A ship fault remote diagnosis method and system based on edge computing

Abstract

The application relates to the technical field of fault diagnosis, and specifically discloses a ship fault remote diagnosis method and system based on edge computing, which comprises the following steps: an edge computing node extracts time-frequency domain features from standardized time sequence monitoring data, removes redundancy through a variance inflation factor to obtain a dimension-reduced feature vector; current working conditions are determined and a feature data set with working condition labels is generated according to Mahalanobis distance and in combination with normal fluctuation intervals as inclusive conditions; deviation features are screened based on the working condition labels, and the deviation degree is determined; when the number of deviation features exceeds a threshold value, feature data in front and rear time windows is extracted to form a sample to be diagnosed; the sample to be diagnosed is matched with rules one by one, and an edge diagnosis result is output; through self-adaptive identification of working conditions, dynamic threshold comparison and fault rule matching on the edge side of a ship, the real-time performance and adaptability of fault diagnosis under variable working conditions are improved, and the application is suitable for remote intelligent monitoring and early warning of key equipment of a ship.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a remote fault diagnosis method and system for ships based on edge computing. Background Technology

[0002] Existing methods for diagnosing ship equipment faults, particularly remote diagnostic solutions based on centralized cloud platforms, lack adaptability to varying ship operating conditions. During actual navigation, operating conditions such as main engine load, sea state, and fuel type change in real time, causing the statistical distribution of equipment monitoring characteristics to drift with these changes. Existing methods typically employ fixed thresholds or discrimination rules trained under a single operating condition, making it difficult to dynamically match the normal fluctuation range under different conditions. For example, alarm thresholds set under low-load, stable conditions are prone to generating numerous false alarms under high-load, variable conditions; conversely, feature benchmarks optimized for high-load conditions may miss early signs of degradation under low-load conditions. This lack of adaptability leads to frequent false alarms or missed alarms in diagnostic systems during scenarios involving sea state changes and maneuvering, reducing the trust of marine engineers in alarm information and increasing unnecessary workload for shore-based expert review.

[0003] Current remote fault diagnosis methods for ships lack online adaptive threshold update mechanisms and lightweight feature selection methods at the edge. Traditional methods rely on batch training of diagnostic models using historical data. Once the model is deployed to the ship's edge equipment, its judgment criteria become fixed, making it difficult to continuously fine-tune threshold parameters using newly generated normal fluctuation data during operation. Meanwhile, the raw features collected by ship sensors have high dimensionality and significant collinearity redundancy. Directly using these features for fault determination not only consumes limited computing and communication resources but also amplifies noise interference, reducing the stability of diagnostic results. When equipment experiences early, minor faults, existing methods struggle to filter out truly sensitive features reflecting state degradation from high-dimensional redundant features, and cannot dynamically adjust the update step size of the tolerance boundary according to the severity of the deviation, leading to lag in tracking progressive faults. Therefore, there is an urgent need to develop a remote fault diagnosis method for ships that can identify operating conditions in real time, dynamically optimize feature dimensions, and adaptively update judgment thresholds at the edge. This would address the problems of high false alarm rates under varying operating conditions, the inability to evolve thresholds online, and interference from redundant features, thereby improving the accuracy and robustness of remote monitoring of critical ship equipment. Summary of the Invention

[0004] This invention provides a remote diagnosis method and system for ship faults based on edge computing, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a remote ship fault diagnosis method based on edge computing, comprising: Edge computing nodes extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain a dimension-reduced feature vector; P2: Based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector in each working condition, and using the normal fluctuation range as the inclusion condition, determine the feature dataset with working condition labels. P3: Based on the feature dataset with working condition labels, filter out the deviation features in each working condition and determine the degree of deviation. When the number of deviation features exceeds the preset deviation counting threshold, extract all feature data within the time window before and after the current moment to form the sample to be diagnosed. P4: Match each sample to be diagnosed to obtain the edge diagnosis results.

[0006] In a preferred embodiment, the edge computing node extracts time-frequency domain features from standardized time-series monitoring data and determines the variance inflation factor of the time-frequency domain features, including: The standardized time-series monitoring data is processed by frame segmentation, and time-domain statistical features and frequency-domain energy features are extracted sequentially from each frame of data to obtain an initial feature matrix; Using one feature column from the initial feature matrix as the dependent variable and the remaining feature columns as independent variables, linear regression is performed to obtain the variance inflation factor.

[0007] In a preferred embodiment, the redundancy removal to obtain the dimensionality-reduced feature vector includes: The feature columns whose variance inflation factor exceeds the preset collinearity tolerance threshold are used as the first set of features for screening. The linear regression and threshold filtering are repeated on the first-selection feature set until the variance inflation factor of all remaining features does not exceed the collinearity tolerance threshold, thus obtaining the dimensionality-reduced feature vector.

[0008] In a preferred embodiment, the step of calculating the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector in each working condition includes: Read the feature mean vector and covariance matrix corresponding to the current comparison condition from the preset multi-condition benchmark feature library; The difference vector between the current dimensionality-reduced feature vector and the feature mean vector is used as the deviation vector; Perform a vector-matrix multiplication operation on the deviation vector and the inverse of the covariance matrix to obtain the intermediate transformation vector; Perform a dot product operation between the intermediate transformation vector and the deviation vector to obtain the squared value of the Mahalanobis distance, and then perform a square root operation to obtain the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector.

[0009] In a preferred embodiment, determining the feature dataset with working condition labels includes: Read the normal fluctuation range corresponding to each operating condition category from the multi-operating condition benchmark feature library; Determine whether each feature value in the current dimensionality-reduced feature vector falls within the normal fluctuation range, and use the ratio of all feature dimensions that fall within the normal fluctuation range to the total dimension of the current dimensionality-reduced feature vector as the range coverage. Based on the Mahalanobis distance and the interval coverage, a comprehensive judgment index is calculated for the current dimensionality-reduced feature vector and each working condition category. The formula for calculating the comprehensive judgment index is as follows: In the formula, The current dimensionality-reduced feature vector and the first The comprehensive judgment index for each working condition category The Mahalanobis distance weights are used. The largest Mahalanobis distance value among all load case categories. The current dimensionality-reduced feature vector and the first Mahalanobis distance for each work condition category For a preset positive small quantity, For interval tolerance weights, For the first Range tolerance for each working condition category; The working condition category with the largest comprehensive judgment index is selected as the working condition label of the current working condition, and the current dimensionality reduction feature vector is associated with the working condition label and stored to generate a feature dataset with working condition label.

[0010] In a preferred embodiment, the step of filtering out deviation features and determining the degree of deviation in each working condition based on the feature dataset with working condition labels includes: Based on the operating condition label, read the upper limit and lower limit of the dynamic tolerance boundary for the current period under the current operating condition from the threshold parameter table stored locally; Each feature value in the current feature data is compared with the corresponding upper and lower limits of the dynamic tolerance boundary. Features whose current feature values ​​exceed the upper and lower limits of the dynamic tolerance boundary are taken as deviation features, and the degree of deviation is determined.

[0011] In a preferred embodiment, the step of extracting all feature data within a time window before and after the current moment to form a sample to be diagnosed when the number of deviation features exceeds a preset deviation counting threshold includes: When the total number of deviation features exceeds the preset deviation count threshold, the current moment is marked as a suspected fault moment; The two segments before and after the suspected fault time are spliced ​​together in chronological order, and the features of the spliced ​​data block are extracted to obtain the sample to be diagnosed.

[0012] In a preferred embodiment, the dynamic tolerance boundary further includes: Calculate the adaptive smoothing coefficient based on the maximum degree of deviation among all deviation characteristics; The formula for calculating the adaptive smoothing coefficient is as follows: In the formula, For adaptive smoothing coefficients, This is the preset lower limit of the smoothing coefficient. This is the preset upper limit of the smoothing coefficient. As a preset sensitivity factor, The maximum degree of deviation; The dynamic tolerance boundary is updated by an exponentially weighted moving average using the adaptive smoothing coefficient, generating an updated dynamic tolerance boundary and replacing the original threshold parameter of the dynamic tolerance boundary.

[0013] In a preferred embodiment, the step of matching the sample to be diagnosed line by line to obtain the edge diagnosis result includes: Each feature value in the sample to be diagnosed is compared with the dynamic tolerance boundary corresponding to the current working condition to generate a feature over-limit flag vector; The feature over-limit flag vector is compared with the feature logic judgment conditions in the preset multiple fault judgment rules one by one, and the fault type corresponding to the rule that satisfies all the conditions is taken as the preliminary matching result. Extract the degree of deviation of the deviation features in the preliminary matching results, analyze the confidence of the fault type, and associate the fault type with the confidence to generate edge diagnosis results.

[0014] To address the aforementioned problems, the present invention also provides a remote ship fault diagnosis system based on edge computing, the system comprising: The edge feature dimensionality reduction module is used by edge computing nodes to extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain dimensionality-reduced feature vectors. The multi-condition adaptive recognition module is used to determine the feature dataset with condition labels based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector of each condition, and with the normal fluctuation range as the inclusion condition. The sample construction module is used to filter out the deviation features in each working condition and determine the degree of deviation based on the feature dataset with working condition labels. When the number of deviation features exceeds the preset deviation counting threshold, all feature data within the time window before and after the current moment are extracted to form the sample to be diagnosed. The fault matching module is used to match the samples to be diagnosed one by one to obtain the edge diagnosis results.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention extracts time-frequency domain features from standardized time-series monitoring data and uses variance inflation factor for redundancy filtering, effectively eliminating collinearity redundancy features in sensor signals and reducing the feature space dimensionality. This reduces the computational load on edge nodes and the bandwidth consumption of satellite communication. Simultaneously, this invention introduces a condition determination mechanism combining Mahalanobis distance and normal fluctuation ranges. Based on the similarity between the current dimensionality-reduced feature vector and the mean vector of each condition's baseline features, as well as the degree to which the feature values ​​encompass the normal fluctuation ranges of each condition, it automatically outputs the most matching condition label. This overcomes the shortcomings of traditional fixed threshold or single-condition models that frequently produce false alarms or missed alarms under different navigation environments. Subsequent deviation screening and fault matching are always performed under the correct condition baseline, significantly improving the diagnostic system's condition tracking capability and alarm reliability. Multi-condition adaptive identification and dynamic feature dimensionality reduction are achieved at the ship's edge, significantly improving diagnostic accuracy and robustness under varying conditions.

[0016] 2. This invention filters out deviation features that exceed the boundary range based on the upper and lower limits of the dynamic tolerance boundary under the current operating condition label, and quantitatively calculates the degree of deviation. When the number of deviation features exceeds a preset threshold, it automatically extracts feature data within the preceding and following time windows to form a sample to be diagnosed. Simultaneously, using the maximum degree of deviation as input, it calculates an adaptive smoothing coefficient through an exponential mapping formula, and uses this coefficient to update the dynamic tolerance boundary using an exponentially weighted moving average, so that the threshold can automatically tighten or loosen according to the actual degradation trend of the equipment without manual intervention. On this basis, this invention matches the sample to be diagnosed with the fault rule base one by one to generate edge-end diagnostic results with confidence. Only the diagnostic results, updated threshold parameters, and a small number of typical samples are transmitted back to the shore-based cloud, which not only ensures the immediate response capability for sudden faults, but also avoids the real-time transmission of the original massive data. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a remote ship fault diagnosis method based on edge computing, provided in an embodiment of the present invention. Figure 2 A functional block diagram of a remote ship fault diagnosis system based on edge computing provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a remote ship fault diagnosis method based on edge computing. The execution entity of this remote ship fault diagnosis method based on edge computing includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, this remote ship fault diagnosis method based on edge computing can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0020] Reference Figure 1 The diagram shown is a flowchart illustrating a remote ship fault diagnosis method based on edge computing according to an embodiment of the present invention. In this embodiment, the remote ship fault diagnosis method based on edge computing includes: P1: Edge computing nodes extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain a dimension-reduced feature vector; In this embodiment of the invention, the edge computing node extracts time-frequency domain features from standardized time-series monitoring data and determines the variance inflation factor of the time-frequency domain features, including: The standardized time-series monitoring data is processed by frame segmentation, and time-domain statistical features and frequency-domain energy features are extracted sequentially from each frame of data to obtain an initial feature matrix; Using one feature column from the initial feature matrix as the dependent variable and the remaining feature columns as independent variables, linear regression is performed to obtain the variance inflation factor.

[0021] The process of redundancy removal to obtain a dimensionality-reduced feature vector includes: The feature columns whose variance inflation factor exceeds the preset collinearity tolerance threshold are used as the first set of features for screening. The linear regression and threshold filtering are repeated on the first-selection feature set until the variance inflation factor of all remaining features does not exceed the collinearity tolerance threshold, thus obtaining the dimensionality-reduced feature vector.

[0022] Edge computing nodes are placed inside the ship's engine room and connected to various sensors to continuously receive standardized time-series monitoring data. The node first divides the continuously collected temperature, pressure, vibration, and other data into frames of fixed time length. Within each frame, it calculates five time-domain statistics for each monitored physical quantity: mean, standard deviation, peak value, skewness, and kurtosis. Simultaneously, it performs a frequency domain transformation on the frame, summing the power spectral densities across different frequency ranges to obtain the energy of each frequency band. This transforms a raw data frame into an initial feature matrix composed of numerous feature columns. For this initial feature matrix, the edge computing node sequentially selects one feature column as the dependent variable and all other feature columns as independent variables. It then uses a standard linear regression to fit the dependency relationship between the dependent and independent variables and calculates the coefficient of determination in the regression results, thereby obtaining the variance inflation factor for that feature column. A larger variance inflation factor indicates significant overlap between that feature and other features.

[0023] When performing redundancy removal, edge computing nodes pre-set a collinearity tolerance threshold. This threshold is determined by analyzing the distribution of feature collinearity in normal historical data of ships offline, and is typically set to a value that can eliminate highly redundant features. The node compares the variance inflation factor of each feature column in the initial feature matrix with this collinearity tolerance threshold. Feature columns with variance inflation factors exceeding the threshold are included in the first-stage screening feature set, indicating that these features are currently high-redundancy candidates for removal. The edge computing node then retrieves this first-stage screening feature set, uses it as a new feature matrix, and recalculates the variance inflation factor for each feature column, repeating the comparison and removal operations until the variance inflation factors of all remaining features do not exceed the collinearity tolerance threshold. The remaining feature columns constitute a dimensionality-reduced feature vector. Each feature in this vector carries relatively independent information, without severe duplication or redundancy, and can more stably reflect the equipment status in subsequent diagnostics.

[0024] The beneficial effects are as follows: Through the aforementioned feature extraction and redundancy removal process, the edge computing nodes directly remove a large number of repetitive sensor features locally on the ship, significantly reducing the amount of data involved in condition identification and fault diagnosis. This reduces the computational burden on the edge nodes and also reduces the amount of effective data that needs to be uploaded to the shore via satellite link. Simultaneously, by eliminating highly collinear redundant features, subsequent condition determination and deviation screening are no longer affected by noise features, significantly improving the stability and reliability of the diagnostic results. This method does not rely on any complex black-box models, but is entirely based on simple calculations using statistical features and linear regression. It is easy to run in real time on resource-constrained shipboard edge devices, meeting the rapid response requirements for fault diagnosis during ship navigation.

[0025] P2: Based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector in each working condition, and using the normal fluctuation range as the inclusion condition, determine the feature dataset with working condition labels. In this embodiment of the invention, the step of determining the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector in each working condition includes: Read the feature mean vector and covariance matrix corresponding to the current comparison condition from the preset multi-condition benchmark feature library; The difference vector between the current dimensionality-reduced feature vector and the feature mean vector is used as the deviation vector; Perform a vector-matrix multiplication operation on the deviation vector and the inverse of the covariance matrix to obtain the intermediate transformation vector; Perform a dot product operation between the intermediate transformation vector and the deviation vector to obtain the squared value of the Mahalanobis distance, and then perform a square root operation to obtain the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector.

[0026] The identified feature dataset with working condition labels includes: Read the normal fluctuation range corresponding to each operating condition category from the multi-operating condition benchmark feature library; Determine whether each feature value in the current dimensionality-reduced feature vector falls within the normal fluctuation range, and use the ratio of all feature dimensions that fall within the normal fluctuation range to the total dimension of the current dimensionality-reduced feature vector as the range coverage. Based on the Mahalanobis distance and the interval coverage, a comprehensive judgment index is calculated for the current dimensionality-reduced feature vector and each working condition category. The formula for calculating the comprehensive judgment index is as follows: In the formula, The current dimensionality-reduced feature vector and the first The comprehensive judgment index for each working condition category The Mahalanobis distance weights are used. The largest Mahalanobis distance value among all load case categories. The current dimensionality-reduced feature vector and the first Mahalanobis distance for each work condition category For a preset positive small quantity, For interval tolerance weights, For the first Range tolerance for each working condition category; The working condition category with the largest comprehensive judgment index is selected as the working condition label of the current working condition, and the current dimensionality reduction feature vector is associated with the working condition label and stored to generate a feature dataset with working condition label.

[0027] Edge computing nodes have a pre-built multi-condition benchmark feature library during the factory or initial voyage calibration phase. This library stores a feature mean vector, a covariance matrix, and a normal fluctuation range for each typical voyage condition. The feature mean vector records the average value of each dimensionality-reduced feature during normal operation of the equipment under that condition. The covariance matrix characterizes the degree of correlation and fluctuation between these features, and the normal fluctuation range describes the allowable range of variation for each feature under normal conditions. When the edge computing node obtains the dimensionality-reduced feature vector at the current moment, it reads the feature mean vector and covariance matrix corresponding to each condition in the library one by one, and subtracts the feature mean vector of the current condition from the feature mean vector of the current condition to be compared dimension by dimension to obtain the deviation vector corresponding to the current condition. This deviation vector reflects the combination of the direction and magnitude of the deviation between the current actual feature and the ideal condition benchmark.

[0028] The edge computing nodes perform the Mahalanobis distance calculation process independently for each working condition. The node first extracts the covariance matrix for that working condition and inverts it to obtain the inverse matrix. Then, it performs a vector-matrix multiplication between the deviation vector and the inverse of the covariance matrix to obtain an intermediate transformation vector. Each component of this intermediate transformation vector represents the weighted result of the deviation vector after being processed by the covariance inverse matrix; its function is to eliminate the influence of correlation between features. Next, the intermediate transformation vector is multiplied by the original deviation vector, that is, the corresponding components of the two vectors are multiplied and then summed to obtain the squared value of the Mahalanobis distance for the current working condition. Finally, the square root of this squared value is performed to obtain the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector of that working condition. The smaller the Mahalanobis distance value, the better the current feature matches the benchmark of that working condition.

[0029] The Mahalanobis distance weight is a pre-set value used to adjust the importance of Mahalanobis distance in the overall judgment. The largest Mahalanobis distance value among all work condition categories is the one with the largest distance value selected by the edge computing node after calculating the Mahalanobis distance between the current feature vector and each work condition. The smaller the Mahalanobis distance between the current dimensionality-reduced feature vector and the currently compared work condition category, the closer the current feature is to the benchmark of that work condition. The preset positive minimum is a very small positive number added after the second parameter to prevent the denominator from being zero, which would prevent the calculation from proceeding. The interval coverage weight is a pre-set value used to adjust the importance of interval coverage in the overall judgment. Interval coverage represents the proportion of features in the current dimensionality-reduced feature vector that fall within the normal fluctuation range of that work condition; this ratio is between zero and one.

[0030] The edge computing node first extracts the largest Mahalanobis distance value among all work condition categories and divides it by the sum of the Mahalanobis distance of the current work condition and a preset positive minimum. This division reflects the closeness of the current work condition's Mahalanobis distance to the worst Mahalanobis distance; the smaller the Mahalanobis distance, the larger the quotient. The node then multiplies the quotient by the Mahalanobis distance weight to obtain the Mahalanobis distance contribution. Simultaneously, the node multiplies the interval coverage weight by the interval coverage of the current work condition to obtain the coverage contribution. Finally, the Mahalanobis distance contribution and the coverage contribution are added to obtain the comprehensive judgment index for the work condition. A larger index indicates a higher degree of matching between the current feature vector and the work condition; the node selects the work condition with the largest index as the final work condition label.

[0031] Edge computing nodes read the pre-stored normal fluctuation ranges for each operating condition category from the multi-operating condition benchmark feature library. Each normal fluctuation range is a continuous numerical range with an upper and lower limit. The node compares each feature value in the current dimensionality-reduced feature vector with the normal fluctuation range corresponding to the currently evaluated operating condition. If the feature value falls between the upper and lower limits of the normal fluctuation range, it is recorded that the feature satisfies the interval coverage condition. After performing the above judgment for all features under the current operating condition, the node counts the number of features falling within the normal fluctuation range and divides this number by the total dimension of the current dimensionality-reduced feature vector. The ratio obtained is the interval coverage of the operating condition. The closer the interval coverage is to one, the more comprehensively the current feature vector falls within the normal fluctuation range of the operating condition.

[0032] Edge computing nodes determine the final operating condition by combining Mahalanobis distance and interval coverage. For each operating condition, the node first finds the maximum Mahalanobis distance among all operating conditions. Then, according to pre-set Mahalanobis distance weights and interval coverage weights, the node multiplies the Mahalanobis distance weight by the quotient of the maximum Mahalanobis distance and the current operating condition's Mahalanobis distance, and adds the interval coverage weight multiplied by the current operating condition's interval coverage to obtain the comprehensive judgment index for that operating condition. The larger the comprehensive judgment index, the higher the degree of matching between the current feature vector and the corresponding operating condition. The node directly uses the operating condition category with the largest comprehensive judgment index as the operating condition label for the current moment, and then binds the current dimensionality-reduced feature vector with this operating condition label together for storage, forming a feature dataset with operating condition labels. Each data item in this dataset clearly knows which navigation operating condition it belongs to, for subsequent deviation analysis and fault diagnosis.

[0033] The beneficial effects are as follows: By simultaneously using Mahalanobis distance to measure the overall similarity between the feature vector and each operating condition benchmark, and using normal fluctuation ranges to determine whether each feature falls within the allowable range, the combined determination of operating conditions overcomes the drawback of a single judgment criterion being prone to misjudgment when the operating condition boundaries are ambiguous. Mahalanobis distance can automatically eliminate the influence of correlation between features, enabling fair comparison even when different feature dimensions differ significantly; normal fluctuation ranges directly provide the physical allowable range for each feature, avoiding erroneous changes to the operating condition label due to slight drift of individual features. With the combination of these two methods, edge computing nodes can accurately identify the ship's current load status, sea state, and other variable operating condition information, providing accurate prerequisites for subsequent dynamic threshold selection and fault matching, and significantly improving the adaptability of the diagnostic system under variable operating conditions.

[0034] P3: Based on the feature dataset with working condition labels, filter out the deviation features in each working condition and determine the degree of deviation. When the number of deviation features exceeds the preset deviation counting threshold, extract all feature data within the time window before and after the current moment to form the sample to be diagnosed. In this embodiment of the invention, the step of filtering out deviation features in each working condition and determining the degree of deviation based on the feature dataset with working condition labels includes: Based on the operating condition label, read the upper limit and lower limit of the dynamic tolerance boundary for the current period under the current operating condition from the threshold parameter table stored locally; Each feature value in the current feature data is compared with the corresponding upper and lower limits of the dynamic tolerance boundary. Features whose current feature values ​​exceed the upper and lower limits of the dynamic tolerance boundary are taken as deviation features, and the degree of deviation is determined.

[0035] When the number of deviation features exceeds a preset deviation counting threshold, all feature data within the time window before and after the current moment are extracted to form a sample to be diagnosed, including: When the total number of deviation features exceeds the preset deviation count threshold, the current moment is marked as a suspected fault moment; The two segments before and after the suspected fault time are spliced ​​together in chronological order, and the features of the spliced ​​data block are extracted to obtain the sample to be diagnosed.

[0036] The dynamic tolerance boundary also includes: Calculate the adaptive smoothing coefficient based on the maximum degree of deviation among all deviation characteristics; The formula for calculating the adaptive smoothing coefficient is as follows: In the formula, For adaptive smoothing coefficients, This is the preset lower limit of the smoothing coefficient. This is the preset upper limit of the smoothing coefficient. As a preset sensitivity factor, The maximum degree of deviation; The dynamic tolerance boundary is updated by an exponentially weighted moving average using the adaptive smoothing coefficient, generating an updated dynamic tolerance boundary and replacing the original threshold parameter of the dynamic tolerance boundary.

[0037] Based on the identified operating condition label, the edge computing node reads the upper and lower limits of the dynamic tolerance boundary for the current cycle under that operating condition from a pre-stored threshold parameter table. The threshold parameter table maintains two sets of dynamic boundary values ​​for each operating condition. The upper limit of the dynamic tolerance boundary is a value slightly higher than the maximum value during normal operation of the equipment under that condition, and the lower limit is a value slightly lower than the minimum value during normal operation. These two boundaries together form an envelope that allows the current characteristic value to fluctuate normally. This envelope is generated from historical normal operation data when the ship is first put into service and is continuously updated and replaced during subsequent operation.

[0038] The edge computing node extracts the upper and lower limits of the dynamic tolerance boundary for each feature value in the current feature data and compares them one by one. If the current feature value is greater than the upper limit or less than the lower limit, the feature is determined to exceed the normal allowable envelope range and is marked as a deviation feature. For each marked deviation feature, the edge computing node determines its degree of deviation according to a unified rule: when the feature value exceeds the upper limit, the degree of deviation is the relative size between the excess portion and the upper limit; when the feature value is below the lower limit, the degree of deviation is the relative size between the lower limit and the excessively low portion. The resulting degree of deviation is a dimensionless positive number, representing the distance of the current feature value from the normal boundary.

[0039] The edge computing node counts the total number of all features marked as deviations at the current moment. This total number is compared with a pre-set deviation counting threshold. The deviation counting threshold is determined by analyzing the upper limit of the number of deviation features generated by occasional fluctuations during normal ship operation. Its value ensures that subsequent actions are only triggered when multiple features simultaneously and significantly exceed the threshold, thus avoiding misjudgments caused by single random noise. When the total number of deviation features exceeds the deviation counting threshold, the edge computing node records the current moment as a suspected fault moment.

[0040] For the suspected fault moment, the edge computing node retrieves two data segments from the circular buffer: one is all standardized time-series monitoring data within a fixed time window preceding that moment, and the other is all standardized time-series monitoring data within the same time window following that moment. These two segments represent the historical trend before the fault and the evolution process after the fault, respectively. The edge computing node directly concatenates these two segments in chronological order to form a continuous data block to be diagnosed. Then, the same feature extraction operation as in claim 1 is performed on this complete data block, namely, framing, extracting time-domain statistical features and frequency-domain energy features, calculating the variance inflation factor and filtering out redundancies, finally obtaining the sample to be diagnosed. This sample contains complete equipment dynamic information before and after the suspected fault, for subsequent rule matching.

[0041] After each deviation feature selection and the determination of the maximum deviation, the edge computing node calculates an adaptive smoothing coefficient based on the maximum deviation among all deviation features. A larger maximum deviation indicates a more severe deviation of the device state from the normal boundary, and the adaptive smoothing coefficient will increase accordingly to make boundary updates more sensitive. Conversely, a small maximum deviation results in a small smoothing coefficient, and boundary updates tend to be smoother. The value of this adaptive smoothing coefficient is limited to a pre-defined lower and upper limit, and the maximum deviation is mapped to a specific value of the smoothing coefficient through a monotonically increasing exponential relationship.

[0042] The preset lower limit of the smoothing coefficient is a value between zero and one, representing the minimum smoothing coefficient used when the device is completely within normal limits. The preset upper limit of the smoothing coefficient is also between zero and one, representing the maximum smoothing coefficient that can be used when the device deviates significantly. The preset sensitivity factor is a constant greater than zero, used to adjust how quickly the smoothing coefficient changes with the maximum deviation; a larger sensitivity factor results in a more rapid increase in the smoothing coefficient. The maximum deviation is the largest of all deviation characteristics; a larger value indicates that the device's current state has more severely exceeded the dynamic tolerance boundary.

[0043] First, the difference between the upper and lower limits of the smoothing coefficient is calculated to obtain a range of allowable fluctuations. Then, the node multiplies a preset sensitivity factor by the maximum deviation, negates the product, and uses it as the exponent of the natural exponential function to calculate its output value. This output value is a number between zero and one; the larger the maximum deviation, the closer the output value is to zero, and the smaller the maximum deviation, the closer the output value is to one. Next, the node subtracts this exponential function output value from one to obtain a new adjustment coefficient between zero and one. The node multiplies the previously calculated smoothing coefficient range by this adjustment coefficient to obtain an offset. Finally, the node adds the lower limit of the smoothing coefficient to this offset to obtain the final adaptive smoothing coefficient. This coefficient is close to the lower limit when the maximum deviation is small, resulting in a slight update effect; it is close to the upper limit when the maximum deviation is large, resulting in a significant update effect.

[0044] After obtaining the adaptive smoothing coefficient, the edge computing nodes use it to update the upper and lower limits of the dynamic tolerance boundary for the current period using exponentially weighted moving averages. Specifically, the currently detected actual feature value is taken as the observation value, the adaptive smoothing coefficient is used as the confidence level of the observation value, and its complement is used as the degree of retention of the old boundary value. The two are then weighted and averaged to obtain the new boundary value. The updated upper and lower limits of the dynamic tolerance boundary immediately replace the original threshold parameters for this operating condition in the edge nodes, becoming the basis for comparison in the next period.

[0045] The beneficial effects are as follows: Through the linkage mechanism of dynamic tolerance boundaries and adaptive smoothing coefficients, edge computing nodes can automatically tighten or widen the normal fluctuation range when ship equipment undergoes slow degradation or operating conditions gradually change, without the need for manual threshold reset. When multiple features simultaneously and significantly exceed the limits, the system immediately extracts complete data within the window before and after the fault for rule matching, ensuring immediate response capability to sudden faults. Simultaneously, the adaptive smoothing coefficient automatically increases with the degree of deviation, matching the boundary update speed with the severity of equipment anomalies. This avoids frequent boundary jitter during normal fluctuations while enabling rapid tracking of severe deviations. The entire process is completed independently at the ship's edge, without cloud intervention, significantly improving the adaptability and reliability of the ship's remote diagnostic system under changing operating conditions and long-term operation.

[0046] P4: Match each sample to be diagnosed to obtain the edge diagnosis results.

[0047] In this embodiment of the invention, the step of matching the sample to be diagnosed line by line to obtain the edge diagnosis result includes: Each feature value in the sample to be diagnosed is compared with the dynamic tolerance boundary corresponding to the current working condition to generate a feature over-limit flag vector; The feature over-limit flag vector is compared with the feature logic judgment conditions in the preset multiple fault judgment rules one by one, and the fault type corresponding to the rule that satisfies all the conditions is taken as the preliminary matching result. Extract the degree of deviation of the deviation features in the preliminary matching results, analyze the confidence of the fault type, and associate the fault type with the confidence to generate edge diagnosis results.

[0048] The edge computing node retrieves the sample to be diagnosed obtained in the previous steps. This sample is a feature matrix that has undergone feature extraction and dimensionality reduction, where each column represents a feature. The node simultaneously reads the dynamic tolerance boundaries corresponding to the current operating condition, including the upper and lower limits of the dynamic tolerance boundaries for that condition. The node compares each row of feature values ​​in the sample to be diagnosed with the upper and lower limits of the dynamic tolerance boundaries for the same feature name. For each feature, if the feature value is greater than the upper limit or less than the lower limit, the corresponding flag is set to an out-of-limit state; if the feature value falls within the boundaries, it is set to a normal state. After all features have been compared, the node arranges all flags in a fixed order according to the features into a feature out-of-limit flag vector. Each component in this vector takes only one of two states, clearly indicating which features in the current sample to be diagnosed exceed the normal operating envelope.

[0049] Edge computing nodes read multiple fault determination rules from a locally stored fault rule base. Each fault determination rule predefines its corresponding fault type and a set of feature logic judgment conditions associated with that fault type. These feature logic judgment conditions specifically describe which features must be in an out-of-limit state, which features must be in a normal state, and the temporal or quantitative relationship between the out-of-limit states of different features when the fault occurs. The node compares the feature out-of-limit flag vector generated in the previous step with each condition in each rule. During the comparison, the node sequentially checks whether each condition listed in the rule is satisfied by the feature out-of-limit flag vector. Only when all feature logic judgment conditions in a rule are satisfied does the node record the fault type corresponding to that rule as a preliminary matching result. For rules that do not satisfy all conditions, the node skips them and continues comparing the next rule.

[0050] Based on the fault type corresponding to the preliminary matching result, the node extracts the deviation degree of the most relevant deviation features from the previously recorded list of deviation features. The deviation degree is a pre-calculated value reflecting the relative extent to which the feature value exceeds the dynamic tolerance boundary. The node comprehensively evaluates these deviation degrees according to a pre-defined confidence analysis method: for multiple related features under the fault type, if their deviation degrees are all large and consistent in direction, a higher confidence level is assigned; if only a few features deviate while other related features are normal, the confidence level is lowered. In this way, the node obtains a confidence level judgment between low and high confidence. Finally, the node associates the fault type corresponding to the preliminary matching result with the calculated confidence level to form a structured edge-end diagnostic result. This diagnostic result includes the specific fault type name and the confidence level of the diagnosis, for reference by shipboard personnel or shore-based experts.

[0051] The beneficial effects are as follows: By comparing the samples to be diagnosed against clearly defined fault determination rules at the edge, the diagnostic process is completely transparent and interpretable, without relying on any black-box models or complex algorithms. The feature over-limit flag vector converts continuous feature values ​​into discrete state flags, greatly simplifying the logical complexity of subsequent rule matching. The rule-by-rule comparison method ensures that only rules that simultaneously meet all conditions are adopted, avoiding misdiagnosis caused by accidental over-limit of a single feature. The confidence level calculated based on the degree of deviation provides operators with a reference for the reliability of the diagnostic results, facilitating decisions on whether to immediately stop maintenance or continue observation. The entire process is completed in real time on the ship's edge equipment, and the diagnostic results can be displayed or alarms can be triggered immediately. At the same time, only a small amount of diagnostic conclusions and confidence level information are transmitted back to the shore via satellite link, greatly saving remote communication bandwidth and meeting the requirements for real-time and self-contained fault diagnosis during ship navigation.

[0052] like Figure 2 The diagram shown is a functional block diagram of a remote ship fault diagnosis system based on edge computing provided in an embodiment of the present invention.

[0053] The remote ship fault diagnosis system 100 based on edge computing described in this invention can be installed in an electronic device. Depending on the functions implemented, the remote ship fault diagnosis system 100 based on edge computing may include an edge feature dimensionality reduction module 101, a multi-condition adaptive identification module 102, a sample construction module 103, and a fault matching module 104. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0054] In this embodiment, the functions of each module / unit are as follows: The edge feature dimensionality reduction module 101 is used by the edge computing node to extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain a dimensionality-reduced feature vector. The multi-condition adaptive recognition module 102 is used to determine the feature dataset with condition labels based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector in each condition, and with the normal fluctuation range as the inclusion condition. The sample construction module 103 is used to filter out deviation features in each working condition and determine the degree of deviation based on the feature dataset with working condition labels. When the number of deviation features exceeds the preset deviation counting threshold, all feature data within the time window before and after the current moment are extracted to form a sample to be diagnosed. The fault matching module 104 is used to match the sample to be diagnosed one by one to obtain the edge diagnosis result.

[0055] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0056] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0057] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0058] 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 present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0059] The embodiments of this application can acquire and process relevant data based on an artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A remote fault diagnosis method for ships based on edge computing, characterized in that, The method includes: P1: Edge computing nodes extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain a dimension-reduced feature vector; P2: Based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector in each working condition, and using the normal fluctuation range as the inclusion condition, determine the feature dataset with working condition labels. P3: Based on the feature dataset with working condition labels, filter out the deviation features in each working condition and determine the degree of deviation. When the number of deviation features exceeds the preset deviation counting threshold, extract all feature data within the time window before and after the current moment to form the sample to be diagnosed. P4: Match each sample to be diagnosed to obtain the edge diagnosis results.

2. The remote ship fault diagnosis method based on edge computing as described in claim 1, characterized in that, The edge computing node extracts time-frequency domain features from standardized time-series monitoring data and determines the variance inflation factor of the time-frequency domain features, including: The standardized time-series monitoring data is processed by frame segmentation, and time-domain statistical features and frequency-domain energy features are extracted sequentially from each frame of data to obtain an initial feature matrix; Using one feature column from the initial feature matrix as the dependent variable and the remaining feature columns as independent variables, linear regression is performed to obtain the variance inflation factor.

3. The remote ship fault diagnosis method based on edge computing as described in claim 2, characterized in that, The process of redundancy removal to obtain a dimensionality-reduced feature vector includes: The feature columns whose variance inflation factor exceeds the preset collinearity tolerance threshold are used as the first set of features for screening. The linear regression and threshold filtering are repeated on the first-selection feature set until the variance inflation factor of all remaining features does not exceed the collinearity tolerance threshold, thus obtaining the dimensionality-reduced feature vector.

4. The remote ship fault diagnosis method based on edge computing as described in claim 1, characterized in that, The step of calculating the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector in each working condition includes: Read the feature mean vector and covariance matrix corresponding to the current comparison condition from the preset multi-condition benchmark feature library; The difference vector between the current dimensionality-reduced feature vector and the feature mean vector is used as the deviation vector; Perform a vector-matrix multiplication operation on the deviation vector and the inverse of the covariance matrix to obtain the intermediate transformation vector; The intermediate transformation vector and the deviation vector are multiplied by a dot product to obtain the squared value of the Mahalanobis distance. Then, the square root of the square root is taken to obtain the Mahalanobis distance between the current dimensionality-reduced feature vector and the feature mean vector.

5. The remote ship fault diagnosis method based on edge computing as described in claim 4, characterized in that, The identified feature dataset with working condition labels includes: Read the normal fluctuation range corresponding to each operating condition category from the multi-operating condition benchmark feature library; Determine whether each feature value in the current dimensionality-reduced feature vector falls within the normal fluctuation range, and use the ratio of all feature dimensions that fall within the normal fluctuation range to the total dimension of the current dimensionality-reduced feature vector as the range coverage. Based on the Mahalanobis distance and the interval coverage, a comprehensive judgment index is calculated for the current dimensionality-reduced feature vector and each working condition category. The formula for calculating the comprehensive judgment index is as follows: In the formula, The current dimensionality-reduced feature vector and the first The comprehensive judgment index for each working condition category The Mahalanobis distance weights are used. The largest Mahalanobis distance value among all load case categories. The current dimensionality-reduced feature vector and the first Mahalanobis distance for each work condition category For a preset positive small quantity, For interval tolerance weights, For the first Range tolerance for each working condition category; The working condition category with the largest comprehensive judgment index is selected as the working condition label of the current working condition, and the current dimensionality reduction feature vector is associated with the working condition label and stored to generate a feature dataset with working condition label.

6. The remote ship fault diagnosis method based on edge computing as described in claim 1, characterized in that, The step of filtering out deviation features and determining the degree of deviation in each working condition based on the feature dataset with working condition labels includes: Based on the operating condition label, read the upper limit and lower limit of the dynamic tolerance boundary for the current period under the current operating condition from the threshold parameter table stored locally; Each feature value in the current feature data is compared with the corresponding upper limit and lower limit of the dynamic tolerance boundary. Features whose current feature values ​​exceed the upper and lower limits of the dynamic tolerance boundary are taken as deviation features, and the degree of deviation is determined.

7. The remote ship fault diagnosis method based on edge computing as described in claim 6, characterized in that, When the number of deviation features exceeds a preset deviation counting threshold, all feature data within the time window before and after the current moment are extracted to form a sample to be diagnosed, including: When the total number of deviation features exceeds the preset deviation count threshold, the current moment is marked as a suspected fault moment; The two segments before and after the suspected fault time are spliced ​​together in chronological order, and the features of the spliced ​​data block are extracted to obtain the sample to be diagnosed.

8. The remote ship fault diagnosis method based on edge computing as described in claim 6, characterized in that, The dynamic tolerance boundary also includes: Calculate the adaptive smoothing coefficient based on the maximum degree of deviation among all deviation characteristics; The formula for calculating the adaptive smoothing coefficient is as follows: In the formula, For adaptive smoothing coefficients, The lower limit of the preset smoothing coefficient, This is the preset upper limit of the smoothing coefficient. As a preset sensitivity factor, The maximum degree of deviation; The dynamic tolerance boundary is updated by an exponentially weighted moving average using the adaptive smoothing coefficient, generating an updated dynamic tolerance boundary and replacing the original threshold parameter of the dynamic tolerance boundary.

9. The remote ship fault diagnosis method based on edge computing as described in claim 1, characterized in that, The step of matching each sample to be diagnosed to obtain the edge diagnosis result includes: Each feature value in the sample to be diagnosed is compared with the dynamic tolerance boundary corresponding to the current working condition to generate a feature over-limit flag vector; The feature over-limit flag vector is compared with the feature logic judgment conditions in the preset multiple fault judgment rules one by one, and the fault type corresponding to the rule that satisfies all the conditions is taken as the preliminary matching result. Extract the degree of deviation of the deviation features in the preliminary matching results, analyze the confidence of the fault type, and associate the fault type with the confidence to generate edge diagnosis results.

10. A remote fault diagnosis system for ships based on edge computing, characterized in that, The system for implementing the remote ship fault diagnosis method based on edge computing as described in claim 1 includes: The edge feature dimensionality reduction module is used by edge computing nodes to extract time-frequency domain features from standardized time-series monitoring data, determine the variance inflation factor of the time-frequency domain features and perform redundancy filtering to obtain dimensionality-reduced feature vectors. The multi-condition adaptive recognition module is used to determine the feature dataset with condition labels based on the Mahalanobis distance between the current dimensionality-reduced feature vector and the mean feature vector of each condition, and with the normal fluctuation range as the inclusion condition. The sample construction module is used to filter out the deviation features in each working condition and determine the degree of deviation based on the feature dataset with working condition labels. When the number of deviation features exceeds the preset deviation counting threshold, all feature data within the time window before and after the current moment are extracted to form the sample to be diagnosed. The fault matching module is used to match the samples to be diagnosed one by one to obtain the edge diagnosis results.

Images