Non-destructive testing and life prediction system for crane welded joints

Abstract

This invention belongs to the field of nondestructive testing and structural health monitoring technology, specifically disclosing a nondestructive testing and life prediction system for crane welded joints. The system includes a multi-source heterogeneous data fusion and acquisition module, a damage evolution modeling module based on a physical information neural network, a multi-scale coupled remaining life prediction module, and an adaptive decision-making and feedback control module. By fusing manufacturing, service, and online inspection data, a digital twin model with embedded physical constraints is constructed to quantify and simulate the damage state. Monte Carlo life prediction is then performed using probabilistic load spectrum extrapolation. Finally, based on the life probability distribution, a graded early warning system is triggered, and feedback is used to optimize the model. This achieves closed-loop management of real-time welded joint status perception, accurate damage evolution, and forward-looking remaining life prediction.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing and structural health monitoring technology, specifically relating to a non-destructive testing and life prediction system for crane welded joints. Background Technology

[0002] As equipment used in modern industrial production and logistics transportation, the structural safety and reliability of lifting machinery are directly related to the smooth implementation of major engineering projects and the safety of personnel and property. Welded joints, as connecting parts in the metal structure of cranes, are subjected to complex alternating loads and heavy impacts over long periods, making them a major potential source of structural fatigue failure and fracture accidents. Therefore, health monitoring and remaining life assessment of welded joints are crucial technical aspects for ensuring the safe operation of cranes.

[0003] Non-destructive testing and life prediction of crane welded joints is an important technical direction for ensuring their structural integrity. This technology aims to assess internal defects and damage levels of welded joints by collecting and analyzing their condition signals without damaging the structure, and based on this, predict their future safe service life, thereby realizing a shift from passive maintenance to proactive predictive maintenance.

[0004] In existing technologies, the monitoring and evaluation of crane welded joints mainly rely on periodic or ad-hoc offline non-destructive testing, such as ultrasonic testing and magnetic particle testing. However, this offline testing mode suffers from lag, making it difficult to promptly capture early hidden dangers such as microcracks that initiate and propagate under alternating loads, and failing to achieve real-time status awareness during service. Furthermore, existing testing methods are typically conducted in isolation, failing to correlate and integrate with the manufacturing process parameters of the welded joint, real-time load spectra during service, and environmental condition data. This results in insufficient depth of damage mechanism analysis and difficulty in establishing accurate damage evolution models. Consequently, the accuracy of life prediction based on testing results is limited, failing to provide precise and forward-looking guidance for maintenance decisions and safe operation of equipment such as heavy-duty cranes, thus constituting a pressing technical problem to be solved in the field of crane safety operation and maintenance. Summary of the Invention

[0005] The purpose of this invention is to provide a non-destructive testing and life prediction system for crane welded joints, in order to solve the problems of lagging offline testing mode, difficulty in fusion of multi-source heterogeneous data, insufficient accuracy of damage evolution model and poor forward-looking life prediction in the prior art.

[0006] This invention provides a non-destructive testing and life prediction system for crane welded joints, comprising: The multi-source heterogeneous data fusion acquisition module is used to continuously collect multi-dimensional status data of the welded joint throughout its entire life cycle and output structured multi-source fusion data packets. The damage evolution modeling module based on physical information neural network is used to receive multi-source fusion data packets from the multi-source heterogeneous data fusion acquisition module and construct a digital twin model of damage evolution that integrates prior physical knowledge and real-time monitoring data. The multi-scale coupled remaining service life prediction module is used to receive the quantized damage state descriptor output by the damage evolution modeling module based on the physical information neural network, and to perform cross-scale prediction from micro-damage evolution to macro-remaining service life. The adaptive decision and feedback control module is used to receive the remaining lifetime probability distribution function output by the multi-scale coupled remaining lifetime prediction module, and generate graded early warning instructions and maintenance decision suggestions.

[0007] Preferably, the multi-source heterogeneous data fusion acquisition module integrates a manufacturing process data acquisition unit, a real-time service condition monitoring unit, and an online non-destructive testing unit; The manufacturing process data acquisition unit is used to extract the original process parameters of the weld joint from the manufacturing execution system; The real-time service condition monitoring unit is used to collect dynamic load spectrum, ambient temperature, ambient humidity and structural vibration acceleration signals of the welded joint in real time during service through a sensor network deployed on the crane structure. The online non-destructive testing unit is used to periodically excite and receive ultrasonic guided wave signals at a preset sampling frequency by using an embedded ultrasonic guided wave sensor array and a distributed fiber optic grating sensor network integrated into the weld joint, while simultaneously monitoring the strain field distribution and temperature gradient changes in the joint area.

[0008] Preferably, the damage evolution modeling module includes a deep physical information neural network, the network structure of which consists of an input layer, multiple hidden layers and an output layer; The input vectors received by the input layer include normalized real-time load spectrum, environmental parameters, ultrasonic guided wave signal feature vectors, and strain field data. The multiple hidden layers embed physical constraint equations derived from fracture mechanics and fatigue damage theory. These equations participate in network training as part of the loss function in the form of partial differential equation residuals. The training process of the damage evolution modeling module is divided into a pre-training phase and an online adaptive training phase. In the pre-training phase, the network is trained in a supervised manner using multi-source data from the historical failure case library and the corresponding final failure modes. In the online adaptive training phase, the multi-source fusion data packets collected in real time are input into the pre-trained model. The predicted state of the model at the previous moment is used as the initial condition. Combined with the measured data input at the current moment, the internal weight parameters of the model are dynamically corrected by solving the residual minimization problem of the embedded physical constraint equation. The output of the damage evolution modeling module is a quantified damage state descriptor of the weld joint at the current moment.

[0009] Preferably, the remaining lifetime prediction module includes a load spectrum extrapolation unit, a damage evolution path simulation unit, and a lifetime distribution calculation unit. The load spectrum extrapolation unit, based on real-time monitoring data of historical service conditions, uses time series analysis and machine learning algorithms to predict the probabilistic load spectrum that the welded joint will bear within a preset prediction period in the future. The damage evolution path simulation unit takes the quantized damage state descriptor at the current moment as the initial state and the future probabilistic load spectrum output by the load spectrum extrapolation unit as the input boundary condition, and drives the digital twin model updated through online adaptive training to perform multiple Monte Carlo simulations. Each simulation generates a damage evolution trajectory from the current state to the preset failure threshold. The lifetime distribution calculation unit performs statistical analysis on all damage evolution trajectories generated by Monte Carlo simulation, calculates the number of cycles or time corresponding to each trajectory reaching the failure threshold, and then fits the probability distribution function of the remaining lifetime.

[0010] Preferably, the adaptive decision and feedback control module has built-in multi-level early warning threshold determination logic; The multi-level early warning threshold determination logic compares the expected remaining lifespan with the preset 3-level early warning threshold in real time, and triggers different levels of early warning signals and decision instructions based on the comparison results. The adaptive decision-making and feedback control module uses the execution effect of early warning decisions at all levels and subsequent actual detection results as feedback data to send back to the damage evolution modeling module based on physical information neural network, which is used to fine-tune the physical constraint equations.

[0011] Preferably, in the online non-destructive testing unit, the arrangement of the embedded ultrasonic guided wave sensor array follows the following principles: The piezoelectric ceramic sensors in the array operate in a transceiver mode and are distributed in a grid-like topology that can cover the heat-affected zone and fusion line of the weld joint. The excitation signal is a sinusoidal pulse modulated by a narrowband Hanning window; The received ultrasonic guided wave signal is first subjected to time-frequency analysis through wavelet packet transform to extract the energy attenuation coefficient, propagation time offset, and nonlinear harmonic component amplitude within the frequency band, which together constitute the characteristic vector of the ultrasonic guided wave signal.

[0012] Preferably, the physical constraint equations embedded in the deep physical information neural network are a hybrid form of the fatigue crack propagation rate equation based on the Paris formula and the damage accumulation equation based on continuous damage mechanics. The hybrid form of the damage accumulation equation uses the amplitude of the nonlinear harmonic component extracted from the feature vector of the ultrasonic guided wave signal as an internal variable characterizing the microscopic plastic deformation and crack closure effect, and uses the strain field gradient monitored by the distributed fiber grating sensor network as a correction parameter for the local stress intensity factor.

[0013] Preferably, the machine learning algorithm used in the load spectrum extrapolation unit is a long short-term memory network; The Long Short-Term Memory Network takes time-series segments of historical load spectra as input and learns the dynamic change patterns of load amplitude, mean, and frequency. During extrapolation prediction, multiple future load spectrum scenarios with different probabilities are generated simultaneously, and these scenarios serve as input boundary conditions for the Monte Carlo simulation.

[0014] Preferably, the threshold in the multi-level early warning threshold determination logic is dynamically adjusted; First, a Level 1 threshold is set, and the Level 2 and Level 3 warning thresholds are adaptively adjusted based on the variance of the remaining life probability distribution function. When the variance of the lifespan prediction increases, the system automatically raises the warning threshold; the magnitude of the threshold adjustment and the change in variance are determined through a preset linear mapping relationship.

[0015] Preferably, the system is deployed on a cloud-edge collaborative computing platform with a microservice architecture; The multi-source heterogeneous data fusion acquisition module and the signal preprocessing part in the online non-destructive testing unit are deployed at the edge computing node close to the crane. The damage evolution modeling module based on physical information neural network, the multi-scale coupled remaining lifetime prediction module, and the adaptive decision and feedback control module are deployed on a cloud server cluster. Data synchronization and command issuance between the cloud and the edge are achieved through an industrial-grade wireless communication protocol.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a multi-source heterogeneous data fusion acquisition module, for the first time deeply integrates the manufacturing process parameters of welded joints, real-time service conditions, and online non-destructive testing signals under a unified time scale, breaking the data isolation in traditional monitoring. This correlational acquisition of data throughout the entire lifecycle provides a solid data foundation for a deep understanding of the complete chain from the emergence of process defects to the evolution of service damage, enabling damage analysis to trace back to the initial state and improving the root cause and comprehensiveness of condition assessment.

[0017] 2. This invention creatively proposes a damage evolution modeling module based on a physical information neural network, embedding prior physical knowledge such as fracture mechanics and fatigue damage theory into a deep neural network in the form of constraint equations. This design not only utilizes the powerful nonlinear fitting capability of neural networks to learn complex patterns from massive monitoring data, but also ensures the scientific and rational nature of the model's evolution direction through the constraints of physical laws. It overcomes the physical unreliability problem that may occur in purely data-driven models when training data is insufficient or when extrapolating working conditions, achieving a complementary advantage between data-driven and physical model-driven approaches, thereby constructing a digital twin of the welded joint with higher accuracy and stronger generalization ability.

[0018] 3. This invention achieves cross-scale probabilistic prediction of remaining lifetime from microscopic damage parameters to macroscopic remaining lifetime through a multi-scale coupled remaining lifetime prediction module. This module employs Monte Carlo simulation, combined with extrapolated future probabilistic load spectra, to generate a large number of possible damage evolution paths, ultimately outputting a probability distribution of remaining lifetime rather than a single deterministic value. This probabilistic prediction method quantitatively reveals the uncertainties in lifetime prediction, providing decision-makers with richer and more reliable risk information, enabling maintenance decisions to be based on scientific risk assessment rather than vague empirical judgments.

[0019] 4. This invention establishes a complete closed loop from prediction to decision-making and model optimization through an adaptive decision-making and feedback control module. The system dynamically triggers tiered early warnings based on the probability distribution of remaining service life and can adaptively adjust the early warning threshold according to prediction uncertainty, achieving intelligent and flexible decision-making. More importantly, the feedback from decision-making effects and actual results is used to continuously optimize the damage evolution model, forming a self-learning and self-improvement reinforcement loop. This allows the system's overall prediction and decision-making capabilities to continuously evolve with service time, maintaining high accuracy and high reliability over the long term. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the principle framework of the damage evolution modeling module based on physical information neural network in this invention; Figure 3This is a logical flowchart of the multi-scale coupled remaining lifetime prediction module in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the adaptive decision-making and feedback control module in this invention. Detailed Implementation Example

[0021] This invention provides a non-destructive testing and life prediction system for crane welded joints, the overall technical architecture of which is shown in the attached figure. Figure 1 As shown, this system achieves high-precision and forward-looking monitoring and management of the entire lifecycle status of welded joints by constructing a closed-loop technology system from real-time perception of microscopic defects to dynamic prediction of macroscopic lifespan. The system includes a multi-source heterogeneous data fusion acquisition module, a damage evolution modeling module based on physical information neural networks, a multi-scale coupled remaining life prediction module, and an adaptive decision-making and feedback control module. These modules are tightly coupled through strictly defined data interfaces and logical timing, forming a highly collaborative and adaptive intelligent monitoring and prediction system.

[0022] The multi-source heterogeneous data fusion acquisition module serves as the data entry point for the entire system, continuously collecting multi-dimensional status data throughout the entire lifecycle of the welded joint. This module is further subdivided into three sub-units: a manufacturing process data acquisition unit, a real-time service condition monitoring unit, and an online non-destructive testing unit. These sub-units correspond to the status information of the welded joint during the manufacturing stage, the initial service stage, and long-term operation, respectively. The manufacturing process data acquisition unit interfaces with the manufacturing execution system through a standardized interface, automatically extracting and recording the original process parameters generated during the manufacturing process of the welded joint. These parameters include, but are not limited to, welding current, welding voltage, welding speed, welding material grade, bevel type, and post-weld heat treatment curve.

[0023] All parameters are stored in a structured data format and include metadata such as timestamps, operator identifiers, and equipment numbers to ensure data traceability and integrity. The real-time service condition monitoring unit uses a sensor network deployed across the crane structure to collect dynamic load spectra, ambient temperature, ambient humidity, and structural vibration acceleration signals from the welded joints during service. The dynamic load spectrum is acquired by a combination of strain gauge arrays and force sensors installed at the connection between the main beam and the outriggers, with a sampling frequency of at least 1000 Hz. Ambient temperature and humidity are collected once per second by industrial-grade temperature and humidity sensors. The structural vibration acceleration signal is acquired by a triaxial accelerometer at a sampling rate of 2000 Hz for subsequent analysis of the structural dynamic response characteristics.

[0024] The multi-source heterogeneous data fusion acquisition module includes an online non-destructive testing unit, employing a composite sensing scheme of an embedded ultrasonic guided wave sensor array and a distributed fiber optic grating sensor network. The piezoelectric ceramic sensors in the ultrasonic guided wave sensor array operate in a transceiver mode, distributed according to a grid-like topology that covers the heat-affected zone and fusion line of the weld joint, ensuring blind-spot-free coverage of the area. The excitation signal uses a narrowband Hanning window modulated sinusoidal pulse with an adjustable center frequency, typically ranging from 50 kHz to 200 kHz, with a pulse width of 5 cycles. The received ultrasonic guided wave signal is first subjected to wavelet packet transform for time-frequency analysis, extracting the energy attenuation coefficient, propagation time offset, and nonlinear harmonic component amplitude within a specific frequency band, collectively forming the ultrasonic guided wave signal feature vector.

[0025] A distributed fiber Bragg grating sensor network is deployed along the weld seam and heat-affected zone, with a density of one sensor point every 10 millimeters, to monitor the strain field distribution and temperature gradient changes in the joint area in real time, achieving a spatial resolution of 1 millimeter and a strain measurement accuracy better than 1 microstrain. This multi-source heterogeneous data fusion acquisition module incorporates a data synchronization and time-stamp alignment engine, employing a Precise Time Protocol (PTP) to perform millisecond-level time alignment of data streams from different sources with different sampling rates. After initial filtering, denoising, and feature extraction at the edge computing nodes, all raw data is encapsulated into structured multi-source fusion data packets and uploaded to a cloud server cluster via an industrial-grade wireless communication protocol for use by subsequent modules.

[0026] The damage evolution modeling module based on a physical information neural network receives multi-source fusion data packets from a multi-source heterogeneous data fusion acquisition module and constructs a digital twin model of damage evolution that integrates prior physical knowledge and real-time monitoring data. This damage evolution modeling module includes a deep physical information neural network, whose network structure consists of an input layer, multiple hidden layers, and an output layer. The input vector received by the input layer includes normalized real-time load spectra, environmental parameters, ultrasonic guided wave signal feature vectors, and strain field data.

[0027] The real-time load spectrum is converted into a stress amplitude sequence and a mean stress sequence; environmental parameters include the current ambient temperature and humidity; the ultrasonic guided wave signal feature vector contains the previously extracted energy attenuation coefficient, propagation time offset, and nonlinear harmonic component amplitude; strain field data uses the local maximum principal strain and its gradient as input. Multiple hidden layers embed physical constraint equations derived from fracture mechanics and fatigue damage theory. These equations participate in network training as part of the loss function in the form of partial differential equation residuals. The training process of this damage evolution modeling module is divided into two stages.

[0028] The first stage is the pre-training stage, which utilizes multi-source data and corresponding final failure modes from a historical failure case library to conduct supervised training on the network, enabling the network to initially learn the complex mapping relationship between damage characteristics and macroscopic failures. The historical failure case library contains at least 500 sets of complete lifecycle data, covering different welding processes, different service environments, and different failure modes.

[0029] The second stage is the online adaptive training stage, in which the multi-source fusion data packets collected in real time are input into the pre-trained model. The predicted state of the model at the previous time step is used as the initial condition. Combined with the measured data input at the current time step, the internal weight parameters of the model are dynamically corrected by solving the residual minimization problem of the embedded physical constraint equation, so as to realize the adaptive synchronous update of the damage state of the welded joint between the digital twin model and the physical entity.

[0030] The output of the damage evolution modeling module is a quantified damage state descriptor of the weld joint at the current moment. The descriptor is a four-dimensional vector, whose components represent the equivalent size, location, orientation of defects such as cracks, and the cumulative percentage of the current damage level relative to the initial state. The principle framework of this damage evolution modeling module is attached. Figure 2 As shown, the complete logical chain of data input, physical constraint embedding, two-stage training mechanism and state output is clearly demonstrated.

[0031] The multi-scale coupled remaining service life prediction module receives a quantified damage state descriptor output from the damage evolution modeling module based on a physical information neural network, and performs cross-scale predictions from microscopic damage evolution to macroscopic remaining service life. This remaining service life prediction module further includes a load spectrum extrapolation unit, a damage evolution path simulation unit, and a life distribution calculation unit. Based on real-time monitoring data of historical service conditions, the load spectrum extrapolation unit uses time series analysis and machine learning algorithms to predict the probabilistic load spectrum that the welded joint will experience within a preset prediction period.

[0032] The load spectrum extrapolation unit employs a Long Short-Term Memory (LSTM) network, which takes time-series segments of historical load spectra as input and learns the dynamic variation patterns of load amplitude, mean, and frequency. During extrapolation prediction, multiple future load spectrum scenarios with different probabilities are simultaneously generated. These scenarios serve as input boundary conditions for Monte Carlo simulations to encompass uncertainties in future service conditions. The damage evolution path simulation unit uses the quantized damage state descriptor at the current moment as the initial state and the future probabilistic load spectrum output by the load spectrum extrapolation unit as input boundary conditions, driving the online adaptively trained and updated digital twin model to perform multiple Monte Carlo simulations. Each simulation generates a damage evolution trajectory from the current state to a preset failure threshold.

[0033] The failure threshold is defined as the equivalent crack size reaching a critical value (e.g., 10 mm) or the cumulative damage percentage reaching 95%. At least 10,000 simulations are performed to ensure the convergence of statistical results. The lifetime distribution calculation unit statistically analyzes all damage evolution trajectories generated by Monte Carlo simulations, calculating the number of cycles or time corresponding to each trajectory reaching the failure threshold, and then fitting a probability distribution function for the remaining lifetime. This probability distribution function outputs the expected value, variance, and lifetime quantiles at different confidence levels of the remaining safe cycles. The logical flow of this remaining lifetime prediction module is attached. Figure 3 As shown, a multi-scale coupled prediction mechanism from load extrapolation, path simulation to lifetime distribution calculation is fully presented.

[0034] The adaptive decision and feedback control module receives the remaining life probability distribution function output by the multi-scale coupled remaining life prediction module and generates tiered early warning instructions and maintenance decision suggestions. This module incorporates multi-level early warning threshold determination logic. It compares the expected remaining life with preset three-level early warning thresholds in real time. The first-level threshold is set at 80% of the design life, the second at 50%, and the third at 20%. When the expected remaining life is greater than the first-level threshold, the system determines the joint is in a safe state and only outputs a routine monitoring report. When the expected remaining life is between the first and second-level thresholds, the system generates a yellow warning signal and suggests shortening the sampling frequency of the online non-destructive testing unit, for example, from once per hour to once every 15 minutes, to increase the monitoring density of the joint. When the expected remaining life is between the second and third-level thresholds, the system generates an orange warning signal and outputs a suggested preventative maintenance window and a list of maintenance measures. When the expected remaining lifespan is less than the Level 3 threshold, the system generates a red alarm signal and immediately outputs a mandatory shutdown and inspection command, while simultaneously triggering the crane's safety interlock mechanism.

[0035] The thresholds in the multi-level early warning threshold determination logic are not fixed values, but dynamically adjusted. First, a Level 1 threshold is set. The Level 2 and Level 3 early warning thresholds are adaptively adjusted based on the variance of the remaining lifetime probability distribution function. When the variance of the lifetime prediction increases, indicating increased uncertainty, the system automatically raises the early warning threshold, triggering a higher-level warning earlier to increase the safety margin. The magnitude of the threshold adjustment and the change in variance are determined through a preset linear mapping relationship. For example, for every 10,000 cycles of increase in variance, the Level 2 threshold is adjusted upwards by 5% of the design lifetime.

[0036] The adaptive decision-making and feedback module uses the execution effects of early warning decisions at each level and subsequent actual detection results as feedback data, which is then fed back to the damage evolution modeling module based on a physical information neural network. This feedback is used to fine-tune the weights or forms of the physical constraint equations, thereby achieving continuous self-optimization of the system's prediction and decision-making capabilities. The multi-level interaction relationships and data flow of this adaptive decision-making and feedback module are shown in the attached figure. Figure 4 As shown, the closed-loop mechanism of early warning judgment, dynamic threshold adjustment and feedback optimization is clearly revealed.

[0037] At the system deployment level, the system is deployed on a cloud-edge collaborative computing platform using a microservice architecture. The multi-source heterogeneous data fusion acquisition module and the signal preprocessing section of the online non-destructive testing unit are deployed on edge computing nodes close to the crane, responsible for real-time acquisition and preliminary processing of high-frequency data, ensuring low latency and high reliability in data processing. The damage evolution modeling module based on physical information neural networks, the multi-scale coupled remaining life prediction module, and the adaptive decision-making and feedback control module are deployed on a cloud server cluster, responsible for performing computationally intensive model training, simulation, and decision-making tasks. Data synchronization and command issuance between the cloud and the edge are achieved through industrial-grade wireless communication protocols (such as 5G private networks or industrial Wi-Fi 6), with communication latency controlled within 50 milliseconds. All data transmissions are end-to-end encrypted using the AES-256 encryption algorithm to ensure data security.

[0038] The physical constraint equation embedded in the physical information neural network is a hybrid form of the fatigue crack propagation rate equation based on the Paris formula and the damage accumulation equation based on continuous damage mechanics. This hybrid form of the damage accumulation equation uses the amplitude of the nonlinear harmonic components extracted from the eigenvectors of the ultrasonic guided wave signal as an internal variable characterizing microscopic plastic deformation and crack closure effects, and uses the strain field gradient monitored by the distributed fiber Bragg grating sensor network as a correction parameter for the local stress intensity factor. This allows the physical constraints to be directly correlated with the real-time signals monitored online, rather than relying on idealized theoretical assumptions. Specifically, the damage evolution rate can be described by the following formula: ; The length of the crack. The number of loops. and For material constants, This represents the stress intensity factor amplitude, the calculation of which takes into account the local stress field corrected by the strain field gradient. The amplitude of the second-order nonlinear harmonic component extracted from the ultrasonic guided wave signal. The coupling coefficient is used to quantify the contribution of microscopic nonlinear effects to the stress intensity factor. This formula is embedded as a physical constraint in the loss function of the neural network, ensuring that the damage evolution path output by the model conforms to fundamental physical laws.

[0039] The load spectrum extrapolation unit employs a Long Short-Term Memory (LSTM) network. Its input is a historical load spectrum time series from the past 30 days, with each time step containing four features: the maximum load, minimum load, average load, and number of load cycles for that day. The network structure consists of two LSTM layers, each with 128 hidden units. The output layer generates a probabilistic load spectrum for each day of the next seven days. In the Monte Carlo simulation, each simulation randomly selects a load spectrum scenario and integrates it forward with the current damage state until the failure threshold is reached. After all simulation results are aggregated, a kernel density estimation method is used to fit the probability distribution function of the remaining lifetime. This probability distribution function accurately characterizes the tail risk in lifetime prediction, providing a basis for high-confidence decision-making.

[0040] In actual operation, the system executes a complete lifespan prediction and decision-making process every 24 hours. At the start of each process, the system first pulls the latest multi-source fusion data package from the edge nodes, updates the state of the digital twin model, then performs load extrapolation and Monte Carlo simulation, and finally generates new warning states and maintenance recommendations. If the system detects any abnormal data, it will automatically start the data verification and fault diagnosis subroutine, attempt to compensate through interpolation of adjacent sensor data or regression of historical data, and mark the data points as invalid if compensation is not possible to avoid polluting the model training process. All operation logs, warning records, and maintenance execution status are persistently stored in a cloud database, forming a complete digital history that can be used for subsequent auditing and model retraining.

[0041] Through the detailed system construction and operation mechanism described above, this invention achieves refined and intelligent management of crane welding joints from manufacturing to service, improving the real-time performance of non-destructive testing, the scientific nature of damage modeling, the foresight of life prediction, and the accuracy of maintenance decisions. Example

[0042] Building upon Example 1, this example enhances and optimizes the system specifically for crane applications in extreme environments. Specifically, in harsh environments such as ports, docks, or open-pit mines characterized by high salt spray, high humidity, and strong vibration, corrosion fatigue and stress corrosion cracking of welded joints become the dominant failure modes. Therefore, this example expands the multi-source heterogeneous data fusion acquisition module by adding a corrosion status monitoring unit.

[0043] The corrosion status monitoring unit consists of a miniature electrochemical noise sensor and a corrosion weight loss probe deployed on the surface of the weld joint. The electrochemical noise sensor acquires fluctuation signals of corrosion current and corrosion potential once per minute, and assesses local corrosion activity by analyzing noise resistance and power spectral density. The corrosion weight loss probe directly measures the amount of metal loss by periodically weighing the metal, providing an absolute benchmark for the corrosion rate. These corrosion-related data are synchronously integrated into a multi-source fusion data package and fed as additional input vectors into a damage evolution modeling module based on a physical information neural network.

[0044] Accordingly, the physical constraint equations in the physical information neural network have also been extended, introducing a coupled damage model of stress corrosion cracking. This coupled damage model uses the product of corrosion current density and local stress level as the driving term for corrosion fatigue damage, thus correcting the original pure mechanical fatigue damage accumulation equation. Specifically, the damage evolution equation is updated as follows: ; As a damage variable, For stress amplitude, The ultimate strength of the material, For corrosion current density, For local stress, These are the parameters to be calibrated. This specific damage evolution equation is also embedded as a physical constraint in the network training process, enabling the model to simultaneously capture the coupling effect of mechanical fatigue and corrosion damage.

[0045] In the multi-scale coupled remaining lifetime prediction module, the load spectrum extrapolation unit incorporates dynamic coupling of environmental corrosion factors. The input to the Long Short-Term Memory (LSTM) network includes not only historical load data but also concurrent environmental corrosion indices (calculated from temperature, humidity, salt spray concentration, pH, etc.). This allows the extrapolated future load spectrum scenario to reflect the potential weakening effect of environmental corrosion on the structural load-bearing capacity. The damage evolution path simulation unit simultaneously considers the stochastic evolution of corrosion damage in each Monte Carlo simulation, thereby generating a more realistic remaining lifetime distribution.

[0046] The adaptive decision-making and feedback control module has also adjusted its dynamic adjustment strategy for the warning thresholds accordingly. In addition to adjusting based on the variance of the life prediction, corrosion rate has been introduced as a second adjustment factor. When the corrosion rate exceeds the preset safety limit (e.g., 0.1 mm / year), the system automatically raises the warning thresholds for levels 2 and 3 to address the risk of accelerated degradation. Simultaneously, anti-corrosion treatment options, such as spraying anti-corrosion coatings and cathodic protection, have been added to the maintenance measures list.

[0047] In terms of system deployment, the edge computing nodes utilize industrial computers with an IP67 protection rating, and all sensor interfaces have been sealed and treated for corrosion protection to ensure long-term stable operation in harsh environments. The communication protocol also features enhanced anti-interference capabilities, employing forward error correction coding and adaptive modulation technology to guarantee data transmission reliability.

[0048] Through the above enhancements, this embodiment enables the system to cope with the challenges of monitoring welded joints in complex corrosive environments, further expanding the application boundaries and engineering value of the present invention.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-destructive testing and life prediction system for crane welded joints, characterized in that, include: The multi-source heterogeneous data fusion acquisition module is used to continuously collect multi-dimensional status data of the welded joint throughout its entire life cycle and output structured multi-source fusion data packets. The damage evolution modeling module based on physical information neural network is used to receive multi-source fusion data packets from the multi-source heterogeneous data fusion acquisition module and construct a digital twin model of damage evolution that integrates prior physical knowledge and real-time monitoring data. The multi-scale coupled remaining service life prediction module is used to receive the quantized damage state descriptor output by the damage evolution modeling module based on the physical information neural network, and to perform cross-scale prediction from micro-damage evolution to macro-remaining service life. The adaptive decision and feedback control module is used to receive the remaining lifetime probability distribution function output by the multi-scale coupled remaining lifetime prediction module, and generate graded early warning instructions and maintenance decision suggestions.

2. The crane welded joint non-destructive testing and life prediction system according to claim 1, characterized in that, The multi-source heterogeneous data fusion acquisition module integrates a manufacturing process data acquisition unit, a real-time service condition monitoring unit, and an online non-destructive testing unit. The manufacturing process data acquisition unit is used to extract the original process parameters of the weld joint from the manufacturing execution system; The real-time service condition monitoring unit is used to collect dynamic load spectrum, ambient temperature, ambient humidity and structural vibration acceleration signals of the welded joint in real time during service through a sensor network deployed on the crane structure. The online non-destructive testing unit is used to periodically excite and receive ultrasonic guided wave signals at a preset sampling frequency by using an embedded ultrasonic guided wave sensor array and a distributed fiber optic grating sensor network integrated into the weld joint, while simultaneously monitoring the strain field distribution and temperature gradient changes in the joint area.

3. The crane welded joint non-destructive testing and life prediction system according to claim 2, characterized in that, The damage evolution modeling module includes a deep physical information neural network, whose network structure consists of an input layer, multiple hidden layers, and an output layer. The input vectors received by the input layer include normalized real-time load spectrum, environmental parameters, ultrasonic guided wave signal feature vectors, and strain field data. The multiple hidden layers embed physical constraint equations derived from fracture mechanics and fatigue damage theory. These equations participate in network training as part of the loss function in the form of partial differential equation residuals. The training process of the damage evolution modeling module is divided into a pre-training phase and an online adaptive training phase. In the pre-training phase, the network is trained in a supervised manner using multi-source data from the historical failure case library and the corresponding final failure modes. In the online adaptive training phase, the multi-source fusion data packets collected in real time are input into the pre-trained model. The predicted state of the model at the previous moment is used as the initial condition. Combined with the measured data input at the current moment, the internal weight parameters of the model are dynamically corrected by solving the residual minimization problem of the embedded physical constraint equation. The output of the damage evolution modeling module is a quantified damage state descriptor of the weld joint at the current moment.

4. The crane welded joint non-destructive testing and life prediction system according to claim 3, characterized in that, The remaining lifetime prediction module includes a load spectrum extrapolation unit, a damage evolution path simulation unit, and a lifetime distribution calculation unit. The load spectrum extrapolation unit, based on real-time monitoring data of historical service conditions, uses time series analysis and machine learning algorithms to predict the probabilistic load spectrum that the welded joint will bear within a preset prediction period in the future. The damage evolution path simulation unit takes the quantized damage state descriptor at the current moment as the initial state and the future probabilistic load spectrum output by the load spectrum extrapolation unit as the input boundary condition, and drives the digital twin model updated through online adaptive training to perform multiple Monte Carlo simulations. Each simulation generates a damage evolution trajectory from the current state to the preset failure threshold. The lifetime distribution calculation unit performs statistical analysis on all damage evolution trajectories generated by Monte Carlo simulation, calculates the number of cycles or time corresponding to each trajectory reaching the failure threshold, and then fits the probability distribution function of the remaining lifetime.

5. The crane welded joint non-destructive testing and life prediction system according to claim 4, characterized in that, The adaptive decision and feedback control module has built-in multi-level early warning threshold determination logic; The multi-level early warning threshold determination logic compares the expected remaining lifespan with the preset 3-level early warning threshold in real time, and triggers different levels of early warning signals and decision instructions based on the comparison results. The adaptive decision-making and feedback control module uses the execution effect of early warning decisions at all levels and subsequent actual detection results as feedback data to send back to the damage evolution modeling module based on physical information neural network, which is used to fine-tune the physical constraint equations.

6. The crane welded joint non-destructive testing and life prediction system according to claim 5, characterized in that, In the online non-destructive testing unit, the arrangement of the embedded ultrasonic guided wave sensor array follows the following principles: The piezoelectric ceramic sensors in the array operate in a transceiver mode and are distributed in a grid-like topology that can cover the heat-affected zone and fusion line of the weld joint. The excitation signal is a sinusoidal pulse modulated by a narrowband Hanning window; The received ultrasonic guided wave signal is first subjected to time-frequency analysis through wavelet packet transform to extract the energy attenuation coefficient, propagation time offset, and nonlinear harmonic component amplitude within the frequency band, which together constitute the characteristic vector of the ultrasonic guided wave signal.

7. The crane welded joint non-destructive testing and life prediction system according to claim 6, characterized in that, The physical constraint equations embedded in the deep physical information neural network are a hybrid form of the fatigue crack propagation rate equation based on the Paris formula and the damage accumulation equation based on continuous damage mechanics. The hybrid form of the damage accumulation equation uses the amplitude of the nonlinear harmonic component extracted from the feature vector of the ultrasonic guided wave signal as an internal variable characterizing the microscopic plastic deformation and crack closure effect, and uses the strain field gradient monitored by the distributed fiber grating sensor network as a correction parameter for the local stress intensity factor.

8. The crane welded joint non-destructive testing and life prediction system according to claim 7, characterized in that, The machine learning algorithm used in the load spectrum extrapolation unit is a long short-term memory network; The Long Short-Term Memory Network takes time-series segments of historical load spectra as input and learns the dynamic change patterns of load amplitude, mean, and frequency. During extrapolation prediction, multiple future load spectrum scenarios with different probabilities are generated simultaneously, and these scenarios serve as input boundary conditions for the Monte Carlo simulation.

9. The crane welded joint non-destructive testing and life prediction system according to claim 8, characterized in that, The thresholds in the multi-level early warning threshold determination logic are dynamically adjusted; First, a Level 1 threshold is set, and the Level 2 and Level 3 warning thresholds are adaptively adjusted based on the variance of the remaining life probability distribution function. When the variance of the lifespan prediction increases, the system automatically raises the warning threshold; the magnitude of the threshold adjustment and the change in variance are determined through a preset linear mapping relationship.

10. The crane welded joint non-destructive testing and life prediction system according to claim 9, characterized in that, The system is deployed on a cloud-edge collaborative computing platform with a microservice architecture; The multi-source heterogeneous data fusion acquisition module and the signal preprocessing part in the online non-destructive testing unit are deployed at the edge computing node close to the crane. The damage evolution modeling module based on physical information neural network, the multi-scale coupled remaining lifetime prediction module, and the adaptive decision and feedback control module are deployed on a cloud server cluster. Data synchronization and command issuance between the cloud and the edge are achieved through an industrial-grade wireless communication protocol.

Images