Diesel engine state monitoring and fault diagnosis system based on dynamic characteristic migration

Abstract

The invention relates to the technical field of diesel engine state monitoring and fault diagnosis, in particular to a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration. The invention discloses a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration, and the system comprises a sensor collection module which obtains the monitoring data of a diesel engine through sensors disposed at different structures of the diesel engine, and carries out the preprocessing and time alignment of the monitoring data, and generates a time sequence monitoring sequence; and the feature extraction module performs single-sensor feature extraction on the time sequence monitoring sequence to generate a feature sequence, performs cross-sensor feature extraction on the time sequence monitoring sequence to generate an associated feature sequence, and constructs a fusion feature vector according to the feature sequence and the associated feature sequence. According to the method, a single-sensor and cross-sensor combined feature extraction strategy is adopted, dimensionality reduction fusion is performed through principal component analysis, hidden static and dynamic correlation features in data are fully mined, and early weak fault signs are effectively captured.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of diesel engine state monitoring and fault diagnosis, and particularly relates to a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration. BACKGROUND

[0002] As a core power device, diesel engines are widely used in commercial vehicles, engineering machinery and ship fields. The operation reliability of diesel engines directly determines the equipment attendance rate and operation cost. According to industry research data, diesel engine failures mainly concentrate on three categories: lubrication system failure, abnormal fuel supply and cooling system failure, and the traditional maintenance mode has significant limitations. Post-maintenance lag: relying on manual inspection or troubleshooting after failure, such as when cylinder pull-out failure occurs, irreversible wear of the piston-cylinder liner has already been caused. Limitations of single sensor monitoring: existing technologies mostly use simple sensor combinations, which are difficult to identify early failures and increase the rate of missed diagnosis.

[0003] Inefficiency of fault feature extraction: diesel engine operating environments have strong vibrations and multiple interference changes in working conditions. Traditional time-domain feature extraction methods are easily disturbed, leading to confusion between fault features and noise signals and reducing the diagnosis rate.

[0004] Complexity of fault location: traditional diagnosis methods can only preliminarily determine the fault type and cannot establish a direct correlation between fault types, fault causes and sensor signals. Fault troubleshooting relies on the experience of technical personnel, takes a long time, and affects the efficiency of equipment restart.

[0005] Therefore, there is an urgent need for a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration, which can solve the problems of high missed diagnosis rate, diagnosis lag and inaccurate fault location of traditional monitoring systems, realize real-time identification of early failures of diesel engine components, break through the bottleneck of feature extraction of multiple source heterogeneous signals, design a feature extraction method with anti-interference and strong generalization, eliminate the influence of working condition changes and environmental interference on the diagnosis results, establish a correlation model of fault type-fault cause-sensor signal, realize the integration of fault location and cause analysis, and shorten the fault troubleshooting time. SUMMARY

[0006] In order to overcome the shortcomings of high missed diagnosis rate and diagnosis lag of traditional monitoring technology, the present application provides a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration.

[0007] The technical implementation scheme of the present application is: a diesel engine state monitoring and fault diagnosis system based on dynamic feature migration, comprising: The sensor acquisition module acquires monitoring data of the diesel engine through sensors deployed at different structures of the diesel engine, pre-processes and time-aligns the monitoring data to generate a time-series monitoring sequence; The feature extraction module extracts single-sensor features from the time-series monitoring sequence to generate a feature sequence, extracts cross-sensor features from the time-series monitoring sequence to generate a correlation feature sequence, and constructs a fusion feature vector based on the feature sequence and the correlation feature sequence. The fault diagnosis module diagnoses based on the fusion feature vector through a fault diagnosis model to generate a model diagnosis result, performs rule reasoning based on the time-series monitoring sequence and the fusion feature vector to generate a rule diagnosis result, and determines a final diagnosis result according to the model diagnosis result and the rule diagnosis result. The risk assessment module performs risk assessment based on the time-series monitoring sequence, the feature sequence, the correlation feature sequence, and the final diagnosis result to obtain a risk index. The maintenance management module generates early warning information and maintenance schemes based on the final diagnosis result and the risk index to assist maintenance personnel in decision-making.

[0008] Preferably, the sensor acquisition module acquires monitoring data of the diesel engine through sensors deployed at different structures of the diesel engine, pre-processes and time-aligns the monitoring data to generate a time-series monitoring sequence, including: the sensors include oil multi-parameter sensors, oil temperature sensors, oil pressure sensors, water temperature sensors, exhaust temperature sensors, speed sensors, and high-pressure common rail pressure sensors, monitoring data with acquisition timestamps is acquired through the sensors, the monitoring data of different sensors is pre-processed, and the monitoring data is time-aligned according to the acquisition timestamps to generate a time-series monitoring sequence.

[0009] Preferably, the feature extraction module extracts single-sensor features from the time-series monitoring sequence to generate a feature sequence, extracts cross-sensor features from the time-series monitoring sequence to generate a correlation feature sequence, and constructs a fusion feature vector based on the feature sequence and the correlation feature sequence, including: fuzzy entropy analysis, trend slope calculation, and fluctuation statistical analysis are performed on the time-series monitoring data to generate a feature sequence, correlation analysis, fitting analysis, and mutual correlation analysis are performed on the time-series monitoring data to generate a correlation feature sequence, the feature sequence and the correlation feature sequence are reduced in dimension through principal component analysis, and a fusion feature vector is constructed.

[0010] Preferably, the fault diagnosis module, based on the fusion feature vector, performs diagnosis through a fault diagnosis model to generate a model diagnosis result, performs rule reasoning based on the time sequence monitoring sequence and the fusion feature vector to generate a rule diagnosis result, and determines a final diagnosis result according to the model diagnosis result and the rule diagnosis result, comprises: performing fault diagnosis through a CNN-LSTM model based on the fusion feature vector to obtain a model diagnosis result, the model diagnosis result comprising a fault type and a fault probability, and performing rule reasoning through a preset rule reasoning library based on the time sequence monitoring sequence and the fusion feature vector to obtain a rule diagnosis result, the rule diagnosis result comprising a number of triggered rules, a fault type and a fault confidence.

[0011] Preferably, the determination of the final diagnosis result according to the model diagnosis result and the rule diagnosis result comprises: performing result consistency test on the model diagnosis result and the rule diagnosis result, if the fault type of the maximum fault probability in the model diagnosis result is consistent with the fault type of the maximum fault confidence in the rule diagnosis result, taking the fault type of the maximum fault probability as the final fault type and calculating a comprehensive score by using a weighted voting method, and if not consistent, calculating an intersection score matrix based on the top three fault types and fault probabilities in the model diagnosis result and the top three fault types and fault confidences in the rule diagnosis result, and performing secondary test based on the intersection score matrix.

[0012] Preferably, the secondary test based on the intersection score matrix comprises: if the fault types corresponding to the row and column of the highest score element in the intersection score matrix are consistent, taking the fault types corresponding to the row and column of the highest score element as the final fault type, and taking the intersection score of the highest score element as the comprehensive score, and if not consistent, further verifying a verification score by combining the typical characteristics of the fault mode library, taking the fault type with the highest verification score as the final fault type, and taking the verification score as the comprehensive score, the final diagnosis result comprising the final fault type and the comprehensive score.

[0013] Preferably, the risk assessment module, according to the time sequence monitoring sequence, feature sequence, associated feature sequence and final diagnosis result, carries out risk assessment to obtain a risk index, comprising: based on the feature sequence analysis feature change rate, carries out weighted summation calculation of feature change rate to obtain a fault development speed factor, based on the final diagnosis result obtains a fault type, according to the fault type from rule reasoning library query obtains a fault grade, based on the time sequence monitoring sequence, feature sequence and associated feature sequence obtains an abnormal sensor category number, total sensor category number, abnormal feature number and total feature number, according to the trigger rule number, abnormal sensor category number, total sensor category number, abnormal feature number and total feature number, obtains a fault influence range factor, according to the fault development speed factor, fault grade and fault influence range factor, obtains a risk index.

[0014] Preferably, the risk index is calculated according to the fault development speed factor, fault grade and fault influence range factor, comprising: according to the trigger rule number, abnormal sensor category number, total sensor category number, abnormal feature number and total feature number, a fault influence range factor is calculated, the fault influence range factor is calculated by a fault influence range factor calculation formula, according to the fault development speed factor, fault grade and fault influence range factor, a risk index is calculated by a risk index calculation formula, the risk index calculation formula is: ; In the formula, is a risk index, is a fault development speed factor, is a fault grade, is a fault influence range factor.

[0015] The fault influence range factor calculation formula is: ; In the formula, is an abnormal sensor category number, is a total sensor category number, is an abnormal feature number, is a total feature number, is a trigger rule number, 、 and are weight coefficients.

[0016] Preferably, the maintenance management module generates early warning information and a maintenance scheme according to the final diagnosis result and the risk index to assist maintenance personnel in decision-making, including: dividing early warning levels according to the risk index and a risk threshold, generating early warning information according to the early warning levels, obtaining maintenance recommendations through a fault mode library according to the final diagnosis result and the risk index, generating a maintenance scheme based on the early warning information and the maintenance recommendations to assist maintenance personnel in maintenance decision-making, recording maintenance information of maintenance personnel, and regularly updating the fault mode library.

[0017] Preferably, the recording of the maintenance information of the maintenance personnel and the regular updating of the fault mode library include: generating a fault maintenance correspondence relationship according to the maintenance information, the model diagnosis result, the rule diagnosis result, the final diagnosis result and the risk index, and regularly updating the fault mode library according to the fault maintenance correspondence relationship.

[0018] The beneficial effects of the present application are: The present application deploys multiple types of sensors and synchronizes them in time to construct a time sequence monitoring sequence that comprehensively reflects the running state of the diesel engine; adopts a feature extraction strategy combining single sensors and cross-sensors, and uses principal component analysis for dimensionality reduction and fusion, which can fully excavate the static and dynamic correlation features hidden in the data and effectively capture early weak fault signs; on this basis, a CNN-LSTM data-driven model and a rule-based reasoning system are used for parallel diagnosis, and the diagnosis results of the two are subjected to consistency test and weighted fusion, which not only takes advantage of the powerful nonlinear feature learning and pattern recognition ability of the deep learning model, but also integrates the interpretability and logical rigor of the domain expert knowledge, thereby overcoming the shortcomings of single method such as being easily disturbed, insufficient generalization ability or incomplete knowledge, and improving the diagnosis accuracy and confidence for different fault types, especially complex faults and new faults.

[0019] The present application dynamically calculates a risk index by comprehensively analyzing the feature change rate, the fault level and the fault influence range factor composed of the proportion of abnormal sensors and features and the number of triggering rules. The risk index can quantitatively represent the severity and development urgency of the current fault state, so that the maintenance personnel can go beyond the simple judgment of whether there is a fault, realize scientific evaluation and priority sorting of the potential impact of the fault, and provide key data support for implementing predictive maintenance.

[0020] The application combines the diagnosis result with the risk index organically through the maintenance management module, automatically generates graded early warning information and targeted maintenance scheme, and assists the on-site personnel in maintenance decision-making, thereby changing from post-maintenance and regular maintenance to on-demand maintenance and predictive maintenance. Meanwhile, through the feedback recording mechanism of the maintenance process and result, the actual maintenance information can be used to continuously update the fault mode library, so that the diagnosis capability and adaptability of the whole system are continuously updated as the use time increases, thereby constructing a continuously developing intelligent operation and maintenance ecology. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A diesel engine state monitoring and fault diagnosis system structure diagram of the application; Figure 2 A fault diagnosis method flowchart of the application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0023] A diesel engine state monitoring and fault diagnosis system based on dynamic feature migration, as shown in Figure 1 and Figure 2 , comprises: A sensor acquisition module acquires monitoring data of the diesel engine through sensors deployed at different structures of the diesel engine, pre-processes the monitoring data, and generates a time sequence monitoring sequence through time alignment. The sensors include an oil multi-parameter sensor, an oil temperature sensor, an oil pressure sensor, a water temperature sensor, an exhaust temperature sensor, a speed sensor, and a high-pressure common rail pressure sensor. The monitoring data with a collection timestamp is acquired through the sensors, the monitoring data of different sensors is pre-processed, and the monitoring data is time-aligned according to the collection timestamp to generate a time sequence monitoring sequence.

[0024] It should be further explained that in the embodiment, the sensor acquisition module acquires monitoring data that can comprehensively reflect the working condition of the diesel engine by monitoring the running state of the diesel engine in real time through multiple sensors deployed at different structural positions of the diesel engine. During the running of the diesel engine, the lubrication system, the fuel supply system, the cooling system, and the power output system are all in a dynamic changing state, and the physical parameter change characteristics corresponding to different systems are different. Therefore, the sensors matched with different structural parts of the diesel engine are respectively arranged to realize the synchronous acquisition of multi-dimensional and multi-source data.

[0025] Specifically, the deployed sensors include an oil multi-parameter sensor, an engine oil temperature sensor, an engine oil pressure sensor, a water temperature sensor, an exhaust temperature sensor, a rotation speed sensor, and a high-pressure common rail pressure sensor. The oil multi-parameter sensor is arranged at a lubricating oil circuit or an oil sump position to collect multi-parameter information reflecting the state of the lubricating oil; the engine oil temperature sensor and the engine oil pressure sensor are respectively arranged at corresponding positions of an engine oil passage to collect temperature changes and pressure changes of the engine oil under different working conditions; the water temperature sensor is arranged at a cooling liquid circulation passage to collect the temperature state of the cooling system of the diesel engine; the exhaust temperature sensor is arranged at an exhaust manifold or an exhaust passage position to collect the exhaust temperature generated in the combustion process of the diesel engine; the rotation speed sensor is arranged at a corresponding position of a crankshaft or a flywheel to collect the running speed of the diesel engine; and the high-pressure common rail pressure sensor is arranged at a high-pressure common rail system of fuel to collect pressure changes of the fuel injection system.

[0026] When collecting and monitoring data, each sensor adds a collection time stamp to each piece of monitoring data according to a unified time reference, and the collection time stamp is used to identify the collection time of the corresponding monitoring data, so as to ensure that the monitoring data from different sensors have clear time attributes. In the above manner, an original monitoring data set containing various monitoring parameters and their corresponding collection time stamps is obtained. After obtaining the monitoring data, the monitoring data of different sensors are respectively preprocessed. The preprocessing includes integrity checking, outlier rejection, and data smoothing processing of the monitoring data, so as to eliminate abnormal data fluctuations caused by sensor jitter, electromagnetic interference, or instantaneous working condition mutations, thereby improving the stability of the monitoring data. After completing the preprocessing, the monitoring data from different sensors are time-aligned according to the collection time stamps carried by each monitoring data. By matching and sorting the collection time stamps, the multi-source monitoring data within the same time point or the same time window are correspondingly combined, so as to eliminate the time deviation problem caused by the inconsistency of the sampling frequencies of different sensors or communication delay. After the time alignment processing, the monitoring data from the oil multi-parameter sensor, the engine oil temperature sensor, the engine oil pressure sensor, the water temperature sensor, the exhaust temperature sensor, the rotation speed sensor, and the high-pressure common rail pressure sensor form a synchronous relationship in the time dimension.

[0027] Based on the preprocessed and time-aligned monitoring data, a time sequence monitoring sequence for subsequent processing is constructed in chronological order. The time sequence monitoring sequence can continuously and completely reflect the multi-parameter evolution of the diesel engine during the running process, and provides basic data support for the analysis and processing of the time sequence monitoring sequence by the subsequent feature extraction module.

[0028] The feature extraction module performs single-sensor feature extraction on the time series monitoring sequence to generate a feature sequence, performs cross-sensor feature extraction on the time series monitoring sequence to generate a correlation feature sequence, and constructs a fusion feature vector according to the feature sequence and the correlation feature sequence; The time series monitoring data is subjected to fuzzy entropy analysis, trend slope calculation and fluctuation statistical analysis to generate a feature sequence, is subjected to correlation analysis, fitting analysis and cross-correlation analysis to generate a correlation feature sequence, and the feature sequence and the correlation feature sequence are subjected to feature dimension reduction through principal component analysis to construct a fusion feature vector.

[0029] It should be further explained that in the embodiment, the feature extraction module is used to process the time series monitoring sequence output by the sensor acquisition module, and extract effective information capable of representing the diesel engine running state and fault evolution features from the multi-source monitoring data. Because the time series monitoring sequence contains multi-dimensional data from different sensors, and each type of data has differences in variation law and physical meaning, a hierarchical processing manner is adopted, and single-sensor feature extraction and cross-sensor feature extraction are respectively performed on the time series monitoring sequence to obtain the feature sequence and the correlation feature sequence.

[0030] For each type of sensor data in the time series monitoring sequence, single-sensor feature extraction is performed to generate a corresponding feature sequence. Single-sensor feature extraction takes the time series monitoring data of a single sensor as the processing object, and is used to depict the complexity variation, development trend and fluctuation of the sensor data itself, thereby reflecting the running state change of the diesel engine in the corresponding physical parameter dimension.

[0031] Specifically, fuzzy entropy analysis is performed on the time series monitoring data, the uncertainty and complexity of the sensor monitoring data in the time dimension are quantified, and a feature value reflecting the internal structure change of the monitoring data is extracted. Fuzzy entropy analysis can effectively represent the complexity difference of the monitoring data between normal state and abnormal state, so that even if the amplitude change of the monitoring data is not obvious in the early stage of fault, the internal structure change can also be reflected through fuzzy entropy analysis. The fuzzy entropy calculation process is as follows: first, a starting data point is selected from the time series monitoring sequence to construct a vector, the local similarity average value of the vector is calculated, and then the fuzzy entropy is calculated according to the local similarity average value of the vector. The local similarity average value calculation formula is as follows: ; In the formula, is the local similarity average value of the dimension vector with the point as the starting point, is the length of the time series monitoring sequence, is the embedding dimension, is the time delay, representing the index interval between adjacent data points in the sequence, is the similarity tolerance, usually taken as 1.5 times of the standard deviation of the sequence data, is the fuzzy membership function, such as exponential function, sigmoid function, is the Euclidean distance between the point starting dimensional vector and the point starting dimensional vector, is the normalization coefficient; fuzzy entropy calculation formula: ; wherein, is the fuzzy entropy value, is the local similarity average value of the point starting dimensional vector.

[0032] At the same time, trend slope calculation is performed on the time series monitoring data, which is used for temperature, pressure and other parameters, and represents the change trend. The change rate in the continuous time interval is calculated. Among them, the time interval is set according to the parameter characteristics: for temperature, pressure and other slowly changing parameters, 10-60 seconds are taken; for rotating speed and other fast changing parameters, 0.1-1 seconds are taken. Trend slope calculation is used to describe the rising or falling trend of key parameters of diesel engine with time, so as to provide a basis for judging whether the fault is in the development stage.

[0033] The fluctuation statistical analysis is performed on the time series monitoring data, which is used for rotating speed, common rail pressure and other parameters. The standard deviation of the sequence is calculated by using the sliding window method (the window length covers several working cycles, for example, 200 sampling points), which reflects the fluctuation amplitude. Fluctuation statistical analysis can be used to identify abnormal oscillation phenomena caused by working condition fluctuation, component wear or system instability. Through the above fuzzy entropy analysis, trend slope calculation and fluctuation statistical analysis, the characteristic sequence for describing the state change of a single sensor is generated. After completing the single sensor feature extraction, further cross-sensor feature extraction is performed on the time series monitoring sequence to generate the associated characteristic sequence.

[0034] Cross-sensor feature extraction is based on synchronized monitoring data in time dimension between different sensors, which is used to extract the change of correlation between different parameters of diesel engine, so as to reflect whether the abnormal state of multi-system coupling occurs. Specifically, the correlation analysis is performed on the time series monitoring data, and the Pearson correlation coefficient between the monitoring data of different sensors is calculated. The Pearson correlation coefficient represents the correlation characteristics of the linear correlation strength between different physical parameters. When the diesel engine is in normal operation state, the correlation between different parameters usually remains relatively stable, while in the process of fault occurrence or development, the correlation will change significantly. At the same time, the least square method is used for fitting analysis of the time series monitoring data, and the model prediction value is obtained. The actual observation value is obtained from the time series monitoring sequence, and the linear fitting goodness is calculated to evaluate the fitting effect, and the correlation characteristics reflecting the stability of the mapping relationship between parameters are extracted. The linear fitting goodness calculation formula is: ; In the formula, is the linear fitting goodness, is the th actual observation value, is the th model prediction value, is the actual observation value mean, and the fitting analysis can be used to judge whether the physical coupling relationship between different parameters is damaged, thereby assisting in identifying the abnormal state caused by performance degradation.

[0035] The cross-correlation analysis is performed on the time series monitoring data, the monitoring sequence of different sensors is selected based on the monitoring data of different sensors, the value of time shift parameter is changed, and the time shift cross-correlation function calculation is performed in time dimension, the correlation characteristics reflecting the change sequence of parameters are extracted, and the time shift cross-correlation function calculation formula is: ; In the formula, is the time shift cross-correlation function value of sequence and sequence when the time shift parameter is is the th value of sequence is the mean value of sequence is the th value of sequence is the mean value of sequence is the th value of sequence is the mean value of sequence is the th value of sequence The total number of sequence values. Cross-correlation analysis can be used to identify the time delay relationship changes between different parameters, providing the basis for the diagnosis of complex faults. Through the above correlation analysis, fitting analysis and cross-correlation analysis, the correlation feature sequence for describing the correlation state changes between multiple sensors is generated.

[0036] After obtaining the feature sequence and the correlation feature sequence, the two are fused. Since the feature sequence and the correlation feature sequence contain multiple dimensional feature information, there is a problem of feature redundancy and high dimension if used directly, therefore the feature sequence and the correlation feature sequence are taken as input, all feature values are sequentially spliced into a high-dimensional vector, and standardized to eliminate the dimension influence. The principal component analysis (PCA dimension reduction processing) is performed on the standardized feature matrix. The principal components with cumulative contribution rate ≥ 90% are selected to form the final fusion feature vector, remove redundant noise and retain the most core fault information. The fusion feature vector can comprehensively reflect the single sensor state change feature and the correlation change feature between multiple sensors, and provide high-quality input data basis for subsequent model-based diagnosis and rule-based reasoning processing.

[0037] The fault diagnosis module performs diagnosis through a fault diagnosis model based on the fusion feature vector, generates a model diagnosis result, performs rule-based reasoning based on the time sequence monitoring sequence and the fusion feature vector, and generates a rule diagnosis result, and determines a final diagnosis result according to the model diagnosis result and the rule diagnosis result; The CNN-LSTM model is used to perform fault diagnosis based on the fusion feature vector to obtain a model diagnosis result, and the model diagnosis result includes a fault type and a fault probability. The rule-based reasoning library is used to perform rule-based reasoning based on the time sequence monitoring sequence and the fusion feature vector to obtain a rule diagnosis result, and the rule diagnosis result includes a number of triggered rules, a fault type and a fault confidence.

[0038] The model diagnosis result and the rule diagnosis result are subjected to result consistency test. If the fault type of the maximum fault probability in the model diagnosis result is consistent with the fault type of the maximum fault confidence in the rule diagnosis result, the fault type of the maximum fault probability is taken as the final fault type and a comprehensive score is calculated by using a weighted voting method. If they are not consistent, a cross score matrix is calculated based on the top three fault types and fault probabilities in the model diagnosis result and the top three fault types and fault confidence in the rule diagnosis result, and secondary test is performed based on the cross score matrix.

[0039] If the row and column corresponding to the highest score element in the cross score matrix are consistent, the row and column corresponding to the highest score element is taken as the final fault type, and the cross score of the highest score element is taken as the comprehensive score; if not, the typical characteristics of the fault mode library are combined to further verify the verification score, the fault type with the highest verification score is taken as the final fault type, and the verification score is taken as the comprehensive score; the final diagnosis result includes the final fault type and the comprehensive score.

[0040] It needs to be further explained that, in the embodiment, as shown in Figure 2 The fault diagnosis module is used for comprehensive analysis of the fusion feature vector output by the feature extraction module and the time sequence monitoring sequence formed by the sensor acquisition module, so as to realize fault diagnosis of the diesel engine running state and output the final diagnosis result with a clear fault type. The model diagnosis and rule reasoning are combined, the same monitoring object is analyzed from different diagnosis mechanisms, and thus the accuracy and stability of the diagnosis result are improved.

[0041] Specifically, based on the fusion feature vector, diagnosis is performed through a fault diagnosis model to generate a model diagnosis result. The fusion feature vector contains multi-dimensional feature information after principal component analysis dimension reduction, and can comprehensively reflect single sensor feature changes and cross-sensor associated feature changes. A CNN-LSTM model is used to process the fusion feature vector, and a pair of fusion feature vector samples is formed after processing historical data of the diesel engine in normal state and multiple known faults (such as cylinder pulling and bearing wear) by a feature extraction module. A cross-entropy loss function and an Adam optimizer are used for supervised training. The performance is monitored by a validation set to prevent overfitting. The model extracts spatial features from the original fault features through a convolutional neural network (CNN). Input data: fusion feature vector. Convolutional layer 1: 32 3x3 convolutional kernels are used, followed by a ReLU activation function. Pooling layer 1: 2x2 max pooling with a step size of 2. Convolutional layer 2: 64 3x3 convolutional kernels are used, followed by a ReLU activation function. Pooling layer 2: 2x2 max pooling with a step size of 2. Convolutional layer 3: 128 2x2 convolutional kernels are used, followed by a ReLU activation function. Flatten layer: flatten the 3-dimensional feature map into a 1-dimensional 256-dimensional vector. The feature sequence extracted by the CNN is input into a long short-term memory network (LSTM) to capture the temporal dependence of the data. LSTM hidden layer 1: contains 64 units, uses 20% Dropout to prevent overfitting, and returns the entire sequence to the next layer. LSTM hidden layer 2: contains 32 units, also uses 20% Dropout, but does not return the sequence, only outputs the final result. The output of the LSTM stage is passed to a fully connected layer for fault classification. Fully connected layer 1: contains 64 neurons, uses a ReLU activation function and 20% Dropout. Fully connected layer 2 (output layer): uses a Softmax activation function to output the probability of each case. By analyzing the combination relationship between different features in the fusion feature vector and the time evolution characteristics, the corresponding model diagnosis result is output. The model diagnosis result includes the fault type and the fault probability corresponding to each fault type.

[0042] Based on the time series monitoring sequence and the fusion feature vector, rule reasoning is performed to generate a rule diagnosis result. Rule reasoning is achieved through a pre-set rule reasoning library, which is obtained through experience summary and engineering practice of personnel in the field. The rule reasoning library pre-stores a plurality of rule information corresponding to typical fault modes of the diesel engine, and each rule sets a trigger condition based on the monitoring data change characteristics in the time series monitoring sequence and the feature state in the fusion feature vector. The rule reasoning library includes rule number, trigger condition, fault type and fault level, and an example of the rule reasoning library is as follows:

[0043] When the timing monitoring sequence and the fusion feature vector satisfy the trigger condition of the corresponding rule, the corresponding rule is triggered and outputs a rule diagnosis result. The rule diagnosis result includes the number of triggered rules, the fault type indicated by the rule, and the fault confidence corresponding to the fault type, and the fault confidence formula is: ; In the formula, is the fault confidence, is the number of sub-conditions of the rule, is the actual value of the th feature, is the th condition threshold, is a normalization function.

[0044] After obtaining the model diagnosis result and the rule diagnosis result respectively, the two are fused to determine the final diagnosis result. The model diagnosis result and the rule diagnosis result are subjected to result consistency test. Specifically, the fault type with the maximum fault probability is selected from the model diagnosis result, and the fault type with the maximum fault confidence is selected from the rule diagnosis result, and the two are compared and judged.

[0045] When the fault type corresponding to the maximum fault probability in the model diagnosis result is consistent with the fault type corresponding to the maximum fault confidence in the rule diagnosis result, the consistent fault type is taken as the final fault type, and a comprehensive score is calculated based on the model diagnosis result and the rule diagnosis result using a weighted voting method. The weighted voting method is weighted calculation according to the fault probability of the model diagnosis result and the fault confidence of the rule diagnosis result, to obtain a comprehensive score reflecting the credibility of the final fault type.

[0046] When the fault type corresponding to the maximum fault probability in the model diagnosis result is inconsistent with the fault type corresponding to the maximum fault confidence in the rule diagnosis result, a secondary test process is further started. Specifically, the fault types with the top three fault probabilities and the corresponding fault probabilities are selected from the model diagnosis result, and the fault types with the top three fault confidences and the corresponding fault confidences are selected from the rule diagnosis result, and a cross score matrix is obtained by weighted calculation based on the above information. The cross score matrix represents the matching degree of the model diagnosis result and the rule diagnosis result under different fault type combinations.

[0047] After obtaining the cross-score matrix, a secondary test is performed based on the cross-score matrix. If the fault types indicated by the row and column corresponding to the highest-score element in the cross-score matrix are consistent, the consistent fault type is taken as the final fault type, and the cross-score corresponding to the highest-score element is taken as the comprehensive score. In this way, in the case where the model diagnosis result and the rule diagnosis result are different, the fault type with the highest matching degree is selected preferentially.

[0048] If the fault types indicated by the row and column corresponding to the highest-score element in the cross-score matrix are still inconsistent, the related fault types are further verified in combination with the typical characteristics of the fault mode library. The fault mode library is constructed based on the clustering analysis of the feature data of historical confirmed cases. According to the historical fault cases, the fusion feature vectors are extracted, and the case data of the same fault type are collected. For each data set of a fault type, the K-means clustering algorithm is used for clustering. For each clustering cluster, the mean of all fusion feature vectors in the cluster is calculated as the fault feature vector of the sub-mode, and the standard deviation of each feature dimension is calculated to determine the feature tolerance range. When a certain number of new confirmed cases are accumulated, the above clustering analysis process is re-executed, and the core features and tolerance range of the existing fault modes are dynamically updated, or new mode entries are formed, thereby realizing the continuous evolution of the fault mode library. The fault mode library stores the typical characteristic information corresponding to various faults, including fault type, fault feature vector (composed of key feature values), fault maintenance corresponding relationship and maintenance suggestion. The time sequence monitoring sequence and the fusion feature vector are subjected to similarity calculation with the typical fault feature vectors in the fault mode library to obtain verification scores for different fault types, and the fault type with the highest verification score is taken as the final fault type, and the corresponding verification score is taken as the comprehensive score. The diagnosis basis is provided for the subsequent risk assessment module and maintenance management module.

[0049] The risk assessment module performs risk assessment based on the time sequence monitoring sequence, the feature sequence, the associated feature sequence and the final diagnosis result to obtain a risk index. Based on the feature sequence analysis feature change rate, the feature change rate is weighted and summed to calculate a fault development speed factor, the fault type is obtained based on the final diagnosis result, the fault level is obtained from the rule reasoning library according to the fault type, the abnormal sensor category number, the total sensor category number, the abnormal feature number and the total feature number are obtained based on the time sequence monitoring sequence, the feature sequence and the associated feature sequence, and the fault influence range factor is calculated based on the trigger rule number, the abnormal sensor category number, the total sensor category number, the abnormal feature number and the total feature number. The risk index is calculated based on the fault development speed factor, the fault level and the fault influence range factor.

[0050] The fault influence range factor is calculated according to the number of trigger rules, the number of abnormal sensor categories, the total number of sensor categories, the number of abnormal features and the total number of features by a fault influence range factor calculation formula, the risk index is calculated according to the fault development speed factor, the fault level and the fault influence range factor by a risk index calculation formula, and the risk index calculation formula is: ; In the formula, is the risk index, is the fault development speed factor, is the fault level, is the fault influence range factor.

[0051] The fault influence range factor calculation formula is: ; In the formula, is the number of abnormal sensor categories, is the total number of sensor categories, is the number of abnormal features, is the total number of features, is the number of trigger rules, , and are weight coefficients.

[0052] It should be further explained that in the embodiment, the risk assessment module is used to quantitatively evaluate the fault risk degree of the current running state of the diesel engine after completing the fault diagnosis, so as to obtain the risk index for representing the comprehensive risk level of the fault. The time sequence monitoring sequence, the feature sequence, the associated feature sequence and the final diagnosis result are comprehensively utilized to analyze the fault risk from multiple dimensions of the fault development speed, the fault severity and the fault influence range, so as to realize the comprehensive evaluation of the running risk of the diesel engine.

[0053] Specifically, the feature change rate is analyzed based on the feature sequence. The feature sequence is obtained by the feature extraction module through fuzzy entropy analysis, trend slope calculation and fluctuation statistical analysis on the time sequence monitoring sequence, and can reflect the change characteristics of the parameter corresponding to a single sensor in the time dimension. The change of each feature in the feature sequence in the continuous time period is analyzed, and the feature change rate corresponding to each feature is calculated, which is used to represent the change speed of the monitoring parameter with time.

[0054] After obtaining the feature change rate corresponding to each feature, the feature change rate is weighted and summed to obtain the fault development speed factor. The fault development speed factor is used to represent the speed of the fault development from the initial state to the severe state.

[0055] According to the fault type, the fault level corresponding to the fault type is obtained from the rule inference library. After obtaining the fault development speed factor and the fault level, the fault influence range is further evaluated. Based on the time sequence monitoring sequence, the feature sequence and the associated feature sequence, the distribution of the abnormal state during the operation of the diesel engine is statistically analyzed. Specifically, by analyzing the abnormality of each sensor monitoring data in the time sequence monitoring sequence, the number of abnormal sensor categories and the total number of sensor categories are obtained; by analyzing the abnormality of each feature in the feature sequence and the associated feature sequence, the number of abnormal features and the total number of features are obtained.

[0056] At the same time, the number of triggered rules in the rule inference process is obtained. The number of triggered rules can reflect the number of abnormal patterns identified based on the rule inference library, and is closely related to the system range and complexity involved in the fault. Based on the number of triggered rules, the number of abnormal sensor categories, the total number of sensor categories, the number of abnormal features and the total number of features, the fault influence range factor is calculated.

[0057] Specifically, the fault influence range factor is calculated by the fault influence range factor calculation formula. In the calculation formula, the ratio of the number of abnormal sensor categories to the total number of sensor categories is used to reflect the coverage range of the abnormality at the sensor level, the ratio of the number of abnormal features to the total number of features is used to reflect the distribution range of the abnormality at the feature level, and the number of triggered rules is involved in the calculation in the form of an index, which is used to enhance the representation ability of the fault influence range in the case of multiple rules triggering at the same time. The weight coefficients in the fault influence range factor formula are determined by the analytic hierarchy process, and a hierarchical structure including a target layer (optimal weight), a criterion layer (index sensitivity, diagnostic contribution, engineering interpretability) and a scheme layer (weight coefficient value) is established. According to the 1-9 scale method, experts in multiple fields compare each level element with each other, and construct the judgment matrix of the criterion layer to the target layer and the scheme layer to each criterion, respectively. Then, the maximum eigenvalue and the corresponding eigenvector of each judgment matrix are calculated, and the eigenvector is normalized to obtain the local weight, which is tested by calculating the consistency ratio (CR) (CR<0.1 is required), and if it does not pass, the experts need to re-evaluate. Finally, the local weight of the scheme layer to each criterion is combined with the weight of the criterion layer to the target to obtain the optimal weight coefficient value. The fault influence range factor calculated by the above method can comprehensively reflect the influence degree of the fault on different systems and different parameters of the diesel engine.

[0058] After obtaining the fault development speed factor, the fault level and the fault influence range factor respectively, the risk index is further calculated according to the risk index calculation formula. In the risk index calculation formula, the risk index is determined by the fault development speed factor, the fault level and the fault influence range factor, which can comprehensively quantify the development speed, severity and influence range of the fault. By substituting the fault development speed factor, the fault level and the fault influence range factor into the risk index calculation formula, the risk index representing the fault risk level of the current diesel engine operating state is calculated. As the quantitative result of the current operating risk of the diesel engine, the risk index can provide a basis for the subsequent maintenance management module to generate warning information and maintenance scheme, so that the maintenance decision can be based on the clear risk assessment result.

[0059] The maintenance management module generates warning information and maintenance scheme according to the final diagnosis result and the risk index, and assists the maintenance personnel to make decisions.

[0060] According to the risk index and the risk threshold, the warning level is divided, the warning information is generated according to the warning level, the repair suggestion is obtained through the fault mode library according to the final diagnosis result and the risk index, the maintenance scheme is generated based on the warning information and the repair suggestion, the maintenance personnel is assisted to make repair decisions, the repair information of the maintenance personnel is recorded, and the fault mode library is updated regularly.

[0061] According to the repair information, the model diagnosis result, the rule diagnosis result, the final diagnosis result and the risk index, a fault maintenance corresponding relationship is generated, and the fault mode library is updated regularly according to the fault maintenance corresponding relationship.

[0062] It needs to be further explained that in the embodiment, the maintenance management module is used to comprehensively manage the fault information of the diesel engine operating state after completing the risk assessment, and provide clear and executable warning information and maintenance scheme to the maintenance personnel to assist the maintenance personnel to make repair decisions.

[0063] Specifically, the current operating state is judged according to the risk index and a preset risk threshold to divide a corresponding warning level. The risk threshold is used to distinguish operating states of different risk levels, and when the risk index is in different threshold intervals, different warning levels are correspondingly matched. The risk threshold is determined by analyzing historical fault cases, each of which contains a finally confirmed fault level (as a true label) and a calculated corresponding risk index. Data cleaning is performed on all risk index samples under each fault level to eliminate obvious outliers. Then, K-means clustering algorithm is applied to cluster analysis of risk index samples of four levels, and each cluster center value is taken as a typical value reference of the risk index of the level; specific quantiles of the risk index distribution of each level are calculated, for example, the risk index of the II level is arranged in ascending order, and the 85% quantile is taken as the initial lower threshold of the II level warning, and the 15% quantile is taken as the initial upper threshold of the I level warning, to establish a three-level threshold to divide four warning levels. By comparing the risk index with the risk threshold, the warning level to which the current fault risk belongs can be determined.

[0064] After determining the warning level, corresponding warning information is generated according to the warning level. The warning information is used to intuitively present the risk state of the current diesel engine to the maintenance personnel, and the content of the warning information corresponds to the warning level and can reflect the emergency degree and processing priority of the fault. Through the warning information, the maintenance personnel can quickly understand that the diesel engine is currently in a normal state, a state that needs attention or a state that needs to be handled immediately.

[0065] At the same time of generating the warning information, further maintenance suggestions matching the current fault state are obtained from the fault mode library according to the final diagnosis result and the risk index. The fault mode library pre-stores association information between multiple fault types and corresponding maintenance measures, and the maintenance management module determines the fault type according to the final diagnosis result and selects the maintenance suggestions corresponding to the current fault state from the fault mode library in combination with the risk degree reflected by the risk index. The maintenance suggestions are used to guide the maintenance personnel to take targeted maintenance measures. After obtaining the warning information and the maintenance suggestions, a maintenance scheme is generated based on the warning information and the maintenance suggestions. The maintenance scheme comprehensively considers the fault type, the risk degree and the content of the maintenance suggestions to form a maintenance decision basis for the maintenance personnel to refer. Through the maintenance scheme, the maintenance personnel can formulate a reasonable maintenance plan on the basis of understanding the fault risk level and the recommended processing mode, thereby improving the maintenance efficiency and reducing the risk of misjudgment.

[0066] After the maintenance personnel complete the actual maintenance operation, the maintenance process is recorded, specifically including recording the maintenance information of the maintenance personnel. The maintenance information is used to describe the actual maintenance measures taken and the corresponding processing results, and is an important data source for subsequent system optimization and knowledge updating. Through the recording of the maintenance information, maintenance data under the actual running environment can be continuously accumulated.

[0067] On the basis of recording the maintenance information, a fault maintenance corresponding relationship is further generated according to the maintenance information, the model diagnosis result, the rule diagnosis result, the final diagnosis result and the risk index. The fault maintenance corresponding relationship is used to describe the actual maintenance mode taken and its effect condition under the condition of a specific fault diagnosis result and risk assessment result, thereby establishing the association between the diagnosis result and the maintenance behavior. The fault mode library is updated regularly according to the fault maintenance corresponding relationship. Through continuous sorting and analysis of the historical maintenance information and the diagnosis result, the maintenance suggestions and the fault modes in the fault mode library can gradually fit the actual running situation, and dynamic optimization is realized. Through the above-mentioned manner, not only can support be provided for the current maintenance decision, but also through continuous updating of the fault mode library, the accuracy and practicality of subsequent fault diagnosis and maintenance management can be continuously improved.

[0068] The above only describes the preferred embodiments of the present application, and it should be noted that for ordinary skilled persons in the art, several improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A diesel engine condition monitoring and fault diagnosis system based on dynamic feature transfer, characterized in that, include: The sensor acquisition module acquires monitoring data of the diesel engine through sensors deployed on different structures of the diesel engine, and performs preprocessing and time alignment on the monitoring data to generate a time-series monitoring sequence. The feature extraction module performs single-sensor feature extraction on the time-series monitoring sequence to generate a feature sequence, performs cross-sensor feature extraction on the time-series monitoring sequence to generate a related feature sequence, and constructs a fusion feature vector based on the feature sequence and the related feature sequence. The fault diagnosis module performs diagnosis based on the fused feature vector through a fault diagnosis model, generates model diagnosis results, performs rule reasoning based on the time-series monitoring sequence and the fused feature vector, generates rule diagnosis results, and determines the final diagnosis result based on the model diagnosis results and the rule diagnosis results. The risk assessment module performs risk assessment based on the time-series monitoring sequence, feature sequence, associated feature sequence, and final diagnosis result to obtain a risk index; The maintenance management module generates early warning information and maintenance plans based on the final diagnostic results and the risk index to assist maintenance personnel in making decisions.

2. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature transfer according to claim 1, characterized in that, The sensor acquisition module acquires monitoring data of the diesel engine through sensors deployed on different structures of the diesel engine, preprocesses and aligns the monitoring data to generate a time-series monitoring sequence, including: the sensors include a multi-parameter oil sensor, an oil temperature sensor, an oil pressure sensor, a coolant temperature sensor, an exhaust temperature sensor, a speed sensor, and a high-pressure common rail pressure sensor; the module acquires monitoring data with acquisition timestamps through the sensors, preprocesses the monitoring data from different sensors, and aligns the monitoring data to generate a time-series monitoring sequence based on the acquisition timestamps.

3. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 1, characterized in that, The feature extraction module performs single-sensor feature extraction on the time-series monitoring sequence to generate a feature sequence, performs cross-sensor feature extraction on the time-series monitoring sequence to generate a related feature sequence, and constructs a fusion feature vector based on the feature sequence and the related feature sequence. This includes: performing fuzzy entropy analysis, trend slope calculation, and fluctuation statistical analysis on the time-series monitoring data to generate a feature sequence; performing correlation analysis, fitting analysis, and cross-correlation analysis on the time-series monitoring data to generate a related feature sequence; and performing feature dimensionality reduction on the feature sequence and the related feature sequence through principal component analysis to construct a fusion feature vector.

4. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 1, characterized in that, The fault diagnosis module performs fault diagnosis based on the fused feature vector using a fault diagnosis model to generate model diagnosis results, performs rule inference based on the time-series monitoring sequence and the fused feature vector to generate rule diagnosis results, and determines the final diagnosis result based on the model diagnosis results and the rule diagnosis results. This includes: obtaining model diagnosis results by performing fault diagnosis using a CNN-LSTM model based on the fused feature vector, where the model diagnosis results include fault type and fault probability; and obtaining rule diagnosis results by performing rule inference based on the time-series monitoring sequence and the fused feature vector using a preset rule inference library, where the rule diagnosis results include the number of triggering rules, fault type, and fault confidence.

5. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 4, characterized in that, The step of determining the final diagnosis result based on the model diagnosis result and the rule diagnosis result includes: performing a consistency test on the model diagnosis result and the rule diagnosis result; if the fault type with the highest failure probability in the model diagnosis result is consistent with the fault type with the highest failure confidence in the rule diagnosis result, then the fault type with the highest failure probability is taken as the final fault type and a weighted voting method is used to calculate the comprehensive score; if they are inconsistent, then a cross-score matrix is ​​calculated based on the top three fault types and failure probabilities in the model diagnosis result and the top three fault types and failure confidence in the rule diagnosis result, and a secondary test is performed based on the cross-score matrix.

6. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 5, characterized in that, The secondary verification based on the cross-score matrix includes: if the fault types corresponding to the row and column of the highest-scoring element in the cross-score matrix are consistent, then the fault type corresponding to the row and column of the highest-scoring element is taken as the final fault type, and the cross-score of the highest-scoring element is taken as the comprehensive score; if they are inconsistent, then the verification score is obtained by combining the typical features of the fault mode library, the fault type with the highest verification score is taken as the final fault type, and the verification score is taken as the comprehensive score. The final diagnosis result includes the final fault type and the comprehensive score.

7. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 6, characterized in that, The risk assessment module performs risk assessment based on the time-series monitoring sequence, feature sequence, associated feature sequence, and final diagnosis result to obtain a risk index. This includes: analyzing the feature change rate based on the feature sequence; calculating the fault development speed factor by weighted summation of the feature change rates; obtaining the fault type based on the final diagnosis result; querying the rule reasoning library to obtain the fault level based on the fault type; obtaining the number of abnormal sensor categories, the total number of sensor categories, the number of abnormal features, and the total number of features based on the time-series monitoring sequence, feature sequence, and associated feature sequence; calculating the fault impact range factor based on the number of triggering rules, the number of abnormal sensor categories, the total number of sensor categories, the number of abnormal features, and the total number of features; and calculating the risk index based on the fault development speed factor, the fault level, and the fault impact range factor.

8. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 7, characterized in that, The step of calculating the risk index based on the fault development speed factor, fault level, and fault impact range factor includes: calculating the fault impact range factor based on the number of triggering rules, the number of abnormal sensor categories, the total number of sensor categories, the number of abnormal features, and the total number of features; calculating the fault impact range factor using the fault impact range factor calculation formula; and calculating the risk index using the risk index calculation formula based on the fault development speed factor, fault level, and fault impact range factor. The risk index calculation formula is as follows: ; In the formula, As a risk index, As the fault development rate factor, Fault level, This is the factor affecting the scope of the fault. The formula for calculating the fault impact range factor is: ; In the formula, The number of abnormal sensor categories, This represents the total number of sensor categories. The number of abnormal features. The total number of features, To the number of rules to trigger, , and These are the weighting coefficients.

9. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 1, characterized in that, The maintenance management module generates early warning information and maintenance plans based on the final diagnosis results and the risk index to assist maintenance personnel in making decisions. This includes: classifying early warning levels based on the risk index and risk threshold; generating early warning information based on the early warning levels; obtaining maintenance suggestions from the fault mode library based on the final diagnosis results and the risk index; generating maintenance plans based on the early warning information and maintenance suggestions to assist maintenance personnel in making maintenance decisions; recording maintenance information of maintenance personnel; and regularly updating the fault mode library.

10. The diesel engine condition monitoring and fault diagnosis system based on dynamic feature migration according to claim 9, characterized in that, The process of recording maintenance personnel's maintenance information and regularly updating the fault mode library includes: generating fault maintenance correspondences based on the maintenance information, model diagnostic results, rule diagnostic results, final diagnostic results, and risk index, and regularly updating the fault mode library based on the fault maintenance correspondences.

Images