Engine fault maintenance strategy generation method based on multi-source data

Abstract

The application discloses an engine fault maintenance strategy generation method based on multi-source data, relates to the technical field of engine fault maintenance, collects engine operation sound signals and rotating speed signals, establishes an engine feature database, counts the historical abnormal frequencies of various features corresponding to fault types, constructs a related feature item set corresponding to various fault types of fixed points, analyzes the number of points of various fault types determined as risk faults, classifies according to the risk point proportion of various fault types, issues a risk warning for an alarm engine fault type, issues a risk warning for a real fault after interference is eliminated, updates and stores the confirmed real fault features, fault types and interference discrimination results into the engine feature database, and synchronously iteratively updates the related feature item set and the determination threshold, so that the application adds an interference discrimination link on the basis of fault classification and determination, avoids false alarms caused by interference, and prevents missed detection caused by interference covering real faults.

Description

Technical Field

[0001] This invention relates to the field of engine fault repair technology, specifically to a method for generating engine fault repair strategies based on multi-source data. Background Technology

[0002] As the core component of a power system, the engine's operating status directly affects the overall reliability and operational safety of the equipment. Timely and accurate identification of early engine faults is crucial for reducing maintenance costs and preventing major accidents. Fault diagnosis methods based on acoustic signatures are widely used in engine condition monitoring due to their advantages such as non-contact acquisition, no need to modify equipment structure, and real-time monitoring. Most existing engine acoustic signature fault diagnosis technologies use single-point acoustic sensors to collect operating sounds, extract spectral features, and match them against preset fault templates, determining the presence of a fault through a fixed threshold. While such solutions can identify typical faults under ideal operating conditions, they have significant technical limitations in real-world, complex operating environments. Furthermore, the complex operating environment within the engine compartment, including background noise, abnormal noises from auxiliary components, and fluctuations in operating load, can interfere with the collected acoustic signature signals, causing shifts in acoustic signature characteristics and a decrease in fault matching accuracy. Existing technologies lack effective interference discrimination mechanisms. When interference causes abnormal characteristics, it is easy to misjudge the interference as a fault and issue false alarms, or cause missed detections because the interference masks the true fault characteristics. The accuracy and reliability of fault diagnosis are difficult to guarantee. The voiceprint information obtained by the single-point acquisition method is limited and cannot cross-verify the authenticity of fault characteristics from a spatial dimension. Local interference can affect the overall judgment result. At the same time, the judgment mode with fixed thresholds cannot adapt to the differences in voiceprint characteristics under different models and operating conditions, nor does it have the ability to continuously optimize with the accumulation of operating data. As the equipment ages and the operating conditions change, the diagnostic accuracy will gradually decrease, making it difficult to meet the monitoring needs of long-term stable operation. Summary of the Invention

[0003] The purpose of this invention is to provide a method for generating engine fault repair strategies based on multi-source data, so as to solve the problems mentioned in the background art.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: a method for generating engine fault repair strategies based on multi-source data, comprising the following steps: S1. Arrange acoustic sensors at multiple points in the engine compartment to collect engine operating sound signals and speed signals. After preprocessing and normalization, extract engine features and establish an engine feature database. S2. Call historical fault records, count the historical abnormal frequency of each feature corresponding to the fault type, filter out the engine features related to the fault, and construct a set of relevant feature items corresponding to each fault type at fixed points; S3. Real-time acquisition of engine operating characteristics at fixed locations, comparison with reference characteristics to identify abnormal features, calculation of risk coefficients for each fault type at the location, and analysis of the number of locations where the fault type is determined to be a risk fault. S4. Classify the fault types according to the proportion of risk points and issue risk warnings for the alarm engine fault types. S5. Extract risk points of undetermined fault types, calculate the feature matching degree between the abnormal features of the risk points and the set of related feature items, and the abnormal overlap degree between the risk points, and determine the real fault based on the feature matching degree and the abnormal overlap degree. S6. After completing an alarm cycle, update and store the confirmed actual fault characteristics, fault type and interference judgment results into the engine feature database, and synchronously iterate and update the relevant feature item set and judgment threshold. Based on historical data, fault-related features are used to establish a mapping between locations and fault features. Fault risk is classified and determined by the proportion of risk locations. An interference verification mechanism is introduced, which combines feature matching degree and location anomaly overlap degree, to dynamically identify fault risks while filtering out interference and false alarms.

[0005] Furthermore, in step S1, after authorization, acoustic sensors are deployed at M fixed points in the engine compartment. Engine data for a monitoring period is collected through the acoustic sensors. The engine data includes engine operating sound signals and speed signals. The engine data is further preprocessed and normalized to obtain engine characteristics and establish an engine characteristic database.

[0006] Further, in step S2, the historical engine feature database is called to obtain N engine fault types judged in the past. For the m-th fixed point and the n-th engine fault type, the engine features of the monitoring time period before the engine fault occurred are called as the fault occurrence features, m=1,2,…,M, n=1,2,…,N. Taking the current time point as the endpoint, the fault records of the most recent P times for the n-th engine fault type are called as reference fault records. In the P reference fault records, the fault occurrence features of the p-th reference fault record are {A(p-1),A(p-2),…,A(pq),…,A(pQ)}, where p=1,2,…,P, and A(pq) represents the p-th reference... The q-th fault occurrence feature in the fault record, where Q is the number of engine features, is then used to obtain the q-th fault occurrence feature of P reference fault records as {A(1-q), A(2-q), ..., A(pq), ..., A(Pq)}. Using the current time point as the endpoint, the engine feature database of the previous preset reference feature update cycle is called. The average value of the engine features at the m-th fixed point where no fault occurred is used as the reference engine feature at the m-th fixed point. The reference engine feature at the m-th fixed point is {A(m-1), A(m-2), ..., A(mq), ..., A(mQ)}, where A(mq) is the q-th reference engine feature at the m-th fixed point. For the m-th fixed point, if | If [A(pq)-A(mq)] / A(mq)| is greater than the preset difference threshold, then the q-th fault in the p-th reference fault record is judged to have an abnormal characteristic. For the difference threshold, engine characteristic data is continuously collected for R monitoring cycles under normal engine operation. For the q-th engine characteristic, the mean and standard deviation of the relative deviation of this characteristic relative to the reference engine characteristic are calculated over the R monitoring cycles. The mean plus three times the standard deviation is used as the initial value of the difference threshold for this characteristic. Three times the standard deviation corresponds to the 3σ principle in statistical process control. Under the assumption of normal distribution, the probability of normal characteristic fluctuations falling within this range is 99.7%. Exceeding this range is judged as abnormal. The difference threshold is calibrated separately for each engine characteristic; otherwise... To determine if the q-th fault occurrence feature of the p-th reference fault record is normal, substitute p=1,2,…,P one by one, and judge the q-th fault occurrence feature of the P-th reference fault records. This will result in the q-th fault occurrence feature of the Pq-th reference fault records being abnormal. The correlation degree B(qn) between the q-th engine feature and the n-th engine fault type is then obtained, where B(qn) = Pq / P. The correlation degree B(qn) is compared with a preset correlation threshold. For the correlation threshold, the B(qn) values ​​of the Q engine features are sorted in ascending order, and the midpoint of adjacent B values ​​is used as a candidate threshold to divide all features into low-correlation and high-correlation groups. The inter-class variance σ between the two groups is then calculated.2 B = ω1ω2(μ1−μ2)², where ω1 and ω2 are the proportions of the two sets of characteristic numbers to the total number, and μ1 and μ2 are the mean values ​​of B in the two sets, respectively. Let σ... 2 The candidate threshold corresponding to the maximum value of B is used as the initial value of the relevance threshold. If B(qn) is greater than the preset relevance threshold, then the q-th engine feature is determined to be a relevant engine feature of the n-th engine fault type; otherwise, the q-th engine feature is determined to be an irrelevant engine feature of the n-th engine fault type. Substituting q=1,2,…,Q one by one, we obtain all relevant engine features of the n-th engine fault type. The number of all relevant engine features of the n-th engine fault type is integrated into the relevant feature set C(mn). Substituting n=1,2,…,N one by one, we obtain the relevant feature set of N engine fault types for the m-th fixed point. Similarly, we obtain the relevant feature set of engine fault types for the other M-1 fixed points. Through statistical analysis of historical fault data, we automatically filter out the voiceprint features that are truly related to various faults, abandoning the subjective bias of manually setting features based on experience, and making the fault judgment basis more objective and accurate. By distinguishing between relevant and irrelevant features, interference from a large number of non-fault-related features can be eliminated. This not only reduces the feature dimensions for subsequent real-time comparisons and improves computational efficiency, but also avoids misleading fault determination results due to irrelevant feature anomalies, effectively improving the accuracy of fault identification. Simultaneously, relevant feature sets corresponding to different fault types are established for different collection points, fully adapting to the differences in acoustic propagation characteristics at different locations. This makes the fault determination criteria for each point more targeted, further enhancing the reliability of multi-point cross-validation.

[0007] Further, in step S3, the engine operation is monitored in real time, and real-time engine features at M fixed points are obtained. For the m-th fixed point, the real-time engine feature {a(m-1), a(m-2), ..., a(mq), ..., a(mQ)} at the m-th fixed point is obtained, where a(mq) is the q-th real-time engine feature at the m-th fixed point. When it is detected that the real-time engine feature a(mq) at the m-th fixed point satisfies |[a(mq)-A(mq)] / A(mq)| is greater than a preset difference threshold, the q-th real-time engine feature at the m-th fixed point is determined to be abnormal, and the q-th real-time engine feature at the m-th fixed point is an abnormal feature; otherwise, the q-th real-time engine feature at the m-th fixed point is determined to be an abnormal feature. Given q real-time engine features that are normal, substitute each of them into q = 1, 2, ..., Q to obtain all abnormal features at the m-th fixed point. Record the number of abnormal features and integrate them into the abnormal feature item set for the m-th fixed point. Denote the number of elements in the relevant feature item set for the n-th engine fault type as En, and denote the number of elements in the intersection of the relevant feature item set for the n-th engine fault type and the abnormal feature item set for the m-th fixed point as Fn. For the n-th engine fault type, obtain the risk coefficient of the n-th engine fault type to the m-th fixed point. The risk coefficient of the n-th engine fault type to the m-th fixed point is D(mn), where D(mn) = Fn / En. Then, compare D(mn) with the preset risk coefficient... Threshold comparisons are performed. For the risk coefficient threshold, for the nth engine fault type, L1 group point feature data when the fault type is confirmed to have occurred is extracted from the historical database as fault samples, and L2 group point feature data when the fault type is not confirmed to have occurred is extracted as normal samples. The risk coefficient D(mn) value of each point in the two groups of samples is calculated. The intersection of the fault sample D value distribution and the normal sample D value distribution is used as the initial value of the risk coefficient threshold. If the two distributions do not intersect, the maximum D value that makes the fault sample recall rate reach 90% is taken as the initial value. 90% is the conventional recall rate lower limit target in the field of fault diagnosis. If D(mn) is greater than the preset risk coefficient threshold, it is determined that the nth fault exists at the mth fixed point. The failure risk of engine failure types is determined by the following: the failure type of the nth engine is a risky failure type at the mth fixed point; otherwise, it is determined that there is no failure risk of the nth engine failure type at the mth fixed point, and the nth engine failure type is a non-risky failure type at the mth fixed point. Substituting n=1,2,…,N one by one, we obtain all the risky failure types at the mth fixed point. Similarly, substituting m=1,2,…,M one by one, we obtain all the risky failure types at the M fixed points. Then, we obtain the number of times the nth engine failure type is a risky failure type, Gn. Substituting n=1,2,…,N one by one, we obtain the number of times the N engine failure types are risky failure types, {G1,G2,…,Gn,…,GN}.Real-time feature-by-feature comparison of operational status with baseline features quantifies risk level by the proportion of the intersection of abnormal features and fault-related features, rather than relying on a single abnormal feature to directly determine a fault. This avoids misjudgments caused by fluctuations in individual features, making single-point risk assessment more scientific and reasonable. Simultaneously, it aggregates risk assessment results from multiple data collection points, statistically analyzing the risk coverage of each fault type from a spatial perspective. This provides a quantitative basis for subsequent grading and judgment, and lays the data foundation for interference verification, thereby improving the overall comprehensiveness and reliability of fault risk assessment.

[0008] Furthermore, in step S4, if Gn=0, it is determined that the engine does not have the risk of the nth engine fault type; if Gn / M>Hn, it is determined that the engine has the risk of the nth engine fault type, and the nth engine fault type is an alarm engine fault type, issuing a risk alarm for the nth engine fault type; if Gn is not 0 and Gn / M≤Hn, the nth engine fault type is regarded as a pending engine fault type, where Hn is a preset point percentage threshold for the nth engine fault type, and the distribution of the percentage Gn / M of the number of points judged as risk fault types in M ​​fixed points when the nth engine fault type is confirmed to have occurred is statistically analyzed. The lower quartile of this distribution is taken as the initial value of Hn, i.e., the 25th percentile value. This ensures that historically, 75% of the faults occurred at locations with a risk percentage higher than Hn, thus triggering an alarm. Substituting each value into n=1,2,…,N, the same processing is applied to N engine fault types to obtain all pending engine fault types and the nth engine fault type. For faults with multiple points consistently showing high risk, an alarm is issued directly, ensuring a fast response to serious faults. Faults with only a few points showing abnormalities are classified as pending and not alarmed temporarily, but are further identified in the subsequent interference verification stage. This mechanism reserves interference filtering space, effectively reducing the probability of false alarms triggered directly by local interference. The tiered mode balances the timeliness of fault warnings with the buffer level for accurate identification, achieving a reasonable balance between sensitivity and accuracy in the overall diagnostic system.

[0009] Furthermore, in step S5, if there is no n such that the nth engine fault type is a pending engine fault type, the interference discrimination process is not entered; if there is n such that the nth engine fault type is a pending engine fault type, the interference discrimination process is entered, the relevant feature set of the nth engine fault type is obtained, engine feature calls are performed on M fixed points, and when the mth fixed point has the qth real-time engine feature anomaly, for the nth pending engine fault type, all points that determine the nth engine fault type as a risk fault type are extracted as risk points, resulting in a total of Gn risk points; for the gth risk point, g = 1, 2, ..., Gn, Obtain the number of intersection elements Fg between the set of abnormal features of the g-th risk point and the set of features related to the n-th engine fault type. Calculate the feature matching degree K(gn) of the g-th risk point, K(gn) = Fg / En. Obtain the average feature matching degree Kn of the n-th undetermined engine fault type, where Kn is the average of the feature matching degrees of the Gn risk points. Obtain the abnormal overlap degree Un between points, where Un is the average number of intersection elements between each pair of abnormal feature sets of the Gn risk points. Set an overlap degree threshold Kth and a feature matching degree threshold Uth. For the overlap degree threshold Kth and the feature matching degree threshold Uth, extract confirmed interference samples from the historical database. The data is compared with real fault samples, where the interference samples refer to historically confirmed pending fault events that are interference. The average feature matching degree Kn and the abnormal overlap degree Un corresponding to each sample are calculated. The 90th percentile of the average feature matching degree in the interference samples is taken as the initial value of Kth, and the 90th percentile of the abnormal overlap degree in the interference samples is taken as the initial value of Uth, ensuring that 90% of the interference samples fall below the threshold and are correctly identified. Simultaneously, it is verified that the proportion of real fault samples with Kn not lower than Kth or Un not lower than Uth is not less than 90%. If it is lower than 90%, Kth and Uth are reduced until this condition is met. If Kn < Kth and Un < Uth, then the nth pending fault event is determined to be the fault. If the engine fault type is an affected engine fault type, no risk alarm is issued; otherwise, the nth pending engine fault type is determined to be a real risk fault, and a risk alarm for the nth engine fault type is issued. Substituting n=1,2,…,N one by one, all real risk faults are obtained. The corresponding maintenance strategy for the real risk fault is called from the database. The maintenance strategy is the maintenance plan with the highest success rate for the corresponding fault type in history. All maintenance strategies for real risk faults are compiled and an overview is generated for the user. A condition-triggered interference discrimination mechanism is adopted, which only performs interference verification on pending faults with ambiguous boundaries, without having to run interference detection algorithms on all fault types, thus balancing discrimination accuracy and computational efficiency.Interference is comprehensively assessed using a dual-dimensional approach: feature matching degree and point anomaly overlap degree. The feature matching degree verifies the degree of agreement between abnormal features and fault features, while the spatial dimension verifies the consistency of abnormal features across multiple points. Feature matching degree reflects the degree of agreement between abnormal features and fault-related features; a lower matching degree indicates that the abnormal features deviate more from the typical manifestations of this type of fault, increasing the likelihood of interference. Point anomaly overlap degree reflects the consistency of abnormal features among multiple collection points; a lower overlap degree indicates that the anomalies lack spatial commonality, making them more likely to be caused by localized interference. These two indicators describe the characteristic manifestations of interference from both the feature and spatial dimensions. Only when both dimensions simultaneously point to interference can a complete chain of evidence be formed, definitively determining it as a false anomaly caused by interference. If only one indicator meets the interference condition while the other does not, the possibility of a genuine fault cannot be sufficiently ruled out. A low feature matching degree but a high point overlap degree may indicate that the fault is in its early stage and the feature manifestations are not yet typical, but multiple points have already shown abnormal trends simultaneously. A low point overlap degree but a high feature matching degree may indicate that the fault occurred on one side, resulting in some points having unclear features.

[0010] Furthermore, in step S6, after completing a risk alert cycle, the engine characteristics, fault type, and interference discrimination results corresponding to the confirmed real fault are updated and stored in the engine characteristic database. At the same time, the relevant feature set and risk coefficient threshold of each fault type are updated based on the engine characteristic database.

[0011] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: On the one hand, by adding an interference discrimination link on the basis of fault classification judgment, through the comprehensive verification of the dual dimensions of feature matching degree and point abnormal overlap degree, it can accurately distinguish between real faults and feature shifts caused by various interferences such as environmental noise, abnormal noise of auxiliary components, and abnormal acquisition link. This not only avoids false alarms caused by interference, but also prevents missed detection caused by interference masking real faults. It can still maintain stable and reliable diagnostic accuracy in complex and noisy actual operating environments, and reduce the probability of false alarms and missed alarms. On the one hand, a distributed acquisition array composed of multi-point acoustic sensors is used to synchronously acquire engine acoustic fingerprint information from different spatial locations. The authenticity of fault characteristics is verified by cross-comparison and consistency verification of multi-point features. This overcomes the limitations of limited information from single-point acquisition and susceptibility to local interference. It can eliminate local abnormal interference from the spatial dimension, making the basis for fault judgment more sufficient and comprehensive.

[0012] On the other hand, a database is set up to update the system regularly. After each diagnosis, the confirmed real fault data and interference discrimination results are stored in the feature library in sync. The fault-related feature set and various judgment thresholds are automatically iterated and optimized. This can adapt to the drift of soundprint characteristics caused by engine equipment aging and changes in operating environment. It breaks through the defect of the traditional fixed parameter scheme that the accuracy decreases over time, so that the system always maintains the optimal diagnostic state and has the ability to continuously improve itself. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the engine fault repair strategy generation method based on multi-source data according to the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Please see Figure 1 This invention provides a technical solution: a method for generating engine fault repair strategies based on multi-source data, comprising the following steps: S1. Arrange acoustic sensors at multiple points in the engine compartment to collect engine operating sound signals and speed signals. After preprocessing and normalization, extract engine features and establish an engine feature database. S2. Call historical fault records, count the historical abnormal frequency of each feature corresponding to the fault type, filter out the engine features related to the fault, and construct a set of relevant feature items corresponding to each fault type at fixed points; S3. Real-time acquisition of engine operating characteristics at fixed locations, comparison with reference characteristics to identify abnormal features, calculation of risk coefficients for each fault type at the location, and analysis of the number of locations where the fault type is determined to be a risk fault. S4. Classify the fault types according to the proportion of risk points and issue risk warnings for the alarm engine fault types. S5. Extract risk points of undetermined fault types, calculate the feature matching degree between the abnormal features of the risk points and the set of related feature items, and the abnormal overlap degree between the risk points, and determine the real fault based on the feature matching degree and the abnormal overlap degree. S6. After completing an alarm cycle, update and store the confirmed actual fault characteristics, fault type and interference judgment results into the engine feature database, and synchronously iterate and update the relevant feature item set and judgment threshold. Based on historical data, fault-related features are used to establish a mapping between locations and fault features. Fault risk is classified and determined by the proportion of risk locations. An interference verification mechanism is introduced, which combines feature matching degree and location anomaly overlap degree, to dynamically identify fault risks while filtering out interference and false alarms.

[0016] In step S1, after authorization, acoustic sensors are deployed at M fixed points in the engine compartment. Engine data for a monitoring period is collected through the acoustic sensors. The engine data includes engine operating sound signals and speed signals. The engine data is further preprocessed and normalized to obtain engine characteristics and establish an engine characteristic database.

[0017] In step S2, the historical engine feature database is invoked to obtain N engine fault types identified in the past. For the m-th fixed point and the n-th engine fault type, the engine features from the monitoring period preceding the occurrence of the engine fault are used as the fault occurrence features, where m = 1, 2, ..., M, n = 1, 2, ..., N. Taking the current time point as the endpoint, the fault records of the most recent P times for the n-th engine fault type are used as reference fault records. In the P reference fault records, the fault occurrence features of the p-th reference fault record are {A(p-1), A(p-2), ..., A(pq), ..., A(pQ)}, where p = 1, 2, ..., P, and A(pq) represents the p-th fault. Referring to the q-th fault occurrence feature in the fault record, where Q is the number of engine features, the q-th fault occurrence feature of P reference fault records is obtained as {A(1-q), A(2-q), ..., A(pq), ..., A(Pq)}. Taking the current time point as the endpoint, the engine feature database of the previous preset reference feature update cycle is called. The average value of the engine features at the m-th fixed point where no fault occurred is used as the reference engine feature at the m-th fixed point. The reference engine feature at the m-th fixed point is {A(m-1), A(m-2), ..., A(mq), ..., A(mQ)}, where A(mq) is the q-th reference engine feature at the m-th fixed point. Given m fixed points, if |[A(pq)-A(mq)] / A(mq)| is greater than a preset difference threshold, then the q-th fault occurrence feature of the p-th reference fault record is determined to be abnormal; otherwise, the q-th fault occurrence feature of the p-th reference fault record is determined to be normal. Substituting p=1,2,…,P into each reference fault record, the q-th fault occurrence feature of the P reference fault records is judged to be abnormal. The correlation degree B(qn) between the q-th engine feature and the n-th engine fault type is obtained, where B(qn)=Pq / P. The correlation degree B(qn) is then compared with a preset correlation threshold. If B(qn) is greater than the preset relevance threshold, then the q-th engine feature is determined to be a relevant engine feature of the n-th engine fault type; otherwise, the q-th engine feature is determined to be an irrelevant engine feature of the n-th engine fault type. Substituting q=1,2,…,Q into each feature, we obtain all relevant engine features of the n-th engine fault type. We then integrate the number of all relevant engine features of the n-th engine fault type into a set of relevant feature items C(mn). Substituting n=1,2,…,N into each feature, we obtain a set of relevant feature items for the m-th fixed point and N engine fault types. Similarly, we obtain a set of relevant feature items for the other M-1 fixed points.

[0018] In step S3, the engine operation is monitored in real time, and real-time engine features at M fixed points are obtained. For the m-th fixed point, the real-time engine features {a(m-1), a(m-2), ..., a(mq), ..., a(mQ)} at the m-th fixed point are obtained, where a(mq) is the q-th real-time engine feature at the m-th fixed point. When it is detected that the real-time engine feature a(mq) at the m-th fixed point satisfies |[a(mq)-A(mq)] / A(mq)| is greater than a preset difference threshold, the q-th real-time engine feature at the m-th fixed point is determined to be a real-time engine feature. If the engine characteristic is abnormal, the q-th real-time engine characteristic at the m-th fixed point is an abnormal characteristic; otherwise, the q-th real-time engine characteristic at the m-th fixed point is considered normal. Substitute q=1,2,…,Q one by one to obtain all abnormal characteristics at the m-th fixed point. Record the number of abnormal characteristics and integrate them into the abnormal characteristic item set of the m-th fixed point. Denote the number of elements in the relevant characteristic item set of the n-th engine fault type as En. Count the number of elements in the intersection of the relevant characteristic item set of the n-th engine fault type and the abnormal characteristic item set of the m-th fixed point. Let Fn be the risk coefficient of the nth engine fault type at the mth fixed point. For the nth engine fault type, obtain the risk coefficient of the nth engine fault type at the mth fixed point, denoted as D(mn), where D(mn) = Fn / En. Compare D(mn) with a preset risk coefficient threshold. If D(mn) is greater than the preset risk coefficient threshold, then it is determined that the mth fixed point has a fault risk associated with the nth engine fault type, and the nth engine fault type is considered a risk fault type for the mth fixed point. Otherwise, it is determined that the mth fixed point does not have a fault risk associated with the nth engine fault type. Given the failure risk of the nth engine failure type, and considering that the nth engine failure type is a non-risk failure type at the mth fixed point, we can substitute each of these values ​​into n=1,2,…,N to obtain all risk failure types at the mth fixed point. Similarly, we can substitute each of these values ​​into m=1,2,…,M to obtain all risk failure types at the M fixed points. This leads to the number of times the nth engine failure type is considered a risk failure type, Gn. Substituting each of these values ​​into n=1,2,…,N, we can obtain the number of times the N engine failure types are considered risk failure types, {G1,G2,…,Gn,…,GN}.

[0019] In step S4, if Gn=0, it is determined that there is no risk of the nth engine fault type; if Gn / M>Hn, it is determined that there is a risk of the nth engine fault type, the nth engine fault type is an alarm engine fault type, and a risk alarm for the nth engine fault type is issued; if Gn is not 0 and Gn / M≤Hn, the nth engine fault type is taken as a pending engine fault type, where Hn is the preset point proportion threshold of the nth engine fault type. Substitute n=1,2,…,N one by one, perform the same processing on N engine fault types to obtain all pending engine fault types and the nth engine fault type.

[0020] In step S5, if there is no n such that the nth engine fault type is a pending engine fault type, the interference discrimination process is not entered; if there is n such that the nth engine fault type is a pending engine fault type, the interference discrimination process is entered, the relevant feature set of the nth engine fault type is obtained, engine feature calls are performed on M fixed points, and when the mth fixed point has the qth real-time engine feature anomaly, for the nth pending engine fault type, all points that determine the nth engine fault type as a risk fault type are extracted as risk points, resulting in a total of Gn risk points; for the gth risk point, g = 1, 2, ..., Gn, the number of intersection elements Fg of the abnormal feature set of the gth risk point and the relevant feature set of the nth engine fault type is obtained, and the feature matching degree K(gn) of the gth risk point is calculated, K(gn) = Fg / En, to obtain The average feature matching degree Kn for the nth pending engine fault type is obtained, where Kn is the average feature matching degree of Gn risk points. The abnormal overlap degree Un between points is obtained, where Un is the average number of intersection elements between each pair of abnormal feature item sets of Gn risk points. An overlap degree threshold Kth and a feature matching degree threshold Uth are set. If Kn < Kth and Un < Uth, the nth pending engine fault type is determined to be an interfered engine fault type, and no risk alarm is issued; otherwise, the nth pending engine fault type is determined to be a real risk fault, and a risk alarm for the nth engine fault type is issued. Substituting n=1,2,…,N one by one, all real risk faults are obtained. The corresponding maintenance strategy for the real risk fault is called from the database. The maintenance strategy is the maintenance plan with the highest success rate for the corresponding fault type in history. All maintenance strategies for real risk faults are sorted out and an overview is generated for the user.

[0021] In step S6, after completing a risk alert cycle, the engine characteristics, fault type, and interference discrimination results corresponding to the confirmed real fault are updated and stored in the engine characteristic database. At the same time, the relevant feature set and risk coefficient threshold of each fault type are updated.

[0022] Example 1: Fault monitoring of a diesel engine in a port yard is illustrated using this example. Four fixed acoustic sensors are positioned at the following locations: left side of the cylinder head, right side of the cylinder head, next to the oil pan, and at the intake manifold. The sampling frequency is set to 16kHz, and the monitoring time period is set to 30 seconds. Engine operating sound signals and speed signals are simultaneously acquired. After preprocessing including framing, windowing, and fast Fourier transform, the speed signal is combined with the signal to convert it into an order spectrum. 32-dimensional standardized acoustic signature features are then extracted and stored in the engine feature database.

[0023] The historical engine feature database was accessed, covering eight common engine faults including valve noise, bearing wear, fuel injection failure, and belt noise. For the left side of the cylinder head, the acoustic signature of the monitoring cycle preceding the fault in the most recent twenty fault records for each fault type was retrieved as the fault occurrence feature. Simultaneously, the average feature value from the three months of fault-free operation was retrieved as the reference engine feature for that location. The frequency of abnormal occurrences of each feature in historical fault records was statistically analyzed, and the correlation between each feature and each type of fault was calculated. A correlation threshold of 0.6 was set; features with a correlation higher than this threshold were considered relevant features for that fault. Statistical filtering revealed 11 relevant features for valve noise at the left side of the cylinder head and 8 relevant features for bearing wear. The relevant feature sets for the four locations and eight fault types were then constructed accordingly.

[0024] During real-time operation, real-time acoustic features from four locations are extracted at each monitoring cycle. Each feature is compared with a reference feature, and a relative deviation exceeding 5% is considered an abnormal feature. For the location on the left side of the cylinder head, six abnormal features were identified in a single monitoring session. After comparison with the 11 relevant features for valve noise faults, five features intersected, resulting in a risk coefficient of approximately 0.45, higher than the preset risk coefficient threshold of 0.3. Therefore, this location was determined to have a valve noise fault risk. Similarly, risk coefficient calculations were performed for all four locations and eight types of faults. Statistics showed that valve noise faults were classified as risk faults at three locations, and bearing wear faults at one location.

[0025] Upon entering the three-level classification judgment stage, the threshold for the percentage of points is set at 0.6. The percentage of risk points for valve noise is 0.75, which is higher than the threshold, so it is directly judged as an engine fault type and a level one risk alarm is issued; the percentage of risk points for bearing wear is 0.25, which is lower than the threshold and not zero, so it is judged as a pending engine fault type and enters the interference discrimination process; the number of risk points for the remaining six types of faults is zero, so they are judged as no risk.

[0026] For the undetermined fault of bearing wear, an abnormal feature set is extracted from a single risk point. The matching degree between the abnormal features at this point and the bearing wear-related features is calculated to be approximately 0.25. Since there is only a single risk point, the abnormal overlap between points is recorded as zero. The feature matching degree threshold for interference judgment is set to 0.4, and the overlap degree threshold is set to 0.3. Since both indicators are lower than the corresponding thresholds, the anomaly is determined to be a false feature shift caused by interference, and no alarm is issued.

[0027] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for generating engine fault repair strategies based on multi-source data, characterized in that: The method includes the following steps: S1. Arrange acoustic sensors at multiple points in the engine compartment to collect engine operating sound signals and speed signals. After preprocessing and normalization, extract engine features and establish an engine feature database. S2. Call historical fault records, count the historical abnormal frequency of each feature corresponding to the fault type, filter out the engine features related to the fault, and construct a set of relevant feature items corresponding to each fault type at fixed points; S3. Real-time acquisition of engine operating characteristics at fixed locations, comparison with reference characteristics to identify abnormal features, calculation of risk coefficients for each fault type at the location, and analysis of the number of locations where the fault type is determined to be a risk fault. S4. Classify the fault types according to the proportion of risk points and issue risk warnings for the alarm engine fault types. S5. Extract risk points of undetermined fault types, calculate the feature matching degree between the abnormal features of the risk points and the set of related feature items, and the abnormal overlap degree between the risk points, and determine the real fault based on the feature matching degree and the abnormal overlap degree. S6. After completing an alarm cycle, update and store the confirmed actual fault characteristics, fault type and interference discrimination results into the engine feature database, and synchronously iterate and update the relevant feature item set and judgment threshold.

2. The engine fault repair strategy generation method based on multi-source data according to claim 1, characterized in that: Based on historical data, fault-related features are used to establish a mapping between locations and fault features. Fault risk is classified and determined by the proportion of risk locations. An interference verification mechanism is introduced, which combines feature matching degree and location anomaly overlap degree, to dynamically identify fault risks while filtering out interference and false alarms.

3. The engine fault repair strategy generation method based on multi-source data according to claim 2, characterized in that: In step S1, after authorization, acoustic sensors are deployed at M fixed points in the engine compartment. Engine data for a monitoring period is collected through the acoustic sensors. The engine data includes engine operating sound signals and speed signals. The engine data is preprocessed and normalized to obtain engine characteristics and establish an engine characteristic database. The engine characteristic database includes historical maintenance records of fault types. The historical maintenance records include fault types, maintenance solutions adopted, and maintenance result feedback.

4. The engine fault repair strategy generation method based on multi-source data according to claim 3, characterized in that: In step S2, the historical engine feature database is invoked to obtain N engine fault types identified in the past. For the m-th fixed point and the n-th engine fault type, the engine features from the monitoring period preceding the occurrence of the engine fault are used as the fault occurrence features, where m = 1, 2, ..., M, n = 1, 2, ..., N. Using the current time point as the endpoint, the fault records of the most recent P times for the n-th engine fault type are used as reference fault records. In these P reference fault records, the fault occurrence features of the p-th reference fault record are {A(p-1), A(p-2), ..., A(pq), ..., A(pQ)}, where p = 1, 2, ..., P, A(pq) represents the q-th fault occurrence feature of the p-th reference fault record, and Q is the number of engine features. This leads to the following... The qth fault occurrence feature of P reference fault records is {A(1-q),A(2-q),…,A(pq),…,A(Pq)}. Taking the current time point as the endpoint, the engine feature database of the previous preset reference feature update cycle is called. The average value of the engine features of the mth fixed point where no fault occurred is used as the reference engine feature of the mth fixed point. The reference engine feature of the mth fixed point is {A(m-1),A(m-2),…,A(mq),…,A(mQ)}, where A(mq) is the qth reference engine feature of the mth fixed point. For the mth fixed point, if |[A(pq)-A(mq)] / A(mq)| is greater than the preset difference threshold, then the qth fault occurrence feature of the pth reference fault record is judged to be abnormal. Otherwise, determine that the qth fault occurrence feature of the p-th reference fault record is normal. Substitute p=1,2,…,P one by one, and judge the qth fault occurrence feature of the P-th reference fault records. This yields the qth fault occurrence feature of the Pq-th reference fault records that is abnormal. The correlation degree B(qn) between the q-th engine feature and the n-th engine fault type is obtained, where B(qn)=Pq / P. The correlation degree B(qn) is compared with a preset correlation threshold. If B(qn) is greater than the preset correlation threshold, then the q-th engine feature is determined to be abnormal. The relevant engine features for the nth engine fault type; otherwise, determine that the qth engine feature is an irrelevant engine feature for the nth engine fault type, and substitute them into q=1,2,…,Q to obtain all relevant engine features for the nth engine fault type. Integrate the number of all relevant engine features for the nth engine fault type into a set of relevant feature items C(mn), and substitute them into n=1,2,…,N to obtain a set of relevant feature items for N engine fault types for the mth fixed point. Similarly, obtain the set of relevant feature items for engine fault types for the other M-1 fixed points.

5. The engine fault repair strategy generation method based on multi-source data according to claim 4, characterized in that: In step S3, the engine operation is monitored in real time, and real-time engine features at M fixed points are obtained. For the m-th fixed point, the real-time engine feature {a(m-1), a(m-2), ..., a(mq), ..., a(mQ)} is obtained, where a(mq) is the q-th real-time engine feature at the m-th fixed point. When it is detected that the real-time engine feature a(mq) at the m-th fixed point satisfies |[a(mq)-A(mq)] / A(mq)| is greater than a preset difference threshold, the q-th real-time engine feature at the m-th fixed point is determined to be abnormal, and the q-th real-time engine feature at the m-th fixed point is considered an abnormal feature; otherwise, the m-th fixed point is determined to be an abnormal feature. If the q-th real-time engine feature at a given point is normal, substitute each of the values ​​q = 1, 2, ..., Q to obtain all abnormal features at the m-th fixed point. Record the number of abnormal features and integrate them into the set of abnormal feature items for the m-th fixed point. Denote the number of elements in the set of relevant feature items for the n-th engine fault type as En. Denote the number of elements in the intersection of the set of relevant feature items for the n-th engine fault type and the set of abnormal feature items for the m-th fixed point as Fn. For the n-th engine fault type, obtain the risk coefficient of the n-th engine fault type for the m-th fixed point. The risk coefficient of the n-th engine fault type for the m-th fixed point is D(mn), where D(mn) = Fn / En.

6. The engine fault repair strategy generation method based on multi-source data according to claim 5, characterized in that: Compare D(mn) with the preset risk coefficient threshold. If D(mn) is greater than the preset risk coefficient threshold, then it is determined that there is a fault risk of the nth engine fault type at the mth fixed point. The nth engine fault type is the risk fault type of the mth fixed point. Otherwise, determine that there is no fault risk of the nth engine fault type at the mth fixed point, and the nth engine fault type is a non-risk fault type at the mth fixed point. Substitute n=1,2,…,N one by one to obtain all risk fault types at the mth fixed point. Similarly, substitute m=1,2,…,M one by one to obtain all risk fault types at the M fixed points. Then obtain the number of times the nth engine fault type is a risk fault type Gn. Substitute n=1,2,…,N one by one to obtain the number of times the N engine fault types are risk fault types {G1,G2,…,Gn,…,GN}.

7. The engine fault repair strategy generation method based on multi-source data according to claim 6, characterized in that: In step S4, if Gn=0, it is determined that there is no risk of the nth engine fault type; if Gn / M>Hn, it is determined that there is a risk of the nth engine fault type, the nth engine fault type is the alarm engine fault type, and a risk alarm for the nth engine fault type is issued. If Gn is not 0 and Gn / M≤Hn, the nth engine fault type is taken as the undetermined engine fault type, where Hn is the preset point proportion threshold of the nth engine fault type. Substitute n=1,2,…,N one by one, perform the same processing on N engine fault types, and obtain all undetermined engine fault types and the nth engine fault type.

8. The engine fault repair strategy generation method based on multi-source data according to claim 7, characterized in that: In step S5, if there is no n such that the nth engine fault type is a pending engine fault type, the interference discrimination process is not entered. If there exists n such that the nth engine fault type is a pending engine fault type, then the interference discrimination process is entered, the relevant feature set of the nth engine fault type is obtained, engine feature calls are performed on M fixed points, and when the mth fixed point has the qth real-time engine feature anomaly, for the nth pending engine fault type, all points that determine the nth engine fault type as a risk fault type are extracted as risk points, resulting in a total of Gn risk points; For the g-th risk point, g = 1, 2, ..., Gn, obtain the number of intersection elements Fg between the abnormal feature set of the g-th risk point and the relevant feature set of the n-th engine fault type. Calculate the feature matching degree K(gn) of the g-th risk point, K(gn) = Fg / En, and obtain the average feature matching degree Kn of the n-th undetermined engine fault type. Kn is the average feature matching degree of the Gn risk points. Obtain the abnormal overlap degree Un between points, where Un is the average number of intersection elements between each pair of abnormal feature sets of the Gn risk points. Set the overlap degree threshold Kth and the feature matching degree threshold Uth. If Kn < Kth and Un < Uth, then determine that the n-th undetermined engine fault type is an interfered engine fault type and do not issue a risk alarm. Otherwise, the nth pending engine fault type is determined to be a real risk fault, and a risk alarm for the nth engine fault type is issued. Substitute n=1,2,…,N one by one to obtain all real risk faults. Call the maintenance strategy corresponding to the real risk fault in the database. The maintenance strategy is the maintenance plan with the highest success rate for the corresponding fault type in history. Organize all the maintenance strategies for real risk faults and generate an overview for the user.

9. The engine fault repair strategy generation method based on multi-source data according to claim 8, characterized in that: In step S6, after completing a risk alert cycle, the engine characteristics, fault type, and interference discrimination results corresponding to the confirmed real fault are updated and stored in the engine characteristic database. At the same time, the relevant feature set and risk coefficient threshold of each fault type are updated.

Images