Intelligent evaluation method for residual strength of marine cryogenic tank box outer container based on deep learning

Abstract

The present application relates to the technical field of strength evaluation, in particular to a method for intelligent evaluation of residual strength of a marine low-temperature tank box outer container based on deep learning, comprising the following steps: collecting three-direction acceleration pressure wall temperature strain and outer container images, identifying crack angle, weld direction, crack depth and weld projection section, determining dangerous anchor points according to load principal vector method direction angle difference and strain peak value, extracting pressure temperature strain drop characteristics and calculating equivalent stress retention, and combining wall thickness stress intensity factor and fracture toughness threshold to output residual strength results. In the present application, the low-temperature lagging pressure peak stress retention of ship motion and the crack geometry are unified and mapped to the dangerous anchor points through the linkage of multi-source time sequence load and image crack information, the weld neighborhood risk is separated from the overall working condition, the timeliness of dangerous crack identification and the adaptability of residual strength judgment to low-temperature continuous action transient impact are improved, and the risk of conservative or dangerous evaluation is reduced.

Description

Technical Field

[0001] This invention relates to the field of strength assessment technology, and in particular to a deep learning-based intelligent assessment method for the residual strength of the outer container of marine cryogenic tanks. Background Technology

[0002] The field of strength assessment technology mainly involves determining the safety status of load-bearing structures under service loads, ambient temperature, internal pressure, material property degradation, weld defects, crack propagation, and boundary constraints. Its core aspects include obtaining structural geometric parameters, determining material yield strength and tensile strength, correcting mechanical properties under cryogenic conditions, calculating thermal stress and internal pressure stress, describing inertial loads caused by ship roll, pitch, and heave, setting the constraint relationship between supports and saddles, analyzing stress concentration at welded joints, recording crack length, depth, and direction, estimating fatigue crack propagation rate, and determining the remaining ultimate bearing capacity. It is commonly used for the service safety analysis of ship cryogenic storage tanks, tank outer containers, pressure vessel shells, and their connecting structures.

[0003] The traditional intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank refers to a method for determining the residual load-bearing state of the outer container of a cryogenic tank in an LNG-powered or hydrogen-powered ship under the combined effects of temperature changes, pressure fluctuations, liquid sloshing, ship motion, and crack damage. Typically, the method first collects data on the outer container's shell thickness, head curvature, weld location, support arrangement, material grade, cryogenic yield strength, fracture toughness, design pressure, measured pressure, internal medium temperature, ambient temperature, ship acceleration, crack length, crack depth, crack tip location, and crack propagation direction. Then, based on classification society specifications, pressure vessel design formulas, finite element shell element or solid element models, SN curves, Paris crack propagation relationships, stress intensity factors, J-integrals, ultimate load criteria, and allowable stress comparison results, the method calculates and assesses the stress distribution, crack propagation stage, and residual ultimate strength of the outer container under different operating conditions.

[0004] Traditional assessment methods often rely on standard formulas and offline finite element calculations in actual operation. Load boundaries are often set according to fixed working conditions. The temporal coupling between temperature, pressure, hull motion, and crack damage is difficult to reflect in a timely manner. The relationship between the local strain peak in the weld neighborhood and crack orientation lacks local screening. The impact of low temperature hysteresis on stress retention and fracture toughness threshold is not adequately expressed, which can lead to delayed identification of dangerous cracks. The residual strength judgment is not sensitive enough to transient impact and continuous low temperature effects. The assessment results are prone to being conservative or dangerous. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a deep learning-based intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank, comprising the following steps:

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank, comprising the following steps:

[0007] S1: Collect longitudinal acceleration data, lateral acceleration data and vertical acceleration data of the target ship, and simultaneously acquire the tank pressure time series, outer container wall temperature time series, weld neighborhood strain peak and marine cryogenic tank outer container image, and input the image into the depth residual network model, output crack tangential angle, weld direction angle, crack depth value and crack weld projection segment.

[0008] S2: Calculate the principal vector of the surface load and the normal direction vector based on the longitudinal acceleration data, the transverse acceleration data and the vertical acceleration data. Combine the weld direction angle, the difference in tangential angle and the strain peak value in the weld neighborhood to screen the cracked weld projection segment and obtain the anchor point of the dangerous location.

[0009] S3: Based on the anchor point of the dangerous location, extract the interval data of the pressure time series inside the tank and the temperature time series of the outer container wall, extract the pressure peak duration, pressure fall time, temperature trough data, and temperature trough duration, and collect the strain peak value and strain fall slope at the crack tip.

[0010] S4: Compare the duration of the temperature trough with the duration of the pressure peak to generate a low-temperature hysteresis gating parameter. Retrieve the low-temperature yield strength reduction factor from the material database and calculate the equivalent stress retention by combining the stress retention and crack depth numerical values.

[0011] S5: Extract the outer surface wall thickness of the container, calculate the Type I stress intensity factor, fracture toughness threshold value and fracture toughness margin difference based on the equivalent stress retention, and make a judgment based on the fracture toughness margin difference, the crack depth value and the outer surface wall thickness of the container to generate the remaining strength assessment result.

[0012] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0013] S101: Collect longitudinal acceleration data, lateral acceleration data and vertical acceleration data of the ship through acceleration sensors, obtain the tank internal pressure time series through pressure sensors, obtain the outer container wall temperature time series through temperature sensors, monitor the strain peak value of the weld neighborhood through strain gauge array, perform time alignment processing on the longitudinal acceleration data, the lateral acceleration data, the vertical acceleration data, the tank internal pressure time series, the outer container wall temperature time series and the weld neighborhood strain peak value to extract synchronization segments and splice them to generate a multi-source correlation matrix;

[0014] S102: Collect images of the outer container of the marine cryogenic tank box around the storage tank through a visual acquisition device, call the multi-source correlation matrix to record time nodes, perform sequence cropping and remove irrelevant images from the images of the outer container of the marine cryogenic tank box, extract and retain pixel values ​​in the images, perform grayscale conversion, and establish a structural mapping image.

[0015] S103: Input the structure mapping image into the deep residual network model for extraction processing, call the built-in convolution kernel of the deep residual network model to perform dot product operation on the pixel gradient to extract contour information, perform dimensionality reduction processing on the contour information through the pooling layer, and perform regression calculation on the dimensionality reduction result based on the fully connected layer to obtain the crack tangential angle, weld direction angle, crack depth value and crack weld projection segment.

[0016] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0017] S201: Perform three-dimensional coordinate system transformation calculation on the longitudinal acceleration data, the lateral acceleration data and the vertical acceleration data, extract the corresponding transformation matrix projected mechanical components in the tangential plane, perform orthogonal rotation operation on the projected mechanical components to separate vertical vector features, and generate surface load principal vector and normal direction vector;

[0018] S202: Perform inverse trigonometric function mapping on the principal vector of the surface load to extract the principal vector azimuth angle, perform subtraction calculation between the principal vector azimuth angle and the weld direction angle to extract the direction deflection, perform three-dimensional solution on the normal direction vector to extract spatial angle parameters, and perform algebraic difference operation between the spatial angle parameters and the crack tangential angle to obtain the tangential angle difference.

[0019] S203: Extract the coordinate sequence of the point set within the projection segment of the cracked weld, call the tangential angle difference to set the angle filtering threshold range, set the stress anomaly discrimination line according to the strain peak value in the weld neighborhood, perform bidirectional verification on each node in the point set coordinate sequence, remove the data of nodes exceeding the limit, aggregate the coordinate nodes that fall within the angle filtering threshold range and cross the inner side of the stress anomaly discrimination line, and obtain the dangerous location anchor point.

[0020] As a further aspect of the present invention, the process of calling the tangential angle difference to set the angle screening threshold range specifically involves: extracting the anisotropic deflection angle parameter set based on the material principal axis deformation rate; multiplying the anisotropic deflection angle parameter with the tangential angle difference to generate an angle fluctuation difference; performing an addition operation on the tangential angle difference and the angle fluctuation difference to generate an upper limit boundary value for the angle; performing a subtraction operation on the tangential angle difference and the angle fluctuation difference to generate a lower limit boundary value for the angle; and combining the upper limit boundary value and the lower limit boundary value to set the angle screening threshold range.

[0021] The process of setting the stress anomaly discrimination line based on the strain peak value in the weld neighborhood specifically involves: extracting the local maximum strain value and the local minimum strain value recorded within the strain peak value in the weld neighborhood; calculating the average value of the local maximum strain value and the local minimum strain value to generate a reference stress line; obtaining the stress sensitivity coefficient value set based on the yield strength reduction ratio of the container metal base material under low temperature service conditions; and multiplying the reference stress line and the stress sensitivity coefficient value to calculate and set the stress anomaly discrimination line.

[0022] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0023] S301: Obtain the anchor point of the danger location, the pressure time series inside the tank and the temperature time series of the outer container wall. Establish a cutoff window based on the time attribute recorded by the anchor point of the danger location. Perform segmentation on the pressure time series inside the tank based on the cutoff window to extract pressure interval data. Extract the duration span of exceeding the warning threshold within the pressure interval data. Calculate the time difference required for the pressure extreme point to drop to the stable baseline. Generate the pressure peak duration and pressure drop time.

[0024] S302: Call the interception window to extract the target temperature segment from the time sequence of the outer container wall temperature, perform size comparison and search on the node values ​​in the target temperature segment to extract the lowest extreme value, count the total number of samples with node values ​​lower than the preset low temperature warning limit, calculate the corresponding time span, and obtain the temperature trough data and the duration of the temperature trough.

[0025] S303: For the anchor point at the dangerous location, activate the strain gauge array to collect dynamic deformation signals, perform peak retrieval on the dynamic deformation signals to extract the maximum peak parameter, extract discrete sampling points for the falling band that crosses the maximum peak parameter, perform first-order differential operation on the sequence of adjacent nodes of the discrete sampling points, and obtain the peak strain value at the crack tip and the strain fall-off slope.

[0026] As a further aspect of the present invention, the process of extracting the duration span exceeding the warning threshold within the pressure range data specifically involves: extracting a historical test data sequence; calculating the statistical mean and variance of the historical test data sequence; obtaining a preset deviation coefficient; summing the product of the statistical mean, the preset deviation coefficient, and the variance to generate the warning threshold; retrieving consecutive sampling nodes exceeding the warning threshold within the pressure range data; and performing a difference operation between the timestamp termination record and the timestamp start record corresponding to the consecutive sampling nodes to extract the duration span.

[0027] The process of calculating the corresponding time span for the total number of samples with statistical node values ​​lower than the preset low-temperature warning limit specifically involves: acquiring a baseline brittle transition temperature record and retrieving a safety reduction ratio parameter; multiplying the baseline brittle transition temperature record with the safety reduction ratio parameter to set the preset low-temperature warning limit; filtering out abnormal temperature nodes with values ​​lower than the preset low-temperature warning limit within the target temperature segment; and multiplying the cumulative total number of abnormal temperature nodes with the sampling interval parameter to obtain the corresponding time span.

[0028] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0029] S401: Call the duration of the temperature trough and the duration of the pressure peak, perform a division operation on the duration of the temperature trough and the duration of the pressure peak to extract the delay ratio, perform interval mapping based on the delay ratio to generate a low temperature hysteresis gating parameter, read the material database, retrieve the temperature trough data in the material database, and obtain the low temperature yield strength reduction coefficient.

[0030] S402: Perform a multiplication operation on the pressure fall-off time and the strain fall-off slope to calculate the deformation fall-off loss, perform a reciprocal operation on the deformation fall-off loss to extract residual parameters, and perform numerical conversion on the preset nominal stress based on the residual parameters to establish the stress retention amount;

[0031] S403: Call the stress retention amount, the low-temperature yield strength reduction factor, the crack depth value, the low-temperature hysteresis gating parameter, and the peak strain at the crack tip; calculate the product of the stress retention amount and the low-temperature yield strength reduction factor to extract the associated stress base; combine the peak strain at the crack tip and the crack depth value to construct a spatial geometric factor; and perform a weighted calculation on the product of the associated stress base and the spatial geometric factor according to the low-temperature hysteresis gating parameter to generate the equivalent stress retention amount.

[0032] As a further aspect of the present invention, the process of generating a cryogenic hysteresis gating parameter based on the interval mapping of the delay ratio specifically involves: obtaining a preset upper delay limit benchmark value and a preset lower delay limit benchmark value, wherein the preset upper delay limit benchmark value and the preset lower delay limit benchmark value are defined by extracting heat transfer delay time samples of similar marine cryogenic tanks under cryogenic test environments; when the delay ratio exceeds the preset upper delay limit benchmark value, extracting a saturation decay coefficient as the cryogenic hysteresis gating parameter, wherein the saturation decay coefficient is set according to the critical strain stagnation rate of ductility loss of the corresponding material in the material database; when the delay ratio is between the preset lower delay limit benchmark value and the preset upper delay benchmark value, calculating the difference between the delay ratio and the preset lower delay benchmark value, and extracting the ratio of the difference to a preset range as the cryogenic hysteresis gating parameter, wherein the preset range is set as the algebraic difference between the preset upper delay benchmark value and the preset lower delay benchmark value; when the delay ratio is lower than the preset lower delay benchmark value, directly using the delay ratio as the cryogenic hysteresis gating parameter.

[0033] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0034] S501: Call the dangerous location anchor point and the equivalent stress retention amount, scan and extract the outer surface wall thickness of the container according to the spatial coordinates of the dangerous location anchor point, perform a product operation on the equivalent stress retention amount and the outer surface wall thickness of the container, extract the square root parameter of the product value to generate the stress coefficient, and establish a type I stress intensity factor.

[0035] S502: Call the temperature trough data and the low temperature yield strength reduction factor to obtain the material database, retrieve the temperature trough data in the material database to extract the low temperature fracture toughness benchmark value, multiply the low temperature fracture toughness benchmark value and the low temperature yield strength reduction factor to establish a fracture toughness threshold value, call the Type I stress intensity factor, calculate the difference between the fracture toughness threshold value and the Type I stress intensity factor value, and obtain the fracture toughness margin difference value.

[0036] S503: Call the fracture toughness margin difference, the crack depth value, and the outer surface wall thickness of the container; divide the crack depth value by the outer surface wall thickness of the container to extract the thickness penetration ratio; set the penetration warning limit and the fracture critical line; compare the thickness penetration ratio with the penetration warning limit to extract the over-limit feature; compare the fracture toughness margin difference with the fracture critical line to extract the instability feature; aggregate the over-limit feature and the instability feature to perform classification mapping and generate the remaining strength assessment result.

[0037] As a further aspect of the present invention, the process of classification and mapping based on the aggregation of over-limit features and instability features specifically involves: obtaining the upper limit of the wall thickness tolerance range as a benchmark proportion coefficient to set the penetration warning limit; calling the lower quantile of the material fracture toughness distribution data to set the fracture critical line; determining the over-limit feature as a penetration indicator when the thickness penetration ratio is greater than the penetration warning limit, and determining the over-limit feature as a tolerance indicator when the thickness penetration ratio is not greater than the penetration warning limit; determining the instability feature as a brittle fracture indicator when the fracture toughness margin difference is less than the fracture critical line, and determining the instability feature as a safety indicator when the fracture toughness margin difference is not less than the fracture critical line; combining the penetration indicator with the brittle fracture indicator and the safety indicator respectively to generate the remaining strength assessment results for the structural failure state and the local repair state; and combining the tolerance indicator with the brittle fracture indicator and the safety indicator respectively to generate the remaining strength assessment results for the degraded operation state and the normal maintenance state.

[0038] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0039] In this invention, the residual strength judgment chain is introduced synchronously with acceleration, pressure, temperature, strain, and image information. The main vector of the load formed by the hull motion is associated with the weld direction and crack tangential relationship. The anchor point of the dangerous position is locked by means of the angle difference and strain peak value. The temporal characteristics such as pressure peak, low temperature trough, and strain fall are extracted around the anchor point. The transient pressure low temperature lag and crack tip response form the same evaluation benchmark. The low temperature yield reduction stress retains the crack depth, wall thickness, and fracture toughness threshold and is uniformly calculated as the fracture toughness margin difference value. This improves the timeliness of dangerous crack identification, enhances the adaptability of residual strength assessment to the continuous action of low temperature and transient impact, and reduces the risk of overly conservative or overly dangerous judgment. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the steps of the present invention;

[0042] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0043] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0044] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0045] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0046] Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0048] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0049] Please see Figure 1 This invention provides a deep learning-based intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank, comprising the following steps:

[0050] S1: Acquire longitudinal, lateral, and vertical acceleration data of the target ship using an accelerometer; obtain the internal pressure time series using a pressure sensor; obtain the external container wall temperature time series using a temperature sensor; acquire the strain peak value in the weld neighborhood using a strain gauge array; acquire the external container image of the marine cryogenic tank using a vision acquisition device; input the external container image of the marine cryogenic tank into a depth residual network model for analysis; and output the crack tangential angle, weld direction angle, crack depth value, and crack weld projection segment.

[0051] S2: Calculate the principal vector of the surface load and the normal direction vector based on longitudinal acceleration data, transverse acceleration data and vertical acceleration data. Compare the principal vector of the surface load with the weld orientation angle. Calculate the tangential angle difference between the crack tangential angle and the normal direction vector. Based on the tangential angle difference and the strain peak value in the weld neighborhood, screen the cracked weld projection segment to obtain the anchor point of the dangerous location.

[0052] S3: Based on the anchor point of the dangerous location, extract the interval data corresponding to the pressure time series inside the tank and the temperature time series of the outer container wall. Extract the pressure time series inside the tank to generate the pressure peak duration and pressure drop time. Extract the temperature time series of the outer container wall to generate temperature trough data and temperature trough duration. Collect the strain peak value and strain drop slope at the crack tip of the anchor point of the dangerous location through the strain gauge array.

[0053] S4: Compare the duration of the temperature trough with the duration of the pressure peak to generate a low-temperature hysteresis gating parameter. Retrieve the low-temperature yield strength reduction factor from the material database based on the temperature trough data. Calculate the stress retention based on the pressure fall-off time and strain fall-off slope. Calculate the equivalent stress retention based on the stress retention, the low-temperature yield strength reduction factor, and the crack depth value, combined with the low-temperature hysteresis gating parameter.

[0054] S5: Extract the outer wall thickness of the container at the anchor point of the dangerous location, calculate the Type I stress intensity factor based on the equivalent stress retention, extract the low temperature fracture toughness benchmark value of the corresponding temperature trough data from the material database, calculate the fracture toughness threshold value by combining the low temperature fracture toughness benchmark value with the low temperature yield strength reduction factor, calculate the difference between the Type I stress intensity factor and the fracture toughness threshold value to generate the fracture toughness margin difference, and make a judgment based on the fracture toughness margin difference, crack depth value and outer wall thickness of the container to generate the remaining strength assessment result;

[0055] Please see Figure 2 The specific steps of S1 are as follows:

[0056] S101: The system collects longitudinal, lateral, and vertical acceleration data of the ship using accelerometers, obtains the internal pressure time series using pressure sensors, obtains the external container wall temperature time series using temperature sensors, and monitors the strain peak value in the weld neighborhood using a strain gauge array. It performs time alignment processing on the longitudinal acceleration data, lateral acceleration data, vertical acceleration data, internal pressure time series, external container wall temperature time series, and weld neighborhood strain peak value to extract synchronization segments and splice them to generate a multi-source correlation matrix.

[0057] A fieldbus communication connection is established with the underlying hardware devices. Following a preset microsecond-level scanning cycle, the triaxial accelerometers of the microelectromechanical system (MEMS) located at designated measuring points on the hull are activated to continuously monitor and acquire longitudinal, lateral, and vertical acceleration data generated by the ship's impact from ocean waves in a Cartesian coordinate system. Simultaneously, a piezoresistive pressure transmitter installed in the gas phase space inside the cryogenic tank is triggered to acquire the tank pressure timing generated by vaporization at a fixed frequency of 50 Hz. A platinum resistance temperature detector attached to the lower part of the outer metal shell is activated to continuously probe the surface heat conduction state at a frequency of 10 Hz, acquiring the outer container wall temperature timing. Dynamic addressing demodulation is performed on the fiber optic strain gauge array deployed in the heat-affected zone of the outer surface weld to extract the strain peak value in the weld neighborhood during the local material yielding stage. To eliminate the inherent physical differences in sampling rate among multi-source heterogeneous sensors, time alignment processing was performed on longitudinal acceleration data, lateral acceleration data, vertical acceleration data, tank pressure time series, outer container wall temperature time series, and weld neighborhood strain peak values. This alignment operation used the 200 Hz high-frequency hardware timestamp of the triaxial accelerometer as the global synchronization reference timeline, and performed linear interpolation on the low-frequency sampled tank pressure time series and outer container wall temperature time series. During interpolation, the data values ​​corresponding to two adjacent valid sampling points were acquired, and the ratio of the difference between these two data values ​​to the corresponding time interval was calculated to extract the slope of change. Based on this slope, the corresponding virtual sampling values ​​were calculated and filled in at the missing time nodes on the global synchronization reference timeline. After frequency unification and completion, a data truncation time window of 10 seconds was set, and synchronization segments were extracted from the time-aligned continuous data stream. The acquired synchronization segments from each type of sensor were spliced ​​and combined column-wise to generate a multi-source correlation matrix containing multi-dimensional features. For example, when extracting a single 10-second synchronization segment, each data type is fixed at a uniform sampling rate of 200 Hz, containing a sequence of 2000 sampling points. The six columns of data—longitudinal acceleration, lateral acceleration, vertical acceleration, pressure, temperature, and peak strain—are horizontally concatenated to construct a multi-source correlation matrix containing 2000 rows and six columns of pure numerical sequences. The execution node of this calculation logic is to establish a unified time-domain calibration benchmark using the generated multi-source correlation matrix, thereby triggering cross-domain joint processing of multimodal data. To clarify the data acquisition configuration specifications for each type of sensor, the associated acquisition parameter details are introduced here.

[0058] Table 1 Configuration Table of Multi-Source Sensor Acquisition Parameters

[0059]

[0060] Table 1 shows the specific parameter constraints for multi-source data acquisition.

[0061] S102: Collect images of the outer container of the marine cryogenic tank around the storage tank using a visual acquisition device, call the multi-source correlation matrix to record time nodes, perform sequence cropping on the images of the outer container of the marine cryogenic tank to remove irrelevant images, extract the pixel values ​​in the retained images, perform grayscale conversion, and establish a structural mapping image.

[0062] The industrial camera hardware interface of the charge-coupled device (CCD) deployed at the nodes of the support frame around the storage tank continuously acquires images of the outer container of the marine cryogenic tank box around the storage tank at a frame rate of 60 frames per second. The multi-source correlation matrix constructed in the above steps is obtained. The extreme values ​​in the multi-source correlation matrix that record the tank pressure time series and the strain peak in the weld neighborhood exceeding the conventional fluctuation tolerance limit are retrieved, and the corresponding millisecond-level time nodes are extracted. The time nodes recorded in the multi-source correlation matrix are used to perform a sequence cropping operation on the images of the outer container of the marine cryogenic tank box. Using the extreme value time node as the central pivot, continuous video frame data with an extension of 2 seconds before and after this time is cropped. Irrelevant images are removed from the cropped continuous video frame data. The Laplacian gradient variance of the brightness values ​​of all pixels in the frame is calculated frame by frame. A variance feature quantity to characterize the image edge sharpness is extracted. A sharpness judgment benchmark with a lower limit of 150 for variance is set. Motion-blurred frames and occluded frames with Laplacian gradient variance values ​​lower than this benchmark are discarded, and effective target images with clear details are retained. Pixel values ​​within the retained image are extracted, and the red, green, and blue color channel components of each pixel are separated. Grayscale conversion calculations are performed on each separated channel component to obtain color channel weighting parameters, where the red channel weight is set to 0.299, the green channel weight to 0.587, and the blue channel weight to 0.114. The extracted channel color values ​​are multiplied and summed with their corresponding weighting parameters to obtain the pixel grayscale value in a single dimension. For example, if the extracted red component value of a specific pixel in the retained image is 150, the green component value is 100, and the blue component value is 50, and these are substituted into the grayscale conversion calculation logic, firstly, the product of 150 and 0.299 is calculated to obtain 44.85, the product of 100 and 0.587 is calculated to obtain 58.7, and the product of 50 and 0.114 is calculated to obtain 5.7. The three product values ​​are then summed to obtain the final converted grayscale value of the pixel, which is 109.25. The original full-color two-dimensional image matrix is ​​replaced with the calculated grayscale values ​​in sequence to establish a structure mapping image. The execution node of this calculation process is to accurately locate the visual abnormality region based on the extreme value of multimodal physical mutation, significantly reduce redundant image data streams, and construct a high-contrast baseline grayscale map layer for the lossless extraction of subsequent contour features.

[0063] S103: Input the structure mapping image into the deep residual network model to perform extraction processing, call the built-in convolution kernel of the deep residual network model to perform dot product operation on the pixel gradient to extract contour information, perform dimensionality reduction processing on the contour information through the pooling layer, and perform regression calculation on the dimensionality reduction result based on the fully connected layer to obtain the crack tangential angle, weld direction angle, crack depth value and crack weld projection segment.

[0064] A deep learning model, pre-optimized with parameters based on a large number of real-world defect maps, is loaded onto the computing platform. The structure mapping image established in the above process is read and input into a deep residual network model for extraction. The input layer of this deep residual network model receives fixed-size two-dimensional grayscale matrix data, which is then passed down to the initial convolutional layer. The built-in convolutional kernel of the deep residual network model is used to perform dot product operations on pixel gradients to extract contour information. The detailed logic of the convolution operation is clarified here: a 7x7 square convolutional kernel is extracted from the first layer. The stride parameter of the sliding window is set to 2. The kernel slides across the surface of the structure mapping image matrix region by region to extract the numerical matrix within the local receptive field. The local receptive field matrix is ​​multiplied one by one with the corresponding values ​​in the weight matrix inside the convolutional kernel. All the multiplied values ​​are then summed, and a constant bias parameter is added as the mapping output value of the central pixel node. A non-linearly modified linear unit activation function is used to filter out all negative response values, retaining only positive gradient responses, thus initially extracting the initial contour edge lines of the light-dark boundary. The data stream then passes through four sets of residual mapping modules connected in series. Each residual mapping module contains two data transmission paths. The main path uses two consecutive 3x3 convolutional kernels to extract deeper abstract shape features and performs batch data normalization. The bypass skip connection path diverts the input data across layers to the output of the main path for bitwise addition, ensuring that the fine crack texture at the bottom layer does not weaken with increasing network depth. After the data stream passes through all residual modules, the contour information is reduced in dimensionality by pooling layers. At the end of the network, a global average pooling layer is called to calculate the arithmetic mean of all pixel values ​​within the two-dimensional feature map of each channel, forcibly compressing the three-dimensional feature tensor into a one-dimensional continuous numerical vector. Finally, regression calculation is performed on the dimensionality reduction result based on fully connected layers. The one-dimensional continuous numerical vector passes through the output terminal containing four independent neurons. Each neuron performs fully connected weight multiplication and linear bias summation calculations on the data stream from the previous layer, corresponding to the solution of the physical quantities of the output environment. The output of the first neuron node value corresponds to the crack tangential angle, the second neuron node value corresponds to the weld direction angle, the third neuron node value corresponds to the crack depth value, and the fourth neuron node value corresponds to the crack length projection segment of the crack weld on the spatial coordinate axis.Regarding the preprocessing operations during the training phase of this deep residual network, 10,000 grayscale images of metal surfaces containing real crack annotations were extracted from the historical database as training data. Gaussian filtering was used to smooth background noise. The mean square error (MSE) between the model's output values ​​and the manually annotated real values ​​was calculated as the loss error evaluation criterion. An adaptive moment estimation algorithm was then called to perform gradient descent iterations. The gradient direction of the partial derivative of the loss error relative to the weights generated in each iteration was calculated, and the convolution kernel values ​​were updated in the reverse direction until the MSE value fell below the stagnation lower bound of 0.05. The execution node of this operational logic lies in automatically quantifying the geometric parameters of crack attributes in complex backgrounds using network weight parameters optimized from a large number of historical samples, directly providing input parameters for crack geometric evaluation.

[0065] Please see Figure 3 The specific steps of S2 are as follows:

[0066] S201: Perform three-dimensional coordinate system transformation calculations on longitudinal acceleration data, lateral acceleration data and vertical acceleration data, extract the corresponding transformation matrix projected mechanical components in the tangential plane, perform orthogonal rotation operations on the projected mechanical components to separate vertical vector features, and generate the surface load principal vector and normal direction vector;

[0067] The system acquires time-series waveform segments retained from the preceding acquisition cycle and performs three-dimensional coordinate system transformation calculations on the longitudinal, lateral, and vertical acceleration data. First, it retrieves the relative pitch, yaw, and roll angle parameters between the three-dimensional coordinate system at the corresponding measurement installation position and the tangential plane of the target shell. Based on these three attitude angle parameters, a spatial rotation transformation matrix is ​​constructed. The longitudinal, lateral, and vertical acceleration data are combined into a three-dimensional column vector. A standard matrix multiplication operation is then performed between the spatial rotation transformation matrix and this three-dimensional column vector. The projected mechanical components of the corresponding transformation matrix in the tangential plane are extracted. The first two numerical elements along the tangential direction in the matrix multiplication output vector are taken as the planar projected shear force data column, and the third numerical element along the direction perpendicular to the tangential plane is taken as the normal compression data column. The system acquires the planar projected shear force data series and the normal compression data series. An orthogonal rotation operation is performed on the projected mechanical components to separate the vertical vector features. For the planar projected shear force data series containing two elements, a 90-degree clockwise orthogonal coordinate rotation is performed with a local origin set in the tangential plane as the rotation center. This forcibly decomposes the originally tilted resultant force vector and maps it to two mutually perpendicular standard coordinate axes, completely separating the in-plane component parallel to the surface direction and the out-of-plane normal component perpendicular to the surface direction, generating the surface load principal vector and the normal direction vector. For example, the projected mechanical components calculated through three-dimensional coordinate system transformation have values ​​of 3 Newtons and 4 Newtons on the two coordinate axes of the tangential plane, respectively. Substituting these values ​​into the two-dimensional Euclidean norm distance calculation logic to solve for the square root of the sum of squares and extract the vector magnitude, the resultant force on the plane is calculated to be 5 Newtons, thus establishing the absolute amplitude of the surface load principal vector in the plane. Simultaneously, the third-dimensional vertical force value is extracted and directly assigned as the absolute amplitude of the normal direction vector. The logic execution node is to strip away the spatial projection error caused by the complex installation posture of the equipment itself, and forcibly standardize the random three-dimensional oscillation force caused by the external environment into the in-plane shear and out-of-plane tension physical model based on the crack occurrence location.

[0068] S202: Perform inverse trigonometric function mapping on the principal vector of the surface load to extract the principal vector azimuth angle, perform subtraction calculation between the principal vector azimuth angle and the weld direction angle to extract the direction deflection, perform three-dimensional calculation on the normal direction vector to extract the spatial angle parameter, and perform algebraic difference operation between the spatial angle parameter and the crack tangential angle to obtain the tangential angle difference.

[0069] Obtain the standard two-dimensional planar vector features generated in the previous step, and perform inverse trigonometric function mapping on the principal vector of the surface load to extract the azimuth angle of the principal vector. Extract the coordinate component values ​​of the principal vector of the surface load in the vertical axis direction and the coordinate component values ​​in the horizontal axis direction in the tangential plane. Divide the vertical axis component value by the horizontal axis component value to calculate the tangent ratio. Call the system's built-in arctangent mathematical mapping function to perform mapping calculation on the tangent ratio, obtain the obtained inverse trigonometric function radian result, and multiply it by the conversion coefficient to convert it into a conventional angle value, which is used as the azimuth angle of the principal vector. Extract the numerical result of the second neuron of the output node of the preceding deep residual network model to obtain the weld direction angle. Subtract the principal vector azimuth angle from the weld direction angle to extract the direction deflection. Directly calculate the absolute value of the algebraic difference generated by subtracting the weld direction angle value from the principal vector azimuth angle value. Spatial position tracing is performed on the separated normal parameters. Three-dimensional calculations are performed on the normal direction vector to extract spatial angle parameters. The cosine of the angle between the normal direction vector and the gravity plumb line in the global three-dimensional coordinate system is calculated, and then the inverse cosine value is obtained to obtain the pure spatial tilt angle as the spatial angle parameter. The numerical result of the first neuron of the output node of the previous deep residual network model is retrieved to obtain the crack tangential angle. The spatial angle parameter and the crack tangential angle are algebraically subtracted to obtain the tangential angle difference. For example, by detecting and extracting the surface load principal vector with a vertical axis component value of 10 and a horizontal axis component value of 10, the tangent ratio division is performed to obtain a ratio of 1. The arctangent function is called to obtain the inverse trigonometric function radian value of 3 / 4.14 (pi divided by 4), which is converted into an angle value to obtain the principal vector azimuth angle of 45 degrees. Subsequently, the weld orientation angle calculated by the depth residual network is extracted to be 30 degrees. Subtracting 30 degrees from 45 degrees yields a directional deflection of 15 degrees. The spatial angle parameter resulting from the extracted normal direction vector is calculated to be 50 degrees, and the crack tangential angle is extracted to be 20 degrees. A simple algebraic difference subtraction operation is performed between the spatial angle parameter 50 degrees and the crack tangential angle 20 degrees, resulting in a final tangential angle difference of 30 degrees. The key point of this logic is that it completely quantifies the complex three-dimensional spatial stress environment into an angular difference relative to the given geometric position of the metal defect, providing a geometrical angular reference parameter for subsequently determining whether the stress environment is sufficient to tear the crack tip.

[0070] S203: Extract the coordinate sequence of the point set within the projection segment of the cracked weld, call the tangential angle difference to set the angle filtering threshold range, set the stress anomaly discrimination line according to the strain peak value in the weld neighborhood, perform bidirectional verification on each node in the point set coordinate sequence, remove the data of nodes exceeding the limit, aggregate the coordinate nodes that fall within the angle filtering threshold range and cross the inner side of the stress anomaly discrimination line, and obtain the dangerous location anchor point.

[0071] The process of setting the angle filtering threshold range by calling the tangential angle difference is as follows: extract the anisotropic deflection angle parameter set based on the material principal axis deformation rate; multiply the anisotropic deflection angle parameter and the tangential angle difference to generate the angle fluctuation difference; perform an addition operation on the tangential angle difference and the angle fluctuation difference to generate the upper limit boundary value of the angle; perform a subtraction operation on the tangential angle difference and the angle fluctuation difference to generate the lower limit boundary value of the angle; combine the upper limit boundary value and the lower limit boundary value of the angle to set the angle filtering threshold range.

[0072] The process of setting the stress anomaly discrimination line based on the strain peak value in the weld neighborhood is as follows: extract the local maximum strain value and local minimum strain value recorded within the strain peak value in the weld neighborhood; calculate the average value of the local maximum strain value and local minimum strain value to generate the reference stress line; obtain the stress sensitivity coefficient value set based on the yield strength reduction ratio of the container metal base material under low temperature service conditions; and calculate and set the stress anomaly discrimination line by multiplying the reference stress line and the stress sensitivity coefficient value.

[0073] The pixel grid array of the structure mapping image is read, and the coordinate sequence of the point set within the projection segment of the cracked weld is extracted. The projection segment line is discretized into a set of two-dimensional coordinate pairs arranged at a fixed pixel spacing. The tangential angle difference is used to set the angle filtering threshold range. This process involves a series of basic parameter calculations. First, the factory material test report of the target container steel is queried to extract the anisotropic deflection angle parameter set based on the material's principal axis deformation rate. The anisotropic deflection angle parameter is multiplied by the tangential angle difference obtained in the previous step to generate the angle fluctuation difference. The tangential angle difference and the angle fluctuation difference are added to generate the upper limit boundary value of the angle; simultaneously, the tangential angle difference and the angle fluctuation difference are subtracted to generate the lower limit boundary value of the angle. The upper and lower limit boundary values ​​are combined to set the angle filtering threshold range, forming a numerical range including both upper and lower limits. A stress anomaly discrimination line is set based on the strain peak value in the weld neighborhood. The process of setting the stress anomaly discrimination line based on the strain peak value in the weld neighborhood specifically includes the following steps: extreme value calculation logic for the acquired time-series data; traversing the historical time period cache sequence to extract the local maximum and minimum strain values ​​recorded within the strain peak value in the weld neighborhood; calculating the sum of the local maximum and minimum strain values, dividing it by 2 to calculate the average value to generate the baseline stress line; obtaining the stress sensitivity coefficient value set based on the yield strength reduction ratio of the container metal base material under low-temperature service conditions; and multiplying the baseline stress line and the stress sensitivity coefficient value to calculate and set the stress anomaly discrimination line. The generated angle filtering threshold range and stress anomaly discrimination line are then used to perform a bidirectional check on each node in the point set coordinate sequence, extracting the stress angle state associated with each node and comparing it with the actual strain state for conditional judgment. During the check, node values ​​that do not meet the set parameters are removed from the data of nodes exceeding the limits. Nodes falling within the angle filtering threshold range and whose corresponding measurement point values ​​cross the inner side of the stress anomaly discrimination line are aggregated to obtain the dangerous location anchor points, which are then output as a three-dimensional coordinate point set containing precise spatial planar positioning. A specific example is introduced to illustrate this. The anisotropic deflection angle parameter is set to 0.1. This 0.1 is multiplied by the tangential angle difference of 30 degrees to generate an angle fluctuation difference of 3 degrees. Adding 30 degrees to 3 degrees generates an upper limit boundary value of 33 degrees, and subtracting 30 degrees from 3 degrees generates a lower limit boundary value of 27 degrees. Thus, the angle screening threshold range is set to 27 degrees to 33 degrees. In the stress discrimination setting, the local maximum strain value is extracted as 400 microstrains, and the local minimum strain value is extracted as 100 microstrains. The sum of the two and the result divided by 2 yields an average value of 250 microstrains, which is used as the baseline stress line. The stress sensitivity coefficient value of the container metal base material is set to 1.2. The 250 microstrains are multiplied by 1.2 to calculate the final value of the stress anomaly discrimination line, which is set to 300 microstrains.When performing a check on a coordinate node, if the node's stress angle is 29 degrees, falling within the 27-33 degree range; and the strain value at that measuring point is 350 microstrains, meeting the requirement of exceeding the 300 microstrain anomaly threshold, then that node is retained and aggregated as a critical location anchor point. The significance of this logical node determination lies in eliminating false anomalies caused by conventional stress concentrations and extracting the location features of the weakest structures most prone to fracture propagation.

[0074] Please see Figure 4 The specific steps of S3 are as follows:

[0075] S301: Obtain the anchor point of the danger location, the pressure time series inside the tank and the temperature time series of the outer container wall. Establish a cutoff window based on the time attribute recorded by the anchor point of the danger location. Perform segmentation on the pressure time series inside the tank based on the cutoff window to extract the pressure interval data. Extract the duration span of exceeding the warning threshold within the pressure interval data. Calculate the time difference required for the pressure extreme point to drop to the stable baseline. Generate the pressure peak duration and pressure drop time.

[0076] The process of extracting the duration span exceeding the warning threshold within the pressure range data specifically involves: extracting historical test data sequences and calculating the statistical mean and variance of the historical test data sequences; obtaining a preset deviation coefficient and summing the product of the statistical mean, the preset deviation coefficient, and the variance to generate the warning threshold; retrieving continuous sampling nodes exceeding the warning threshold within the pressure range data and performing a difference operation between the timestamp termination record and the timestamp start record corresponding to the continuous sampling nodes to extract the duration span.

[0077] The process involves acquiring the three-dimensional coordinate parameters of the previously confirmed hazardous location anchor points, as well as the tank pressure time series and outer container wall temperature time series under a unified clock node. A cutoff window is established based on the time attributes of the hazardous location anchor point records, starting with the microsecond-level timestamp mapped when the anchor point is determined to be in an abnormal hazardous state, and extending outwards to generate a fixed-duration 60-second data reading window. Based on the cutoff window, the tank pressure time series is segmented to extract pressure interval data, copying only the set of local pressure sampling fluctuation values ​​within this 60-second period from the global long-sequence pressure data stream. Within the pressure interval data, the duration span exceeding the warning threshold is extracted, and the time difference required for the pressure extreme point to drop to the stable baseline is calculated. This difference calculation process performs a traversal scan of the descent slope after reaching the maximum peak, generating two key time-domain feature parameters: the pressure peak duration and the pressure drop time. The process of extracting the duration span exceeding the warning threshold within the pressure interval data specifically includes rigorous statistical threshold establishment and clock difference calculation logic. First, extract the normalized historical test data sequence under normal operating conditions where no structural abnormality alarms occurred within the previous 72 hours, and calculate the statistical average value of the historical test data sequence. Then, calculate the square of the absolute difference between the value of each sampling point in the historical test data sequence and the statistical average value. Sum all the squares and divide by the total number of sampling points to calculate the variance value. Obtain a preset deviation coefficient, and directly multiply the preset deviation coefficient by the variance value to obtain the deviation compensation amount. Sum the statistical average value with the product term containing the variance value to calculate the warning threshold. Within the pressure range data, sequentially search for consecutive sampling nodes whose sampling node values ​​are greater than the warning threshold, marking the initial node that triggered the threshold exceedance state and the last node that ended the exceedance state and returned to below the threshold. Perform a difference operation on the timestamp termination record and timestamp start record corresponding to the consecutive sampling nodes, calculate the difference between the two absolute time points, and extract the duration span. For example, the statistical average value of the historical stable test data sequence extracted from the tank is 2 MPa. The variance of the sum of the squares of the differences between all historical sampling points and this 2 MPa is calculated to be 0.04. A preset deviation coefficient is set to 3 (this parameter is set based on the statistical principle of tolerating fluctuations of 3 standard deviations under normal conditions). The product of the preset deviation coefficient 3 and the variance 0.04 is calculated to obtain 0.12 MPa. The average value of 2 MPa is summed with the deviation compensation product 0.12 MPa to generate the final warning threshold set at 2.12 MPa. Within the pressure range data, a continuous high-pressure impact pulse segment with values ​​greater than 2.12 MPa is retrieved. The end timestamp of this impact segment (152000 milliseconds) and the beginning timestamp (151000 milliseconds) are extracted, and the difference between them is calculated to give a duration of 1000 milliseconds, which is taken as the pressure peak duration.The key to this step is that instead of using extreme points to measure the degree of pressure damage, it precisely captures and quantifies the effective time range of the continuous energy accumulation effect of high-pressure load on weak nodes.

[0078] S302: Call the interception window to extract the target temperature segment from the time sequence of the external container wall temperature. Perform size comparison and search on the node values ​​within the target temperature segment to extract the lowest extreme value. Calculate the corresponding time span by counting the total number of samples whose node values ​​are lower than the preset low temperature warning limit, and obtain the temperature trough data and the duration of the temperature trough.

[0079] The process of calculating the corresponding time span for the total number of samples with statistical node values ​​lower than the preset low-temperature warning limit is as follows: obtain the baseline brittle transition temperature record and retrieve the safety reduction ratio parameter; multiply the baseline brittle transition temperature record and the safety reduction ratio parameter to set the preset low-temperature warning limit; filter out abnormal temperature nodes with values ​​lower than the preset low-temperature warning limit within the target temperature segment; multiply the cumulative total number of abnormal temperature nodes with the sampling interval parameter to obtain the corresponding time span.

[0080] The previous step establishes a time-domain sliding window reference with a fixed extended time span. The intercept window is invoked to extract the target temperature segment from the external container wall temperature time series, reducing the massive global temperature data stream to a limited set of local temperature monitoring nodes within a 60-second range of the intercept window. The lowest extreme value is extracted by comparing the node values ​​within the target temperature segment. Bubble sorting or sequential search is performed on the platinum resistance feedback values ​​at each time point within the segment to filter out high-temperature interference points and obtain abnormally cold node data with the lowest absolute value. The total number of samples with node values ​​below the preset low-temperature warning limit is calculated to determine the corresponding time span. A loop counter is incremented to accumulate the number of samples falling into this extreme cold temperature range, obtaining the temperature trough data and its duration. The process of calculating the corresponding time span for the total number of samples with node values ​​below the preset low-temperature warning limit specifically includes parameter transformation calculations for the material's brittle physical boundary. The composition analysis report and performance parameter test table of the specific model of container metal shell material are queried to obtain the reference brittle transition temperature record for the metal reaching the glass transition brittleness critical state. The safety reduction ratio parameter, mandated by the container safety design specifications, is retrieved. The baseline brittle transition temperature record value is multiplied by the safety reduction ratio parameter using a purely numerical operation to lower the threshold and set a preset low-temperature warning lower limit. The system traverses and scans within the target temperature segment, filtering to determine if the currently measured temperature value is lower than the preset low-temperature warning lower limit, and counts and archives any abnormal temperature nodes that meet the criteria. The cumulative total of abnormal temperature nodes is multiplied by the fixed sampling interval parameter set by the equipment to obtain the corresponding time span. To illustrate the specific parameter generation details, the baseline brittle transition temperature record value extracted from the material property table is set to -160 degrees Celsius, and the specification safety reduction ratio parameter is retrieved as 0.95. Multiplying the baseline brittle transition temperature record value of -160 degrees Celsius by the safety reduction ratio parameter 0.95 calculates that the preset low-temperature warning lower limit value has been raised to a threshold of -152 degrees Celsius. Within the extracted target temperature segment of 60 seconds, temperature measurement points with values ​​below -152 degrees Celsius were selected point by point, resulting in a cumulative total of 150 abnormal temperature nodes. The fixed sampling interval parameter of the temperature monitoring channel was set to 0.1 seconds (corresponding to a 10 Hz sampling frequency). The total number of abnormal temperature nodes (150) was multiplied by the sampling interval parameter (0.1 seconds) to calculate the corresponding time span of 15 seconds. This lowest temperature extreme and its sustained 15-second duration were used as the corresponding temperature trough data and its duration. The key to this calculation is the accurate acquisition of the true cold penetration time effect of extreme low temperatures on material brittleness degradation, rather than transient surface temperature fluctuations. A reduction factor was used to prevent missed detections due to temperature conduction hysteresis.

[0081] S303: For anchor points at dangerous locations, the strain gauge array is activated to collect dynamic deformation signals. Peak retrieval is performed on the dynamic deformation signals to extract the maximum peak parameter. Discrete sampling points are extracted for the falling bands that cross the maximum peak parameter. First-order differential operation is performed on the sequence of adjacent nodes of the discrete sampling points to obtain the peak strain value and strain fall-off slope at the crack tip.

[0082] Based on the previously extracted and located three-dimensional weak coordinate set, strain gauge arrays are activated to collect dynamic deformation signals at anchor points in hazardous locations. The system sends a high-frequency wake-up command to the grating sensors attached to the physical area adjacent to the anchor points in hazardous locations, collecting analog quantities of the metal elongation and contraction variables of the container shell under high-pressure cryogenic conditions. These are then converted from analog to digital via photoelectric conversion to generate a continuous, high-density digitally sampled dynamic deformation signal data stream. Peak retrieval is performed on the dynamic deformation signal to extract the maximum peak parameter. The entire timeframe of the load-induced abrupt response is scanned, and the magnitudes of each strain value are compared to locate and extract the maximum reading value that produces the local elastic deformation peak. Discrete sampling points are extracted for the energy release decline band after crossing the maximum peak parameter node. First-order differential operations are performed on the sequence of adjacent nodes of the discrete sampling points. The strain value at the discrete sampling point corresponding to the next sampling time is obtained and subtracted from the strain value at the discrete sampling point corresponding to the previous adjacent sampling time to obtain the difference in strain value variation. Simultaneously, the timestamp value corresponding to the next sampling time is obtained and subtracted from the timestamp value corresponding to the previous sampling time to obtain the time interval duration. The ratio of the strain value variation difference to the time interval duration is calculated to obtain the rate of change. The peak strain at the crack tip caused by shell tension and the strain fall slope reflecting the rate of strain recovery are obtained. For example, the maximum peak parameter, i.e., the peak strain at the crack tip, is found to be 1200 microstrains. Two adjacent discrete sampling points are extracted in the falling band, where the strain value measured at 2.1 seconds is 1100 microstrains and the strain value measured at 2.2 seconds is 1050 microstrains. The difference between the latter point (1050) and the former point (1100) is calculated to be -50 microstrains. The difference between the previous time point of 2.2 seconds and the previous time point of 2.1 seconds is calculated to be 0.1 seconds. Dividing this difference of -50 microstrain by the time interval of 0.1 seconds yields a first-order differential result of -500 microstrain per second. The absolute value of this rate of change, 500 microstrain per second, is extracted as the strain rate slope characterizing the material's ability to recover elasticity under cryogenic conditions. This execution node transcends the limitations of static stress analysis, reflecting the hindered recovery phenomenon caused by elastic decay of metallic materials under complex loads in ultra-low temperature environments through a dynamic slope, providing dynamic parameter basis for strength assessment.

[0083] Please see Figure 5 The specific steps of S4 are as follows:

[0084] S401: Call the duration of temperature trough and the duration of pressure peak, perform a division operation on the duration of temperature trough and the duration of pressure peak to extract the delay ratio, perform interval mapping based on the delay ratio, generate low temperature hysteresis gating parameters, read the material database, retrieve temperature trough data in the material database, and obtain the low temperature yield strength reduction factor.

[0085] The process of generating cryogenic hysteresis gating parameters based on interval mapping of delay ratios is as follows: First, a preset upper and lower delay baseline value is obtained. These values ​​are defined by extracting heat transfer delay time samples from similar marine cryogenic tanks under cryogenic testing conditions. Second, when the delay ratio exceeds the preset upper delay baseline value, a saturation decay coefficient is extracted as the cryogenic hysteresis gating parameter. This coefficient is set based on the critical strain stagnation rate for ductility loss of the corresponding material in the material database. Third, when the delay ratio is between the preset lower and upper delay baseline values, the difference between the delay ratio and the lower delay baseline value is calculated. The ratio of this difference to a preset range is extracted as the cryogenic hysteresis gating parameter. The preset range is set as the algebraic difference between the preset upper and lower delay baseline values. Fourth, when the delay ratio is below the preset lower delay baseline value, the delay ratio is directly used as the cryogenic hysteresis gating parameter.

[0086] The process involves calling the values ​​describing the duration of each physical field's effect obtained from previous processing, specifically the durations of temperature troughs and pressure peaks. A division operation is then performed on these two durations to extract the delay ratio parameter. Based on this calculated delay ratio parameter, a multi-level threshold interval mapping judgment rule is executed to generate a low-temperature hysteresis gating parameter characterizing the tailing effect of the ductile-brittle transition in metals caused by cryogenic environments. Further, by reading a background standard material database storing the fracture properties of various alloy steels under different environmental conditions, the process retrieves the previously located temperature trough data values ​​within this database based on unique key values. This data is then matched to find the constitutive characteristic data of the metal under these extremely low temperature conditions, and the low-temperature yield strength reduction coefficient is extracted. The process of generating the low-temperature hysteresis gating parameter based on interval mapping using the delay ratio specifically includes branch judgment execution logic. First, preset upper and lower delay benchmark values ​​are obtained. These benchmark values ​​are specifically set by extracting test samples of heat transfer delay times recorded by test probes from different batches of similar marine cryogenic tanks under extremely cold gas-phase cyclic testing conditions established in a standardized laboratory. These samples represent the time intervals during which the inner wall temperature suddenly drops to the outer wall temperature in a synchronized step response. Gaussian fitting is then performed on the distribution of this large sample test time series to define the distribution region. In extreme hysteresis cases where the delay ratio exceeds the preset upper delay benchmark value, a fixed calibrated saturation decay coefficient is extracted as the cryogenic hysteresis gating parameter output. The saturation decay coefficient is a constant set based on the critical strain stagnation rate of ductility loss exhibited by the corresponding metallic material at its worst temperature before plastic deformation occurs and fractures due to loss of ductility. When the delay ratio is within the normal fluctuation range, i.e., between the preset lower and upper delay benchmark values, the difference between the current delay ratio and the preset lower delay benchmark value is calculated. The ratio of this calculated difference to the preset range is extracted as the cryogenic hysteresis gating parameter. The preset range is fixed as the simple algebraic difference between the preset upper and lower delay benchmark values. In cases of extremely sensitive conduction where the delay ratio is lower than the preset lower delay benchmark value, the current delay ratio is directly retained as the cryogenic hysteresis gating parameter. Using a specific example and parameter values, the preset upper delay benchmark value is set to 0.8, the preset lower delay benchmark value is set to 0.3, and the preset range is 0.8 minus 0.3, which equals 0.5. The duration of the temperature trough is extracted as 15 seconds, and the duration of the pressure peak is extracted as 25 seconds; the two are divided to extract a delay ratio of 0.6. Since 0.6 falls between 0.3 and 0.8, the intermediate interval mapping condition branch is triggered. The difference between the delay ratio 0.6 and the preset lower limit of delay baseline 0.3 is calculated to obtain 0.3. The difference 0.3 is extracted and divided by the preset range span 0.5 to calculate the ratio, resulting in a low-temperature hysteresis gating parameter value of 0.6.If the calculated delay ratio is 0.9, exceeding the upper limit of 0.8, then a constant set by the critical strain stagnation rate for ductility loss in the material library, such as 1.0, is directly extracted as the saturation attenuation coefficient output. If the calculated delay ratio is 0.2, below the lower limit of 0.3, then 0.2 is directly output. Simultaneously, for the temperature trough data of -152 degrees Celsius, the material library is searched to obtain the low-temperature yield strength reduction factor of 0.85 at this temperature. The execution node of this calculation logic objectively quantifies the time difference misalignment phenomenon when the external low-temperature environment and the internal high-pressure impact coincide on the outer surface of the shell, avoiding the amplified stress assessment error caused by the direct linear addition of asynchronous environmental factors.

[0087] S402: Perform a multiplication operation on the pressure fall-off time and strain fall-off slope to calculate the deformation fall-off loss, perform a reciprocal operation on the deformation fall-off loss to extract residual parameters, and perform numerical conversion on the preset nominal stress based on the residual parameters to establish the stress retention amount;

[0088] The preceding steps extract values ​​reflecting the strain recovery rate and time domain width. A pure numerical multiplication operation is performed on the pressure fallback time and strain fallback slope to calculate the deformation fallback loss, representing the unrecovered energy dissipation state. Since this increase in deformation fallback loss indicates permanent and irreversible damage caused by internal micro-lattice dislocation slip, the reciprocal operation is performed on the deformation fallback loss value (i.e., the ratio of the constant 1 to this value) to extract residual parameters. This inversely maps the negatively lost deformation loss to the ratio of residual potential energy retained within the metal material that cannot be released. Based on the calculated residual parameters, a numerical multiplication is performed on the pre-determined nominal stress, which is determined by looking up a table based on the container's rated working pressure and design dimensions under ideal conditions, to establish and obtain the stress retention amount for subsequent residual mechanics calculations. In a specific example, the pressure fallback time after the pressure peak subsides is retrieved as 3 seconds, and the absolute value of the material strain fallback slope parameter is obtained as 500 microstrains per second. Multiplying 3 and 500 yields a deformation recovery loss of 1500 microstrain. Taking the reciprocal of 1500 (1 divided by 1500) extracts a residual parameter of approximately 0.00067. The pre-set nominal stress constant for the hazardous area, as defined in the equipment design specifications, is 300 MPa. Multiplying this residual parameter of 0.00067 with the pre-set nominal stress of 300 MPa, the final calculated unreleased stress retention at this moment is 0.2 MPa. The significance of this node operation lies in abandoning the traditional approach of relying solely on static pressure gauge readings to extrapolate stress. It quantifies the secondary failure driving force parameter latent within the microcrack tip due to multiple impacts by utilizing the hysteresis characteristics of micro-deformation recovery.

[0089] S403: Call the stress retention amount, low temperature yield strength reduction factor, crack depth value, low temperature hysteresis gating parameter and crack tip strain peak value, calculate the product of stress retention amount and low temperature yield strength reduction factor to extract the associated stress base, combine crack tip strain peak value and crack depth value to construct spatial geometric factor, and perform weighted calculation on the product of associated stress base and spatial geometric factor according to low temperature hysteresis gating parameter to generate equivalent stress retention amount;

[0090] The process involves obtaining the comprehensive environmental characterization parameters derived independently from multiple preceding stages. This includes accessing the stress retention value, low-temperature yield strength reduction factor, crack depth value, low-temperature hysteresis gating parameter, and peak strain at the crack tip. The product of the stress retention value and the low-temperature yield strength reduction factor is calculated to extract the associated stress baseline reflecting the shrinkage of the material's physical boundaries. The peak strain at the crack tip and the crack depth value are combined to construct a spatial geometric factor characterizing the amplification effect of local stress concentration. Based on the low-temperature hysteresis gating parameter, a weighted calculation is performed on the product term constructed from the associated stress baseline and the spatial geometric factor, proportionally shrinking to generate the equivalent stress retention amount remaining under combined severe working conditions. Calculations are performed using the parameter values ​​to obtain a stress retention value of 0.2 MPa and a low-temperature yield strength reduction factor of 0.85. Multiplying 0.2 and 0.85 yields an associated stress baseline of 0.17 MPa. The crack depth is set to 3 mm, and the representative parameter for the peak strain at the crack tip is set to 1.5. The combined calculation multiplies 1.5 and 3 to construct a spatial geometric factor value of 4.5. Based on the extracted low-temperature hysteresis gating parameter value of 0.6, the associated stress base of 0.17 is multiplied by the spatial geometric factor 4.5 to obtain a product term of 0.765. Finally, the low-temperature hysteresis gating parameter 0.6 is used to perform a weighted multiplication of this product term value of 0.765, generating the final equivalent stress retention value of 0.459 MPa. The execution of this node logic thoroughly integrates all multimodal data (deformation retention potential energy, material embrittlement effect coefficient, cryogenic environment time hysteresis characteristics, and visual crack depth size) into a one-dimensional scalar, ensuring that the final strength parameter fully absorbs the influence weights of various heterogeneous defect factors.

[0091] Please see Figure 6 The specific steps of S5 are as follows:

[0092] S501: Call the anchor point of the dangerous location and the equivalent stress holding amount, scan and extract the outer surface wall thickness of the container according to the spatial coordinates of the anchor point of the dangerous location, perform a product operation on the equivalent stress holding amount and the outer surface wall thickness of the container, extract the square root parameter of the product value to generate the stress coefficient, and establish the Type I stress intensity factor.

[0093] The final strength parameters are constructed based on unified spatial anchor point coordinate data. The three-dimensional coordinate parameters of the anchor points at critical locations are used, along with the equivalent stress retention value calculated in the above process. An ultrasonic thickness gauge is used to scan and extract the actual remaining wall thickness of the container's outer surface in the corresponding area, based on the spatial coordinate parameters of the anchor points at critical locations. The equivalent stress retention value and the actual remaining wall thickness of the container's outer surface are multiplied. The square root parameter mathematical operation is performed on the resulting product value to generate a comprehensive stress coefficient reflecting the macroscopic fracture propagation dynamics. Finally, a Type I stress intensity factor characteristic value specifically for characterizing this type of open-type defect state is established and obtained. An example is provided: the equivalent stress retention value is 0.459 MPa. Based on the anchor point positioning, the ultrasonic ranging probe detects and extracts the actual wall thickness of the container's outer surface after corrosion or wear, which is 20 mm. A simple multiplication operation is performed between the equivalent stress retention value of 0.459 and the container's outer surface wall thickness of 20 mm, yielding a result of 9.18. The square root of the product result 9.18 was extracted to calculate a stress coefficient of 3.03 (rounded to two decimal places). This comprehensive stress coefficient value of 3.03, along with its implied physical units, was converted to standard fracture mechanics dimensions to establish a Type I stress intensity factor of 3.03 MPa √m under this monitoring condition. This process, by linking thickness loss data with stress accumulation data through physical location, intuitively reflects the concentrated growth trend of destructive energy caused by residual stress accumulating as the container wall thins.

[0094] S502: Call the temperature trough data and the low temperature yield strength reduction factor to obtain the material database. Search the temperature trough data in the material database to extract the low temperature fracture toughness benchmark value. Multiply the low temperature fracture toughness benchmark value and the low temperature yield strength reduction factor to establish the fracture toughness threshold value. Call the Type I stress intensity factor to calculate the difference between the fracture toughness threshold value and the Type I stress intensity factor value to obtain the fracture toughness margin difference value.

[0095] Based on the extreme temperature threshold, a baseline threshold for the material's remaining ability to resist fracture damage is established. The parameters of the extreme cold state temperature trough established in the previous steps, along with the low-temperature yield strength reduction factor reflecting performance degradation, are retrieved. A database access link is established to obtain the backend interface of the metallic material database. The temperature layer corresponding to the temperature trough data is searched in the internal index tree of the material database, and the low-temperature fracture toughness benchmark value obtained from the original material standard test under this environmental condition is extracted. The low-temperature fracture toughness benchmark value and the low-temperature yield strength reduction factor are multiplied to establish a baseline parameter for the fracture toughness threshold value for the current specific defect node. The dynamic real-time value of the Type I stress intensity factor, which was just constructed and output in the previous step, is retrieved. The difference between the fracture toughness threshold value baseline parameter value and the real-time value of the Type I stress intensity factor is calculated to obtain the fracture toughness margin difference used to assess the safety margin of the container. For example, based on the temperature trough data of -152 degrees Celsius, the low-temperature fracture toughness benchmark value for the corresponding special low-temperature steel at this time is found to be 45 MPa √m. The low-temperature yield strength reduction factor is obtained as 0.85. Multiplying the low-temperature fracture toughness baseline value of 45 by the low-temperature yield strength reduction factor of 0.85, the fracture toughness threshold value after attenuation under severe working conditions is calculated to be reduced to 38.25 MPa √m. The currently monitored Type I stress intensity factor value is retrieved as 3.03 MPa √m. The difference between the material resistance threshold value of 38.25 and the current failure driving force value of 3.03 is calculated, yielding a fracture toughness margin difference value of 35.22 MPa √m. This execution logic clearly defines the red line of structural failure, providing a quantitative distance margin value for subsequent judgment on whether it has entered the unstable brittle fracture edge region.

[0096] S503: Call the fracture toughness margin difference, crack depth value and container outer surface wall thickness, divide the crack depth value by the container outer surface wall thickness to extract the thickness penetration ratio, set the penetration warning limit and fracture critical line, compare the thickness penetration ratio with the penetration warning limit to extract the over-limit feature, compare the fracture toughness margin difference with the fracture critical line to extract the instability feature, aggregate the over-limit feature and instability feature to perform classification mapping, and generate the remaining strength assessment result;

[0097] The process of classification and mapping based on the aggregation of over-limit and instability features is as follows: First, the upper limit of the wall thickness tolerance range is obtained as a benchmark proportion coefficient to set the penetration warning limit. Second, the lower quantile of the material fracture toughness distribution data is used to set the fracture critical line. Third, when the thickness penetration ratio is greater than the penetration warning limit, the over-limit feature is determined as a penetration indicator; when the thickness penetration ratio is not greater than the penetration warning limit, the over-limit feature is determined as a tolerance indicator. Fourth, when the fracture toughness margin difference is less than the fracture critical line, the instability feature is determined as a brittle fracture indicator; when the fracture toughness margin difference is not less than the fracture critical line, the instability feature is determined as a safety indicator. Fifth, the penetration indicator is combined with the brittle fracture indicator and the safety indicator to generate the remaining strength assessment results for the structural failure state and the local repair state, respectively. Sixth, the tolerance indicator is combined with the brittle fracture indicator and the safety indicator to generate the remaining strength assessment results for the degraded operation state and the normal maintenance state, respectively.

[0098] All assessment parameters are converged, and a health status classification is implemented. The parameters representing the redundancy resistance to damage, including the fracture toughness margin difference, crack depth extracted visually, and the actual wall thickness of the container's outer surface transmitted in real-time by the probe, are used. The crack depth is divided by the actual wall thickness to calculate the ratio, generating a thickness penetration ratio parameter representing the physical penetration state. The container's factory inspection standards are consulted to obtain tolerance requirements and set penetration warning limits. Simultaneously, a fracture critical line is set based on large-sample failure statistics. The extracted thickness penetration ratio is compared with the set penetration warning limits to identify over-limit or within-safety limits. The fracture toughness margin difference is compared with the set fracture critical line to identify instability features indicating a risk of instantaneous bursting. Based on the over-limit and instability features, binary logic is used for aggregation and mapping classification to generate and output a residual strength assessment result status report for direct presentation at the maintenance terminal. The process of classification and mapping based on the aggregation of over-limit and instability features specifically involves setting stringent parameter limits and combination logic. The upper limit of the allowable wall thickness tolerance range for equipment manufacturing (e.g., the maximum allowable proportion of grinding loss of 0.2) is obtained from the factory drawings and directly used as the benchmark proportion coefficient to set the penetration warning limit to 0.2. The lower quantile of the material's fracture toughness distribution attenuation statistics over its entire life cycle (e.g., set as an extremely weak resistance value of 5 MPa √m) is called to set the fracture critical line to 5 MPa √m. During feature judgment, if the calculated thickness penetration ratio is greater than the set penetration warning limit value, the system will determine and assign the over-limit feature to the measuring point as a penetration marker; if the calculated thickness penetration ratio is not greater than the penetration warning limit value, the system will determine and assign the over-limit feature as a safe tolerance marker. When the extracted fracture toughness margin difference is less than the set fracture critical line value, the instability characteristic is determined to be brittle fracture due to the imminent risk of disintegration; when the extracted fracture toughness margin difference is not less than the set fracture critical line value, the instability characteristic is determined to be safe. The extracted penetration markers are matched with potential brittle fracture markers and safety markers in pairs. For the combination of "penetration marker superimposed on brittle fracture marker," a residual strength assessment result corresponding to the highest risk "structural failure state" is generated; for the combination of "penetration marker superimposed on safety marker," a residual strength assessment result corresponding to the "partial repair state" is generated. The tolerance markers are matched with brittle fracture markers and safety markers in pairs. For the combination of "tolerance marker superimposed on brittle fracture marker," a residual strength assessment result corresponding to the "degraded operation state" is generated; for the combination of "tolerance marker superimposed on safety marker," a residual strength assessment result corresponding to the "normal maintenance state" is generated. Using the data obtained from previous calculations as a complete example, the current crack depth is extracted to be 3 mm, and the outer surface wall thickness of the container is extracted to be 20 mm.Dividing 3 mm by 20 mm, the thickness penetration ratio parameter is calculated to be 0.15. The penetration warning limit is set to 0.2. Since the thickness penetration ratio of 0.15 is not greater than the penetration warning limit of 0.2, the over-limit feature is triggered as a tolerance indicator. The fracture toughness margin difference value obtained from the previous calculation is retrieved, which is 35.22 MPa √m. The fracture critical line is set to 5 MPa √m. Since the fracture toughness margin difference of 35.22 is not less than the fracture critical line of 5, the instability feature is triggered as a safety indicator. The tolerance indicator and the safety indicator are logically aggregated and matched in a matching rule table, and the remaining strength assessment result is directly mapped and output as "normal maintenance state".

[0099] Table 2 Residual Strength Assessment Status Classification Mapping Rules

[0100]

[0101] Table 2 shows the evaluation and classification output obtained by aggregating various discrimination indicators. The fundamental significance of this evaluation and grading logic lies in its accurate mapping of the state by integrating the toughness margin quantitative index constructed from parameters of the cryogenic high-pressure hidden environment, rather than solely relying on the external crack length evaluation equipment. Comparative verification results from 500 actual tests show that the early detection rate of containers in a partially repaired state using this aggregation classification mapping process is 28% higher than the traditional method of simple visual reproduction, and no cases of brittle fracture failure are missed.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A deep learning-based intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank, characterized in that, Includes the following steps: S1: Collect longitudinal acceleration data, lateral acceleration data and vertical acceleration data of the target ship, and simultaneously acquire the tank pressure time series, outer container wall temperature time series, weld neighborhood strain peak and marine cryogenic tank outer container image, and input the image into the depth residual network model, output crack tangential angle, weld direction angle, crack depth value and crack weld projection segment. S2: Calculate the principal vector of the surface load and the normal direction vector based on the longitudinal acceleration data, the transverse acceleration data and the vertical acceleration data. Combine the weld direction angle, the difference in tangential angle and the strain peak value in the weld neighborhood to screen the cracked weld projection segment and obtain the anchor point of the dangerous location. The specific steps of S2 are as follows: S201: Perform three-dimensional coordinate system transformation calculation on the longitudinal acceleration data, the lateral acceleration data and the vertical acceleration data, extract the corresponding transformation matrix projected mechanical components in the tangential plane, perform orthogonal rotation operation on the projected mechanical components to separate vertical vector features, and generate surface load principal vector and normal direction vector; S202: Perform inverse trigonometric function mapping on the principal vector of the surface load to extract the azimuth angle of the principal vector, perform three-dimensional calculation on the normal direction vector to extract the spatial angle parameter, and perform algebraic difference operation on the spatial angle parameter and the crack tangential angle to obtain the tangential angle difference. S203: Extract the coordinate sequence of the point set within the projection segment of the cracked weld, call the tangential angle difference to set the angle filtering threshold range, set the stress anomaly discrimination line according to the strain peak value in the weld neighborhood, perform bidirectional testing on each node in the point set coordinate sequence, remove the data of nodes exceeding the limit, aggregate the coordinate nodes that fall within the angle filtering threshold range and cross the inner side of the stress anomaly discrimination line, and obtain the dangerous location anchor point; S3: Based on the anchor point of the dangerous location, extract the interval data of the pressure time series inside the tank and the temperature time series of the outer container wall, extract the pressure peak duration, pressure fall time, temperature trough data, and temperature trough duration, and evaluate the peak strain and strain fall slope at the crack tip. The specific steps for S3 are as follows: S301: Obtain the anchor point of the danger location, the pressure time series inside the tank and the temperature time series of the outer container wall. Establish a cutoff window based on the time attribute recorded by the anchor point of the danger location. Perform segmentation on the pressure time series inside the tank based on the cutoff window to extract pressure interval data. Extract the duration span of exceeding the warning threshold within the pressure interval data. Calculate the time difference required for the pressure extreme point to drop to the stable baseline. Generate the pressure peak duration and pressure drop time. S302: Call the interception window to extract the target temperature segment from the time sequence of the outer container wall temperature, perform size comparison and search on the node values ​​in the target temperature segment to extract the lowest extreme value, count the total number of samples with node values ​​lower than the preset low temperature warning limit, calculate the corresponding time span, and obtain the temperature trough data and the duration of the temperature trough. S303: For the anchor point at the dangerous location, the strain gauge array is activated to collect dynamic deformation signals. The maximum peak parameter is extracted by peak retrieval of the dynamic deformation signal. Discrete sampling points are extracted for the falling band that crosses the maximum peak parameter. The first-order differential operation is performed on the sequence of adjacent nodes of the discrete sampling points to obtain the peak strain value at the crack tip and the strain fall-off slope. S4: Compare the duration of the temperature trough with the duration of the pressure peak to generate a low-temperature hysteresis gating parameter. Retrieve the low-temperature yield strength reduction factor from the material database and calculate the equivalent stress retention by combining the stress retention and crack depth numerical values. The specific steps of S4 are as follows: S401: Call the duration of the temperature trough and the duration of the pressure peak, perform a division operation on the duration of the temperature trough and the duration of the pressure peak to extract the delay ratio, perform interval mapping based on the delay ratio to generate a low temperature hysteresis gating parameter, read the material database, retrieve the temperature trough data in the material database, and obtain the low temperature yield strength reduction coefficient. S402: Perform a multiplication operation on the pressure fall-off time and the strain fall-off slope to calculate the deformation fall-off loss, perform a reciprocal operation on the deformation fall-off loss to extract residual parameters, and perform numerical conversion on the preset nominal stress based on the residual parameters to establish the stress retention amount; S403: Call the stress retention amount, the low-temperature yield strength reduction factor, the crack depth value, the low-temperature hysteresis gating parameter, and the crack tip strain peak value; calculate the product of the stress retention amount and the low-temperature yield strength reduction factor to extract the associated stress base; combine the crack tip strain peak value and the crack depth value to construct a spatial geometric factor; and perform a weighted calculation on the product of the associated stress base value and the spatial geometric factor according to the low-temperature hysteresis gating parameter to generate the equivalent stress retention amount. S5: Extract the outer surface wall thickness of the container, calculate the Type I stress intensity factor, fracture toughness threshold value and fracture toughness margin difference based on the equivalent stress retention, and make a judgment based on the fracture toughness margin difference, the crack depth value and the outer surface wall thickness of the container to generate the remaining strength assessment result.

2. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 1, characterized in that, The specific steps of S1 are as follows: S101: The longitudinal acceleration data, lateral acceleration data and vertical acceleration data of the ship are collected by an accelerometer; the pressure time series inside the tank is obtained by a pressure sensor; the temperature time series of the outer container wall is obtained by a temperature sensor; and the strain peak value in the weld neighborhood is monitored by a strain gauge array. Time alignment processing is performed on the longitudinal acceleration data, the lateral acceleration data, the vertical acceleration data, the pressure time series inside the tank, the temperature time series of the outer container wall and the strain peak value in the weld neighborhood to extract synchronization segments and splice them to generate a multi-source correlation matrix. S102: Collect images of the outer container of the marine cryogenic tank box around the storage tank through a visual acquisition device, call the multi-source correlation matrix to record time nodes, perform sequence cropping and remove irrelevant images from the images of the outer container of the marine cryogenic tank box, extract and retain pixel values ​​in the images, perform grayscale conversion, and establish a structural mapping image. S103: Input the structure mapping image into the deep residual network model for extraction processing, call the built-in convolution kernel of the deep residual network model to perform dot product operation on the pixel gradient to extract contour information, perform dimensionality reduction processing on the contour information through the pooling layer, and perform regression calculation on the dimensionality reduction result based on the fully connected layer to obtain the crack tangential angle, weld direction angle, crack depth value and crack weld projection segment.

3. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 1, characterized in that, The process of setting the angle filtering threshold range by calling the tangential angle difference specifically involves: extracting the anisotropic deflection angle parameter set based on the material principal axis deformation rate; multiplying the anisotropic deflection angle parameter with the tangential angle difference to generate an angle fluctuation difference; performing an addition operation on the tangential angle difference and the angle fluctuation difference to generate an upper limit boundary value for the angle; performing a subtraction operation on the tangential angle difference and the angle fluctuation difference to generate a lower limit boundary value for the angle; and combining the upper limit boundary value and the lower limit boundary value to set the angle filtering threshold range. The process of setting the stress anomaly discrimination line based on the strain peak value in the weld neighborhood specifically involves: extracting the local maximum strain value and the local minimum strain value recorded within the strain peak value in the weld neighborhood; calculating the average value of the local maximum strain value and the local minimum strain value to generate a reference stress line; obtaining the stress sensitivity coefficient value set based on the yield strength reduction ratio of the container metal base material under low temperature service conditions; and multiplying the reference stress line and the stress sensitivity coefficient value to calculate and set the stress anomaly discrimination line.

4. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 1, characterized in that, The process of extracting the duration span exceeding the warning threshold within the pressure range data specifically involves extracting historical test data sequences and calculating the statistical mean and variance of the historical test data sequences. Obtain a preset deviation coefficient, sum the product of the statistical average value, the preset deviation coefficient, and the variance value to generate the warning threshold; retrieve continuous sampling nodes with values ​​greater than the warning threshold within the pressure range data, perform a difference operation between the timestamp termination record and the timestamp start record corresponding to the continuous sampling nodes, and extract the duration span; The process of calculating the corresponding time span for the total number of samples with statistical node values ​​lower than the preset low-temperature warning limit specifically involves: acquiring a baseline brittle transition temperature record and retrieving a safety reduction ratio parameter; multiplying the baseline brittle transition temperature record with the safety reduction ratio parameter to set the preset low-temperature warning limit; filtering out abnormal temperature nodes with values ​​lower than the preset low-temperature warning limit within the target temperature segment; and multiplying the cumulative total number of abnormal temperature nodes with the sampling interval parameter to obtain the corresponding time span.

5. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 1, characterized in that, The process of generating cryogenic hysteresis gating parameters by performing interval mapping based on delay ratio specifically involves: obtaining a preset upper delay limit benchmark value and a preset lower delay limit benchmark value, wherein the preset upper delay limit benchmark value and the preset lower delay limit benchmark value are defined by extracting heat transfer delay time samples of similar marine cryogenic tanks under cryogenic test environments; when the delay ratio exceeds the preset upper delay limit benchmark value, a saturation decay coefficient is extracted as the cryogenic hysteresis gating parameter, wherein the saturation decay coefficient is set according to the critical strain stagnation rate of ductility loss of the corresponding material in the material database; when the delay ratio is between the preset lower delay limit benchmark value and the preset upper delay limit benchmark value, the difference between the delay ratio and the preset lower delay limit benchmark value is calculated, and the ratio of the difference to a preset range span is extracted as the cryogenic hysteresis gating parameter, wherein the preset range span is set as the algebraic difference between the preset upper delay limit benchmark value and the preset lower delay limit benchmark value; If the delay ratio is lower than the preset lower limit benchmark value, the delay ratio will be directly used as the low temperature hysteresis gating parameter.

6. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 1, characterized in that, The specific steps of S5 are as follows: S501: Call the dangerous location anchor point and the equivalent stress retention amount, scan and extract the outer surface wall thickness of the container according to the spatial coordinates of the dangerous location anchor point, perform a product operation on the equivalent stress retention amount and the outer surface wall thickness of the container, extract the square root parameter of the product value to generate the stress coefficient, and establish a type I stress intensity factor. S502: Call the temperature trough data and the low temperature yield strength reduction factor to obtain the material database, retrieve the temperature trough data in the material database to extract the low temperature fracture toughness benchmark value, multiply the low temperature fracture toughness benchmark value and the low temperature yield strength reduction factor to establish a fracture toughness threshold value, call the Type I stress intensity factor, calculate the difference between the fracture toughness threshold value and the Type I stress intensity factor value, and obtain the fracture toughness margin difference value. S503: Call the fracture toughness margin difference, the crack depth value, and the outer surface wall thickness of the container; divide the crack depth value by the outer surface wall thickness of the container to extract the thickness penetration ratio; set the penetration warning limit and the fracture critical line; compare the thickness penetration ratio with the penetration warning limit to extract the over-limit feature; compare the fracture toughness margin difference with the fracture critical line to extract the instability feature; aggregate the over-limit feature and the instability feature to perform classification mapping and generate the remaining strength assessment result.

7. The intelligent assessment method for the residual strength of the outer container of a marine cryogenic tank based on deep learning as described in claim 6, characterized in that, The process of classification and mapping based on the aggregation of over-limit features and instability features specifically involves: obtaining the upper limit of the wall thickness tolerance range as a benchmark proportion coefficient to set the penetration warning limit; calling the lower quantile of the material fracture toughness distribution data to set the fracture critical line; determining the over-limit feature as a penetration identifier when the thickness penetration ratio is greater than the penetration warning limit, and determining the over-limit feature as a tolerance identifier when the thickness penetration ratio is not greater than the penetration warning limit. When the difference in fracture toughness margin is less than the fracture critical line, the instability feature is determined to be a brittle fracture indicator; when the difference in fracture toughness margin is not less than the fracture critical line, the instability feature is determined to be a safety indicator. The penetration indicator is combined with the brittle fracture indicator and the safety indicator respectively to generate the remaining strength assessment results for the structural failure state and the partial repair state. The tolerance indicator is combined with the brittle fracture indicator and the safety indicator respectively to generate the remaining strength assessment results for the degraded operation state and the normal maintenance state.

Images