A projection-based visualization method for abnormal component inspection

Abstract

The present invention discloses a projection-based visualization method for abnormal component inspection. First, the aviation gas turbofan engine data is preprocessed, including key steps such as clustering, dimensionality reduction, and anomaly detection; then, visual mapping is performed using visual channels such as position, transparency, and color according to data attributes. A component projection view is designed, which uses scatter plots and contour plots to present an overview of the distribution of training data, and a special glyph is designed for abnormal component instances in the view to find similar patterns of abnormal data; a data inspection view is designed, and the timing working mode of the component under working conditions is encoded with five colors and the connecting lines therebetween in the upper part of the view, and the real data of the monitoring parameters is displayed in the lower part of the view with an area chart. The present invention can assist predictive maintenance personnel to intuitively and effectively discover anomalies in the data, and analyze the causes and patterns of the anomalies, thereby guiding the deployment of subsequent component work to extend the service life of the component.

Description

Technical Field

[0001] The invention relates to the technical field of information visualization and visual analysis, and in particular to a projection-based abnormal component inspection visualization method. Background Art

[0002] With the advent of Industry 4.0, the current modern industrial system has shown a new trend of large-scale, automated and complex development. Due to the increase in system complexity, the failure of key equipment will bring huge economic losses and even cause casualties. Taking aircraft as an example, due to the uncertainty of the working environment, the engine will have problems such as inconsistent exhaust temperature or rotor speed, surge, etc., that is, the subsystem or component is in an abnormal state. How to find abnormal components from massive data and analyze their abnormal patterns and causes is the focus of predictive maintenance personnel.

[0003] In existing aviation systems, the abnormal component inspection method is often run independently as a black box and integrated into an alarm system. Analysts can only obtain the results of model calculations, and it is difficult to conduct abnormal discovery and attribution analysis on instance-level data. In recent years, in the field of artificial intelligence and data mining, although many anomaly detection methods have been proposed, they mostly focus on improving the accuracy of the model and lack consideration of the causes of the anomaly from the perspective of data. In addition, statistical analysis methods have the disadvantages of poor interpretability of results and complex analysis methods.

[0004] The emergence of information visualization and visual analysis methods has provided predictive maintenance personnel with new solutions. Analysts can use the intuitiveness of visualization and the flexible interaction between views to explore the laws and patterns behind the data. Information visualization methods convert data into perceptible graphics, symbols, colors, etc. to enhance data recognition efficiency and convey effective information. At the same time, visual analysis methods can combine human wisdom and machine intelligence, use interactive visualization interfaces to introduce human knowledge and experience into the analysis process, and assist analysts in comprehensively and intuitively completing data analysis and reasoning decisions.

[0005] Visual analysis provides effective help for data insight in anomaly discovery. In the specific practice of abnormal component identification, although existing visual analysis can identify abnormal data and issue timely alarms, it only considers the comparative analysis of abnormal and normal data, lacks the exploration and attribution of similar patterns of abnormal data, while the study of similar patterns is of reference value for similar fault diagnosis, abnormal component identification, and high-quality data set construction. Summary of the invention

[0006] In response to the above problems, this study intends to provide a projection-based abnormal component inspection visualization method, which can not only help predictive maintenance personnel quickly find abnormal components in the data, but also support the discovery of abnormal patterns of component data instances and abnormal attribution; it can also guide the deployment of subsequent component work to extend the service life of the component. The specific technical solution is as follows:

[0007] A projection-based abnormal component inspection visualization method comprises the following steps:

[0008] S1: Data acquisition and processing

[0009] After obtaining the aviation gas turbofan engine data, the data is cleaned and processed, and organized into structured data that can be directly used in a visual view;

[0010] S2: Visual Mapping

[0011] Visually map the data obtained in step S1 through the visual channel:

[0012] Design component projection views, presenting an overview of data distribution using scatter plots and contour plots, with proximity indicating similarity in component degradation patterns; using a special glyph of radar charts nested in multiple layers of rings to compare similar patterns in abnormal data;

[0013] Design data inspection views, use horizontal bar charts to map the number of times components operate under different working conditions; use area charts to display the real data of monitored variables, and stack the area charts vertically to help analysts discover the relationship between the variables;

[0014] S3: Visualization layout and implementation

[0015] Visualize and layout the visual modules that complete the mapping in S2:

[0016] In the component projection view, the component is mapped to each point in the scatter plot, and the glyph retains the relative position between the scatter points. For the glyph, there is an improved radar chart in the middle to record the operation of the component, and multiple superimposed rings on the outside.

[0017] In the data inspection view, the timeline layout method is adopted. The point-line graph at the top of the view shows the operating sequence running in the current time period, and the timeline module with an area graph is at the bottom of the view;

[0018] S4: Interaction Design

[0019] In the scatter plot view, hovering the mouse over a data point will highlight the point, allowing users to zoom in and out to view the local area of ​​the scatter plot in detail; by drawing a closed irregular shape on the scatter plot, the scatter points are replaced with a glyph for in-depth exploration of the data; the slider controls the length of the time slice encoded in each circle in the glyph.

[0020] Furthermore, in step S1, data processing includes:

[0021] (1) Working condition identification:

[0022] The working conditions are divided according to three operating conditions: altitude, Mach number, and sea level temperature, and the working conditions are clustered using k-means.

[0023] (2) Data Dimensionality Reduction:

[0024] The components are represented by multi-dimensional time series data, that is:

[0025] X k =[t1,t2…,t i ,…,t T ],k=1,2,…,n,t i ∈R M

[0026] Where: n is the number of components; M is the number of selected features; T is the length of the life cycle; t i is the value of each feature at time i; k is the component number;

[0027] Aggregate the data using a sliding time window and get:

[0028]

[0029] Where: w is the time window size; x i is the value of each feature of window w_i after aggregation;

[0030] For the obtained x w_i Using t-distributed random neighborhood embedding, x w_i The vectors are projected into one-dimensional space, and the components are represented as:

[0031] X k ∈R T-w+1 ,k=1,2,…,n

[0032] For the degradation curve obtained after dimensionality reduction, t-SNE is used to reduce the dimension to two-dimensional space, and X k It is expressed as:

[0033] X k ∈R 2 ,k=1,2,…,n

[0034] Where: n is the number of components;

[0035] (3) Calculation of outliers:

[0036] The Pearson coefficient is used to define anomalies. i and k j The two time series of the sth monitoring variable of the component Its relevance is defined as follows:

[0037]

[0038]

[0039] Where: For number k i The value of the sth monitoring variable of the component at time point t; is component k j The value of the sth monitoring variable at time point t; TS is the length of the current time interval; for The average value in the time interval TS;

[0040] The outliers are defined as:

[0041]

[0042] Where: n is the number of components.

[0043] Furthermore, in step S2, the visualization mapping is specifically:

[0044] S21: Map the shape, position and color of the scatter plot in the projection view of the component: use the position of the scatter plot to map the dimension reduction result of the component; use the color to map the clustering result of the component; and use the transparency of the point to map the life cycle length of the component.

[0045] S22: Perform color and area mapping on the contour map in the component projection view: use specific colors to fill the blank spaces between the contour lines; use the depth of the filling color to map the density of the projected components in the local area; use the area of ​​the petals to map the number of behaviors;

[0046] S23: Shape, size and color mapping of the improved radar chart with special glyphs in the component projection view: Use the color of the points to map different working condition categories; Use the size of the center circle to map the life cycle length of the component; Use the shape of the radar chart to encode the number of operations of the component under six working conditions, and compare the operation of the component during the degradation period by comparing the shapes of the hexagons;

[0047] S24: Map the position, arc and color of the improved radar chart of special glyphs in the component projection view: different ring positions encode different monitoring indicators; the arc encodes the length of the time slice; the color of the ring encodes the abnormality of the time slice;

[0048] S25: Color and position mapping of data inspection view: six symbols are used to map six working conditions; the lines between the origins of different colors represent the timing relationship of the components working under different working conditions; for highly correlated monitoring variables, cubic Bezier curves are used to connect them on the left side of the area chart.

[0049] Furthermore, the visualization layout scheme of the component projection view is specifically as follows:

[0050] S3a: For scatter plots and contour plots, according to the screen width and height allocated by the system, the scatter plots and contour lines are mapped onto the screen in a linear scale. The density and transparency of the lines in the contour plots are used to map the distribution density of components in different location areas, thus completing the basic layout of the view.

[0051] S3b: For the component glyph view, select each slice range and the monitoring variables to be displayed in the control bar, and calculate the glyph size according to the number of monitoring variables;

[0052] S3c: After the size is determined, the glyphs are discretized according to the collision detection algorithm based on force-guided layout to solve the visual occlusion problem caused by dense data in local areas and sparse data in distant areas;

[0053] S3d: Count the number of operations of each component under six working conditions, and determine the distance from each point to the center point in the improved radar chart through a linear scale;

[0054] S3e: For the superimposed ring chart, the rings are drawn from the inside to the outside according to the selection order of the monitoring variables, and the rings are cut into multiple segments according to the length of the time slice. The drawing of each segment is completed according to the size of the outlier in each time slice.

[0055] Furthermore, in S3a, the density and transparency of the lines in the contour map are used to map the distribution density of components in different location areas, specifically including:

[0056] S3a1: 2D Scalar Field Computation

[0057] 1a) For a given set of data points (x1, x2…, x n ; y1,y2,…,y n ), determine its minimum and maximum observed values, that is, (x1, y1) and (x n ,y n ), where n is the number of components;

[0058] 1b) Estimate the grid group distance h that should be used when performing two-dimensional grid division, so as to obtain the grid boundary points (a0, a1, ..., a Ve ) and (b0,b1,…,b He ), where a ve+1 -a ve =h,b he+1 -b he =h,ve=0,1,…,Ve-1,he=0,1,…,He-1; Ve is the number of grid lines in the vertical direction; He is the number of grid lines in the horizontal direction;

[0059] The grid group distance h is calculated as follows:

[0060]

[0061] Where: IQR(x) is the difference between the upper quartile and the lower quartile of the sample; n is the number of components;

[0062] 1c) Count the frequency of scattered data in each grid area, that is, (g 0,0 ,g 0,1 ,…,g Ve,He );

[0063] S3a2: Contour calculation

[0064] 2a) Determine an initial threshold, compare the data value in the grid with the set threshold, mark the grid with a value greater than the threshold as 1, and mark the grid value with a value less than the threshold as 0, thereby obtaining a binary image;

[0065] 2b) Assume that a 2*2 pixel block in the binary image is a contour unit; construct a binary index based on the four values ​​of each cell corner, scan around the cell in a clockwise direction, use a bitwise OR operation, move from the highest bit in the upper left corner to the lowest bit, and generate a four-bit index;

[0066] 2c) Use the cell index to access a pre-built lookup table that lists the edges needed to represent the cell;

[0067] 2d) Apply linear interpolation between the original data values ​​to find the exact location of the contour line along the cell edge;

[0068] S3a3: Scale Mapping and Color Mapping

[0069] Contour fill color: The contour color curve smoothly transitions from white to dark; the system represents area density as a value from 0 to 1.

[0070] Furthermore, in S3c, the specific method of separating the glyphs is as follows:

[0071] S3c1: Get the maximum and minimum xy coordinates of all data selected by the lasso, map the difference between the maximum and minimum values ​​to the width and height of the canvas, recalculate the real coordinates of the data in the canvas according to the linear scale, and reproject the data points to the new canvas;

[0072] S3c2: Assume the data points on the canvas to be particles with the same mass and radius, and make the particles move by adding different mechanical models;

[0073] S3c3: Add collision force to each particle and use Verlet to calculate the position and velocity of the particle after Δt time;

[0074] S3c4: Repeat the iterations many times to obtain the position of the particle when its final velocity is 0, which is the position of the final data point. At this time, the particles are correlated and separated and tend to be stable.

[0075] Furthermore, the S3d is specifically:

[0076] S3d1: Map the running statistics under each working condition to the distance cRadius from the center of the radar chart to the axis. The calculation process is as follows:

[0077]

[0078] Where: is the maximum number of operations under working condition o; is the minimum number of operations under working condition o; sum o is the number of times the current component operates under working condition o; radarRadius is the radius of the radar chart; O is the total number of working condition categories;

[0079] S3d2: Calculate the X coordinate xpos of the point in the figure based on the distance from the point to the center of the circle o and the Y-axis coordinate ypos o , as follows:

[0080]

[0081] S3d3: The size of the central circle maps the life cycle length of the component. The size of the scattered points on the radar coordinate axis maps the number of operations under the working condition, which is consistent with the position from the center of the circle. It belongs to secondary coding and is calculated as follows:

[0082]

[0083]

[0084] In the formula: rul min The shortest life cycle of a component;max is the maximum life cycle of the component; R is the maximum radius of the center circle; r is the maximum radius of the coordinate axis circle; rulRadius is the radius of the center circle; dotArea o is the distance from the point represented by working condition o to the center of the circle;

[0085] S3d4: The subsequence outliers of each monitoring variable are mapped to the length of the arc, and the angle occupied by the arc is calculated as follows:

[0086]

[0087] Where: RULLen is the life cycle length of component degradation; represents the length of the sth monitoring variable of component k in the time interval TS;

[0088] S3d5: The distance from the center of the circle represented by the sth monitoring variable is:

[0089] rRadius s =rDis+(rGap+rBandWidth)·(s-1),s=1,2,3,…,M

[0090] Where: M is the number of selected features; rDis is the inner ring radius; rGap is the interval distance between rings; rBandWidth is the width of the ring.

[0091] Furthermore, in step S3, the specific process of visual layout and implementation of the data inspection view is as follows:

[0092] S31: Determine the life cycle range to be displayed, design an area chart, and connect two highly correlated monitoring variables using a cubic Bezier curve based on calculations;

[0093] S32: Determine the color corresponding to each working condition, determine the display form of each dot, including the layout position and the line drawing method, and layout and implement the point-line diagram.

[0094] The beneficial effects of the present invention are:

[0095] 1) The present invention aims to make up for the shortcomings of existing methods in analyzing component abnormal data patterns. Traditional alarm systems often cannot obtain the specific reasons for data abnormalities. This visualization method can intuitively discover abnormal data in the data through dimensionality reduction projection. For the subset of abnormal data, a special glyph is designed to locate the cause of its abnormality. Through the juxtaposition of special glyphs, the abnormal pattern can be clearly and intuitively displayed.

[0096] 2) The present invention makes up for the shortcomings of traditional systems that can only view raw data, have a large cognitive burden, and are difficult to locate abnormalities. The visualization method subdivides the working conditions of components into six categories. Users can directly view the working condition sequence of components and explain component abnormalities caused by abnormal working condition sequences. At the same time, the three-dimensional Bezier connection can assist users in discovering highly correlated monitoring indicators, making it easier to locate abnormalities. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 It is a schematic diagram of the overall process framework of the projection-based abnormal component inspection visualization method of the present invention.

[0098] Figure 2 It is a visual coding and visualization layout diagram of the component projection view in the present invention.

[0099] Figure 3 It is a schematic diagram of the visual encoding and visualization layout of the contour map in the present invention.

[0100] Figure 4 It is a schematic diagram of the visual encoding and visualization layout of the glyphs in the present invention.

[0101] Figure 5 It is a schematic diagram of the implementation effect and overall layout of the data inspection view in the present invention. DETAILED DESCRIPTION

[0102] The present invention is further described in detail below with reference to the accompanying drawings and specific data implementations.

[0103] The present invention uses an effective information visualization method, combined with a multi-view linkage strategy and flexible interactive means, to achieve multi-angle analysis of anomalies in aviation gas turbofan engine data, helping predictive maintenance personnel to discover anomalies in the data and analyze their abnormal patterns and causes. The technical solution includes: data acquisition and processing, visual mapping, visual layout and implementation, and interactive design. The specific steps are as follows:

[0104] Step 1: Data acquisition and processing

[0105] Based on the aviation gas turbofan engine data provided by NASA, effective information is screened and data is stored.

[0106] 1. Data acquisition: The FD004 data set in the data was selected for use, which contains data of 249 engines running from the initial state to complete failure. The maximum life cycle of the engine is 543, and the minimum life cycle is 128. The specific data contains 26 dimensions, of which 1-2 dimensions are the machine number and time point; 3-5 dimensions are the working conditions of the machine: altitude, Mach number and sea level temperature; 6-26 dimensions are the monitoring indicators of the engine: total fan inlet temperature, low-pressure compressor temperature, high-pressure compressor temperature, low-pressure turbine temperature, fan inlet pressure, bypass delivery pipe pressure, high-pressure compressor air pressure, physical fan speed, physical core speed, engine pressure ratio, high-pressure compressor static pressure, fuel volume and high-pressure turbine static pressure ratio, corrected fan speed, corrected core speed, bypass ratio, fuel-air ratio in the combustion chamber, exhaust valve heat content, required fan speed, corrected required fan speed, high-pressure turbine coolant emission, low-pressure turbine coolant emission.

[0107] 2. Data processing: Data processing includes working condition identification and data dimension reduction operations.

[0108] (1) Working condition identification: Since the engine works in different working conditions, the working conditions can be divided according to three operating conditions: altitude, Mach number, and sea level temperature. The working conditions are clustered using k_means. The main process of the algorithm is: 1) specify k, that is, divide the data into k categories; 2) randomly select k points from the data as the centroid of each cluster; 3) measure the distance between the data point and the centroid by a certain distance calculation method, and divide the data point to the nearest cluster centroid; 4) recalculate the centroid of each cluster; 5) if one of the three conditions is met: the centroid of the newly formed cluster does not change, the point remains in the same cluster, and the maximum number of iterations is reached, then the iteration ends, otherwise repeat steps 3 to 5. Experiments show that the best effect is when k = 6, and the engine working conditions are divided into working conditions 1, working conditions 2, working conditions 3, working conditions 4, working conditions 5, and working conditions 6.

[0109] (2) Data dimensionality reduction: Components can be represented by multi-dimensional time series data, namely:

[0110] X k =[t1,t2…,t i ,…,t T ],k=(1,2,…,n),t j ∈R M

[0111] Where: n is the number of components; M is the number of selected features; T is the length of the life cycle; t i is the value of each feature at time i; k is the component number.

[0112] In order to reduce the amount of calculation and the deviation caused by inaccurate data collection, a sliding time window is used to aggregate the data, and the following is obtained:

[0113]

[0114] Where: w is the time window size; x w_i is the value of each feature of window w_i after aggregation;

[0115] For the obtained x w_i Using t-distributed random neighborhood embedding, the algorithm steps are as follows:

[0116] a) Use Euclidean distance to characterize the similarity between vectors and express the similarity as conditional probability:

[0117]

[0118] Where: w_i is the variance.

[0119] b) Calculate the similarity between the points in the two-dimensional space to be projected, using conditional probability representation:

[0120]

[0121] Where: y w_i is the location of the data point in the low-dimensional space.

[0122] c) Kullback-Leibler divergence is often used to measure the distance between two probability distributions, minimizing the sum of KL divergences:

[0123]

[0124] Where: n is the number of components.

[0125] d) To alleviate the problem of data point crowding in low-dimensional space, t-distribution is used to calculate the similarity between data points.

[0126]

[0127] Where: y w_i is the location of the data point in the low-dimensional space.

[0128] After the above steps, x w_i The vectors are projected into one-dimensional space, so the components can be represented as:

[0129] X k ∈R T-w+1 ,k=(1,2,…,n)

[0130] For the one-dimensional time series after dimensionality reduction, the above t-SNE is used to reduce the dimension to two-dimensional space, X k It can be expressed as:

[0131] X k ∈R 2 ,k=(1,2,…,n)

[0132] Where: n is the number of components.

[0133] (3) Calculation of outliers: Use the Pearson coefficient to define outliers. i ,k j Two time series of the sth monitoring variable Its relevance is defined as follows:

[0134]

[0135]

[0136] Where: For number k i The value of the sth monitoring variable of the component at time point t; is component k j The value of the sth monitoring variable at time point t; TS is the length of the current time interval; for The average value in the time interval TS;

[0137] The outliers are defined as:

[0138]

[0139] Where: n is the number of components.

[0140] Step 2: Visual Mapping

[0141] After data acquisition and processing, the component projection view (such as Figure 2 as shown) and data checking (as Figure 5 As shown in the figure, a visual mapping scheme is designed.

[0142] 1. Component projection view

[0143] (1) Scatter plot

[0144] The shape, position and color of the scatter plot in the component projection view are mapped: the position of the scatter plot is used to map the dimension reduction result of the component; the color is used to map the clustering result of the component; the transparency of the point is used to map the life cycle length of the component. The specific implementation example is as follows:

[0145] Position: A point in the scatter plot maps a component, the two-dimensional coordinates of the point in the view represent the dimension reduction result of the component, and the distance between the points maps the similarity between the degradation trajectories of the components;

[0146] Color: Components are clustered into 6 categories according to their degradation trajectories, and each category is marked with a different color. Points of the same color are clustered together, indicating a good clustering effect;

[0147] Transparency: The transparency of a point represents the length of the component life cycle. The longer the life cycle, the lower the transparency.

[0148] (2) Contour map

[0149] Color map the contour map in the component projection view: fill the blank spaces between the contour lines with a specific color; use the depth of the fill color to map the density of the component projected in the local area.

[0150] The area of ​​the petals is used to map the number of behaviors. The specific implementation example is as follows:

[0151] Contour line color: For the blank spaces between contour lines, the system fills them with green. In a local area, if the density of projected components is higher, the filling color will be darker, which can help analysts understand the real situation of data distribution;

[0152] Area: Use the area of ​​the petals to map the number of behaviors. The larger the area, the more times it happens, and the smaller the area, the fewer times it happens.

[0153] (3) Special fonts

[0154] This part adopts radial layout, including InnerArea and OuterArea. Figure 4 (a) In the InnerArea section, an improved radar chart is used to map the number of components operating under each working condition. In the OuterArea section, an overlay ring chart is used to represent the abnormality of different monitoring indicators in each time period.

[0155] InnerArea

[0156] The improved radar chart of special glyphs in the component projection view is mapped in shape, size, and color: the color of the point is used to map different working condition categories; the size of the center circle is used to map the life cycle length of the component; the shape of the radar chart is used to encode the number of operations of the component under six working conditions, and the operation of the component during the degradation period is compared by comparing the shapes of the hexagons. The specific embodiments are as follows:

[0157] Color: 6 colors are used to map different working conditions.

[0158] Size: The size of the center circle encodes the length of the component's life cycle.

[0159] Shape: It contains a mirrored arrangement of six axes inside. The distance between the points on the axes and the axis center is used to encode the number of operations under six working conditions. The points on the axes are connected end to end to form an irregular hexagon. By comparing the shapes of the hexagons, the operating conditions of the components during the degradation period can be roughly compared.

[0160] OuterArea

[0161] The improved radar chart of special glyphs in the component projection view is mapped to position, arc, and color: different circular ring positions encode different monitoring indicators; the arc encodes the length of the time slice; the color of the circular ring encodes the abnormality of the time slice. The specific implementation example is as follows:

[0162] Location: OuterArea contains several circles, each representing a monitoring indicator. Figure 4 In (a), the four rings from the inside to the outside are the monitoring data of the fan inlet total temperature, high-pressure compressor temperature, low-pressure turbine temperature, and fan inlet pressure. The number of rings and the monitoring indicators they represent can be selected by the user.

[0163] Radians: Time slices the monitored data. The length of the time slice is selected by the user. The radians map the length of the time slice. Figure 4 In (a), the degradation period is divided into five time slices;

[0164] Color: Each donut slice has a different color, using color to encode the degree of anomaly for that time slice.

[0165] 2. Data inspection view

[0166] Color and position mapping of data inspection view: six symbols are used to map six working conditions; the lines between the origins of different colors represent the timing relationship of the components working under different working conditions; for highly correlated monitoring variables, cubic Bezier curves are used to connect on the left side of the area chart. The specific implementation is as follows:

[0167] (1) Point-line graph

[0168] Color: Six contrasting colors of red, orange, yellow, green, cyan and blue are used to map each working condition.

[0169] Length: Use the length of the rectangle to encode the number of times the component has been run under this condition.

[0170] (2) Area chart

[0171] Height: The height of the area chart reflects the size of the real data.

[0172] Position: The two ends of the red line are highly correlated monitoring variables. When the Pearson coefficient is greater than 0.5, the system will automatically connect the corresponding variables.

[0173] Step 3: Visualization layout and implementation

[0174] 1. Visual layout and implementation of component projection view

[0175] When there are too many components, the scattered points in the scatter plot will be blocked, which makes it impossible for analysts to effectively obtain the data density in the current area. Figure 3 The density and transparency of the center line (as shown in the figure) map the distribution density of components in different locations. Users can clearly find dense or sparse areas in the data, which is of great significance for discovering outliers in the data. The implementation method is as follows:

[0176] (1) Two-dimensional scalar field calculation

[0177] a) For a given set of data points (x1, x2…, x n ; y1,y2,…,y n ), determine its minimum and maximum observed values, that is, (x1, y1) and (x n ,y n ), where n is the number of components.

[0178] b) Estimate the grid group distance h that should be used when performing two-dimensional grid division, so as to obtain the grid boundary points (a0, a1, ..., a Ve ) and (b0,b1,…,b He ), where a ve+1 -a ve =h,b he+1 -b he =h,ve=0,1,…,Ve-1,he=0,1,…,He-1; Ve is the number of vertical grid lines; He is the number of horizontal grid lines. h is calculated as follows:

[0179]

[0180] Where: IQR(x) is the difference between the upper quartile and the lower quartile of the sample; n is the number of components.

[0181] c) Count the frequency of scattered data in each grid area, that is, (g 0,0 ,g 0,1 ,…,g Ve,He ).

[0182] (2) Contour line calculation

[0183] a) Step 1: Determine an initial threshold, compare the data value in the grid with the set threshold, mark the grid with a value greater than the threshold as 1, otherwise mark the grid value as 0, thus obtaining a binary image.

[0184] b) Assume that a 2*2 pixel block in the binary image is a contour unit. Build a binary index based on the four values ​​of each cell corner, scan around the cell in a clockwise direction, use the bitwise OR operation, move from the highest bit in the upper left corner to the lowest bit, and generate a four-bit index.

[0185] c) Use the cell index to access a pre-built lookup table containing 16 entries listing the edges needed to represent the cell.

[0186] d) Apply linear interpolation between the original data values ​​to find the exact position of the contour line along the edge of the cell. Through the above four steps, we get the contour line of a certain threshold, and repeat the above four steps with num% as the density interval to get num contour lines.

[0187] (3) Scale mapping and color mapping

[0188] According to the screen width and height allocated by the system, the scatter plot and contour lines are mapped to the screen in a linear scale to complete the basic layout of the view.

[0189] Contour fill color: The contour color curve smoothly transitions from white to dark green. The system represents the area density as a value from 0 to 1, and then obtains the final color representation through the following formula:

[0190]

[0191] Where: hexColor is the hexadecimal value of the RGB color; data i is the density of the current area; data min is the minimum density; data max is the maximum density.

[0192] After the user interactively selects a set of components, the view retains the relative positions between the components and expands into a special glyph map, such as Figure 2 The right half is shown. By checking the glyphs, the time period and monitoring amount of abnormal component degradation process can be found. At the same time, the degradation pattern of abnormal components can be found by judging the similarity between glyphs. The implementation process of special glyphs is as follows:

[0193] (1) Spatial layout implementation

[0194] After selecting the abnormal component to be analyzed, the glyph design of the canvas display component set needs to be regenerated according to the circled area. Since the shape selected by the user's lasso cannot be obtained, when the local area data is dense and the distant area is sparse, it will cause serious visual occlusion in the view, such as Figure 2 This paper implements a collision detection algorithm based on force-guided layout to avoid the intersection of glyphs. The implementation is as follows:

[0195] a) Get the maximum and minimum xy coordinates of all the data selected by the lasso, map the difference between the maximum and minimum values ​​to the width and height of the canvas, recalculate the real coordinates of the data in the canvas according to the linear scale, and reproject the data points to the new canvas.

[0196] b) Assume that the data points on the canvas are charged particles with the same mass and radius, and make the particles move by adding different mechanical models.

[0197] c) Add a certain collision force to each particle and use Verlet to calculate the position and velocity of the particle after Δt time.

[0198] d) Repeat the iteration several times to obtain the position of the particle when the final velocity is 0, which is the position of the final data point. At this time, the particles are correlated and separated and tend to be stable. At this time, it can be found that the relative positions between the scattered points are well preserved, and the glyphs overlap.

[0199] (2) Font design and implementation

[0200] For a single component, the data includes the operating statistics of the component under various working conditions when it degrades, and the specific definitions are as follows:

[0201] D={I o |o∈[1,O]}

[0202] Where: O is the type of working condition; I o is the number of operations under working condition o.

[0203] like Figure 4 As shown in (b), the running statistics under each working condition are mapped to the distance cRadius from the center of the radar chart to the axis. The calculation process is as follows:

[0204]

[0205] Where: is the maximum number of operations under working condition o; is the minimum number of operations under working condition o; sum o is the number of times the current component operates under working condition o; radarRadius is the radius of the radar chart; O is the total number of working condition categories.

[0206] Calculate the X coordinate xpos of the point in the figure based on the distance from the point to the center of the circle o and the Y-axis coordinate ypos o , as follows:

[0207]

[0208] The size of the center circle maps the life cycle length of the component, and the size of the scattered points on the radar coordinate axis maps the number of operations under the working condition, which is consistent with the position from the center of the circle. It belongs to secondary coding and is calculated as follows:

[0209]

[0210]

[0211] In the formula: rul min The shortest life cycle of a component; max is the maximum life cycle of the component; R is the maximum radius of the center circle; r is the maximum radius of the coordinate axis circle; rulRadius is the radius of the center circle; dotArea o is the distance from the point represented by working condition o to the center of the circle.

[0212] The subsequence outliers of each monitoring variable are mapped to the length of the arc, and the angle occupied by the arc is calculated as follows:

[0213]

[0214] Where: RULLen is the life cycle length of component degradation; represents the length of the sth monitoring variable of component k in the time interval TS;

[0215] The distance between ring k and the center of the circle is:

[0216] rRadius s =rDis+(rGap+rBandWidth)·(s-1),s=1,2,3,…,M

[0217] Where: M is the number of selected features; rDis is the inner ring radius; rGap is the interval distance between rings; rBandWidth is the width of the ring.

[0218] 2. Data inspection view visualization layout and implementation:

[0219] Connecting line drawing: The highly correlated variables are marked with red connecting lines on the left side of the area chart, such as Figure 5When the Pearson coefficient is greater than 0.5, the system will draw a cubic Bezier curve to automatically connect the corresponding variables. Four points are needed to draw a cubic Bezier curve, namely: the starting point spoint, the end point epoint, and two control points cpoint1 and cpoint2. Based on these four points, the result after drawing is as follows: Figure 4 The coordinates of the two control points are determined as follows:

[0220]

[0221]

[0222] Where: spoint x is the x-axis coordinate of the starting point; epoint x is the x-axis coordinate of the end point; epoint y is the y-axis coordinate of the end point; spoint y is the y-axis coordinate of the starting point.

[0223] Step 4: Interaction Design

[0224] In the component projection view, it includes highlighting, zooming, lassoing, time slicing, and reconfiguring operations:

[0225] Highlight: In the scatter view, in order to help users quickly understand the basic information of the component, the system provides a highlight interaction method. When the analyst focuses the mouse on a data point, the point will be highlighted with color, supplemented by a prompt box to display the component number, life cycle length, working condition operation statistics and other information.

[0226] Zoom: Considering that more data points may cause overlap between points, causing visual confusion, users can zoom in and out to view the local area in detail, which also helps in the subsequent selection of data points.

[0227] Lasso: After finding anomalies or points of interest, users often need to check this part of the data. This method supports drawing a closed irregular shape on the scatter plot. If a data point is in the middle of the shape when the user completes the lasso, the point will be selected and the glyph corresponding to the data point will be displayed in the pop-up view.

[0228] Time Slices: The analyst can control the range of the time slices using the slider above. This paper provides greater flexibility, allowing users to select overlapping and unequal time slices to explore the correlation of monitored variables during the degradation process of the component.

[0229] Reconfiguration: This visualization method allows users to switch between scatter plots and contour plots through tabs. When the component life cycle is too long, the transparency of the points will be too low, resulting in obstruction between points. At the same time, due to uneven data distribution or excessive data volume, the view may be severely blocked locally. At this time, switching to contour plots through reconfiguration can help alleviate visual confusion.

Claims

1. A projection-based abnormal component inspection visualization method, characterized in that: The following steps are involved: S1: Data acquisition and processing After obtaining the aviation gas turbofan engine data, the data is cleaned and processed, and organized into structured data that can be directly used in a visual view; S2: Visual Mapping Visually map the data obtained in step S1 through the visual channel: Design component projection views, presenting an overview of data distribution using scatter plots and contour plots, with proximity indicating similarity in component degradation patterns; using a special glyph of radar charts nested in multiple layers of rings to compare similar patterns in abnormal data; Design data inspection views, use horizontal bar charts to map the number of times components operate under different working conditions; use area charts to display the real data of monitored variables, and stack the area charts vertically to help analysts discover the relationship between the variables; S3: Visualization layout and implementation Visualize and layout the visual modules mapped in S2: In the component projection view, the component is mapped to each point in the scatter plot, and the glyph retains the relative position between the scatter points. For the glyph, there is an improved radar chart in the middle to record the operation of the component, and multiple superimposed rings on the outside. In the data inspection view, the timeline layout method is adopted. The point-line graph at the top of the view shows the operating sequence running in the current time period, and the timeline module with an area graph is at the bottom of the view; S4: Interaction Design In the scatter plot view, hovering the mouse over a data point highlights the point, allowing users to zoom in and out to view the local area of ​​the scatter plot in detail; by drawing a closed irregular shape on the scatter plot, the scatter point is replaced with a glyph to explore the data in depth; the length of the time slice encoded by each ring in the glyph is controlled by the slider; In step S2, the visualization mapping is specifically as follows: S21: Map the shape, position and color of the scatter plot in the projection view of the component: use the position of the scatter plot to map the dimension reduction result of the component; use the color to map the clustering result of the component; and use the transparency of the point to map the life cycle length of the component. S22: Perform color and area mapping on the contour map in the component projection view: use specific colors to fill the blank spaces between the contour lines; use the depth of the filling color to map the density of the projected components in the local area; use the area of ​​the petals to map the number of behaviors; S23: Map the shape, size and color of the improved radar chart with special glyphs in the component projection view: use the color of the points to map different working condition categories; use the size of the center circle to map the life cycle length of the component; The shape of the radar chart is used to encode the operation quantity of the component under the six working conditions, and the operation of the component during the degradation period is compared by comparing the shapes of the hexagons; S24: Map the position, arc and color of the improved radar chart of special glyphs in the component projection view: different ring positions encode different monitoring indicators; the arc encodes the length of the time slice; the color of the ring encodes the abnormality of the time slice; S25: Data inspection view color and position mapping: six symbols are used to map six working conditions; the lines between the origins of different colors represent the timing relationship of the components working under different working conditions; for highly correlated monitoring variables, cubic Bezier curves are used to connect them on the left side of the area chart; The visualization layout scheme of the component projection view is specifically as follows: S3a: For scatter plots and contour plots, according to the screen width and height allocated by the system, the scatter plots and contour lines are mapped onto the screen in a linear scale. The density and transparency of the lines in the contour plots are used to map the distribution density of components in different location areas, thus completing the basic layout of the view. S3b: For the component glyph view, select each slice range and the monitoring variables to be displayed in the control bar, and calculate the glyph size according to the number of monitoring variables; S3c: After the size is determined, the glyphs are discretized according to the collision detection algorithm based on force-guided layout to solve the visual occlusion problem caused by dense data in local areas and sparse data in distant areas; S3d: Count the number of operations of each component under six working conditions, and determine the distance from each point to the center point in the improved radar chart through a linear scale; S3e: For the superimposed ring chart, the rings are drawn from the inside to the outside according to the selection order of the monitoring variables, and the rings are cut into multiple segments according to the length of the time slice. The drawing of each segment is completed according to the size of the outlier in each time slice.

2. The projection-based abnormal component inspection visualization method according to claim 1, characterized in that: In step S1, data processing includes: (1) Working condition identification: The working conditions are divided according to three operating conditions: altitude, Mach number, and sea level temperature, and the working conditions are clustered using k-means. (2) Data Dimensionality Reduction: The components are represented by multi-dimensional time series data, that is: X k =[t1,t2…,t i ,…,t T ],k=1,2,…,n,t i ∈R M Where: n is the number of components; M is the number of selected features; T is the length of the life cycle; t i is the value of each feature at time i; k is the component number; Aggregate the data using a sliding time window and get: Where: w is the time window size; x w_i is the value of each feature of window w_i after aggregation; For the obtained x w_i Using t-distributed random neighborhood embedding, x w_i The vectors are projected into one-dimensional space, and the components are represented as: X k ∈R T-w+1 ,k=1,2,…,n For the degradation curve obtained after dimensionality reduction, t-SNE is used to reduce the dimension to two-dimensional space, and X k It is expressed as: X k ∈R 2 ,k=1,2,…,n Where: n is the number of components; (3) Calculation of outliers: The Pearson coefficient is used to define anomalies for k. i and k j The two time series of the sth monitoring variable of the component Its relevance is defined as follows: Where: For number k i The value of the sth monitoring variable of the component at time point t; For number k j The value of the sth monitoring variable of the component at time point t; TS is the length of the current time interval; for The average value in the time interval TS; The outliers are defined as: Where: n is the number of components.

3. The projection-based abnormal component inspection visualization method according to claim 1, characterized in that: In S3a, the density and transparency of the lines in the contour map are used to map the distribution density of components in different location areas, specifically including: S3a1: 2D Scalar Field Computation 1a) For a given set of data points (x1, x2…, x n ; y1,y2,…,y n ), determine its minimum and maximum observed values, that is, (x1, y1) and (x n ,y n ), where n is the number of components; 1b) Estimate the grid group distance h that should be used when performing two-dimensional grid division, so as to obtain the grid boundary points (a0, a1, ..., a Ve ) and (b0,b1,…,b He ), where a ve+1 -a ve =h,b he+1 -b he =h,ve=0,1,…,Ve-1,he=0,1,…,He-1; Ve is the number of grid lines in the vertical direction; He is the number of grid lines in the horizontal direction; The grid group distance h is calculated as follows: Where: IQR(x) is the difference between the upper quartile and the lower quartile of the sample; n is the number of components; 1c) Count the frequency of scattered data in each grid area, that is, (g 0,0 ,g 0,1 ,…,g Ve,He ); S3a2: Contour calculation 2a) Determine an initial threshold, compare the data value in the grid with the set threshold, mark the grid with a value greater than the threshold as 1, and mark the grid value with a value less than the threshold as 0, thereby obtaining a binary image; 2b) Assume that a 2*2 pixel block in the binary image is a contour unit; construct a binary index based on the four values ​​of each cell corner, scan around the cell in a clockwise direction, use a bitwise OR operation, move from the highest bit in the upper left corner to the lowest bit, and generate a four-bit index; 2c) Use the cell index to access a pre-built lookup table that lists the edges needed to represent the cell; 2d) Apply linear interpolation between the original data values ​​to find the exact position of the contour line along the cell edge; S3a3: Scale Mapping and Color Mapping Contour fill color: The contour color curve smoothly transitions from white to dark; the system represents area density as a value from 0 to 1.

4. The projection-based abnormal component inspection visualization method according to claim 1, characterized in that: In S3c, the specific method of separating the glyphs is as follows: S3c1: Get the maximum and minimum xy coordinates of all data selected by the lasso, map the difference between the maximum and minimum values ​​to the width and height of the canvas, recalculate the real coordinates of the data in the canvas according to the linear scale, and reproject the data points to the new canvas; S3c2: Assume the data points on the canvas to be particles with the same mass and radius, and make the particles move by adding different mechanical models; S3c3: Add collision force to each particle and use Verlet to calculate the position and velocity of the particle after Δt time; S3c4: Repeat the iterations many times to obtain the position of the particle when its final velocity is 0, which is the position of the final data point. At this time, the particles are correlated and separated and tend to be stable.

5. The projection-based abnormal component inspection visualization method according to claim 1, characterized in that: The S3d is specifically: S3d1: Map the running statistics under each working condition to the distance cRadius from the center of the radar chart to the axis. The calculation process is as follows: Where: is the maximum number of operations under working condition o; is the minimum number of operations under working condition o; sum o is the number of times the current component operates under working condition o; radarRadius is the radius of the radar chart; O is the total number of working condition categories; S3d2: Calculate the X coordinate xpos of the point in the figure based on the distance from the point to the center of the circle o and the Y-axis coordinate ypos o , as follows: S3d3: The size of the central circle maps the life cycle length of the component. The size of the scattered points on the radar coordinate axis maps the number of operations under the working condition, which is consistent with the position from the center of the circle. It belongs to secondary coding and is calculated as follows: In the formula: rul min The shortest life cycle of a component; max is the maximum life cycle of the component; R is the maximum radius of the center circle; r is the maximum radius of the coordinate axis circle; rulRadius is the radius of the center circle; dotArea o is the distance from the point represented by working condition o to the center of the circle; S3d4: The subsequence outliers of each monitoring variable are mapped to the length of the arc, and the angle occupied by the arc is calculated as follows: Where: RULLen is the life cycle length of component degradation; represents the length of the sth monitoring variable of component k in the time interval TS; S3d5: The distance from the center of the circle represented by the sth monitoring variable is: rRadius s =rDis+(rGap+rBandWidth)·(s-1),s=1,2,3,…,M Where: M is the number of selected features; rDis is the inner ring radius; rGap is the interval distance between rings; rBandWidth is the width of the ring.

6. The projection-based abnormal component inspection visualization method according to claim 1, characterized in that: In step S3, the specific process of visual layout and implementation of the data inspection view is as follows: S31: Determine the life cycle range to be displayed, design an area chart, and connect two highly correlated monitoring variables using a cubic Bezier curve based on calculations; S32: Determine the color corresponding to each working condition, determine the display form of each dot, including the layout position and the line drawing method, and layout and implement the point-line diagram.

Images