Engine oil burning diagnosis and maintenance system

Abstract

The invention provides an engine oil burning diagnosis and maintenance system, and relates to engine oil burning diagnosis. The method comprises a multi-source data acquisition model, a heterogeneous data deep fusion module, a multi-level diagnosis module and a prediction maintenance module. The multi-source data acquisition model is used for acquiring multi-source data in the operation process of an engine in real time through an integrated sensor array, and the multi-source data comprises vibration data, acoustic data, exhaust component data and temperature distribution data; the heterogeneous data deep fusion module is used for performing data conversion on the multi-source data to obtain unified multi-source feature representation; the multi-level diagnosis module is used for performing multi-level diagnosis on the multi-source feature representation to obtain a multi-level diagnosis result; and the prediction maintenance module is used for carrying out maintenance analysis according to the multi-layer diagnosis result to obtain a maintenance decision suggestion containing a predicted life value, a fault evolution path and optimal maintenance time. By means of the scheme, the comprehensiveness and accuracy of engine oil burning diagnosis and maintenance are improved.

Description

Technical Field

[0001] This application relates to oil burning diagnosis, and particularly to an oil burning diagnosis and maintenance system. Background Art

[0002] With the development of the automotive industry and technological progress, the performance of vehicles has been significantly improved. However, this is accompanied by some complex technical problems, such as the phenomenon of engine oil burning. Oil burning not only affects the normal operation of the engine but also leads to excessive emissions, and in severe cases, it may even cause engine damage. Currently, the diagnosis of such problems in the market mainly relies on traditional experience and manual inspection methods, mainly including judging whether there is oil burning by observing the exhaust color, measuring the crankcase pressure, checking the spark plug carbon deposition, etc. In addition, there are some more advanced methods, such as using a single sensor to monitor the oil consumption. However, these diagnostic methods all have a common technical defect: due to relying on a single data source for fault judgment, the diagnostic results are easily interfered by environmental factors, making it difficult to accurately identify and locate faults. For example, observing the exhaust color alone may be affected by weather and light, and the monitoring data of a single sensor may also be interfered by noise, thus affecting the accuracy of diagnosis and delaying the maintenance time. Summary of the Invention

[0003] This application provides an oil burning diagnosis and maintenance system to improve the comprehensiveness and accuracy of oil burning diagnosis and maintenance.

[0004] In a first aspect, this application provides an oil burning diagnosis and maintenance system, which includes: A multi-source data acquisition model, a heterogeneous data deep fusion module, a multi-level diagnosis module, and a predictive maintenance module; The multi-source data acquisition model is used to collect multi-source data during the operation of the engine in real time through an integrated sensor array, and the multi-source data includes vibration data, acoustic data, exhaust gas component data, and temperature distribution data; The heterogeneous data deep fusion module is used to perform data conversion on the multi-source data to obtain a unified multi-source feature representation; The multi-level diagnosis module is used to perform multi-level diagnosis on the multi-source feature representation to obtain multi-level diagnosis results; The predictive maintenance module is used to perform maintenance analysis based on the multi-level diagnosis results to obtain maintenance decision suggestions including predicted life values, fault evolution paths, and optimal maintenance times.

[0005] In the above technical solution, vibration data, acoustic data, exhaust gas component data, and temperature distribution data during the operation of the engine are collected through a multi-source data acquisition model, realizing comprehensive monitoring of the engine state and effectively avoiding the problem that a single data source is easily affected by environmental interference. The heterogeneous data deep fusion module performs data conversion on the multi-source data to obtain a unified multi-source feature representation, which not only solves the problem of inconsistent time series of heterogeneous data but also extracts the deep correlation features between the data. Combining with the multi-level diagnosis module to perform multi-level diagnosis on the unified multi-source feature representation, through the progressive process of feature extraction, time series analysis, knowledge reasoning, and diagnostic decision-making, the accuracy and interpretability of fault diagnosis are improved. Finally, the predictive maintenance module performs maintenance analysis based on the multi-level diagnosis results. Through the comprehensive evaluation of the predicted life value, fault evolution path, and optimal maintenance time, the transformation from passive maintenance to active predictive maintenance is realized, which not only reduces the maintenance cost but also improves the maintenance efficiency. This complete technical solution from data acquisition, feature fusion to diagnostic prediction effectively solves the technical problem of low accuracy of traditional single-data-source diagnosis methods.

[0006] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: Vibration data, acoustic data, exhaust gas component data, and temperature distribution data during the operation of the engine are collected through a multi-source data acquisition model, realizing comprehensive monitoring of the engine state and effectively avoiding the problem that a single data source is easily affected by environmental interference. The heterogeneous data deep fusion module performs data conversion on the multi-source data to obtain a unified multi-source feature representation, which not only solves the problem of inconsistent time series of heterogeneous data but also extracts the deep correlation features between the data. Combining with the multi-level diagnosis module to perform multi-level diagnosis on the unified multi-source feature representation, through the progressive process of feature extraction, time series analysis, knowledge reasoning, and diagnostic decision-making, the accuracy and interpretability of fault diagnosis are improved. Finally, the predictive maintenance module performs maintenance analysis based on the multi-level diagnosis results. Through the comprehensive evaluation of the predicted life value, fault evolution path, and optimal maintenance time, the transformation from passive maintenance to active predictive maintenance is realized, which not only reduces the maintenance cost but also improves the maintenance efficiency. This complete technical solution from data acquisition, feature fusion to diagnostic prediction effectively solves the technical problem of low accuracy of traditional single-data-source diagnosis methods. Description of the Drawings

[0007] Figure 1 It is a schematic structural diagram of an oil burning diagnosis and maintenance system provided by an embodiment of the present application; Figure 2 It is a schematic structural diagram of a heterogeneous data deep fusion module provided by an embodiment of the present application; Figure 3This is a schematic structural diagram of a multi-level diagnostic module provided by an embodiment of the present application. Detailed implementation manners

[0008] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments.

[0009] In the description of the embodiments of the present application, words such as "for example" or "for illustration" are used to represent examples, illustrations or explanations. Any embodiment or design solution described as "for example" or "for illustration" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "for example" or "for illustration" is intended to present related concepts in a specific manner.

[0010] In the description of the embodiments of the present application, the meaning of the term "a plurality" refers to two or more. For example, a plurality of systems refers to two or more systems, and a plurality of screen terminals refers to two or more screen terminals. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly indicating the technical features indicated. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. The terms "include", "comprise", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0011] To facilitate the understanding of the method and system provided by the embodiments of the present application, before introducing the embodiments of the present application, the background of the embodiments of the present application will be introduced first.

[0012] With the rapid development of Internet technology and the increasingly wide spread of information, the diagnosis of oil burning in news articles has become an important research topic in content security management. At present, there are mainly two types of technical routes for oil burning diagnosis: one is the method based on dictionary matching, which realizes recognition by establishing a preset sensitive word library and performing string matching between the text to be recognized and the sensitive words in the word library; the other is the method based on machine learning, which extracts features from the text and trains a model to establish a classifier to judge whether the text contains sensitive content. Although these traditional methods show certain recognition effects in specific scenarios, due to the lack of in-depth modeling and effective utilization of the potential semantic association relationships between sensitive words, the recognition results often cannot accurately capture sensitive content with similar semantics but different expression forms, resulting in difficulties in ensuring the comprehensiveness and accuracy of recognition. Through the above background introduction, those skilled in the art can understand the problems existing in the prior art. Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments.

[0013] Based on the above background technology, further, please refer to Figure 1 , Figure 1 FIG. is a schematic structural diagram of an oil burning diagnosis and maintenance system provided by an embodiment of the present application. This system can be implemented depending on a computer program or run as an independent tool application. Specifically, in the embodiment of the present application, this method can be applied to a server, but can also be applied to electronic devices such as a server. An oil burning diagnosis and maintenance system includes the following modules: a multi-source data acquisition model 1, a heterogeneous data deep fusion module 2, a multi-level diagnosis module 3, and a predictive maintenance module 4; The multi-source data acquisition model 1 is used to collect multi-source data during the operation of the engine in real time through an integrated sensor array. The multi-source data includes vibration data, acoustic data, exhaust component data, and temperature distribution data; Specifically, the multi-source data acquisition model realizes data acquisition by setting an integrated sensor array at key parts of the engine. Specifically, an MEMS vibration sensor array is set at key positions such as the engine block, cylinder head, and crankcase, with a sampling frequency of 10 - 20 kHz, for collecting the vibration data of the engine. The vibration data can reflect the internal mechanical motion state of the engine. A broadband acoustic sensor is set around the engine, with a frequency response range of 20 Hz - 20 kHz, for collecting the acoustic data of the engine. The acoustic data can reflect the characteristic sounds generated during abnormal operation of the engine. A zirconia oxygen sensor and a NOx sensor are set at the exhaust pipe for collecting the exhaust component data of the engine. The exhaust component data can reflect the combustion condition of the engine. An infrared thermal imager with a resolution of not less than 640×480 pixels is arranged on the engine surface for collecting the temperature distribution data of the engine. The temperature distribution data can reflect the thermal load conditions of various parts of the engine. Through this multi-source data acquisition method, comprehensive monitoring of the engine operation state is realized, overcoming the problem that data collected by a single sensor is easily affected by the environment, and providing a comprehensive and reliable data basis for subsequent fault diagnosis.

[0014] Based on the above embodiment, as an optional embodiment, the multi-source data acquisition model 1 further includes: a vibration data acquisition module 11, an acoustic data acquisition module 12, an exhaust data acquisition module 13, and a thermal data acquisition module 14; The vibration data acquisition module 11 is used to collect the vibration data of the engine through the MEMS vibration sensor array set at key parts of the engine; The acoustic data acquisition module 12 is used to collect the acoustic data of the engine through the broadband acoustic sensor; The exhaust data acquisition module 13 is used to collect the exhaust data of the engine through the zirconia oxygen sensor and the NOx sensor; The temperature data acquisition module 14 is used to collect the temperature data of the engine through the infrared thermal imager with a resolution of not less than 640×480 pixels.

[0015] Specifically, the specific process of the multi-source data acquisition model 1 for comprehensively monitoring the engine operating state is as follows: The vibration data acquisition module 11 arranges the MEMS vibration sensor array at key parts such as the engine cylinder block, cylinder head, and crankcase. The MEMS vibration sensor array is a high-precision sensor manufactured using microelectromechanical system technology, with a sampling frequency of 10 - 20 kHz, achieving precise acquisition of the engine mechanical vibration state and obtaining vibration data. The acoustic data acquisition module 12 sets broadband acoustic sensors around the engine. The frequency response range of the broadband acoustic sensors is 20 Hz - 20 kHz, covering the main acoustic frequency bands generated during engine operation, achieving the capture of abnormal engine sounds and obtaining acoustic data. The exhaust data acquisition module 13 installs a zirconia oxygen sensor and a NOx sensor at the exhaust pipe. The zirconia oxygen sensor is used to detect the oxygen content in the exhaust, and the NOx sensor is used to detect the nitrogen oxide content, achieving real-time monitoring of the engine emission components and obtaining exhaust data. The temperature data acquisition module 14 arranges an infrared thermal imager with a resolution of not less than 640×480 pixels on the engine surface. The infrared thermal imager converts the detected infrared radiation intensity into a temperature distribution image, achieving precise measurement of the engine thermal distribution and obtaining temperature data. Through the collaborative work of these four acquisition modules, multi-dimensional monitoring of the engine operating state is achieved, providing a comprehensive and reliable data basis for subsequent fault diagnosis.

[0016] The heterogeneous data deep fusion module 2 is used to perform data conversion on the multi-source data to obtain a unified multi-source feature representation; Specifically, the specific process of the heterogeneous data deep fusion module for data conversion of multi-source data is as follows: First, the multi-scale time-frequency analysis module processes the vibration data and acoustic data through wavelet packet decomposition. Wavelet packet decomposition decomposes the signal into sub-signals of different frequency bands, obtaining time-frequency feature data, which contains the complete feature information of vibration and acoustic signals in the time domain and frequency domain. Then, the data synchronization and alignment module uses the dynamic time warping algorithm to perform temporal alignment of heterogeneous data for the time-frequency feature data, temperature data, and exhaust data. The dynamic time warping algorithm calculates the optimal alignment path between different data sequences through dynamic programming, obtaining the aligned multi-modal data. Next, the data feature selection module performs feature selection on the aligned multi-modal data based on the mutual information criterion. The mutual information criterion screens the most representative features by calculating the correlation between features and fault categories, obtaining multi-modal feature data. Finally, the deep feature fusion module performs data fusion on the multi-modal feature data through a graph attention network. The graph attention network regards the features of different modalities as nodes in the graph and realizes the adaptive fusion of features by calculating the attention weights between nodes, obtaining a unified multi-source feature representation. This multi-step data fusion process solves the problem of temporal inconsistency of heterogeneous data, extracts the deep correlation features between data, and provides high-quality feature input for subsequent fault diagnosis.

[0017] Based on the above embodiments, as an alternative embodiment, please refer to Figure 2 , the heterogeneous data deep fusion module 2 further includes: a multi-scale time-frequency analysis module 21, a data synchronization and alignment module 22, a data feature selection module 23, and a deep feature fusion module 24; The multi-scale time-frequency analysis module 21 is used to obtain time-frequency feature data for the vibration data and the acoustic data through wavelet packet decomposition; Specifically, the specific process of the multi-scale time-frequency analysis module 21 for processing vibration data and acoustic data is as follows: First, the vibration data and acoustic data are input into the wavelet packet decomposition algorithm. Wavelet packet decomposition decomposes the signal at different scales, where the dual-tree complex wavelet transform is used as the basis function, and the decomposition level is 3 - 5 layers. In each layer of decomposition, the signal is divided into two parts: approximation coefficients and detail coefficients. Through wavelet packet decomposition, the vibration data is decomposed into sub-signals reflecting the mechanical motion characteristics of different frequency bands, and the acoustic data is decomposed into sub-signals characterizing the acoustic characteristics of different frequency bands. These sub-signals are reconstructed and combined to form time-frequency feature data containing complete information in the time domain and frequency domain. This multi-scale time-frequency analysis method not only retains the time-domain characteristics of the signal but also extracts the frequency-domain characteristics, effectively improving the recognition ability of the abnormal state of the engine.

[0018] The data synchronization and alignment module 22 is used to perform heterogeneous data time series alignment on the time-frequency feature data, the temperature data, and the exhaust data through the dynamic time warping algorithm to obtain aligned multi-modal data; Specifically, the specific process of the data synchronization and alignment module 22 for performing heterogeneous data time series alignment on the time-frequency feature data, the temperature data, and the exhaust data is as follows: First, a distance matrix of the time series data is constructed through the dynamic time warping algorithm, and this distance matrix reflects the similarity between data at different time points. Then, based on this distance matrix, the optimal alignment path is calculated using the dynamic programming method, and the optimal alignment path represents the time series correspondence between different data sequences. Next, according to the calculated optimal alignment path, time series mapping and resampling are performed on the time-frequency feature data, the temperature data, and the exhaust data to unify data with different sampling frequencies and different time scales to the same time base. Finally, the aligned data is combined into a multi-modal data sequence, achieving the time series consistency of heterogeneous data. Through this data alignment process, the problem of data asynchronization caused by inconsistent sampling frequencies of different types of sensors is solved, providing a time series aligned data basis for subsequent feature selection and fusion.

[0019] The data feature selection module 23 is used to receive the multi-modal data and perform feature selection on the aligned multi-modal data based on the mutual information criterion to obtain multi-modal feature data; Specifically, the specific process of the data feature selection module 23 for performing feature selection on the aligned multi-modal data is as follows: First, the multi-modal data is received, and the mutual information value between each feature in the multi-modal data and the fault label is calculated. The mutual information value reflects the statistical correlation between the feature and the fault category. Then, the features are sorted based on the mutual information criterion. The mutual information criterion evaluates the importance of features by calculating the conditional mutual information to remove redundant and irrelevant features. Next, a feature importance threshold is set, and a feature subset with a mutual information value higher than the threshold is selected to retain the key features with a high contribution degree to fault diagnosis. Finally, the selected features are combined to form multi-modal feature data, which contains the most discriminative features in each modality. Through this feature selection method based on the mutual information criterion, the redundancy of the data is reduced, the key features related to fault diagnosis are highlighted, and the efficiency and accuracy of subsequent feature fusion are improved.

[0020] The deep feature fusion module 24 is used to perform data fusion on the multi-modal feature data through a graph attention network to obtain the unified multi-source feature representation.

[0021] Specifically, the specific process of the deep feature fusion module 24 for fusing multi-modal feature data is as follows: First, the multi-modal feature data is constructed into a graph structure, where the nodes in the graph structure represent features of different modalities, and the edges between the nodes represent the correlation relationships between the features. Then, a graph attention network is used for feature fusion. The graph attention network includes 3-5 graph attention layers, and the number of attention heads in each layer is 4-8. The weight relationships between different feature nodes are calculated through the attention mechanism. Next, based on the calculated attention weights, different modality features are adaptively weighted to achieve dynamic fusion of features. Finally, the fused features are mapped to a unified feature space to obtain a unified multi-source feature representation. Through this feature fusion method based on the graph attention network, the correlation information between different modality features is fully utilized, realizing deep fusion of features and providing high-quality feature input for subsequent fault diagnosis.

[0022] The multi-level diagnosis module 3 is used to perform multi-level diagnosis on the multi-source feature representation to obtain multi-level diagnosis results; Specifically, the specific process of the multi-level diagnosis module for performing multi-level diagnosis on the unified multi-source feature representation is as follows: First, the feature extraction module extracts features from the unified multi-source feature representation through an improved ResNet structure. Among them, the residual feature extraction unit processes the data through multiple layers of convolution, the channel attention unit calculates the attention weights of each channel for feature enhancement, and the feature dimensionality reduction unit reduces the feature dimensionality through a 1×1 convolution kernel to obtain fault feature data. Then, the time series analysis module uses a BiLSTM network for time series modeling. Among them, the bidirectional LSTM encoding unit encodes the fault feature data forward and backward, the attention calculation unit calculates the time series attention weights through the softmax function, and the feature fusion unit performs weighted fusion to obtain time series feature data. Then, the knowledge reasoning module performs fault reasoning through a causal graph network. Among them, the causal relationship construction unit constructs a fault causal graph according to the time series feature data and a preset fault knowledge base, the node state calculation unit performs probability reasoning on the node states in the fault causal relationship data, and the fault location unit analyzes the fault propagation path to obtain fault reasoning data. Finally, the diagnosis decision module performs fusion analysis through an ensemble learning method. Among them, the base classifier training unit trains multiple base classifier models using different learning algorithms, the ensemble strategy unit fuses the prediction results through a weighted voting method, and the credibility analysis unit calculates the credibility score of the diagnosis result to obtain multi-level diagnosis results including the fault type, fault location, and fault severity. This progressive diagnosis process fully utilizes the advantages of deep learning and knowledge reasoning, improving the accuracy and interpretability of fault diagnosis.

[0023] Based on the above embodiments, as an alternative embodiment, the multi-level diagnosis module 3 further includes: a feature extraction module 31, a timing analysis module 32, a knowledge reasoning module 33, and a diagnosis decision-making module 34; The feature extraction module 31 is configured to extract features from the unified multi-source feature representation through an improved ResNet structure to obtain fault feature data; Specifically, the specific process of the feature extraction module 31 for extracting features from the unified multi-source feature representation is as follows: First, the residual feature extraction unit processes the unified multi-source feature representation through a multi-layer convolutional network. Each layer of convolution includes a common convolutional layer and a residual connection. The residual connection directly transmits the shallow features to the deep network through a skip connection method. Then, the channel attention unit receives the residual feature data, calculates the importance weights of each feature channel, adaptively weights each channel in the residual feature data to obtain weighted feature data, thereby highlighting the contribution of important feature channels. Finally, the feature dimension reduction unit uses a 1×1 convolutional kernel to perform dimension reduction processing on the weighted feature data. The dimension reduction process reduces data redundancy while retaining key feature information to obtain fault feature data. Through this feature extraction method based on the improved ResNet structure, the problem of difficult training of deep networks is solved, the effectiveness of feature extraction is enhanced, and high-quality feature input is provided for subsequent timing analysis.

[0024] Based on the above embodiments, as an alternative embodiment, the feature extraction module further includes: a residual feature extraction unit, a channel attention unit, and a feature dimension reduction unit; The residual feature extraction unit is configured to extract features from the unified multi-source feature representation through multi-layer convolution to obtain residual feature data; The channel attention unit is configured to calculate the attention weights of each channel according to the residual feature data and weight the residual feature data to obtain weighted feature data; The feature dimension reduction unit is configured to perform dimension reduction processing on the weighted feature data through a 1×1 convolutional kernel to obtain the fault feature data.

[0025] Specifically, the specific process of the feature extraction module for extracting features from the unified multi-source feature representation is as follows: First, the residual feature extraction unit adopts a multi-layer convolutional network structure. Each layer of convolution includes an ordinary convolutional layer and a residual connection. The ordinary convolutional layer extracts features through convolution operations, and the residual connection enables the deep network to directly learn feature residuals by establishing a direct connection between the nth layer and the (n + 2)th layer, so as to extract features from the unified multi-source feature representation and obtain residual feature data. Then, the channel attention unit receives the residual feature data, obtains the global information of each channel through global average pooling, calculates the channel attention weights using a two-layer fully connected network, and performs a weighting operation on each channel in the residual feature data to obtain weighted feature data. Finally, the feature dimensionality reduction unit processes the weighted feature data using a 1×1 convolutional kernel. The 1×1 convolutional kernel realizes the recombination and dimensionality reduction of information between channels through dot product operations in the spatial dimension, and obtains fault feature data with a lower dimension. Through this process of feature extraction and dimensionality reduction, the problem of network degradation is solved, the contribution of important feature channels is highlighted, and at the same time, the data redundancy is reduced, providing an efficient feature representation for fault diagnosis.

[0026] The time series analysis module 32 is used to perform time series modeling on the fault feature data through a BiLSTM network to obtain time series feature data; Specifically, the specific process of the time series analysis module 32 for performing time series modeling on the fault feature data is as follows: First, the bidirectional LSTM encoding unit encodes the fault feature data using two LSTM networks in the forward and backward directions. The forward LSTM network processes from the starting position of the sequence to capture the forward time series dependence, and the backward LSTM network processes from the end position of the sequence to capture the backward time series dependence. The encoding results in both directions form bidirectional time series encoded data. Then, the attention calculation unit calculates the importance weights of each time step in the bidirectional time series encoded data through the softmax function to obtain time series weight data, which reflects the contribution degree of different time steps to the fault features. Finally, the feature fusion unit performs weighted fusion on the bidirectional time series encoded data according to the time series weight data, highlighting the time series features with importance to obtain time series feature data. Through this time series modeling method based on the BiLSTM network, a two-way analysis of the fault development process is realized, the ability to capture time series patterns is enhanced, and a feature representation containing complete time series information is provided for subsequent fault reasoning.

[0027] Based on the above embodiments, as an optional embodiment, the time series analysis module further includes: a bidirectional LSTM encoding unit, an attention calculation unit, and a feature fusion unit; The bidirectional LSTM encoding unit is used to perform forward and backward encoding on the fault feature data to obtain bidirectional time series encoded data; The attention calculation unit is used to calculate the temporal attention weights of the bidirectional temporal encoding data through the softmax function to obtain temporal weight data; The feature fusion unit is used to perform weighted fusion on the bidirectional temporal encoding data according to the temporal weight data to obtain the temporal feature data.

[0028] Specifically, the specific process of the temporal analysis module for performing temporal modeling on the fault feature data is as follows: First, the bidirectional LSTM encoding unit uses two LSTM networks, a forward LSTM network and a backward LSTM network, to encode the fault feature data. Each LSTM network includes three control units: an input gate, a forget gate, and an output gate. The forward LSTM network processes from the starting position of the time series to capture the forward temporal dependence, and the backward LSTM network processes from the end position of the time series to capture the backward temporal dependence. The encoding results in both directions are combined to form bidirectional temporal encoding data. Then, the attention calculation unit evaluates the importance of the features at each time step in the bidirectional temporal encoding data, and maps the features to the 0-1 interval through the softmax function to obtain normalized attention coefficients, forming temporal weight data reflecting the importance of each time step. Finally, the feature fusion unit applies the temporal weight data to the bidirectional temporal encoding data, assigns larger weight values to the features at time steps with high importance, assigns smaller weight values to the features at time steps with low importance, and obtains the temporal feature data through weighted summation. Through this temporal modeling method based on bidirectional LSTM and attention mechanism, the comprehensive capture of fault temporal features and the highlighting of important information are achieved, providing a high-quality temporal feature representation for subsequent fault reasoning.

[0029] The knowledge reasoning module 33 is used to perform fault reasoning on the temporal feature data through a causal graph network to obtain fault reasoning data; Specifically, the specific process of the knowledge reasoning module 43 for fault reasoning with respect to the time-series feature data is as follows: First, the causal relationship construction unit constructs a fault causal graph based on the time-series feature data and a preset fault knowledge base. The nodes in the graph represent faulty components and symptom features, and the edges represent fault propagation relationships. In this way, expert experience and data features are combined to form fault causal relationship data. Then, the node state calculation unit calculates the states of the nodes in the fault causal relationship data through a probabilistic reasoning method. The calculation process is based on the Bayesian reasoning principle, combined with historical fault cases and current observation data, to obtain node state data reflecting the fault probabilities of the nodes. Finally, the fault location unit analyzes the fault propagation path based on the node state data, determines the root cause fault and the affected range by tracing the fault propagation link, and obtains fault reasoning data including the fault location result. Through this reasoning method based on the causal graph network, the combination of data-driven and knowledge-driven improves the interpretability and accuracy of fault diagnosis, providing a reliable reasoning basis for subsequent diagnostic decisions.

[0030] Based on the above embodiments, as an alternative embodiment, the knowledge reasoning module further includes: a causal relationship construction unit, a node state calculation unit, and a fault location unit; The causal relationship construction unit is configured to construct a fault causal graph according to the time-series feature data and a preset fault knowledge base, and obtain fault causal relationship data; The node state calculation unit is configured to calculate the states of the nodes in the fault causal relationship data through probabilistic reasoning, and obtain node state data; The fault location unit is configured to analyze the fault propagation path based on the node state data, and obtain the fault reasoning data.

[0031] Specifically, the specific process of the knowledge reasoning module for fault reasoning on the time-series feature data is as follows: First, the causal relationship construction unit constructs a fault causal graph by using the expert experience and fault cases stored in the preset fault knowledge base and combining with the time-series feature data. The nodes in this causal graph represent fault components and symptom features, and the connecting edges between the nodes represent the propagation relationship and influence intensity of the faults. In this way, the fault causal relationship data is obtained in the form of a graph structure. Then, the node state calculation unit performs state calculation on each node in the fault causal relationship data, adopts the conditional probability reasoning method of the Bayesian network, and calculates the probability distribution of each fault node according to the observed symptom features to obtain the node state data reflecting the fault states of each node. Finally, based on the node state data, the fault location unit traces the propagation link of the fault by analyzing the connection relationship and state probability between the nodes, identifies the root cause component and the affected range of the fault, and obtains the fault reasoning data including the fault source, propagation path, and influence degree. Through this fault diagnosis method that combines expert knowledge and probability reasoning, the interpretability and accuracy of fault diagnosis are improved, providing a reliable basis for maintenance decision-making.

[0032] The diagnosis and decision-making module 34 is used to perform fusion analysis on the fault reasoning data through an ensemble learning method to obtain the multi-level diagnosis results, where the multi-level diagnosis results include the fault type, fault location, and fault severity.

[0033] Specifically, the specific process of the diagnosis and decision-making module 34 for performing fusion analysis on the fault reasoning data is as follows: First, the base classifier training unit trains the fault reasoning data by using different learning algorithms, including decision trees, support vector machines, and random forests, etc. Multiple base classifier models are constructed through the combination of multiple classification methods, and each base classifier model learns the fault features from different perspectives. Then, the ensemble strategy unit fuses the prediction results of multiple base classifier models through a weighted voting method. Different weights are assigned in the weighted voting process according to the performance of each base classifier to obtain the fused diagnosis data. Finally, the credibility analysis unit calculates the credibility score of the diagnosis result according to the fused diagnosis data, and generates the multi-level diagnosis results including the fault type, fault location, and fault severity in combination with the credibility score and the fused diagnosis data. Through this decision-making method based on ensemble learning, the limitations of a single classifier are reduced, and the reliability and robustness of the diagnosis result are improved, providing an accurate diagnosis basis for subsequent maintenance decision-making.

[0034] The predictive maintenance module 4 is used to perform maintenance analysis according to the multi-level diagnosis results to obtain maintenance decision-making suggestions including the predicted life value, fault evolution path, and optimal maintenance time.

[0035] Specifically, the specific process of the predictive maintenance module performing maintenance analysis on the multi-layer diagnosis results is as follows: First, the remaining life prediction module predicts the life of the multi-layer diagnosis results through deep survival analysis. The deep survival analysis uses the survival analysis method to establish an equipment life prediction model to obtain a predicted life value. Next, the fault evolution analysis module performs fault trend analysis on the multi-layer diagnosis results through a dynamic Bayesian network. The dynamic Bayesian network characterizes the fault development law by establishing a time series probability graph model to obtain a fault evolution path. Then, the maintenance decision optimization module optimizes and analyzes the predicted life value and the fault evolution path through a reinforcement learning algorithm, wherein the state space construction unit constructs the maintenance decision state space according to the predicted life value and the fault evolution path to obtain state feature data, the action strategy generation unit analyzes the state feature data through a double Q learning algorithm to generate a maintenance action strategy to obtain strategy evaluation data, and the reward calculation unit calculates the comprehensive reward value of maintenance cost and equipment reliability according to the strategy evaluation data, and determines the optimal maintenance time according to the comprehensive reward value, and finally generates a maintenance decision recommendation. This maintenance decision method based on predictive analysis realizes the transition from passive maintenance to active predictive maintenance, reduces maintenance costs, and improves maintenance efficiency.

[0036] Based on the above embodiment, as an optional embodiment, the predictive maintenance module further includes: a remaining life prediction module, a fault evolution analysis module and a maintenance decision optimization module; The remaining life prediction module is used to perform life prediction on the multi-layer diagnosis results through deep survival analysis to obtain a predicted life value; Specifically, the specific process of the remaining life prediction module for life prediction of multi-layer diagnosis results is as follows: First, the multi-layer diagnosis results are input into the deep survival analysis model, which uses the Weibull-Cox mixed model as the basic framework, in which the Weibull distribution is used to model the benchmark life distribution, and the shape parameter range is set to 1.5-2.5. Then, the deep neural network is used to perform nonlinear mapping on the characteristics such as fault type, fault location and fault severity to obtain a feature representation reflecting the current state of the equipment. Next, the feature representation is combined with the Weibull-Cox mixed model to calculate the conditional failure probability distribution under the current fault state, which takes into account multiple influencing factors such as the equipment's usage time, fault state and operating environment. Finally, the average remaining life is calculated based on the conditional failure probability distribution to obtain the predicted life value. Through this life prediction method based on deep survival analysis, accurate prediction of the remaining life of the equipment is achieved, providing an important time reference for maintenance decisions.

[0037] The fault evolution analysis module is used to perform fault trend analysis on the multi-layer diagnosis results through a dynamic Bayesian network to obtain a fault evolution path; Specifically, the specific process of the fault evolution analysis module for performing fault trend analysis on multi-layer diagnosis results is as follows: First, the multi-layer diagnosis results are input into a dynamic Bayesian network as observation data. The dynamic Bayesian network models the fault state through time slices, and each time slice contains a set of nodes reflecting the current fault state. Then, based on historical fault data, a state transition probability matrix is learned. This matrix describes the evolution law of the fault state between different time slices, and the current state vector is calculated through the forward algorithm. Next, the particle filter algorithm is used to perform multi-step prediction on the future state. The particle filter approximates the state distribution by maintaining a set of weighted particles, and multiple possible paths of fault development are predicted. Finally, according to the probability weights of each path, the most likely fault evolution sequence is extracted to form a fault evolution path. Through this fault trend analysis method based on the dynamic Bayesian network, accurate modeling and prediction of the fault development law are achieved, providing a basis for formulating preventive maintenance strategies.

[0038] The maintenance decision optimization module is used to perform optimization analysis on the predicted life value and the fault evolution path through a reinforcement learning algorithm to obtain the optimal maintenance time, and generate a maintenance decision recommendation by combining the predicted life value, the fault evolution path, and the optimal maintenance time.

[0039] Specifically, the specific process of the maintenance decision optimization module for performing optimization analysis on the predicted life value and the fault evolution path is as follows: First, the state space construction unit constructs a maintenance decision state space based on the predicted life value and the fault evolution path. This state space includes features in multiple dimensions such as the device health state, maintenance cost, and reliability, forming state feature data. Then, the action policy generation unit analyzes the state feature data using the double Q-learning algorithm, where the discount factor is set to 0.85 - 0.95, and two independent Q-networks are alternately updated to avoid overestimation of the decision value, generating a maintenance action policy including the maintenance timing and maintenance method to obtain policy evaluation data. Finally, the reward calculation unit calculates the comprehensive reward value according to the policy evaluation data, combines the maintenance cost and the device reliability, determines the optimal maintenance time by maximizing the long-term cumulative reward, and generates a maintenance decision recommendation including specific maintenance measures and execution times by combining the predicted life value, the fault evolution path, and the optimal maintenance time. Through this decision optimization method based on reinforcement learning, the adaptive optimization of the maintenance strategy is achieved, minimizing the maintenance cost while ensuring the device reliability.

[0040] Based on the above embodiments, as an alternative embodiment, the maintenance decision optimization module further includes: a state space construction unit, an action policy generation unit, and a reward calculation unit; The state space construction unit is used to construct a maintenance decision state space based on the predicted life value and the fault evolution path, and obtain state feature data; The action strategy generation unit is used to analyze the state feature data through a double Q-learning algorithm, generate a maintenance action strategy, and obtain policy evaluation data; The reward calculation unit is used to calculate a comprehensive reward value of maintenance cost and equipment reliability according to the policy evaluation data, and determine the optimal maintenance time according to the comprehensive reward value and generate the maintenance decision suggestion.

[0041] Specifically, the specific process of the maintenance decision optimization module for optimizing the maintenance decision is as follows: First, the state space construction unit takes the predicted life value and the fault evolution path as inputs, constructs a maintenance decision state space including multi-dimensional features such as equipment health status, remaining life, and fault development trend, and obtains state feature data through feature vectorization and normalization processing. Then, the action strategy generation unit uses a double Q-learning algorithm to analyze the state feature data. The double Q-learning algorithm constructs two independent Q networks through a maintenance action set, a state transition probability, and an immediate reward function, and alternately updates the network parameters to avoid over-optimization of policy evaluation, generates a maintenance action strategy including maintenance timing and maintenance plan, and obtains policy evaluation data. Finally, the reward calculation unit comprehensively calculates the cost expenditure and equipment reliability improvement brought by the maintenance behavior according to the policy evaluation data, combines these two indicators in a weighted manner to form a comprehensive reward value, determines the optimal maintenance time that can balance the maintenance cost and equipment reliability based on this comprehensive reward value, and generates a maintenance decision suggestion including specific maintenance measures and execution time. Through this multi-level maintenance decision optimization process, the precise formulation and dynamic adjustment of the maintenance strategy are realized, ensuring the optimization of the maintenance effect.

[0042] The above are only exemplary embodiments of the present disclosure and should not be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure still fall within the scope covered by the present disclosure. Those skilled in the art will easily think of other implementation schemes of the present disclosure after considering the specification and the disclosure of the practical truth.

[0043] This application aims to cover any variations, uses, or adaptive changes of the present disclosure, which follow the general principles of the present disclosure and include common general knowledge or conventional technical means in the technical field not recorded in the present disclosure. The specification and the embodiments are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. An oil consumption diagnosis and maintenance system, characterized in that The system includes: a multi-source data acquisition model, a heterogeneous data deep fusion module, a multi-level diagnosis module, and a predictive maintenance module; The multi-source data acquisition model is used to collect multi-source data during the operation of the engine in real time through an integrated sensor array. The multi-source data includes vibration data, acoustic data, exhaust component data, and temperature distribution data; The heterogeneous data deep fusion module is used to perform data conversion on the multi-source data to obtain a unified multi-source feature representation; The multi-level diagnosis module is used to perform multi-level diagnosis on the multi-source feature representation to obtain multi-level diagnosis results; The predictive maintenance module is used to perform maintenance analysis based on the multi-level diagnosis results to obtain maintenance decision suggestions including predicted life values, fault evolution paths, and optimal maintenance times.

2. The system according to claim 1, wherein The multi-source data acquisition model further includes: a vibration data acquisition module, an acoustic data acquisition module, an exhaust data acquisition module, and a thermal data acquisition module; The vibration data acquisition module is used to collect the vibration data of the engine through a MEMS vibration sensor array arranged at key parts of the engine; The acoustic data acquisition module is used to collect the acoustic data of the engine through a broadband acoustic sensor; The exhaust data acquisition module is used to collect the exhaust data of the engine through a zirconia oxygen sensor and a NOx sensor; The temperature data acquisition module is used to collect the temperature data of the engine through an infrared thermal imager with a resolution of not less than 640×480 pixels.

3. The system according to claim 2, characterized in that, The heterogeneous data deep fusion module further includes: a multi-scale time-frequency analysis module, a data synchronization and alignment module, a data feature selection module, and a deep feature fusion module; The multi-scale time-frequency analysis module is used to obtain time-frequency feature data for the vibration data and the acoustic data through wavelet packet decomposition; The data synchronization and alignment module is used to perform heterogeneous data time series alignment on the time-frequency feature data, the temperature data, and the exhaust data through the dynamic time warping algorithm to obtain aligned multi-modal data; The data feature selection module is used to receive the multi-modal data and perform feature selection on the aligned multi-modal data based on the mutual information criterion to obtain multi-modal feature data; The deep feature fusion module is used to perform data fusion on the multi-modal feature data through a graph attention network to obtain the unified multi-source feature representation.

4. The system according to claim 1, wherein The multi-level diagnosis module further includes: a feature extraction module, a time series analysis module, a knowledge reasoning module, and a diagnosis decision module; The feature extraction module is used to extract features from the unified multi-source feature representation through an improved ResNet structure to obtain fault feature data; The time series analysis module is used to perform time series modeling on the fault feature data through a BiLSTM network to obtain time series feature data; The knowledge reasoning module is used to perform fault reasoning on the time series feature data through a causal graph network to obtain fault reasoning data; The described diagnostic decision-making module is used to perform fusion analysis on the fault inference data through an integrated learning method to obtain the multi-layer diagnostic results, where the multi-layer diagnostic results include fault type, fault location, and fault severity.

5. The system according to claim 4, wherein The feature extraction module further includes: a residual feature extraction unit, a channel attention unit, and a feature dimensionality reduction unit; The residual feature extraction unit is used to perform feature extraction on the unified multi-source feature representation through multi-layer convolution to obtain residual feature data; The channel attention unit is used to calculate the attention weights of each channel according to the residual feature data, and weight the residual feature data to obtain weighted feature data; The feature dimensionality reduction unit is used to perform dimensionality reduction processing on the weighted feature data through a 1×1 convolutional kernel to obtain the fault feature data.

6. The system according to claim 4, wherein The time series analysis module further includes: a bidirectional LSTM encoding unit, an attention calculation unit, and a feature fusion unit; The bidirectional LSTM encoding unit is used to perform forward and backward encoding on the fault feature data to obtain bidirectional time series encoding data; The attention calculation unit is used to calculate the time series attention weights of the bidirectional time series encoding data through the softmax function to obtain time series weight data; The feature fusion unit is used to perform weighted fusion on the bidirectional time series encoding data according to the time series weight data to obtain the time series feature data.

7. The system according to claim 4, characterized in that The knowledge inference module further includes: a causal relationship construction unit, a node state calculation unit, and a fault location unit; The causal relationship construction unit is used to construct a fault causal graph according to the time series feature data and a preset fault knowledge base to obtain fault causal relationship data; The node state calculation unit is used to calculate the states of each node in the fault causal relationship data through probability inference to obtain node state data; The fault location unit is used to analyze the fault propagation path according to the node state data to obtain the fault inference data.

8. The system according to claim 4, characterized in that The diagnostic decision-making module further includes: a base classifier training unit, an integration strategy unit, and a credibility analysis unit; The base classifier training unit is used to train the fault inference data using different learning algorithms to obtain multiple base classifier models; The integration strategy unit is used to fuse the prediction results of the multiple base classifier models through a weighted voting method to obtain fusion diagnostic data; The credibility analysis unit is used to calculate the credibility score of the diagnostic result according to the fusion diagnostic data, and generate the multi-layer diagnostic result by combining the credibility score and the fusion diagnostic data.

9. The system according to claim 1, wherein The predictive maintenance module further includes: a remaining life prediction module, a fault evolution analysis module, and a maintenance decision optimization module; The remaining life prediction module is used to perform life prediction on the multi-layer diagnostic results through deep survival analysis to obtain a predicted life value; The fault evolution analysis module is used to perform fault trend analysis on the multi-layer diagnostic results through a dynamic Bayesian network to obtain a fault evolution path; The maintenance decision optimization module is used to optimize and analyze the predicted life value and the fault evolution path through a reinforcement learning algorithm to obtain the optimal maintenance time, and generate maintenance decision suggestions in combination with the predicted life value, the fault evolution path, and the optimal maintenance time.

10. The system according to claim 9, characterized in that The maintenance decision optimization module further includes: a state space construction unit, an action policy generation unit, and a reward calculation unit; The state space construction unit is used to construct a maintenance decision state space based on the predicted life value and the fault evolution path to obtain state feature data; The action policy generation unit is used to analyze the state feature data through a double Q learning algorithm to generate a maintenance action policy and obtain policy evaluation data; The reward calculation unit is used to calculate a comprehensive reward value of maintenance cost and equipment reliability according to the policy evaluation data, and determine the optimal maintenance time and generate the maintenance decision suggestion according to the comprehensive reward value.

Images