signals, with a weighting coefficient of 1.5-3 times due to differences in fault sensitivity. Vibration signals include instantaneous impact information of the equipment, such as high-frequency components >5kHz generated by bearing spalling, which have a response speed 2-3 orders of magnitude faster than oil parameters for early faults. Dynamic weight adjustment mechanism: Under heavy load conditions, the upper limit (3 times) is taken because the mechanical stress increases dramatically, leading to an increase in the entropy of vibration characteristic information; under steady-state light load conditions, the lower limit (1.5 times) is taken to avoid vibration noise interfering with the alignment process. Maximum time offset constraint: Threshold setting of 1 / 3 of the oil sampling period; typical value, oil sampling 1Hz → maximum offset 333ms; To ensure the integrity of the equipment rotation cycle, the rotation angle of the low-speed shaft (10 rpm) of the wind turbine gearbox within 1 / 3 of the sampling cycle should be ≤2° to prevent the tooth surface contact stress distribution from being cut. Suppress spurious correlations caused by overstretching, such as incorrectly associating cooler leaks with gear wear; Dynamic programming path optimization:

[0039] In the formula, The vibration weight represents the importance or degree of influence of vibration on the optimization objective; The oil weight represents the importance or degree of influence of the oil in the optimization objective; In time arrive The vibration-weighted Euclidean distance function between the two points is used to calculate the impact of vibration in path planning; In time arrive The weighted Euclidean distance function between the oil and the path is used to calculate the influence of the oil in the path planning. Path search affected Hard constraints This represents the maximum time constraint for path search, indicating that path planning must be completed within this timeframe. Adaptive interpolation strategy, signal characteristic-driven differential interpolation:

[0040] Core logic of oil interpolation:

[0041] In the formula, These are the interpolated oil parameter values, i.e., the estimated oil parameters calculated based on the current operating condition vector S; is the weight of the i-th historical similar working condition, used to calculate the interpolation result; k is the number of historical similar working conditions; Let S be the oil parameter value under the i-th historical similar operating condition; S is the current operating condition vector, including parameters such as load, temperature, and speed. Let i be the working condition vector of the i-th historical similar working condition; Anomaly detection and self-check linkage defense: 2 times sampling cycle missing trigger rule, for example, oil sampling at 1Hz → continuous missing for >2 seconds; From a physical perspective, the problem exceeds the circulation cycle of lubricating oil within the gearbox (typically 1.5-3 seconds), making reconstruction through temporal correlation impossible; from an algorithmic perspective, the problem exceeds the effective support domain of cubic spline interpolation. Self-test linkage process: Missing length > 2T → Mark data abnormality → Sensor diagnosis → Check power supply voltage, verify communication CRC, compare redundant sensors → Generate diagnostic report.

[0042] Information loss control: A 5% threshold was set and verified based on the ISO18436-4 standard dataset. When information loss > 5%, the bearing failure detection rate dropped sharply from 98% to 72%. Formula for calculating the change in entropy ΔH:

[0043] In the formula, The feature entropy before alignment contains the original information. The feature entropy after alignment reflects the information loss; Entropy calculation is optimized by introducing permutation entropy to replace traditional Shannon entropy, thereby improving the sensitivity to transient characteristics of vibration signals; The closed-loop control with dynamic compensation activates a third-level response when ΔH > 5%.

[0044] The feature fusion module specifically includes: The pre-trained embedding model adopts a Transformer-based bidirectional encoder structure, and the high-dimensional feature space is set to 256-1024 dimensions. The feature unification mapping unit directly performs standardized mapping on numerical oil physicochemical data, and extracts time-frequency domain features from time-series vibration signals before embedding and mapping, ensuring the consistency of the distribution of different types of data in the high-dimensional space.

[0045] The self-attention calculation in the two-stage attention mechanism unit adopts the scaled dot product attention algorithm, which calculates the similarity matrix between similar feature vectors and obtains the normalized internal association weights by combining the softmax function; the self-attention calculation process also introduces position encoding information to preserve the temporal sequence characteristics of time-series data.

[0046] The cross-attention mechanism enables the semantic association mining of different types of features by constructing a heterogeneous feature interaction matrix. The element values ​​of the interaction matrix are obtained by weighted calculation of cosine similarity and mutual information values. The generation of the enhanced fusion features adopts a combination of weighted summation and feature concatenation, and the overfitting risk is reduced by dropout regularization during the fusion process.

[0047] It should be noted that the high-dimensional feature space is constructed as follows: The Transformer encoder's domain adaptation capability employs a bidirectional encoder structure (12 hidden layers + 16 attention heads) and learns a general representation of oil monitoring data through unsupervised pre-training. The input layer employs piecewise normalization mapping for numerical physicochemical data (viscosity / acid value): parameters with huge dimensional differences are reduced to the [-1,1] interval to eliminate magnitude interference; First, multi-scale time-frequency feature extraction is performed on the vibration time-series signal: In the time domain, the kurtosis index captures impact failures; Frequency domain analysis of bearing fault characteristic frequency bands using wavelet packet energy entropy; Energy ratio of IMF components in the Hilbert-Huang transform in the time-frequency domain; The dynamic space is 256-1024 dimensions. The low-dimensional region (256-512 dimensions) stores basic physical properties, such as the viscosity-temperature relationship conforming to the Arrhenius equation; the high-dimensional region (768-1024 dimensions) encodes complex correlation patterns, such as the nonlinear mapping between abrasive morphology and vibration modulation sidebands. The distribution consistency guarantee mechanism narrows the feature distance of similar equipment data through comparative learning and pushes away abnormal operating condition samples, thereby increasing the "normal-abnormal" class distance in high-dimensional space by 2.3 times.

[0048] Two-stage attention mechanism: Self-attention layer, scaling dot product attention calculation formula:

[0049] In the formula, Q, K, and V are the query, key, and value matrices of the same type of feature vectors, respectively; The scaling factor suppresses gradient vanishing caused by excessively large dot product values ​​in high-dimensional space; Position encoding involves injecting rotating machinery phase encoding into vibration signal processing to convert the equipment's rotational speed pulse signal into an angular position vector.

[0050] In the formula, For position encoding, it represents the 2i-th encoded value of the device at a specific position (pos); pos is the position, representing the current angular position of the device; i is the index, representing the position in the encoding sequence; dmodel is the dimension or parameter of the model, which may be related to the model of rotating machinery phase encoding. This allows attention weights to carry information about the equipment's rotation cycle, precisely linking vibration phase with changes in oil parameters. Cross-attention layer: A method for constructing heterogeneous feature interaction matrices, where matrix element Mij represents the correlation strength between the i-th type of oil parameter and the j-th type of vibration feature, calculated as follows:

[0051] In the formula, is a matrix element representing the correlation strength between the i-th type of oil parameter and the j-th type of vibration characteristic; α and β are weighting coefficients used to balance the influence of two different correlation measurement methods, with default values ​​of 0.6 and 0.4 respectively; represents the vectors of oil parameters and vibration characteristics, respectively; I(·) is a mutual information function based on kernel density estimation, used to capture nonlinear dependencies, such as the implicit correlation between additive degradation rate and vibration envelope spectrum kurtosis.

[0052] Fusion generation mechanism: Weighted-stitching collaborative fusion strategy, key path:

[0053] In the formula, γ represents the final feature obtained through a weighted-concatenation collaborative fusion strategy; γ is the dynamic gating coefficient, a weighting parameter used to balance the contributions of intra-class and cross-domain features in the fused feature. Its value is determined based on the confidence level of the current operating condition, taking 0.7 for steady-state conditions and 0.3 for transient conditions. This is the weight matrix, used to weight the features within a class; For intra-class features extracted by self-attention, such as the set of resonant frequency bands in the vibration spectrum; For splicing cross-domain features, These are cross-domain features discovered through cross-attention; Physical interpretability is guaranteed by imposing tribological constraints on the weighted results to ensure that the fusion characteristics conform to the mechanical principle of wear amount ∝ load × sliding distance; Dropout regularization combats overfitting by inserting random spatial channels after the feature concatenation layer. 15% of the feature channels are randomly masked with probability p=0.3; During the training phase, an additional 10% of spatial location features are discarded (simulating a sensor failure scenario).

[0054] The intelligent evaluation model specifically includes: The feature encoding layer adopts a hybrid structure of stacked convolutional neural networks (CNN) and gated recurrent units (GRU). The CNN extracts spatial correlation features of enhanced fusion features, and the GRU captures temporal dependencies. The multi-task prediction layer includes an oil quality grade assessment sublayer and a fault warning sublayer, which adopt fully connected neural networks and lightweight Transformer architectures, respectively. The oil quality grade is divided into four levels: excellent, good, medium and poor. The fault warning sublayer can output the probability of fault occurrence and the expected time window.

[0055] The federated learning optimization unit adopts an improved federated averaging algorithm, which adjusts the aggregation ratio of parameters of each device by introducing a device condition weight coefficient. The condition weight coefficient is dynamically calculated based on the device's operating load, working time, and historical failure frequency. At the same time, a differential privacy protection module is configured to add Gaussian noise before the parameters are uploaded to ensure local data privacy and security, and the noise intensity is adaptively adjusted according to the parameter sensitivity.

[0056] It should be noted that the feature encoding layer uses a CNN-GRU hybrid structure for spatial feature extraction. The stacked CNN employs three depthwise separable convolutional layers. The first layer (5×1 convolution kernel) extracts the cross-dimensional correlation between oil parameters; Sub-layer (3×3 convolution kernel) to mine local frequency band resonance modes of vibration spectrum; The final layer (1×1 convolutional kernel) achieves cross-modal feature compression, reducing the 256-dimensional input to 64-dimensional key features; Hollow convolution embedding, with an expansion rate of 2 set in the third layer, expands the receptive field to 1.5 times the vibration signal period, capturing long-term wear trends; Bidirectional GRU fusion: The forward layer captures the cumulative degradation of lubrication performance, while the reverse layer identifies sudden precursors to failure.

[0057] The visualization output module specifically includes: It supports multi-dimensional display modes, where oil quality assessment results are presented as dynamic trend curves combined with numerical dashboards, and fault warning information adopts a hierarchical reminder mechanism, with different color indicators and audible and visual alarms corresponding to level one to level three warnings; the feature association heat map can be drilled down to specific feature dimensions according to user operation, and it also supports time interval query and comparative analysis of historical assessment data.

[0058] It should be noted that the oil quality assessment uses a spatiotemporal fusion trend curve and a dual Y-axis adaptive coordinate system: the left side displays physicochemical parameters (viscosity / acid value), and the right side maps vibration characteristics (RMS value / kurtosis), thus resolving curve distortion caused by dimensional differences. The operating condition marking overlay technology overlays color bands (red = heavy load, green = light load) for load rate, speed, and other operating conditions above the curve, intuitively revealing the correlation between parameter changes and operating conditions; Smart Dashboard Design:

[0059] A graded response mechanism for fault early warning, including audible and visual alarms: Level 1 warning: 85dB pulse siren and 360° rotating red light (ISO7731 standard); Level 2 warning: 70dB intermittent beeping and flashing yellow light; Level 3 warning: Silent screen pop-up (to avoid frequent interference).

[0060] Feature association heatmap: drill-down association matrix, multi-layer association architecture Level 1: Top-level dimensions include wear correlation, contamination correlation, and decay correlation; Level 2: Wear correlation includes Fe-Cu correlation, vibration modulation index, and additive consumption rate; contamination correlation includes moisture-dielectric constant, particle size-viscosity, and silicon trend. Level 3: Fe-Cu correlation includes time-domain correlation coefficient, frequency-domain coherence function, and mutual information entropy; Dynamic drilling technology: Click "Wear Correlation" to drill down to the "Fe-Cu Correlation" dimension, showing the time-varying curve of the iron-copper concentration ratio and its fitting with the Weibull distribution; continue drilling down to the "Mutual Information Entropy" layer, showing the pixel-level correlation hot zone between the vibration envelope spectrum and the ferrographic abrasive image; Historical data comparison, cross-temporal and spatial comparison of three modalities:

[0061] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A big data based oil quality assessment system, characterized in that, The application relates to an oil quality evaluation system based on heterogeneous data fusion, which comprises the following modules: a data acquisition module configured with a heterogeneous sensor array for synchronously collecting oil physical and chemical data, equipment vibration signals, environmental parameters and equipment operation state data, wherein the heterogeneous sensor array comprises an oil liquid multi-parameter integrated sensor and a high-frequency vibration sensor, and a preset timestamp synchronous triggering mechanism; a space-time calibration module internally provided with a dynamic time warping-interpolation collaborative algorithm for time alignment processing of the vibration signals and the oil component data, wherein an optimal time matching path is generated by constructing a time distance matrix and adopting a dynamic programming algorithm, and missing data is supplemented by combining an adaptive interpolation strategy, so that information loss caused by timestamp differences is controlled below a preset threshold value; a feature fusion module comprising a feature uniform mapping unit and a two-stage attention mechanism unit, wherein the feature uniform mapping unit converts the heterogeneous data to the same high-dimensional feature space through a pre-training embedding model, and the two-stage attention mechanism unit first calculates the internal correlation weight of the same type of features through self-attention, and then mines the semantic correlation between different types of features through cross-attention to generate enhanced fusion features; an intelligent evaluation model based on a deep neural network, wherein the deep neural network comprises a feature encoding layer, a multi-task prediction layer and a federated learning optimization unit, is used for receiving the enhanced fusion features, and outputs oil quality grade evaluation results and equipment fault warning information; a visual output module for real-time display of the evaluation results, the fault warning information and the feature correlation heat map.

2. The big data based oil quality assessment system as claimed in claim 1, wherein, The feature fusion module specifically comprises: the pre-training embedding model adopts a bidirectional encoder structure based on a Transformer, and the dimension of the high-dimensional feature space is set to 256-1024 dimensions; the feature uniform mapping unit directly performs standardized mapping on the numerical oil physical and chemical data, and first extracts time-frequency domain features and then performs embedding mapping on the time-series vibration signals, so that the distribution consistency of different types of data in the high-dimensional space is ensured.

3. The big data based oil quality assessment system as claimed in claim 1 wherein, The feature fusion module specifically comprises: the self-attention calculation in the two-stage attention mechanism unit adopts a scaled dot-product attention algorithm, similarity matrices between the same type of feature vectors are calculated, and normalized internal correlation weights are obtained by combining a softmax function; position encoding information is also introduced in the calculation process of the self-attention to retain the time sequence features of the time-series data.

4. The big data based oil quality assessment system as claimed in claim 1, wherein, The feature fusion module specifically comprises: the cross-attention realizes semantic correlation mining of different types of features by constructing a heterogeneous feature interaction matrix, and element values of the interaction matrix are obtained by weighted calculation of cosine similarity and mutual information values; the generation of the enhanced fusion features adopts a combination of weighted summation and feature splicing, and a dropout regularization process is performed in the fusion process to reduce the risk of overfitting.

5. The big data based oil quality assessment system as claimed in claim 1 wherein, The space-time calibration module specifically comprises: the construction of the time distance matrix adopts weighted Euclidean distance measurement, and the weight of the vibration signal feature in the weighted Euclidean distance is 1.5-3 times that of the oil component data feature weight; the dynamic programming algorithm is provided with a maximum time offset constraint, and the maximum allowed time offset is 1 / 3 of the oil component data acquisition period, so as to limit the excessive stretching and compression of nonlinear time alignment.

6. The big data based oil quality assessment system as claimed in claim 1 wherein, The space-time calibration module specifically comprises: The adaptive interpolation strategy comprises adopting cubic spline interpolation for missing data of the vibration signal and adopting weighted average interpolation based on historical similar working conditions for missing data of the oil composition data; when the length of single-section continuous missing data exceeds 2 times of the oil composition data acquisition period, an abnormality marker is started and a sensor state self-check is triggered; the preset threshold is 5%, and the information loss amount is calculated by a change amount of characteristic entropy values before and after alignment.

7. The big data based oil quality assessment system as claimed in claim 1 wherein, The intelligent evaluation model specifically comprises: The feature encoding layer adopts a hybrid structure of stacked convolutional neural network (CNN) and gated recurrent unit (GRU), extracts spatial correlation features of enhanced fusion features through CNN, and captures time sequence dependency relationship by using GRU; the multi-task prediction layer comprises an oil quality grade evaluation sublayer and a fault early warning sublayer, and adopts a fully connected neural network and a lightweight Transformer architecture respectively, wherein the oil quality grade is divided into four grades of excellent, good, medium and poor, and the fault early warning sublayer can output a fault occurrence probability and a predicted occurrence time window.

8. The big data based oil quality assessment system as claimed in claim 1 wherein, The intelligent evaluation model specifically comprises: The federal learning optimization unit adopts an improved federal average algorithm, adjusts the aggregation proportion of parameters of each device end by introducing a device working condition weight coefficient, and dynamically calculates the working condition weight coefficient based on device running load, working time length and historical fault frequency; at the same time, a differential privacy protection module is configured, Gaussian noise is added before parameter uploading, local data privacy safety is ensured, and the noise intensity is adaptively adjusted according to the parameter sensitivity.

9. The big data based oil quality assessment system as claimed in claim 1 wherein, The data acquisition module specifically comprises: The oil multi-parameter integrated sensor is used to collect viscosity, moisture content, acid value, dielectric constant, particle size and metal element concentration data of the oil; the high-frequency vibration sensor is used to collect vibration acceleration, speed and displacement signals of the device; and the timestamp synchronous triggering mechanism is realized based on a hardware synchronous pulse, so that the timestamp error of each sensor in the heterogeneous sensor array is controlled within ±1 ms.

10. The big data based oil quality assessment system as claimed in claim 1 wherein, The visualization output module specifically comprises: Supporting multi-dimensional display modes, wherein the oil quality evaluation result is presented in the form of a dynamic trend curve combined with a numerical instrument panel, the fault early warning information adopts a hierarchical reminder mechanism, and first to third level warnings correspond to different color identifiers and sound-light alarms; the feature correlation heat map can be drilled to a specific feature dimension according to user operation, and supports time interval query and comparative analysis of historical evaluation data.

Images