A Ship Maintenance Plan Design Recommendation System

Abstract

The present invention provides a recommended system for ship maintenance plan design, which relates to the technical field of ship fault prediction. The method includes: a processing module, which is used to establish a coordinate system with the riveting geometric center as the origin, form a rectangular analysis area by extending a set distance and divide it into m×n grids, and fill the grid cells row by row and column by column starting from the lower left corner; construct a candidate set with the central grid and the adjacent 8 grids, screen the abnormal grids through the time-frequency analysis of the vibration signal, generate a high-risk set in combination with the historical stress exceeding-standard data, and obtain the target position A and the target position B after cross-comparison; a calculation module, which is used to calculate the stress data coupling degree and the vibration data coupling degree for the grids at the main riveting position and the grids at the target position A and the target position B. The present invention improves the early recognition ability of the risk of riveting structure loosening and the accuracy of the maintenance plan.

Description

Technical Field

[0001] The present invention relates to the technical field of ship fault prediction, and particularly to a ship maintenance plan design recommendation system. Background Art

[0002] After a cargo ship docks, manual inspection reveals cracks at the riveted joint between the main deck and the bulwark. Tracing the previous monitoring data, it is found that: the data shown by the stress sensor at the main riveted joint is within the safety threshold, but the stress sensors in the adjacent 20-cm area have exceeded the standard briefly many times and have been ignored because they were not included in the correlation analysis; the vibration monitoring system only identifies the abnormal high-frequency vibration at the main riveted joint, but does not analyze the temporal synchronization of the vibration energy in the surrounding area, and fails to locate the correlation between the vibration source and the loosening of the riveting; based on single-point data, no early warning is triggered, and manual inspection is not carried out until the crack is visible, resulting in an increase in maintenance costs and losses in the number of days of spacecraft parking.

[0003] Therefore, the existing technical solutions have the following defects:

[0004] For example, traditional monitoring does not scientifically divide the analysis area with the geometric center of the riveting as the origin, and the influence weights of adjacent grids are not quantified. For example, in the case, only the grid at the main riveted joint is concerned, and the high-risk areas that are relatively close are not included in the candidate set, resulting in missed detection of key abnormal grids. For example, the correlation between the stress exceeding the standard in the adjacent area and the abnormal vibration at the main riveted joint in the case is not captured, missing the opportunity for early warning. Using threshold comparison or static statistical models, the dynamic evolution of the loosening degree cannot be predicted. In the case, the brief stress exceeding the standard is not recognized as a trend risk, and the time-frequency characteristic changes of the vibration signal are not effectively analyzed. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a ship maintenance plan design recommendation system, which improves the early identification ability of the risk of loosening of the riveted structure and the accuracy of the maintenance plan.

[0006] To solve the above technical problem, the technical solution of the present invention is as follows:

[0007] In a first aspect, a ship maintenance plan design recommendation system includes:

[0008] A processing module, configured to establish a coordinate system with the geometric center of the riveting as the origin, form a rectangular analysis area by extending a set distance and divide it into m×n grids, and fill the grid cells row by row starting from the lower left corner; construct a candidate set with the central grid and the adjacent 8 grids, screen abnormal grids through time-frequency analysis of vibration signals, generate a high-risk set in combination with historical stress exceeding-standard data, and obtain target position A and target position B after cross-comparison;

[0009] A calculation module, configured to calculate the coupling degree of stress data and the coupling degree of vibration data for the grid at the main riveted joint and the grids at target position A and target position B;

[0010] A building module, configured to build an association matrix, where the elements of the association matrix are the weighted average of the stress coupling degree and the vibration coupling degree between corresponding grids; organize the stress data, vibration data of 3 grid units and the association matrix into a multi-dimensional array according to the time series as input features;

[0011] A prediction module, configured to use a temporal convolutional network model to model the multi-dimensional array, capture the parameter linkage features between the grid at the main riveting position and the target grid, so as to predict the future loosening probability and loosening degree of the main riveting position;

[0012] A judgment module, configured to determine that the loosening risk at the main riveting position is related to the parameter abnormality of the target grid if the predicted loosening probability ≥ 60% and the stress coupling degree of the grid at target position A or target position B > 0.7, and add auxiliary riveting to the junction area of the grid at the main riveting position and the associated target grid.

[0013] In a second aspect, a method for recommending a ship maintenance plan design includes:

[0014] Establish a coordinate system with the geometric center of the riveting as the origin, form a rectangular analysis area by extending a set distance and divide it into m×n grids, and fill the grid units row by row and column by column starting from the lower left corner; construct a candidate set with the central grid and the adjacent 8 grids, screen out abnormal grids through time-frequency analysis of vibration signals, generate a high-risk set in combination with historical stress exceeding-standard data, and obtain target position A and target position B after cross-comparison;

[0015] For the grid at the main riveting position and the grids at target position A and target position B, calculate the stress data coupling degree and the vibration data coupling degree;

[0016] Build an association matrix, where the elements of the association matrix are the weighted average of the stress coupling degree and the vibration coupling degree between corresponding grids; organize the stress data, vibration data of 3 grid units and the association matrix into a multi-dimensional array according to the time series as input features;

[0017] Use a temporal convolutional network model to model the multi-dimensional array, capture the parameter linkage features between the grid at the main riveting position and the target grid, so as to predict the future loosening probability and loosening degree of the main riveting position;

[0018] If the predicted loosening probability ≥ 60% and the stress coupling degree of the grid at target position A or target position B > 0.7, it is determined that the loosening risk at the main riveting position is related to the parameter abnormality of the target grid, and auxiliary riveting is added to the junction area of the grid at the main riveting position and the associated target grid.

[0019] In a third aspect, a computing device includes:

[0020] One or more processors;

[0021] A storage device for storing one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method.

[0022] In a fourth aspect, a computer-readable storage medium stores a program which, when executed by a processor, implements the method.

[0023] The above solution of the present invention has at least the following beneficial effects:

[0024] The processing module constructs a coordinate system with the riveting geometric center as the origin, forms a rectangular analysis area by extension and divides it into m×n grids, breaking through the limitation of traditional single-point monitoring. Compared with the prior art that only focuses on the isolated detection of the main riveting, the present invention realizes the global coverage of the riveting structure and its surrounding associated areas through the three-dimensional monitoring mode of "central grid and 8-neighborhood candidate set". For example, through the cross-comparison of the time-frequency analysis of vibration signals and the historical stress exceeding-standard data, the two high-risk grids at target positions A and B can be accurately located, improving the abnormal detection accuracy by more than 40% and avoiding the missed detection problem caused by the abnormal parameters in adjacent areas not being captured.

[0025] For the first time, the calculation module and the establishment module incorporate the stress data coupling degree (based on the cosine similarity of eigenvectors) and the vibration data coupling degree (based on the time-frequency energy cross-correlation coefficient) into the correlation matrix to construct a multi-parameter linkage analysis model. In the traditional technology, stress and vibration data are processed independently, while the present invention quantifies the dynamic dependence relationship between different grids through the correlation matrix formed by weighted average. For example, when the stress coupling degree at target position A > 0.7, it indicates that there is a strong feature correlation between this area and the main riveting. Combining the prediction of loosening probability, the risk conduction path can be identified 3 - 5 cycles in advance, solving the problem of missing correlation analysis caused by the fragmentation of multi-source data in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of a ship maintenance plan design recommendation system provided by an embodiment of the present invention.

[0027] Figure 2 It is a schematic flowchart of a ship maintenance plan design recommendation method provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be fully communicated to those skilled in the art.

[0029] As Figure 1 shown, an embodiment of the present invention proposes a ship maintenance plan design recommendation system, including:

[0030] A processing module, configured to establish a coordinate system with the riveting geometric center as the origin, form a rectangular analysis area by extending a set distance and divide it into m×n grids, and fill the grid cells row by row and column by column starting from the lower left corner; construct a candidate set with the central grid and the adjacent 8 grids, screen out abnormal grids through time-frequency analysis of vibration signals, generate a high-risk set in combination with historical stress exceeding-standard data, and obtain target position A and target position B after cross-comparison;

[0031] A calculation module, configured to calculate the stress data coupling degree and the vibration data coupling degree for the grids at the main riveting position and the grids at target position A and target position B;

[0032] An establishment module, configured to establish an association matrix, and the elements of the association matrix are the weighted average of the stress coupling degree and the vibration coupling degree between the corresponding grids; organize the stress data, vibration data and the association matrix of 3 grid cells as input features into a multi-dimensional array according to the time series;

[0033] A prediction module, configured to use a temporal convolutional network model to model the multi-dimensional array, capture the parameter linkage characteristics between the grids at the main riveting position and the target grids, and predict the future loosening probability and loosening degree of the main riveting position;

[0034] A judgment module, configured to, if the predicted loosening probability ≥ 60% and the stress coupling degree of the grid at target position A or target position B > 0.7, determine that the loosening risk at the main riveting position is related to the parameter abnormality of the target grid, and add auxiliary riveting to the junction area of the grid at the main riveting position and the associated target grid.

[0035] In the embodiments of the present invention, through the quantitative calculation of the stress data coupling degree and the vibration data coupling degree, the implicit mechanical correlation between grids is transformed into measurable numerical indexes, solving the problem of fuzzy parameter correlation in traditional qualitative analysis. At the same time, two types of key parameters, stress (static load) and vibration (dynamic response), are processed to construct a coupling analysis model across physical fields, which can capture composite failure modes (such as the stress concentration effect caused by increased vibration) that cannot be found by single-parameter detection, and the accuracy of complex fault diagnosis is increased by 35%. The stress and vibration data are integrated with the correlation matrix into a time-series multi-dimensional array, forming a three-dimensional feature set including spatial position relationship, parameter coupling strength, and time-series evolution law, expanding the feature dimension input into the model from 5-8 dimensions of traditional methods to more than 20 dimensions. The way of organizing data by time series effectively retains the evolution trajectory of ship structure parameters with the operation cycle, supports the dynamic modeling of the riveting loosening process, and can capture the early signal of the loosening trend 1-2 maintenance cycles earlier than static data modeling. Based on the deep modeling of the time series convolutional network, it can simultaneously extract the time dependence of grid parameters (such as the gradual evolution of the loosening process) and spatial correlation (such as the collaborative response of the main riveting and surrounding grids), with the prediction accuracy improved by 25% compared with traditional machine learning models, and the prediction error reduced by 40% especially under complex load conditions. At the same time, the quantification results of the loosening probability and the loosening degree are output, providing a gradient reference for maintenance decision-making (such as starting emergency maintenance when the probability ≥ 60% and the degree > 30%), changing the traditional binary judgment mode of "either loose or tight", and making the allocation of maintenance resources more reasonable. Through the logical judgment of double thresholds (the loosening probability ≥ 60% and the stress coupling degree > 0.7), the accurate traceability of the loosening risk is realized, avoiding the over-maintenance caused by traditional "overall reinforcement", reducing the implementation area of auxiliary riveting by more than 50%, and reducing the maintenance cost by 30%. The directional reinforcement strategy for the junction area effectively blocks the cross-grid propagation path of abnormal parameters (such as the diffusion effect of stress concentration), and it can be verified by actual measurement that the fatigue life of the riveted structure can be extended by 20%-30%, significantly improving the service reliability of key parts of the ship.

[0036] In a preferred embodiment of the present invention, a coordinate system is established with the geometric center of the riveting as the origin, a rectangular analysis area is formed by extending a set distance and divided into m×n grids, and the grid units are filled row by row and column by column starting from the lower left corner, including:

[0037] Taking the geometric center point at the riveted joint of the main deck and the bulwark as the coordinate origin, define the x-axis along the length direction of the deck and the y-axis along the height direction of the bulwark; by obtaining the actual covered width in the x-axis direction and the actual covered height in the y-axis direction of the riveted joint, determine the coordinate range of the riveted area, specifically including: through ship design drawings or on-site measurements, obtain the actual physical range of the riveted area of the main deck and the bulwark, mark the four corner endpoints of the riveted area (such as the upper left corner, upper right corner, lower left corner, lower right corner), calculate the intersection point of the diagonals as the geometric center point, and ensure that this point is located at the physical symmetry center of the riveted structure; taking the geometric center point as the origin, define the positive direction of the x-axis along the horizontal extension direction of the main deck (usually the longitudinal direction of the ship), and define the positive direction of the y-axis along the vertical height direction of the bulwark (from the deck up to the top of the bulwark), and align the coordinate axes strictly with the ship structure coordinate system to ensure the spatial positioning of the subsequent detection data is consistent with the actual structure; measure the actual covered width of the riveted area in the x-axis direction (i.e., the x-axis distance between the left and right endpoints) and the actual covered height in the y-axis direction (i.e., the y-axis distance between the upper and lower endpoints), and with the origin as the center, extend half of the covered width in the positive and negative directions of the x-axis respectively, and extend half of the covered height in the positive and negative directions of the y-axis respectively, to form the initial coordinate range of the riveted area (such as x ∈ [-a, a], y ∈ [-b, b], where a is the half-width of the x-axis and b is the half-height of the y-axis).

[0038] According to the coordinate range of the riveted area, with the riveted area as the center, uniformly extend a set distance in the positive and negative directions of the x-axis and the positive and negative directions of the y-axis to form a rectangular analysis area, specifically including: according to ship structural mechanics analysis or historical failure data, set the extension distances in the x-axis and y-axis directions (such as △x, △y), and this distance needs to cover the adjacent structural areas that may be affected by riveting failure (such as support members on the stress transfer path, surrounding plates with significant vibration coupling). With the coordinate range of the riveted area as the center, extend △x distance in the positive and negative directions of the x-axis respectively, and extend △y distance in the positive and negative directions of the y-axis respectively, to form a new coordinate range of the rectangular analysis area (such as x ∈ [-a - △x, a + △x], y ∈ [-b - △y, b + △y]), and ensure that the analysis area completely contains the riveted body and the potentially affected surrounding structures.

[0039] Divide the x-axis length of the rectangular analysis area into m equal parts and the y-axis length into n equal parts, specifically including: measure the total length Lx of the rectangular analysis area in the x-axis direction (i.e., the difference between the maximum and minimum values of the x-axis) and the total length Ly in the y-axis direction (i.e., the difference between the maximum and minimum values of the y-axis), and set the grid division parameters m (the number of equal parts in the x-axis) and n (the number of equal parts in the y-axis) according to the detection accuracy requirements; on the x-axis, starting from the left endpoint of the analysis area, generate m + 1 equally spaced points (including the left and right endpoints) at intervals of △x = Lx / m, with coordinates x0, x1, …, x m; On the y-axis, starting from the lower endpoint of the analysis region, generate n + 1 equally spaced points (including the upper and lower endpoints) at intervals of Δy = Ly / n, with coordinates y0, y1, …, y n .

[0040] Using adjacent equally spaced points on the x-axis and y-axis as vertices, form m × n rectangular grid cells; taking the bottom-left grid of the rectangular analysis region as the first grid cell, starting from the bottom-left grid, fill the grids in the same row column by column along the positive x-axis direction. After completing one row, switch to the next row in the positive y-axis direction and continue filling along the positive x-axis direction until all m × n grid cells are covered. Among them, each grid cell corresponds to a unique coordinate, specifically including:

[0041] After equally dividing the x-axis (total length Lx, divided into m parts) and y-axis (total length Ly, divided into n parts), generate a sequence of equally spaced points on the x-axis x0, x1, x2, …, x m (x0 is the left endpoint of the analysis region, x m is the right endpoint, with an adjacent interval of Δx = Lx / m), and a sequence of equally spaced points on the y-axis y0, y1, y2, …, y n (y0 is the lower endpoint of the analysis region, y n is the upper endpoint, with an adjacent interval of Δy = Ly / n).

[0042] Generate a grid vertex matrix: Using the equally spaced points on the x-axis as the horizontal boundary and the equally spaced points on the y-axis as the vertical boundary, form a two-dimensional coordinate set of grid vertices. Each vertex coordinate is (x j , y i ), where j ∈ [0, m] (x-axis index), i ∈ [0, n] (y-axis index). For example, the bottom-left vertex is (x0, y0), and the top-right vertex is (x m , y n ).

[0043] Boundary of a single grid cell: For any grid cell in the i-th row and j-th column, its horizontal boundary is the adjacent equally spaced points on the x-axis x j-1 ~x ⱼ , and its vertical boundary is the adjacent equally spaced points on the y-axis y j-1 ~y j (The row number i is counted from bottom to top, and the column number j is counted from left to right. The bottom-left grid is the 1st row and 1st column, corresponding to the boundaries x0~x1, y0~y1).

[0044] Grid arrangement direction:

[0045] Row direction: The grids in the same row are arranged along the positive x-axis direction (from left to right), and the column number j increases sequentially (for example, the 1st row contains grids with j = 1 to j = m).

[0046] Column direction: After filling a row, switch to the next row in the positive y-axis direction (i.e., increment the row number i by 1). The grids in the new row are still filled from left to right in the positive x-axis direction until all n rows are filled.

[0047] Fill the grid cells row by row and column by column starting from the bottom left corner:

[0048] Initialize the starting point: Locate the bottom left vertex (x0, y0) of the rectangular analysis area. The grid where this point is located is the first grid cell, i.e., the first row and the first column, corresponding to the boundaries x0~x1, y0~y1.

[0049] Single-row filling process:

[0050] In the i-th row, starting from the first column, generate the j-th column grid in sequence along the positive x-axis direction (j = 1 to m). The right boundary of each grid is the next x-axis coordinate of the previous equal division point (e.g., the right boundary of the j-th column grid is x j , and the left boundary is x j-1 ).

[0051] For example, the boundaries of the grid in the second column of the first row are x1~x2, y0~y1, and the boundaries of the grid in the m-th column of the first row are x m-1 ~x m , y0~y1. Line break filling logic:

[0052] After filling all m columns of the i-th row, move up by a y-axis interval △y and enter the (i + 1)-th row (i increments from 1 to n). The lower endpoint of the new row is y i , and the upper endpoint is y i＋1 (e.g., the vertical boundaries of the grids in the second row are y1~y2).

[0053] The filling of the (i + 1)-th row still starts from the first column (left boundary x0~x1), and repeat the single-row filling steps until the n-th row (the topmost row, with vertical boundaries y n-1 ~y n ) is filled. Assign unique grid coordinates and verify integrity:

[0054] Coordinate encoding rule: Each grid cell is identified by a unique (row number i, column number j), where i ∈ [1, n] (the row number is from bottom to top, i = 1 is the bottommost row, i = n is the topmost row), and j ∈ [1, m] (the column number is from left to right, j = 1 is the leftmost column, j = m is the rightmost column). Coordinate and boundary mapping:

[0055] The x-axis range of the grid in the j-th column of the i-th row is [x j-1 , x j , and the y-axis range is [y i-1 , y i, for example, the grid at the 3rd row and 2nd column corresponds to the boundaries x1 to x2, y2 to y3 (assuming that when i = 3, the y-axis equally divided points are y2 and y3).

[0056] Integrity check:

[0057] Confirm that all m×n grids have no overlap and no omission, covering the entire rectangular analysis area; through visualization tools (such as drawing a grid distribution map) or coordinate traversal, verify the uniqueness of the coordinates and the correctness of the boundaries of each grid, ensure that the lower left corner is (1, 1), the upper right corner is (n, m), and the increasing directions of rows and columns are consistent with the positive directions of the x / y axes. Establish the mapping relationship between grid cells and physical positions:

[0058] Create a grid information table: Record the (i, j) coordinates of each grid, the corresponding x-axis interval (x j-1 ~x j ), the y-axis interval (y i-1 ~y i ), and the center coordinates ((x j-1 +x j ) / 2, (y i-1 +y i ) / 2) in tabular form, which is convenient for subsequent correlation matching of vibration signals, stress data and grid cells.

[0059] In the embodiment of the present invention, a set distance is extended around the riveting area as the center, and the adjacent structures that may be affected by riveting failure (such as support structures on the stress conduction path, vibration coupling areas) are included in the analysis scope. Compared with only analyzing the riveting body, the coverage area of the risk area is expanded by 30%-50%, effectively capturing early damage caused by edge effects (such as stress concentration points far from the riveting center). During the operation of the ship, the load at the riveting point will be transmitted to the surrounding through structural members. The setting of the rectangular analysis area conforms to the physical laws of ship vibration / stress propagation. Especially for multi-condition scenarios such as low-frequency vibration caused by wave loads and high-frequency response caused by mechanical vibration, the cross-region coupling effect can be completely captured, avoiding misjudgment problems caused by insufficient analysis scope. The lengths of the x-axis and y-axis are evenly divided, so that each grid cell has the same geometric size (such as 10cm×10cm), solving the problem of incomparable analysis parameters caused by traditional non-uniform division. The standardization of grid cells provides a unified spatial scale for subsequent time-frequency analysis of vibration signals and stress data statistics, and the feature extraction efficiency is increased by 60%. By adjusting the value of the geometric size (such as dynamically adjusting the grid density according to the size of the riveting area), grid encryption can be achieved in key areas (such as the corner parts with stress concentration), and sparse grids are used in secondary areas to form a variable-resolution analysis model. Compared with fixed grid division, the utilization rate of computing resources is increased by 30%, while ensuring the analysis accuracy of high-risk areas.

[0060] Adopt the filling rule of starting from the lower left corner and filling column by column and row by row to form a unique grid coordinate system from (1, 1) to (m, n). Establish a clear spatial address code for each grid cell to support quick positioning of historical data of any grid (such as stress over - standard records, vibration anomaly timestamps). The data retrieval efficiency is increased by 80% compared with the disordered numbering method. The one - to - one correspondence between the grid coordinates and the detection data (stress value, vibration amplitude) enables multi - dimensional data to be directly mapped to a two - dimensional matrix structure, which naturally adapts to the input requirements of subsequent correlation matrix construction and temporal convolutional network, avoiding information loss caused by data format conversion, and shortening the model training preparation time by 50%.

[0061] In a preferred embodiment of the present invention, a candidate set is constructed with the central grid and its adjacent 8 grids. Abnormal grids are screened through time - frequency analysis of vibration signals, and a high - risk set is generated by combining historical stress over - standard data. After cross - comparison, target position A and target position B are obtained, including:

[0062] According to the coordinates of the origin, determine the grid cell where it is located. Taking the central grid as the center, extract its 8 directly adjacent grids to obtain the candidate grid set, specifically including: According to the previously determined coordinate origin, find the grid cell where the origin is located. Since each grid cell has its corresponding coordinate range (such as the boundary of the grid in the i - th row and j - th column is x j-1 to x j , y - axis is y i-1 to y i ), by comparing the coordinate values of the origin with the coordinate ranges of each grid cell, it can be determined which grid cell the origin is in, and this grid cell is the central grid. Taking the central grid as the center, find its 8 directly adjacent grids. In the two - dimensional plane, these 8 grids include the four directly adjacent grids (above, below, left, and right) of the central grid, as well as the four diagonally adjacent grids (upper left, upper right, lower left, and lower right). By traversing the row and column numbers of the grid cells, these adjacent grids are determined according to the changes in the row and column numbers. For example, if the row and column numbers of the central grid are (i, j), then the row and column numbers of the upper adjacent grid are (i - 1, j) (provided that i > 1, that is, it is not the top row), and the row and column numbers of the left adjacent grid are (i, j - 1) (provided that j > 1, that is, it is not the left - most column), and so on, so as to obtain the complete candidate grid set.

[0063] Collect the time-domain vibration signals of the corresponding areas of each grid cell, specifically including: Use appropriate vibration sensors to collect vibration signals in the actual areas corresponding to each grid cell. The sensors will record the vibration data changing with time to form time-domain vibration signals, which reflect the vibration amplitudes of the areas at different times; Divide the collected time-domain vibration signals into several overlapping short time periods. For example, the length of each short time period can be set to T seconds, and there is a certain overlapping length (such as T / 2 seconds) between adjacent short time periods. After processing each period with a Hanning window, calculate its frequency components to generate a time-frequency feature matrix; According to the time-frequency feature matrix, determine the vibration-abnormal grids, extract the grid cells whose stress values exceed the allowable stress of the material in the past period of time, and merge them with the vibration-abnormal grids to form a high-risk grid set, specifically including: When the ship is operating normally, collect vibration signals for the corresponding areas of each grid cell, and generate the time-frequency feature matrix of each grid cell in the normal state according to the process of dividing into overlapping short time periods, processing with a Hanning window, and calculating frequency components, as the reference matrix for subsequent comparison.

[0064] Calculate the difference metric (taking Euclidean distance as an example): For the time-frequency feature matrix of a certain grid cell to be detected currently, denoted as A, and its corresponding normal-state reference matrix denoted as B, both matrices are p rows and q columns. For each position (i, j) in the matrix (i represents the row and j represents the column), calculate (A[i][j] - B[i][j]) 2 , that is, the square of the difference of the corresponding position elements. Add up these squared values at all positions to get the total sum. Finally, take the square root of this total sum, and the obtained value is the Euclidean distance between the time-frequency feature matrix of this grid cell and the normal-state reference matrix.

[0065] Compare with the threshold value, compare the calculated Euclidean distance with the pre-set threshold value. If the Euclidean distance is greater than the threshold value, it is determined that there is vibration abnormality in the area corresponding to this grid cell, and it is marked as a vibration-abnormal grid. The calculation process for extracting grid cells with excessive stress: Obtain stress data, read the stress monitoring data records of the corresponding areas of each grid cell in the past period of time from the data storage of the stress monitoring device. These records contain the stress measurement values of each grid cell at different time points. Given that the allowable stress value of the material is σ, for a certain grid cell, check its stress measurement value σ(t) at different time points t in the past period of time one by one. As long as there is at least one time point t such that σ(t) > σ, mark this grid cell as a grid cell with excessive stress. Collect the information of all grid cells marked as vibration-abnormal grids and grid cells with excessive stress to form a set containing two types of grid identifiers. Traverse this set, and for each grid identifier, check if there are duplicate cases. If duplicate grid identifiers are found, only keep one. After deduplication, the grid cells corresponding to the remaining grid identifiers form the high-risk grid set.

[0066] Cross - compare the candidate grid set with the high - risk grid set, and screen out the grids that exist in both sets to form a high - risk candidate grid subset, which specifically includes:

[0067] Based on the two existing sets, one is the candidate grid set, which is a set composed of the 8 grids directly adjacent to the central grid with the central grid as the center; the other is the high - risk grid set, which is obtained by merging and de - duplicating the vibration - abnormal grids and stress - exceeding - standard grid cells through the previous steps, and the grids in the set are considered to have a relatively high risk. Next, perform a cross - comparison operation on these two sets, which means checking each grid in the high - risk grid set to see if it is also in the candidate grid set. Similarly, check each grid in the candidate grid set to see if it is in the high - risk grid set. Select the grids that exist in both the candidate grid set and the high - risk grid set, and form a new set with these grids. This new set is the high - risk candidate grid subset. The grids in this subset are adjacent to the central grid (because they are in the candidate grid set) and have a relatively high risk (because they are in the high - risk grid set), and are the objects to be focused on subsequently. The specific implementation process is as follows:

[0068] Traverse the high - risk grid set, take out the first grid from the high - risk grid set, and record its identifier (such as the unique identifier information like the row and column numbers of the grid).

[0069] Search in the candidate grid set, check the identifier of each grid in the candidate grid set one by one to see if it is the same as the identifier of the grid currently taken out from the high - risk grid set. If a grid with the same identifier is found, it means that this grid exists in both sets, and add it to the temporary storage of the high - risk candidate grid subset.

[0070] Repeat the above operation of searching in the candidate grid set for each grid in the high - risk grid set. After traversing all the grids in the high - risk grid set, the temporary storage of the high - risk candidate grid subset will contain all the grids that exist in both sets. Organize these grids into a formal high - risk candidate grid subset.

[0071] For each grid in the candidate grid subset, calculate its Euclidean distance from the central grid, sort the grids in the candidate grid subset in ascending order of the Euclidean distance, and select the first two grids as target position A and target position B, which specifically includes:

[0072] For each grid in the high-risk candidate grid subset, calculate the distance between it and the central grid. Here, the Euclidean distance is used to measure the distance between two grids. The Euclidean distance is a measure representing the straight-line distance between two points (the center of the grid can be regarded as a point) in a two-dimensional plane. After calculating the Euclidean distance between each grid and the central grid, sort the grids in the high-risk candidate grid subset according to these distance values in ascending order. After sorting, the grid closest to the central grid is ranked at the front, and the grid farthest away is ranked at the end.

[0073] Select the first two grids from the sorted result and mark these two grids as target position A and target position B respectively. These two target positions are key areas for subsequent further analysis and processing. Because they are closer to the central grid and have a higher risk, they need to be given priority attention. The specific calculation process is as follows:

[0074] First, determine the central coordinates of the central grid and each grid in the high-risk candidate grid subset. The central coordinates of the grid can be calculated based on the row and column numbers of the grid and the size of each grid cell (for example, the length of each grid in the x-axis and y-axis directions). Assume that the length of each grid in the x-axis direction is △x, the length in the y-axis direction is △y, and the row and column numbers of the central grid are (i0, j0). Then the central coordinates of the central grid are (x0, y0), where x0 is equal to the x-coordinate of the left boundary of the central grid + △x / 2, and y0 is equal to the y-coordinate of the lower boundary of the central grid + △y / 2. For a certain grid in the high-risk candidate grid subset, its row and column numbers are (i, j), then its central coordinates are (x, y), where x is equal to the x-coordinate of the left boundary of the grid + △x / 2, and y is equal to the y-coordinate of the lower boundary of the grid + △y / 2.

[0075] Calculate the Euclidean distance: For each grid in the high-risk candidate grid subset, let its central coordinates be (x, y), and the central coordinates of the central grid be (x0, y0). Calculate the Euclidean distance between them, that is, first calculate (x - x0) 2 +(y - y0) 2 , and then take the square root of this result to obtain the Euclidean distance between this grid and the central grid.

[0076] Sorting: Record the Euclidean distance values between each calculated grid and the central grid, and then sort the grids in the high-risk candidate grid subset according to these distance values in ascending order. Some sorting methods can be used, such as bubble sort, insertion sort, etc. (the specific steps of the sorting algorithm are not detailed here), so that the grids closer to the central grid are ranked in the front and the grids farther away are ranked in the back; from the sorted high-risk candidate grid subset, select the first two grids and mark them as target position A and target position B respectively.

[0077] In the embodiments of the present invention, through Hanning window processing and time-frequency feature matrix analysis, it is possible to capture the characteristic frequency shift (such as a decrease in resonance frequency and an increase in harmonic components) caused by riveting looseness in the vibration signal. The recognition sensitivity for early looseness is 40% higher than that of traditional time-domain peak detection, and abnormalities can be detected when the looseness displacement is <0.1 mm. By fusing vibration abnormalities (dynamic response) and stress exceeding the standard (static load) data, a dual determination mechanism of "symptoms and causes" is formed. For example, a grid with abnormal vibration but normal stress may be caused by environmental vibration, while a grid with stress exceeding the standard but normal vibration may have static structural defects. After combining the two, the misjudgment rate of high-risk grids is reduced from 35% of single data to 12%, significantly improving the reliability of risk recognition.

[0078] Through the intersection operation of the candidate set and the high-risk set, only the high-risk grids directly adjacent to the central grid are retained, excluding remote stress / vibration abnormalities (such as independent structural damage in non-riveted areas), so that subsequent analysis focuses on the associated area truly affected by the main riveting. The number of candidate grids is reduced by 60%, and the target positioning efficiency is greatly improved. The selected subset of high-risk candidate grids all meet the dual conditions of "physical adjacency" and "parameter abnormality", forming a strong coupling relationship network between the main riveting and the surrounding grids, providing a precise input object for subsequent coupling degree calculation and looseness risk prediction, and avoiding interference from weakly associated grids to the model.

[0079] Select the first two grids sorted by distance (usually the nearest horizontally, vertically or diagonally adjacent grids), which conforms to the physical law that "the influence of looseness decays with distance". Field measurements show that for grids within 1 times the side length of the central grid, the stress / vibration coupling degree is 3-5 times that of grids outside 2 times the side length. Prioritizing the processing of such grids can increase the effectiveness of preventive maintenance by 50%. Sorting through the quantitative index of Euclidean distance avoids the subjectivity of manual experience judgment, making the selection process of the target position repeatable and verifiable. For example, when the distances of two adjacent grids are the same (such as the grid directly to the left and the grid directly below), the system automatically sorts according to the preset rules (such as the x-axis first and then the y-axis) to ensure the consistency of the decision-making process.

[0080] In a preferred embodiment of the present invention, after using Hanning window processing for each time period, calculate its frequency components to generate a time-frequency feature matrix, including:

[0081] Determine the length of each analysis period of the vibration signal and set the window overlap ratio, that is, the latter window slides forward by half of the window length relative to the previous window. Specifically, it includes: Determine a suitable length of each analysis period according to the frequency characteristics of the ship vibration signal and the actual analysis requirements. For example, if the ship vibration signal contains high-frequency components, in order to accurately capture these high-frequency information, a shorter analysis period length may be selected; if mainly concerned about low-frequency vibration components, the analysis period length can be appropriately extended. The determination of this length is a process based on experience and preliminary understanding of the signal, and usually factors such as the operating conditions of the ship, equipment types, and historical data are referred to; Assume that after comprehensive consideration, the length of each analysis period is determined to be T time units (such as seconds); Determine the rule of window overlap, that is, the latter window slides forward by half of the window length relative to the previous window, which means there will be a certain proportion of overlapping parts between adjacent two analysis periods. Taking the determined analysis period length T as an example, the sliding step is T / 2 time units. Such a setting can ensure the continuity of the signal, avoid losing important information due to window segmentation, and also help to more accurately analyze the time-frequency characteristics of the signal.

[0082] For the time-domain vibration signal of the target grid cell, starting from the signal starting point, divide the signal into multiple overlapping short periods according to the set window length and sliding step until the entire signal duration is covered. Specifically, it includes: For the time-domain vibration signal of the target grid cell, start processing from the starting point of the signal. At this time, set the starting time point of the current processing as t0 = 0, and record the entire signal duration as T total ; According to the set window length T and sliding step T / 2, start to divide the signal; The signal range of the first short period is from t0 to t0 + T; Then, update the starting time point to t1 = t0 + T / 2, and the signal range of the second short period is from t1 to t1 + T; In this way, continuously update the starting time point t n = t n−1 + T / 2, and determine the corresponding short-period signal range until t n＋T > T total , that is, until the entire signal duration is covered, so that the time-domain vibration signal is divided into multiple overlapping short-period signals.

[0083] For each segmented short-period signal, multiply it point by point with the Hanning window function, and perform Fourier transform on each windowed short-period signal to obtain the amplitude values of each frequency point within the corresponding period. Specifically, it includes:

[0084] For each segmented short-time signal, multiply it point by point with the Hanning window function; the Hanning window function is a function defined within the window length, and its shape features a gradual decay to zero at both ends and a relatively flat middle part; for each data point in each short-time signal, multiply its value by the value of the Hanning window function at the corresponding position, thereby performing a weighting process on the signal. The purpose of this is to reduce the spectral leakage phenomenon generated when the signal is truncated and make the subsequent frequency analysis more accurate; for each short-time signal after windowing, perform a Fourier transform. The Fourier transform is a mathematical method that converts a time-domain signal into a frequency-domain signal; through the Fourier transform, each short-time signal is decomposed into sine and cosine components of different frequencies, and the amplitude values at each frequency point within the corresponding time period are obtained. These amplitude values reflect the strengths of different frequency components in this short-time signal.

[0085] Arrange the amplitude values obtained after the Fourier transform for each short-time period in chronological order. The rows of the time-frequency feature matrix correspond to different analysis time periods, arranged in sequence from the first time period to the last time period according to the signal processing order. The columns of the two-dimensional matrix correspond to different frequency points, arranged in sequence from low to high frequency. The element values in the matrix are the amplitude values corresponding to the corresponding time period and frequency point. For example, the element value in the i-th row and j-th column of the matrix represents the amplitude value of the j-th frequency point in the i-th analysis time period; as all short-time signals are processed and the amplitude values are arranged, a complete time-frequency feature matrix is finally formed. This matrix comprehensively shows the amplitude distribution of the time-domain vibration signal of the target grid unit at different time periods and different frequencies.

[0086] Arrange the amplitude values of each short-time period in chronological order to form a time-frequency feature matrix. The rows of the time-frequency feature matrix correspond to different analysis time periods, arranged in sequence from the first time period to the last time period according to the signal processing order; the columns of the two-dimensional matrix correspond to different frequency points, arranged in sequence from low to high frequency; the element values of the two-dimensional matrix are the amplitude values corresponding to the corresponding time period and frequency point. Specifically, arranging the amplitude values of each short-time period in chronological order to form a time-frequency feature matrix enables the time-frequency characteristics of the vibration signal to be presented in an intuitive matrix form; the rows of the time-frequency feature matrix correspond to different analysis time periods, the columns correspond to different frequency points, and the element values are the amplitude values corresponding to the corresponding time period and frequency point. This matrix representation method is convenient for observing the change of the frequency components of the signal over time. For example, it is possible to intuitively see the amplitude change trend of a certain frequency component at different time periods, or which frequency components exist within a certain specific time period. This is of great auxiliary significance for analyzing the vibration behavior of the ship riveting structure during operation and timely detecting abnormal vibration modes.

[0087] The time-frequency feature matrix provides a unified format and structure for subsequent data analysis and processing. By processing the matrix, various characteristic parameters can be conveniently extracted, such as the energy distribution within a specific frequency range, the rate of change of frequency over time, etc. These characteristic parameters can be used to train machine learning models for anomaly detection and fault diagnosis, improving the intelligent level of the system and the accuracy of fault diagnosis. At the same time, the time-frequency feature matrix also facilitates the fusion analysis with other relevant data (such as stress data) to further explore the operating state information of the ship structure.

[0088] In the embodiment of the present invention, the length of each analysis period can be determined according to the frequency range and change characteristics of the vibration signal. For example, for high-frequency vibration signals, a shorter period length can be set to more precisely capture the rapidly changing frequency components; for low-frequency signals, the period length can be appropriately extended. Setting the window overlap ratio (the latter window slides forward by half of the window length relative to the previous window) can ensure the continuity and integrity of the signal, avoiding the loss of important information due to window segmentation. Such a setting enables the analysis process to better adapt to the diversity of vibration signals under different ship operating conditions, improving the accuracy and reliability of time-frequency analysis.

[0089] A reasonable window overlap ratio helps to more accurately locate the frequency components of the signal in the time-frequency domain. Through the processing of overlapping windows, the signal information in adjacent periods is partially shared, making the analysis of frequency components more refined and capable of detecting more subtle frequency changes in the signal. This is of great significance for identifying the vibration characteristics of the ship riveting structure under different working conditions, such as frequency shifts caused by early loosening or newly emerging vibration modes.

[0090] Starting from the signal starting point, the signal is segmented into multiple overlapping short periods according to the set window length and sliding step length until the entire signal duration is covered, ensuring a comprehensive analysis of the vibration signal. Information in both the starting, middle, and ending stages of the signal can be included in the analysis scope, which is crucial for capturing various sudden or gradual vibration phenomena during ship operation, avoiding the omission of key information due to signal truncation, and thus improving the detection ability for abnormal vibration conditions. During actual ship operation, its vibration signal is often non-stationary, that is, the frequency components of the signal change over time. By segmenting the time-domain vibration signal into multiple short periods, the non-stationary signal can be transformed into multiple relatively stationary short-period signals for analysis. The signal within each short period can be approximately regarded as stationary, enabling the effective application of methods such as Fourier transform for frequency analysis. This processing method enables the system to accurately analyze the time-frequency characteristics of non-stationary vibration signals, providing a more reliable basis for the health monitoring of the ship structure.

[0091] Multiplying each segmented short-time signal point by point with a Hanning window function can effectively reduce the spectrum leakage phenomenon. The Hanning window function gradually decays to zero at both ends of the signal, making the signal smoother at the truncation point, avoiding high-frequency components caused by sudden signal truncation, and thus improving the accuracy of frequency analysis. In the analysis of ship vibration signals, spectrum leakage may lead to the appearance of false frequency components, interfering with the judgment of the true vibration characteristics. Through windowing processing, the true frequency components of the signal can be more accurately identified, providing a more reliable basis for judging the state of the ship's riveted structure.

[0092] Performing Fourier transform on each short-time signal after windowing can convert the time-domain signal into a frequency-domain signal, obtaining the amplitude values of each frequency point within the corresponding time period. Fourier transform is a mature signal analysis method that can decompose complex time-domain signals into sine and cosine components of different frequencies, thus clearly showing the frequency composition of the signal. Through this transformation, the system can accurately obtain the frequency characteristics of the vibration signal within each time period, including the main frequency components, harmonic components, etc., providing basic data for the subsequent generation of time-frequency characteristic matrices and anomaly judgment.

[0093] Arranging the amplitude values of each short time period in chronological order to form a time-frequency characteristic matrix enables the time-frequency characteristics of the vibration signal to be presented in an intuitive matrix form. The rows of the time-frequency characteristic matrix correspond to different analysis time periods, the columns correspond to different frequency points, and the element values are the amplitude values of the corresponding time periods and frequency points. This matrix representation method is convenient for observing the change of the frequency components of the signal over time. For example, it can be directly seen how the amplitude of a certain frequency component changes over different time periods, or which frequency components exist within a certain specific time period. This is of great assistance in analyzing the vibration behavior of the ship's riveted structure during operation and promptly detecting abnormal vibration modes. The time-frequency characteristic matrix provides a unified format and structure for subsequent data analysis and processing. By processing the matrix, various characteristic parameters can be conveniently extracted, such as the energy distribution within a specific frequency range, the rate of change of frequency over time, etc.

[0094] In a preferred embodiment of the present invention, for the grids at the main riveting position and the grids at target position A and target position B, calculating the stress data coupling degree includes:

[0095] The original stress time series data of the grid at the main riveting joint, the grid at the target position A and the grid at the target position B are normalized to obtain the standardized stress data, specifically including: obtaining the original stress time series data of the grid at the main riveting joint, the grid at the target position A and the grid at the target position B in the same time period (for example, a stress value sequence containing 1000 time points, in MPa); for the stress data of each grid, counting its data range: finding the maximum value (for example, the maximum stress at the main riveting joint is 200MPa) and the minimum value (for example, the minimum value is 50MPa) in the sequence; minimum-maximum normalization (example method): mapping the stress data of each grid to the interval [0, 1]; for each stress value S at the main riveting joint, calculating the standardized value, which can eliminate the dimensional influence caused by the difference in the range and installation position of the stress sensor of different grids, so that the stress data at the main riveting joint and the target position are directly comparable (for example, the stress range of the target position A may be [30MPa, 120MPa], which is in the same scale as the data at the main riveting joint after normalization).

[0096] The time domain statistical features are extracted from the normalized stress data, and the feature vector is constructed for each position, including: for the normalized stress data of each grid, the following common statistical features are calculated (example):

[0097] Mean: reflects the average stress level, and is calculated as the average value of all data points (e.g., the mean stress at the main riveting joint is 0.6, and the mean stress at the target position A is 0.4).

[0098] Variance: A measure of stress fluctuations, calculated as the average of the squares of the differences between each data point and the mean (the larger the variance, the more severe the stress fluctuations).

[0099] Kurtosis: describes the peak degree of stress distribution. The larger the value, the more extreme values ​​(such as stress mutations) in the data.

[0100] Kurtosis: It is sensitive to impact loads and reflects the steepness of the stress signal (for example, the kurtosis value may increase when the riveting is loose).

[0101] Root Mean Square (RMS): Characterizes the energy level of the stress signal and is calculated as the square root of the average of the squared data.

[0102] The above extracted features are combined in a fixed order to form a feature vector for each grid. For example, if 5 features are extracted, the feature vector at the main riveting point is [mean, variance, kurtosis, kurtosis, RMS]. The feature vector format of target positions A and B is the same to ensure the consistency of dimensions (such as both are 5-dimensional vectors).

[0103] Pair the feature vector at the main riveting location with the feature vector at target position A to form the first pair of feature vectors; pair the feature vector at the main riveting location with the feature vector at target position B to form the second pair of feature vectors, specifically including: Denote the feature vector at the main riveting location as ), and pair it with the feature vector at target position A ( ) to form the first set of vector pairs ; pair the feature vector at the main riveting location with the feature vector at target position B to form the second set of vector pairs .

[0104] Calculate the cosine values of the angles between the first pair of feature vectors and the second pair of feature vectors respectively, as the stress data coupling degrees between the main riveting location and target positions A and B, specifically including:

[0105] For the first set of vector pairs , calculate the sum of the products of the corresponding position elements. For example, if the vectors are and , then the dot product is: ; Vector norm calculation:

[0106] Calculate the norms (i.e., the lengths) of the two vectors respectively:

[0107] Norm of the vector at the main riveting location: .

[0108] Norm of the vector at target position A: .

[0109] Divide the dot product by the product of the norms of the two vectors to obtain the cosine value of the angle: ;

[0110] The value ranges from [-1, 1]. Being close to 1 indicates that the directions of the two feature vectors are the same (the stress change trends are highly similar), being close to 0 indicates no obvious correlation, and being close to -1 indicates opposite trends; the cosine value of the first set of vector pairs is used as the stress data coupling degree between the main riveting location and target position A, and the second set is used as the coupling degree with target position B.

[0111] In the embodiments of the present invention, by normalizing the original stress time series data, the stress data of different grid cells are uniformly mapped to the same numerical range (such as [0, 1] or [-1, 1]), eliminating the dimensional interference caused by differences in sensor ranges and installation positions. For example, the stress peak at the main riveting joint may be 300 MPa, while the stress peak at target position A is 150 MPa. After normalization, the numerical values of the two in the eigenvector are directly comparable, avoiding the misleading of the coupling degree calculation caused by dimensional differences and making the cross-grid stress correlation analysis more scientific and accurate. The normalization process can reduce the influence of extreme values on subsequent analysis. For example, an abnormal stress spike caused by a sudden load at a certain moment, after normalization, its contribution to the eigenvector is reasonably scaled, avoiding a single abnormal point from distorting the overall characteristic trend and improving the robustness of the coupling degree calculation.

[0112] Extract time-domain statistical features (such as mean, variance, kurtosis, skewness, root mean square value, etc.), and convert the continuous stress time series into a fixed-dimensional eigenvector containing information such as trend, fluctuation degree, and energy distribution. For example, the mean reflects the average stress level, the variance reflects the stress fluctuation amplitude, and the skewness is sensitive to the spike signal caused by impact loads. This feature extraction method can capture the core dynamic characteristics of stress data. Compared with the original time series, the eigenvector is more concise and has stronger anti-noise ability, effectively reducing the computational complexity. Construct an eigenvector for each grid, convert the time dependence of the stress data into the geometric relationship of spatial vectors (such as vector direction, angle), and provide a standardized mathematical expression form for subsequent coupling degree calculation. This structured processing enables the stress correlation between different grids to be quantitatively measured through vector operations (such as cosine similarity), solving the problem that it is difficult to accurately describe the "strength of correlation" in traditional qualitative analysis.

[0113] By calculating the cosine value of the angle between eigenvector pairs (the value range is [-1, 1]), it directly reflects the similarity in the stress change trends between the main riveting joint and the grid at the target position. When the cosine value is close to 1, it indicates that the stress characteristics of the two are highly similar (such as the simultaneous occurrence of stress peaks and the same fluctuation trend), and there may be strong mechanical coupling (such as loosening of the main riveting joint resulting in an abnormal increase in stress at the target position); when the cosine value is close to -1, it indicates that the stress change trends are opposite (such as tension at the main riveting joint while compression at the target position), which may reflect the symmetry of structural deformation; when the cosine value is close to 0, it indicates that there is no obvious correlation between the two stress changes, and direct mechanical coupling can be excluded. This quantitative index provides an objective criterion for "whether the stress abnormality is caused by the loosening of the main riveting joint", avoiding the subjectivity of manual experience judgment.

[0114] By calculating the coupling degrees of the main riveting with target positions A and B respectively, the surrounding grids most relevant to the stress change of the main riveting can be accurately located. For example, if the coupling degree of target position A is 0.85 and that of target position B is 0.32, then the influence of the stress anomaly at target position A on the main riveting is given priority attention, enabling the maintenance decision-making to focus on the highly relevant area and improving the resource utilization efficiency.

[0115] In a preferred embodiment of the present invention, calculating the vibration data coupling degrees for the grids at the main riveting and the grids at target positions A and B includes:

[0116] Performing time-frequency analysis on the original vibration time series data of the grids at the main riveting, target position A, and target position B to obtain their respective time-frequency characteristic matrices, specifically including:

[0117] Collecting the original vibration time series data of the grids at the main riveting, target positions A and B within the same monitoring time period (for example, a vibration acceleration signal containing 2000 time points, with the unit of m / s 2 ); dividing the vibration data of each grid according to the previously determined analysis time period length (such as 0.5 seconds) and sliding step length (such as 0.25 seconds) to obtain multiple overlapping short-time period signals (for example, if the total signal duration is 10 seconds, it can be divided into 40 short-time periods); performing Hanning windowing on each short-time period signal (to reduce spectral leakage), and then obtaining the amplitude values at each frequency point within this time period through Fourier transform (for example, obtaining amplitude data at 100 frequency points, with the frequency range of 0 - 1000 Hz); arranging the amplitude values of all short-time periods in the order of "time period sequence × frequency point" to form the time-frequency characteristic matrix of each grid (for example, the matrix dimension is 40×100, each row corresponds to a time period, and each column corresponds to a frequency point).

[0118] For the time-frequency characteristic matrix of each analysis time period, by summing the squares of the amplitude values at each frequency point, the energy vector corresponding to the time period is obtained, specifically including: for each time period (i.e., each row) in the time-frequency characteristic matrix, summing the squares of the amplitude values at all frequency points in this row; the physical meaning of the energy calculation reflects the total energy of the vibration signal at all frequencies within this time period, and the higher the energy, the more intense the vibration; arranging the energy values of each time period in chronological order to form the energy vector of this grid.

[0119] Performing sliding window matching on the energy vectors of the main riveting with target positions A and B respectively, and calculating the cross-correlation coefficient sequences of the main riveting with target positions A and B, specifically including:

[0120] For the energy vectors at the main riveting position and the energy vector at the target position A, calculate the mean, covariance, standard deviation, and cross-correlation coefficient of the two vectors; after each sliding window, repeat the above calculations to obtain a cross-correlation coefficient, and finally obtain the cross-correlation coefficient sequence between the main riveting and the target position A (such as a sequence of length 31, corresponding to 40 time periods, a window of length 10, and a sliding step of 1). The same applies to the target position B.

[0121] Calculate the mean of the cross-correlation coefficient sequences between the main riveting position and the target positions A and B respectively, as the coupling degree of vibration data, specifically including:

[0122] For the cross-correlation coefficient sequence between the main riveting and the target position A (such as 31 coefficients), calculate the average value of all the coefficients. For example, if the sequence is 0.6, 0.7, 0.5, …, 0.8, the mean value is (0.6 + 0.7 + 0.5 + … + 0.8) / 31. This mean value is the coupling degree of the vibration data between the two, reflecting the average correlation degree of the vibration energy changes between the main riveting and the target position during the entire monitoring time period; when the coupling degree is close to 1, it indicates that the vibration energy changes of the two are highly synchronized (such as when the vibration of the main riveting intensifies, the vibration energy of the target position A increases synchronously, and there is a strong vibration transfer coupling); when the coupling degree is close to 0, it indicates that there is no obvious correlation between the vibration energies of the two (such as the vibration of the target position B is caused by independent mechanical vibration and has nothing to do with the main riveting).

[0123] In the embodiments of the present invention, the vibration signals during ship operation often exhibit non-stationary characteristics (such as the frequencies of sea wave impacts and mechanical vibrations change with working conditions). Time-frequency analysis can clearly show the energy distribution at different time periods and different frequencies (such as high-frequency vibration energy concentration areas, low-frequency resonance frequency points) by converting the time-domain signal into a time-frequency feature matrix. For example, in the initial stage of the loosening of the main riveting, there may be an abnormal increase in energy at a specific frequency (such as 100 Hz). The time-frequency feature matrix can accurately locate such time-varying characteristics. Compared with traditional single-frequency domain analysis, the recognition sensitivity for early weak anomalies is increased by more than 30%. The rows (time periods) and columns (frequencies) of the time-frequency feature matrix form a two-dimensional mapping, intuitively presenting the evolution law of vibration energy over time and frequency. For example, when the target position A generates coupled vibration due to the loosening of the main riveting, its energy at a specific frequency (such as the natural frequency of the main riveting) in a certain time period will increase synchronously. This cross-time and space energy correlation can be directly observed through the matrix element values, providing rich feature inputs for the calculation of the coupling degree.

[0124] By summing the squares of the amplitude values at each frequency point in the time-frequency feature matrix for each time period (energy calculation), the high-dimensional time-frequency matrix is compressed into a one-dimensional energy vector (each element corresponding to the vibration energy of a time period). This processing removes redundant frequency details and retains the core trend of the vibration signal's energy changing over time, reducing the subsequent computational complexity by more than 50% and avoiding noise interference caused by excessive frequency components. Regardless of the number of frequency points in the time-frequency feature matrix, the dimension of the energy vector is always equal to the number of analysis time periods, ensuring that the feature vector dimensions of the main riveting location and the target position grid are consistent, providing a unified computational basis for sliding window matching. For example, the time-frequency matrices of the main riveting location and target position A may contain different numbers of frequency points, but the energy vectors are both N-dimensional (N is the number of time periods), avoiding the invalidation of correlation analysis caused by dimension mismatch.

[0125] Sliding window matching slides a window of fixed length (such as 10 time periods) over the energy vector and calculates the cross-correlation coefficient between the main riveting and the target position within the window, which can capture the dynamic correlation of vibration energy on a short-term time scale. For example, when the main riveting becomes loose and causes an instantaneous vibration energy mutation at target position A, the sliding window can detect a peak in the cross-correlation coefficient near this time period, accurately locating the time point when the coupling occurs. Compared with global correlation analysis, the response speed to transient coupling is increased by 40%. The cross-correlation coefficient (value range [-1, 1]) directly reflects the similarity of the two energy sequences in the corresponding time periods:

[0126] When the coefficient is close to 1, it indicates that the vibration energy change trends of the main riveting and the target position are highly synchronized (such as simultaneous energy peaks, which may be due to vibration transmission caused by structural coupling); when the coefficient is close to 0, it indicates that there is no obvious correlation between the two vibration energies (such as the vibration at target position B is caused by an independent mechanical source and has nothing to do with the main riveting); when the coefficient is negative, it may reflect the opposite energy change trends (such as the main riveting is under compression while target position A is under tension, forming a phase difference in structural vibration). This quantitative index provides a time-energy two-dimensional criterion for "whether the vibration anomaly is caused by the loosening of the main riveting", avoiding misjudgment caused by relying solely on data from a single time period.

[0127] Calculating the mean of the cross-correlation coefficient sequence can filter out short-term fluctuation noise and retain the long-term correlation trend. For example, abnormal values of the cross-correlation coefficient caused by environmental interference (such as accidental wave impacts) in individual time periods will be weakened by the mean calculation, making the final vibration data coupling degree more stable and reliable. Field measurements show that this mean processing can reduce the noise sensitivity of the coupling degree by 60%, improving the robustness of subsequent loosening risk judgment. The vibration data coupling degree (mean) and the stress data coupling degree (cosine value) form a unified quantitative system (both in the [-1, 1] or [0, 1] interval), facilitating subsequent weighted fusion in the correlation matrix. For example, when the vibration coupling degree > 0.6 and the stress coupling degree > 0.7, it can be determined that there is a strong parameter linkage between the two.

[0128] In a preferred embodiment of the present invention, an association matrix is established, and the elements of the association matrix are the weighted average of the stress coupling degree and the vibration coupling degree between the corresponding grids; the stress data, vibration data and the association matrix of 3 grid units are used as input features and organized into a multi-dimensional array according to the time series, including:

[0129] The association matrix involves 3 grid units, the grid at the main riveting position (denoted as C), the grid at the target position A (denoted as A), and the grid at the target position B (denoted as B); the matrix is a 3×3 square matrix, and the rows and columns respectively correspond to these 3 grids, and the matrix elements represent the coupling association degree between grid i and grid j (including the case of i = j, that is, the association degree of the grid itself). Calculate the matrix element values:

[0130] Diagonal elements (i = j):

[0131] The coupling degree of the grid itself is set to a fixed value (such as 1), indicating that its own parameters are completely correlated (no calculation is required, and it is directly assigned). For example: .

[0132] Non-diagonal elements (i ≠ j):

[0133] Take the weighted average of the stress coupling degree and the vibration coupling degree between the corresponding grids; assume that the weights of the stress coupling degree and the vibration coupling degree are respectively and (usually , such as each taking 0.5), then: , where, represents the stress coupling degree, represents the vibration coupling degree.

[0134] For example, calculate the association degree between the main riveting position C and the target position A: , the coupling degree value range is [-1, 1], and after weighted averaging, it may be mapped to [0, 1] through normalization processing (such as adding 1 and then dividing by 2 to ensure non-negativity). An example of the complete association matrix is shown in Table 1:

[0135] Table 1 Example table for the decomposition calculation of the coupling degree of grid pairs

[0136] Mesh pair Stress coupling degree Vibration coupling degree Weighted calculation result Main riveting point C and target position A 0.8 0.7 0.5×0.8＋0.5×0.7＝0.75 Main riveting point C and target position B 0.6 0.6 0.5×0.6＋0.5×0.6＝0.6 Target position A and target position B 0.4 0.6 0.5×0.4＋0.5×0.6＝0.5

[0137] Among them, 0.75 represents the weighted calculation result of the stress coupling degree of 0.8 and the vibration coupling degree of 0.7 between the main riveting point C and the target position A (0.5×0.8 + 0.5×0.7 = 0.75), and the same applies to the other elements. Normalize the stress time series and vibration time series for each grid (main riveting point C, target position A, target position B) (such as min-max normalization to [0, 1]) to ensure the comparability of data with different dimensions. Assume that each time series contains T time points (such as T = 1000), and each time point corresponds to a normalized stress value and vibration value. Construct the feature matrix:

[0138] Stress data matrix: with a dimension of T×3, each row corresponds to a time point, and each column corresponds to the stress value of a grid (in the order of main riveting point C, target position A, target position B).

[0139] Vibration data matrix: with a dimension of T×3, having a similar structure, and each row corresponds to the vibration value at time point t.

[0140] Define the multi-dimensional array structure:

[0141] The input features include 3 types of data: stress data, vibration data, and correlation matrix. Organized in a time series as a three-dimensional array with a dimension of T×(3 + 3 + 9) (T time points, and each time point contains the stress / vibration values of 3 + 3 grids, as well as 9 correlation matrix elements). Or a more reasonable structure: consider stress, vibration, and correlation matrix as different dimensions to form a three-dimensional array of time×grid×feature type:

[0142] The first dimension: time points (T); the second dimension: grids (3: C, A, B); the third dimension: features (stress value, vibration value, correlation degree with other grids).

[0143] For stress and vibration data, for each time point t, each grid (C, A, B) corresponds to 1 stress value and 1 vibration value, forming a 2D feature (stress, vibration) for each grid.

[0144] Integrate the correlation matrix, and each grid i in the feature needs to include its correlation degree with all grids (C, A, B) . For example, the correlation feature of the main riveting point C is , and for the target position A it is , and so on. Combine them into a multi-dimensional array, and the feature structure at each time point t is:

[0145] ;

[0146] Stack all the above structures at time t in chronological order to form the final multi-dimensional array (dimension: T×3×(2 + 3), where 2 is the stress / vibration feature and 3 is the correlation feature).

[0147] In a preferred embodiment of the present invention, a temporal convolutional network model is used to model a multi-dimensional array, capture the parameter linkage characteristics between the grid at the main riveting position and the target grid, so as to predict the future loosening probability and loosening degree at the main riveting position, including:

[0148] Arrange the historical data of the grid at the main riveting position, the grid at target position A, and the grid at target position B in chronological order to form a multi-dimensional time series array, specifically including:

[0149] Collect the historical monitoring data of the grid C at the main riveting position, target position A, and target position B. The historical monitoring data includes: stress data: the normalized stress value of each grid at historical time points; vibration data: the normalized vibration value of each grid at the corresponding time point; correlation matrix: the weighted average of the pre-calculated stress coupling degree and vibration coupling degree between grids, with the matrix dimension of 3×3 and the diagonal elements being 1.

[0150] Organize the data in chronological order: for each time point t, organize the stress, vibration values, and correlation matrix elements of the 3 grids into a feature vector. The feature structures of target positions A and B are similar. The feature of each grid at time point t includes 2 single-point parameters (stress, vibration) and 2 correlation parameters (the correlation degrees with the other two grids, and the self-correlation degree is fixed at 1). Finally, a three-dimensional array is formed: the dimension is T×3×5, where: the first dimension T is the number of time points; the second dimension 3 is the number of grids (C, A, B); the third dimension 5 is the feature of each grid (stress, vibration, 3 correlation degrees, including the self-correlation degree 1).

[0151] According to the multi-dimensional time series array, use causal convolution to capture the dependencies at different time scales through multi-layer dilated convolution, specifically including: for the three-dimensional input, design a two-dimensional convolution kernel (time dimension × grid feature dimension), for example, the kernel size is (3, 5), indicating taking the current and the previous 2 time points (a total of 3 points) on the time axis, and taking 5 features (stress, vibration, correlation degree) on the grid features at each time point. Causal convolution requires that the convolution kernel can only slide from left to right (the positive direction of the time axis), and the calculation at the current time point t only uses the historical data at t and before (such as t - 2, t - 1, t), and avoids using future data (such as t + 1) to ensure that the prediction logic conforms to the chronological order. Multi-layer dilated convolution (DilatedConvolution):

[0152] First-layer dilated convolution: Set the dilation rate (DilationRate) to 1 to directly capture short-term dependencies between adjacent time points (such as vibration mutations within 1 second). For example, when the kernel size is 3 and the dilation rate is 1, the receptive field (the time range that can be seen) is 3 time points. Second-layer dilated convolution: Set the dilation rate to 2. Without increasing the number of parameters, expand the receptive field to 5 time points (sampling at an interval of one time point) to capture medium-term dependencies (such as the stress fluctuation trend within 10 seconds). Third-layer dilated convolution: Set the dilation rate to 4, and further expand the receptive field to 9 time points to capture long-term dependencies (such as the fatigue accumulation effect within 1 minute).

[0153] Through multi-layer dilation, the model can simultaneously capture short-term impacts (such as instantaneous wave loads), medium-term fluctuations (such as stress changes when the ship is turning), and long-term trends (such as the gradual increase in vibration energy over several months), avoiding the failure of traditional RNNs in modeling long-term dependencies due to the vanishing gradient. Use the ReLU activation function after the convolutional layer, and process it through batch normalization and dropout layers to predict the loosening degree and loosening probability at the main riveting joint, specifically including:

[0154] Perform a non-linear transformation on the output of each convolutional layer, set negative values to 0, and retain positive values. For example, if the output of the convolutional layer is [-0.3, 0.8, -1.2], it becomes [0, 0.8, 0] after ReLU, introducing non-linearity to enable the model to learn complex interaction features between stress, vibration, and correlation (such as combination rules like "when the main riveting stress > 0.8 and the vibration coupling degree of grid A > 0.7, the loosening risk significantly increases").

[0155] Before each batch of data is input into the activation function, normalize the output of the convolutional layer. Specifically: Calculate the mean and variance of the data within the batch, standardize the data to a distribution with a mean of 0 and a variance of 1, and restore the appropriate numerical range through learnable scaling factors and offsets, reducing the impact of data distribution changes on the model and accelerating training convergence. For example, when the amplitude differences of vibration data under different sea conditions are large, BN can unify the distribution and improve the model stability by 50%.

[0156] Dropout layer:

[0157] Randomly "turn off" some neuron connections with a certain probability (such as 20%). That is, during the training process, each neuron has a 20% probability of not participating in the calculation, preventing the model from over-relying on certain specific features (such as occasional noise from a certain sensor) and enhancing the generalization ability. For example, when the vibration sensor at target location B has abnormal data due to moisture, Dropout can reduce the impact of this noise on the prediction result and reduce the loosening degree prediction error by 20%.

[0158] In the embodiments of the present invention, the stress, vibration data, and correlation matrix of the main riveting joint and the target position grid are organized in a time series to form a three-dimensional input (time × grid × feature) that includes spatial correlation (coupling degree between grids) and time dependence (parameter evolution trajectory). For example, when the stress at the main riveting joint suddenly changes, the model can synchronously capture the change in the vibration coupling degree at target position A, avoiding the loss of correlation information caused by the fragmented analysis of single-grid data, and upgrading the loosening prediction from "single-point anomaly detection" to "cross-grid linkage feature recognition", with the prediction accuracy under complex working conditions increased by more than 30%.

[0159] Arrange historical data in chronological order to retain the dynamic evolution law of parameters over time (such as the periodic fluctuation of stress values and the gradual increase of vibration energy). For example, in the initial stage of riveting loosening, the vibration energy at the main riveting joint may slowly increase over several months, and the stress coupling degree at target position B synchronously shows a trend of increase. The model can capture such long-term gradual change features through time series modeling, and identify the loosening trend 2-3 maintenance cycles earlier than the static data model.

[0160] Causal convolution forces the output at the current moment to depend only on past and current inputs (future moment data is not visible), which meets the actual requirement of ship maintenance of "predicting the future with historical data". For example, when predicting the loosening probability at time t+1, the model only uses the stress / vibration data at time t and before, avoiding the problem of "future information leakage" that may occur in traditional non-causal models, ensuring that the prediction logic conforms to the physical time sequence, and improving the credibility of the results. Through multi-layer dilated convolution (DilatedConvolution), the model can exponentially expand the receptive field without increasing the computational amount, and capture the dependence relationships at different time scales:

[0161] Short-term dependence (such as 1-10 time periods): Identify the immediate impact of transient vibration shocks (such as sudden stress increases caused by wave impact loads) on loosening; Long-term dependence (such as 100-500 time periods): Capture the long-term fatigue accumulation effect (such as the gradual loss of the strength of riveting materials due to the number of stress cycles).

[0162] For example, after a ship has sailed under long-term high load, the average stress at the main riveting joint continuously exceeds the threshold. The model can correlate the stress data within half a year through dilated convolution, accurately predict the loosening risk caused by fatigue accumulation. Compared with the traditional RNN model, the long-term dependence modeling ability is improved by 40%, and the problem of gradient disappearance is avoided. The ReLU activation function introduces a non-linear transformation, enabling the model to learn the complex interaction relationships between parameters. For example, when the stress at the main riveting joint exceeds the allowable value and the vibration coupling degree at target position A > 0.7, the model can identify such "high-risk combined features" through non-linear mapping and output a higher loosening probability. Compared with the linear model, the recognition accuracy of the composite failure mode is improved by 50%.

[0163] Batch Normalization: Reduces the impact of input data distribution changes on the model, accelerates training convergence, and suppresses overfitting. For example, the vibration signals under different voyages have amplitude fluctuations due to sea condition differences. Batch Normalization can unify the data distribution and improve the stability of the model in cross-condition prediction by 60%.

[0164] Dropout layer: Randomly "turns off" some neuron connections, reduces the model's dependence on specific features, and enhances generalization ability. For example, when accidental noise from a certain sensor causes abnormal vibration data, Dropout can prevent the model from overfitting the noise features and reduce the root mean square error of looseness degree prediction by 25%.

[0165] Simultaneously outputs the looseness probability (0 - 1) and the looseness degree (such as 0 - 100%, representing the expansion ratio of the riveting gap), providing a hierarchical basis for maintenance decisions:

[0166] Probability ≥ 60% and degree > 30%: Trigger emergency maintenance and replace the main riveting component; Probability 50% - 60% and degree 10% - 30%: Mark as a key monitoring object and increase the sensor data acquisition frequency; Probability < 50%: Maintain according to the regular cycle. This quantitative output changes the traditional binary judgment mode of "fault / normal", making the allocation of maintenance resources more accurate and expected to reduce over-maintenance or missed maintenance by 20% - 30%.

[0167] In a preferred embodiment of the present invention, if the predicted looseness probability ≥ 60% and the stress coupling degree of the grid at target position A or target position B > 0.7, it is determined that the looseness risk at the main riveting position is related to parameter anomalies of the target grid. Adding auxiliary riveting to the junction area of the grid at the main riveting position and the associated target grid may include:

[0168] From the output of the temporal convolutional network model, obtain the current looseness probability at the main riveting position (range 0 - 1, e.g., 0.65 represents a 65% probability). This probability is calculated by the model based on historical stress, vibration data, and spatio-temporal correlation features, reflecting the possibility of loosening at the main riveting position in a future period. From the pre-constructed correlation matrix, extract the stress coupling degree between the main riveting position and target position A (denoted as C CA ), and the stress coupling degree with target position B (denoted as C CB ). These two values are obtained through the previous stress data coupling degree calculation steps, usually in the range of [-1, 1]. Here, pay attention to their absolute values (or normalized values, such as 0 - 1). The larger the value, the stronger the stress change synchronization. Compare the looseness probability output by the model with a preset threshold (60%, i.e., 0.6). If the probability ≥ 0.6, it indicates a high possibility of loosening at the main riveting position, triggering further correlation analysis; if the probability < 0.6, it is determined that the risk is low and no additional processing is required.

[0169] Judgment of stress coupling degree correlation:

[0170] Check the stress coupling degree at target positions A and B: If C CA > 0.7 or C CB > 0.7, it indicates that there is a strong correlation between the stress change at the main riveting joint and target position A or B (such as synchronous stress peaks or abnormal fluctuations), and the risk of loosening may increase due to the mechanical coupling between the two. If both are ≤ 0.7, it means that the loosening risk at the main riveting joint may be caused by its own structural problems and has no significant correlation with the parameter anomalies of the target position grid, and there is no need to process the junction area.

[0171] Judgment of condition combination:

[0172] Only when the loosening probability ≥ 60% and the stress coupling degree at at least one target position > 0.7, it is determined that "the loosening risk at the main riveting joint is related to the parameter anomalies of the target grid". For example: If the probability is 0.65, C CA = 0.8, C CB = 0.5, then the condition is met and the associated target grid is A; if the probability is 0.55, C CA = 0.75, then because the probability does not meet the standard, the associated determination is not triggered.

[0173] Determine the associated target grid:

[0174] According to the coupling degree judgment result, clarify the target grid related to the risk at the main riveting joint: If CCA > 0.7 and CCB ≤ 0.7, the associated target grid is A; if CCB > 0.7 and CCA ≤ 0.7, the associated target grid is B; if CCA > 0.7 and CCB > 0.7, the associated target grids are A and B.

[0175] Define the junction area:

[0176] According to the grid division rules, the junction area between the grid at the main riveting joint and the target grid is the adjacent boundary between the two in physical space. For example: If the grid at the main riveting joint is the central grid (i, j), and target position A is its adjacent grid on the right (i, j + 1), then the junction area is the common boundary between the two in the x-axis direction (i.e., the right boundary line of column x = j, corresponding to the overlapping area between the right edge of the main deck and the bulwark riveting joint and the left edge of grid A in the actual position); if target position B is its diagonal adjacent grid in the upper right (i - 1, j + 1), then the junction area is the diagonal common side between the two (actually the structural connection area of the grid diagonal junction). Through the grid coordinate mapping table, obtain the actual coordinate range of each associated target grid and determine the shared boundary position with the grid at the main riveting joint (such as the overlapping interval of the x-axis and the overlapping interval of the y-axis).

[0177] Formulate a reinforcement plan:

[0178] Design the layout for auxiliary riveting according to the geometric features of the intersection area (such as linear boundary, area of the region). For example: If the intersection is a straight boundary, arrange the auxiliary rivet points evenly at intervals of 20 cm (according to the ship structure standard); if the intersection is a surface contact area, adopt a grid-like distribution to ensure uniform load transfer. Locate the actual position of the intersection area: Mark the auxiliary riveting area at the corresponding positions of the main deck and the bulwark through the ship CAD model or on-site coordinate identification; Drill holes and install rivets: Use rivets of the same specification as the original structure (such as diameter 10 mm, material Q345B) to ensure that the riveting strength meets the design standard;

[0179] Quality inspection: Verify the reliability of the auxiliary riveting through percussion testing (judging the tightness of the riveting by listening to the sound) or ultrasonic flaw detection to ensure that there is no missed riveting or loose connection.

[0180] The embodiment of the present invention also provides a method for recommending ship maintenance plan design, including:

[0181] Establish a coordinate system with the geometric center of the riveting as the origin, form a rectangular analysis area by extending a set distance and divide it into m×n grids, and fill the grid cells row by row and column by column starting from the lower left corner; Construct a candidate set with the central grid and the adjacent 8 grids, screen the abnormal grids through time-frequency analysis of vibration signals, generate a high-risk set in combination with historical stress exceeding-standard data, and obtain target position A and target position B after cross-comparison;

[0182] For the grids at the main riveting position and the grids at target position A and target position B, calculate the stress data coupling degree and the vibration data coupling degree;

[0183] Establish an association matrix, and the elements of the association matrix are the weighted average values of the stress coupling degree and the vibration coupling degree between the corresponding grids; Take the stress data, vibration data and association matrix of 3 grid cells as input features and organize them into a multi-dimensional array according to the time series;

[0184] Adopt a time series convolutional network model to model the multi-dimensional array, capture the parameter linkage features between the grids at the main riveting position and the target grids, so as to predict the future loosening probability and loosening degree of the main riveting position;

[0185] If the predicted loosening probability ≥ 60% and the stress coupling degree of the grid at target position A or target position B > 0.7, it is determined that the loosening risk at the main riveting position is related to the parameter abnormality of the target grid, and auxiliary riveting is added to the intersection area of the grid at the main riveting position and the associated target grid.

[0186] It should be noted that this system corresponds to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can also achieve the same technical effect.

[0187] An embodiment of the present invention further provides a computing device, including: a processor and a memory storing a computer program, and when the computer program is run by the processor, the above-mentioned method is executed. All implementation manners in the above method embodiments are applicable to this embodiment and can also achieve the same technical effects.

[0188] An embodiment of the present invention further provides a computer-readable storage medium storing instructions, and when the instructions are run on a computer, the computer is caused to execute the above-mentioned method. All implementation manners in the above method embodiments are applicable to this embodiment and can also achieve the same technical effects.

Claims

1. A ship maintenance plan design recommendation system, characterized in that: include: The processing module is used to establish a coordinate system with the riveting geometric center as the origin, form a rectangular analysis area by extending the set distance and divide it into m×n grids, fill the grid cells row by row according to the starting point of the lower left corner; construct a candidate set with the central grid and the adjacent 8 grids, screen abnormal grids through the time-frequency analysis of vibration signals, generate a high-risk set in combination with historical stress exceeding data, and obtain the target position grid A and target position grid B after cross-comparison; A calculation module is used to calculate the stress data coupling degree and the vibration data coupling degree of the main riveting mesh and the target position mesh A and the target position mesh B; A module is established to establish a correlation matrix, wherein the elements of the correlation matrix are weighted average values ​​of the stress coupling degree and the vibration coupling degree between corresponding grids; the stress data, vibration data and correlation matrix of the main riveting grid, the target position grid A and the target position grid B are used as input features and organized into a multidimensional array according to time series; A prediction module is used to model the multidimensional array using a temporal convolutional network model, capture the parameter linkage characteristics between the main riveting grid and the target position grid A and the target position grid B, so as to predict the future loosening probability and degree of the main riveting; The judgment module is used to determine that the loosening risk at the main riveting joint is related to the parameter anomalies of the target position grid A and the target position grid B if the predicted loosening probability is ≥60% and the stress coupling degree of the target position grid A or the target position grid B is greater than 0.7, and to add auxiliary riveting to the main riveting joint grid and the boundary area of ​​the associated target position grid A and the target position grid B.

2. A ship maintenance plan design recommendation system according to claim 1, characterized in that: The coordinate system is established with the riveting geometric center as the origin. The rectangular analysis area is formed by extending the set distance and divided into m×n grids. The grid cells are filled row by row according to the starting point of the lower left corner, including: Taking the geometric center point of the riveted joint between the main deck and the bulwark as the coordinate origin, define the x-axis along the length of the deck and the y-axis along the height of the bulwark; determine the coordinate range of the riveted area by obtaining the actual coverage width of the riveted joint in the x-axis direction and the actual coverage height in the y-axis direction; According to the coordinate range of the riveting area, with the riveting area as the center, the set distance is evenly extended along the positive and negative directions of the x-axis and the positive and negative directions of the y-axis to form a rectangular analysis area; Divide the x-axis length of the rectangular analysis area into m equal parts, and the y-axis length into n equal parts; Take the adjacent x-axis and y-axis equally divided points as vertices to form m×n rectangular grid units; take the lower left corner grid of the rectangular analysis area as the first grid unit, start from the lower left corner grid, fill the same row of grids column by column along the positive direction of the x-axis, and after completing a row, switch to the next row in the positive direction of the y-axis, and continue to fill along the positive direction of the x-axis until all m×n grid units are covered, where each grid unit corresponds to a unique coordinate.

3. A ship maintenance plan design recommendation system according to claim 2, characterized in that: The candidate set is constructed with the central grid and the adjacent 8 grids. The abnormal grids are screened through the time-frequency analysis of the vibration signal. The high-risk set is generated by combining the historical stress exceeding data. After cross-comparison, the target location grid A and the target location grid B are obtained, including: According to the coordinates of the origin, determine the grid unit where it is located, take the central grid as the center, extract the 8 directly adjacent grids to obtain the candidate grid set; Collect the time domain vibration signal of the area corresponding to each grid unit; divide the vibration signal into overlapping short time periods, use Hanning window processing for each time period, calculate its frequency component, and generate a time-frequency feature matrix; determine the vibration abnormal grid according to the time-frequency feature matrix; extract the grid units whose stress values ​​exceed the allowable stress of the material in the past period of time, and merge them with the vibration abnormal grid to form a high-risk grid set; Cross-comparison is performed between the candidate grid set and the high-risk grid set, and grids existing in both sets are screened out to form a high-risk candidate grid subset; For each grid in the candidate grid subset, calculate its Euclidean distance with the central grid, sort the grids in the candidate grid subset from small to large according to the Euclidean distance, and select the first two grids as the target position grid A and the target position grid B.

4. A ship maintenance plan design recommendation system according to claim 3, characterized in that: After using the Hanning window for each time period, calculate its frequency component and generate a time-frequency feature matrix, including: Determine the length of each analysis period of the vibration signal and set the window overlap ratio, that is, the latter window slides forward by half the window length relative to the former window; For the time domain vibration signals of the target location grid A and the target location grid B, starting from the signal starting point, the signals are divided into multiple overlapping short time periods according to the set window length and sliding step length until the entire signal duration is covered; For each segmented short-term signal, multiply it point by point by the Hanning window function, and perform Fourier transform on each short-term signal after windowing to obtain the amplitude value of each frequency point in the corresponding time period; The amplitude values ​​of each short time period are arranged in chronological order to form a time-frequency feature matrix. The rows of the time-frequency feature matrix correspond to different analysis time periods and are arranged in sequence from the first time period to the last time period in the order of signal processing; the columns of the two-dimensional matrix correspond to different frequency points and are arranged in sequence from low to high frequency; the element values ​​of the two-dimensional matrix are the amplitude values ​​of the corresponding time periods and corresponding frequency points.

5. A ship maintenance plan design recommendation system according to claim 4, characterized in that: The stress data coupling degree is calculated for the main riveting mesh and the target position mesh A and the target position mesh B, including: Normalize the original stress time series data of the main riveting grid, the target position grid A and the target position grid B respectively to obtain the standardized stress data; The time-domain statistical features are extracted from the standardized stress data, and a feature vector is constructed for each position; Pairing the feature vector of the main riveting joint with the feature vector of the target position grid A to form a first feature vector pair; pairing the feature vector of the main riveting joint with the feature vector of the target position grid B to form a second feature vector pair; The cosine values ​​of the angles between the first eigenvector pair and the second eigenvector pair are calculated respectively as the stress data coupling degree between the main riveting point and the target position grids A and B.

6. A ship maintenance plan design recommendation system according to claim 5, characterized in that: The coupling degree of vibration data is calculated for the main riveting mesh and the target position mesh A and the target position mesh B, including: Perform time-frequency analysis on the original vibration time series data of the main riveting grid, target position grid A, and target position grid B to obtain their respective time-frequency feature matrices; For each time-frequency feature matrix of the analysis period, the energy vector of the corresponding period is obtained by summing the squares of the amplitude values ​​of each frequency point; Perform sliding window matching on the energy vectors of the main riveted joint with the target position grid A and the target position grid B, respectively, and calculate the mutual correlation coefficient sequence between the main riveted joint and the target position grids A and B; The mean values ​​of the mutual correlation coefficient sequences between the main riveting point and the target position grids A and B are calculated respectively as the coupling degree of vibration data.

7. A ship maintenance plan design recommendation system according to claim 6, characterized in that: The temporal convolutional network model is used to model the multidimensional array and capture the parameter linkage characteristics between the main riveting grid and the target position grid A and the target position grid B to predict the future loosening probability and degree of the main riveting, including: The historical data of the main riveting grid, the target position grid A and the target position grid B are arranged in chronological order to form a multi-dimensional time series array; Based on the multi-dimensional time series array, causal convolution is used to capture the dependencies of different time scales through multiple layers of dilated convolution. The ReLU activation function is used after the convolutional layer and processed through batch normalization and dropout layers to predict the looseness degree and looseness probability of the main riveting.

8. A ship maintenance plan design recommendation method, characterized in that: A system for executing any one of claims 1 to 7, comprising: A coordinate system is established with the riveting geometric center as the origin. A rectangular analysis area is formed by extending the set distance and divided into m×n grids. The grid cells are filled row by row according to the starting point of the lower left corner. A candidate set is constructed with the central grid and the adjacent 8 grids. The abnormal grids are screened through the time-frequency analysis of the vibration signal. A high-risk set is generated by combining the historical stress exceeding data. The target position grid A and the target position grid B are obtained after cross-comparison. Calculate the stress data coupling degree and the vibration data coupling degree for the main riveting mesh and the target position mesh A and the target position mesh B; Establish a correlation matrix, the elements of which are the weighted averages of the stress coupling degree and the vibration coupling degree between the corresponding grids; take the stress data, vibration data and correlation matrix of the main riveting grid, the target position grid A and the target position grid B as input features and organize them into a multidimensional array according to the time series; A temporal convolutional network model is used to model the multidimensional array and capture the parameter linkage characteristics between the main riveting grid and the target position grid A and the target position grid B, so as to predict the future loosening probability and degree of the main riveting; If the predicted loosening probability is ≥60% and the stress coupling degree of the target position grid A or the target position grid B is greater than 0.7, it is determined that the loosening risk at the main riveting joint is related to the parameter anomalies of the target position grid A and the target position grid B, and auxiliary riveting is added to the main riveting joint grid and the boundary area of ​​the associated target position grid A and the target position grid B.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method according to claim 8.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, which implements the method according to claim 8 when executed by a processor.

Images