System for monitoring steel structure state of key part of wharf loading and unloading equipment
By deploying sensor arrays and edge computing gateways on dock loading and unloading equipment, and combining dynamic normal modeling and anomaly decoupling technology, the problem that existing monitoring technologies cannot distinguish between aging and damage has been solved, enabling efficient and reliable monitoring and predictive maintenance of steel structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-12
- Publication Date
- 2026-04-03
AI Technical Summary
Existing monitoring technologies cannot effectively distinguish between gradual aging and sudden damage to the steel structures of dock loading and unloading equipment, resulting in low reliability of monitoring results and difficulty in achieving predictive maintenance.
By deploying sensor arrays and edge computing gateways, and combining dynamic normal modeling and anomaly decoupling technologies, damage can be identified through dynamic health benchmarks and morphological matching, enabling real-time monitoring and autonomous diagnosis of the steel structure's condition.
It achieves highly sensitive detection of early minor damage, reduces false alarm rate, improves the long-term reliability and confidence of monitoring results, and supports predictive maintenance without human intervention.
Smart Images
Figure CN121786410A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of structural health monitoring technology, and in particular to a steel structure condition monitoring system for key components of dock loading and unloading equipment. Background Technology
[0002] Terminal loading and unloading equipment, especially large container cranes, is the core equipment for port operations. Their critical load-bearing steel structures are subjected to complex and repetitive dynamic loads over long periods in the high-salt-spray, high-humidity marine environment, making them high-risk areas for fatigue damage and sudden structural failure. To ensure operational safety, existing technologies typically employ vibration analysis and strain monitoring methods to monitor the condition of the steel structures.
[0003] However, existing monitoring technologies generally face a fundamental and long-standing technical challenge. These technologies typically rely on a fixed "health status" baseline established at the factory or during the initial installation of the monitoring system. However, over decades of service, steel structures inevitably undergo gradual physical property evolution due to factors such as material fatigue accumulation, minute plastic deformation, and environmental erosion. This evolution includes a slow decrease in structural stiffness or changes in damping characteristics. This "normal aging" process also generates responses in monitoring signals, its characteristics intertwined and coupled with signal distortions caused by early "sudden damage" (such as the initiation and propagation of fatigue cracks). Due to the static nature of their baseline, existing monitoring systems cannot adapt to this dynamic evolution of the structure. This leads to a dilemma after long-term operation: either frequent false alarms occur due to the inability to filter out the background effects of "normal aging," or the alarm threshold is raised to reduce false alarms, resulting in insensitivity to genuine early minute damage signals and missed detections. This technical challenge ultimately leads to low reliability of long-term monitoring results from existing monitoring systems, making it difficult to achieve truly predictive maintenance. Summary of the Invention
[0004] This invention provides a steel structure condition monitoring system for key parts of dock loading and unloading equipment to solve the technical problem that existing monitoring technologies, which use static health benchmarks, cannot adapt to the normal physical characteristics evolution of structures during long-term service, thus making it impossible for the monitoring system to effectively distinguish between "gradual aging" and "sudden damage", ultimately resulting in low long-term reliability of monitoring results.
[0005] In view of the above problems, the present invention provides a steel structure condition monitoring system for key parts of dock loading and unloading equipment, characterized in that it includes:
[0006] Sensor arrays deployed at key locations of the steel structure;
[0007] The data acquisition unit is connected to the sensor array;
[0008] and an edge computing gateway, wherein the software layer running on the edge computing gateway includes:
[0009] The feature extraction module for working condition coupling is used to respond to external preset work event signals, extract and process high signal-to-noise ratio event features that are strongly correlated with the preset work event from the continuous vibration data stream of the data acquisition unit;
[0010] The unsupervised modeling module for dynamic normality is used to train and maintain a dynamic normality model that can reconstruct the features of normal events based on a set of historical normal event features. The dynamic normality model is updated as the structural state evolves normally.
[0011] An anomaly decoupling and quantization module is used to input the current event features output by the feature extraction module into the dynamic normal model, calculate the residual signal that only represents the anomaly information, and compare the energy of the residual signal and the form of the residual signal in the transform domain with a preset damage mode template to quantize and obtain a comprehensive anomaly score.
[0012] The decision release module is used to compare the anomaly score output by the anomaly decoupling and quantification module with an alarm threshold dynamically generated based on the statistical distribution of anomaly scores corresponding to historical normal event characteristics, so as to output a health status conclusion or early warning signal of the steel structure.
[0013] The technical solution provided in this application has at least the following technical effects or advantages:
[0014] By establishing a dynamic health benchmark that can "age" synchronously with the monitored physical entity and introducing a morphology-matching-based damage identification mechanism, the coupling interference problem between "normal aging" signals and "sudden damage" signals is fundamentally solved, achieving effective decoupling between the two. This enables the system to maintain high detection sensitivity for early, minor damage throughout the entire equipment lifecycle. Simultaneously, its dynamically adaptive alarm threshold keeps the false alarm rate extremely low, significantly improving the long-term reliability of monitoring results and the confidence level of alarm signals. This provides a feasible technical approach for achieving true, non-manual predictive maintenance. Attached Figure Description
[0015] Figure 1 The structural block diagram of the steel structure condition monitoring system for key parts of dock loading and unloading equipment provided in the embodiments of the present invention. Detailed Implementation
[0016] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.
[0017] For examples, please refer to Figure 1 A monitoring system for the condition of key steel structures in dock loading and unloading equipment, including:
[0018] Sensor arrays deployed at key locations of the steel structure;
[0019] The data acquisition unit is connected to the sensor array;
[0020] and an edge computing gateway, wherein the software layer running on the edge computing gateway includes:
[0021] The feature extraction module for working condition coupling is used to respond to external preset work event signals, extract and process high signal-to-noise ratio event features that are strongly correlated with the preset work event from the continuous vibration data stream of the data acquisition unit;
[0022] The unsupervised modeling module for dynamic normality is used to train and maintain a dynamic normality model that can reconstruct the features of normal events based on a set of historical normal event features. The dynamic normality model is updated as the structural state evolves normally.
[0023] An anomaly decoupling and quantization module is used to input the current event features output by the feature extraction module into the dynamic normal model, calculate the residual signal that only represents the anomaly information, and compare the energy of the residual signal and the form of the residual signal in the transform domain with a preset damage mode template to quantize and obtain a comprehensive anomaly score.
[0024] The decision release module is used to compare the anomaly score output by the anomaly decoupling and quantification module with an alarm threshold dynamically generated based on the statistical distribution of anomaly scores corresponding to historical normal event characteristics, so as to output a health status conclusion or early warning signal of the steel structure.
[0025] In this specific embodiment, the preset operation event signal refers to a specific timing action that causes the steel structure to generate the instantaneous maximum or second-largest impact load during the normal operation cycle of the dock loading and unloading equipment (such as a quay crane).
[0026] Specifically, in this embodiment, the moment when the container spreader completes container placement is preferably selected as the trigger point for the preset operation event.
[0027] The mapping relationship is as follows: The feature extraction module receives and identifies electrical signals from the Terminal Operating System (TOS) or the hoisting and trolley mechanism PLC (Programmable Logic Controller) indicating that the load transfer is complete or the action is in place. This electrical signal is the aforementioned "preset operation event signal". The system uses this signal as a timestamp to accurately extract data packets from the vibration data stream within a specific time window before and after the event (e.g., 0.5 seconds before the event to 1.5 seconds after the event), thereby obtaining event features strongly correlated with the load impact.
[0028] The monitoring system disclosed in this invention begins its entire operation with a system initialization and baseline modeling phase. The technical objective of this phase is to autonomously establish an initial dynamic health baseline for a dock loading / unloading equipment at its current lifecycle stage, without relying on any pre-set damage samples. Upon initial power-on or receipt of an external initialization command, the system automatically enters "initial learning mode" and executes the following process in sequence:
[0029] System Deployment and Interface Establishment: The first step in this stage is system deployment and interface establishment. Specifically, this involves rigidly fixing at least one broadband piezoelectric vibration sensor to a critical part of the steel structure to be monitored, such as the lower flange plate at the mid-span of the main beam, the heat-affected zone of the weld at the root of the boom, or the connecting lugs of the front and rear tie rods. The signal output of the vibration sensor is connected to the input port of a multi-channel synchronous data acquisition unit via a shielded cable. The data acquisition unit is connected to an edge computing gateway via an industrial bus or Ethernet interface to transmit continuous, digitized raw vibration data streams to the memory buffer of the edge computing gateway in real time. Simultaneously, the software system deployed on the edge computing gateway actively establishes a TCP / IP-based communication link with the Port Equipment Control System (TOS) through its network communication module. After successfully establishing the link, the software system subscribes to one or more predefined key event signals strongly correlated with the operational cycle from the TOS system. A preferred embodiment is subscribing to a "container contact confirmation signal" that identifies when a container contacts the chassis or ground and the locking pin is engaged. At this point, the system has completed the physical and communication layer deployment, preparing for subsequent data acquisition. As an alternative technical solution, the key event signals can also be generated locally through image recognition or point cloud analysis by installing independent machine vision sensors or LiDAR sensors near the lifting device, thereby reducing the coupling dependency on the TOS system.
[0030] Initial Event Feature Accumulation: After successful interface establishment, the system process automatically enters the initial event feature accumulation phase. The technical objective of this phase is to collect a sufficient number of high-quality sample data that can represent the normal behavior of the structure. Specifically, the feature extraction module for operational condition coupling within the system is activated and begins continuously monitoring the subscribed "container landing confirmation signal." During this accumulation phase, each time the module detects the arrival of this signal, it immediately performs a feature extraction operation. The computer execution steps for this operation include:
[0031] First, the time when the signal is received is used as the timestamp T0;
[0032] Second, read and extract a raw vibration data segment from the memory buffer of the edge computing gateway, from 1 second before T0 to 4 seconds after T0, for a total duration of 5 seconds;
[0033] Third, apply a third-order Butterworth bandpass filter to the truncated data segment to filter out signal noise outside the frequency range;
[0034] Fourth, the filtered data segments are stored as a standardized data matrix. This matrix is defined as an "event feature data packet". During this stage, the system will repeatedly perform this extraction operation, storing each generated event feature data packet in the non-volatile memory of the edge computing gateway, along with a timestamp and operating parameters, but without invoking any diagnostic or alarm algorithms. This accumulation process will continue until the number of stored event feature data packets reaches a preset threshold, such as 10,000. The preset threshold (e.g., 10,000) is set to ensure that the sample size is sufficient to cover various normal operating condition combinations of the device under no-load, half-load, full-load, and different wind speed levels, to prevent the model from overfitting due to training sample bias. At this point, the system determines that it has obtained sufficient statistical samples for initial modeling and automatically ends this accumulation process, moving to the next stage.
[0035] Training of the initial dynamic normal state model: After the accumulation process of initial event features is completed, the system automatically switches to the training phase of the initial dynamic normal state model. The technical goal of this phase is to generate a mathematical model that can accurately describe and reconstruct the dynamic response behavior of the steel structure in its current "youthful" healthy state, using a large amount of collected normal sample data. The specific implementation path of this process is as follows: the unsupervised modeling module for dynamic normal state within the system is activated, and the following computer steps are executed:
[0036] Read all accumulated event feature data packets, totaling 10,000, from the non-volatile memory of the edge computing gateway, and combine this dataset into the initial training set;
[0037] Construct a deep autoencoder neural network model, the network structure of which includes an encoder for data dimensionality reduction and a decoder for data dimensionality upscaling and reconstruction;
[0038] Specifically, the model employs a fully connected symmetric structure, where the number of nodes in the input and output layers is determined by the dimension of the event feature vector (e.g., the input layer is N-dimensional). The encoder contains at least two hidden layers, with the number of nodes decreasing layer by layer (e.g., NN / 2-N / 4) until the intermediate bottleneck layer. The bottleneck layer's nodes are used to extract low-dimensional core manifold features. The decoder structure is mirror-symmetric to the encoder, with the number of nodes increasing layer by layer. The inter-layer activation function preferably uses non-linear ReLU or Tanh functions to capture the non-linear features in the vibration signal, and the output layer uses a Sigmoid or Linear function.
[0039] Third, the initial training set is input into the deep autoencoder neural network model for unsupervised training. The training process defines the mean squared error between the decoder's reconstructed output and the original input as the loss function. The training employs a gradient descent-based optimization algorithm to iteratively adjust the network weight parameters to minimize the value of the loss function. This mean squared error serves as the loss function for the training process. The training process continues until the loss function converges, resulting in an initial dynamic normal model, which is then saved by the system for use in subsequent real-time monitoring phases.
[0040] Initial alarm threshold generation: After successfully generating the initial dynamic normal model, the system process then proceeds to the initial alarm threshold generation stage. The technical objective is to establish a statistically reasonable initial decision-making benchmark for subsequent anomaly assessments. The specific implementation path of this process is as follows:
[0041] The system inputs all 10,000 initial event feature data packets previously used for training into the solidified initial dynamic normal model one by one, without training.
[0042] For each input event feature, the system calls the algorithm of the anomaly decoupling and quantization module, which will be described later, to calculate its corresponding comprehensive anomaly score.
[0043] Third, after calculating the scores of all 10,000 samples, the system uses these 10,000 scores as training data and applies a statistical method, such as one based on the interquartile range (IQR) or Gaussian distribution, to calculate an upper limit that includes 99.9% of the "normal" scores. This upper limit is defined as the initial alarm threshold and is stored together with the initial dynamic normal model. At this point, the entire system initialization and baseline modeling phase is complete, and the system will automatically exit the "initial learning mode" and prepare to enter the next stage of real-time monitoring and autonomous diagnosis.
[0044] Real-time Monitoring and Autonomous Diagnosis Phase: After the system initialization and baseline modeling phases are completed, the system automatically enters the real-time monitoring and autonomous diagnosis phase. The technical objective of this phase is to perform real-time health status assessments for each work cycle during the equipment's daily operation, accurately identify response signal distortions caused by sudden abnormal damage, and autonomously update and adapt to normal structural aging over time. Once in this phase, the system will continuously operate in "real-time monitoring mode," with its regular workflow beginning with the capture of real-time event characteristics.
[0045] Real-time event feature capture: The technical implementation path of this step is consistent with the feature extraction operation in the initial accumulation stage, the difference being the execution frequency and purpose. Specifically, the feature extraction module coupled with the operating conditions within the system will perform a complete feature extraction operation in real time and periodically every time the "container landing confirmation signal" is detected, generating a current event feature representing the dynamic response of the latest operation cycle.
[0046] Decoupling and Quantization of Anomaly Scores: After acquiring the current event characteristics, the system process seamlessly transitions to the decoupling and quantization of anomaly scores. This step is the core diagnostic step of this invention, and its technical objective is to decouple and quantify the weak residual signals contained in the current event characteristics, which are caused by sudden damage.
[0047] a. Calculation of residual signal: The specific implementation path is as follows:
[0048] First, the system loads the initial dynamic normal model generated during the baseline modeling phase;
[0049] Second, the current event features are input as a vector X into the dynamic normal model;
[0050] Third, the dynamic normal model performs a forward propagation calculation on the input vector to obtain a reconstructed output vector M(X). In a physical sense, this reconstructed output vector M(X) represents the model's "most reasonable and healthiest" reproduction of the current input based on the "normal" knowledge it has learned.
[0051] Fourth, calculate the vector difference between the current event feature vector X and the reconstructed output vector M(X) to obtain the residual signal vector E=XM(X). Since the model has been trained to reconstruct "normal" features with extremely low error, the obtained residual signal vector E no longer contains "normal" response information in a physical sense. Instead, it retains to the maximum extent the signal distortion caused by "abnormal" factors (such as sudden injury) that the model cannot understand.
[0052] b. Generation of Frequency Domain Distortion Vector: After calculating the residual signal vector containing only anomalous information, the decoupling and quantization of the anomalous score is followed by the generation of the frequency domain distortion vector. The technical objective of this step is to transform the residual signal from the time domain to the frequency domain to reveal the distribution characteristics of its energy in different frequency ranges, i.e., the "morphology" of the signal. A preferred technical implementation path for this process is to use wavelet packet transform. The specific computer execution steps include: First, selecting a suitable wavelet basis function, such as the Daubechies 4 wavelet; Second, setting a decomposition level, such as 3 levels, at which point the residual signal will be decomposed into 2^3 eight non-overlapping frequency bands; Third, performing the 3-level wavelet packet transform on the residual signal vector to obtain eight wavelet packet decomposition coefficient sequences corresponding to different frequency bands; Fourth, calculating the energy of each coefficient sequence, for example, calculating its root mean square value or energy norm; Fifth, organizing these eight calculated energy values into an 8-dimensional vector in order from low frequency to high frequency. This vector is the frequency domain distortion vector. Each element quantifies the energy intensity of the residual signal in the corresponding frequency band, thus fully characterizing the frequency domain morphological features of this anomaly.
[0053] c. Template Matching of Damage Patterns: After successfully generating the frequency domain distortion vector representing the current anomaly morphology, the process enters the template matching step for damage patterns. The technical objective of this step is to determine the degree of similarity between the current anomaly morphology and known anomaly morphologies caused by actual structural damage. The specific implementation path is as follows: First, load one or more predefined damage pattern templates from the storage unit of the edge computing gateway. The damage mode template is a normalized vector with the same dimension as the frequency domain distortion vector. Its content is obtained in advance through high-precision finite element simulation or physical experiments, and is used to characterize the distribution of residual signal energy in each frequency band most likely caused by a typical structural damage (e.g., fatigue crack). Preferably, the preset damage mode template is obtained as follows: based on the finite element model (FEM) of the quay crane of this type, simulated cracks of different lengths (e.g., 5mm, 10mm, 20mm) are pre-set at key welds, transient dynamic simulation is performed, and the energy distribution feature vector of the residual signal in the frequency domain is extracted and normalized and stored. Alternatively, destructive test data from decommissioned equipment of the same type can be used as the template. Second, the latest frequency domain distortion vector is calculated using a cosine similarity algorithm. With the damage mode template The cosine of the angle between the vectors is the morphological similarity. This cosine value is defined as the morphological similarity, and its value ranges from [-1, 1]. The closer the value is to 1, the more it matches the morphology of the current anomaly in the frequency domain with that of a typical lesion, and the higher the probability that it is a real lesion.
[0054] d. Calculation of the comprehensive anomaly score: After obtaining the residual signal norm representing the "quantity" of the anomaly and the morphological similarity representing the "property" of the anomaly, the process proceeds to the calculation of the comprehensive anomaly score. The calculation is performed using the following formula:
[0055] in, That is, the Euclidean norm of the residual signal vector calculated in the aforementioned steps, which represents the anomalous energy value; This refers to the calculated morphological similarity; α and γ are preset non-negative exponential parameters used for nonlinear amplification, for example, they can be 2.0 and 3.0 respectively. The physical meaning of this formula is that the final anomaly score depends on both the energy and morphology of the anomaly. An anomaly signal will only have its final anomaly score exponentially amplified when its energy exceeds a certain threshold and its morphology highly matches that of a typical lesion. This enables highly sensitive identification of real lesion signals and effective suppression of noise signals with inconsistent morphology.
[0056] Let's illustrate this with a numerical example: Suppose in a certain calculation, the residual signal norm is 0.18, the morphological similarity is 0.98, and the parameters α=2.0 and γ=3.0. Then the comprehensive anomaly score is calculated as follows: In contrast, if in another calculation the residual norm is also 0.18, but because its morphology is completely unrelated to the damage pattern, resulting in a morphological similarity of 0, then the final anomaly score will be calculated as follows: In this way, the system achieves precise decoupling and quantization of abnormal signals.
[0057] It should be noted that the specific values of the preset parameters α and γ are determined based on the on-site noise level. For example, in ports with strong sea breezes and high environmental noise, the value of γ can be increased (e.g., set to 4.0) to enhance the dependence on morphological matching and suppress false alarms caused by random environmental noise; while in inland river ports, the value of α can be appropriately decreased (e.g., set to 1.5) to improve the sensitivity to early energy release from microcracks.
[0058] Alarm Decision Issuance: After calculating the comprehensive anomaly score in the decoupling and quantification phase of the anomaly score, the system process enters the alarm decision issuance phase. The technical objective of this phase is to make a final judgment on the health status of the structure based on the anomaly score in a statistically reliable and engineering-robust manner. The specific implementation path is as follows: First, the decision issuance module within the system loads the latest dynamic alarm threshold generated during the baseline modeling phase or updated in subsequent evolutionary processes. Second, the latest calculated comprehensive anomaly score is compared with the dynamic alarm threshold. Third, a preset decision logic is executed to avoid false alarms caused by single, accidental fluctuations. A preferred decision logic is the "sliding window counting method." For example, the system checks whether the anomaly score has exceeded the threshold three times in the most recent 10 work cycles. Only when the triggering condition of the decision logic is met will the decision issuance module determine that the structural health status is abnormal and, through the network communication module, issue a high-confidence warning signal containing information such as a timestamp, anomaly score, and relevant operating parameters to the port's upper-level monitoring system or a designated maintenance personnel terminal. If the triggering conditions are not met, the system will continue with the next round of real-time monitoring.
[0059] Periodic Evolution of Model and Thresholds: To address the technical challenge of the slow, continuous aging of the steel structure's physical properties and ensure the system maintains monitoring accuracy over decades of service, this invention also includes a periodic evolution process for the model and thresholds. This process is the core mechanism for the system's autonomous evolution and dynamic adaptation, specifically driven by a preset "model lifecycle management" strategy. This strategy can be defined based on time or data volume; a preferred embodiment is to trigger the evolution process "every 50,000 accumulated event features judged as normal, or every 30 days of operation, whichever comes first." When the triggering condition is met, the system will automatically execute the following computer steps:
[0060] First, a new training set is formed by calling up the event feature data packets that were all judged to be "normal" in the most recent evolution cycle;
[0061] Second, the system uses this new training set, which better reflects the current state of the structure, to incrementally train or fully retrain the deep autoencoder neural network in the unsupervised modeling module of the dynamic normal state until its loss function converges again, thereby generating a new version of the dynamic normal state model that can accurately reconstruct the current normal features.
[0062] Third, the system inputs all event features from the new training set into the new model one by one, recalculates the comprehensive abnormal score corresponding to each normal feature, and generates a new dynamic alarm threshold that matches the statistical distribution of these new scores.
[0063] Fourth, the system will overwrite the old version with the new model and thresholds, and use this as the diagnostic and decision-making benchmark for the next monitoring cycle. Through this periodic and autonomous evolution process, the present invention ensures that its monitoring benchmark can evolve in sync with the natural aging process of the physical structure, thereby maintaining extremely high detection sensitivity and extremely low false alarm rate for "sudden abnormal damage" throughout the entire equipment life cycle.
[0064] This detailed embodiment elaborates on the internal operating logic of each major functional module, aiming to provide a detailed basis and explanation for those skilled in the art to understand and implement it. It should be emphasized that the above description constitutes a specific, preferred embodiment, but the concept of the present invention is not limited thereto. Any equivalent transformations, modifications, or improvements based on the core spirit of the present invention, without departing from the technical principles and scope disclosed in this specification, should be considered to fall within the scope of protection claimed by the present invention, as long as they achieve the same or similar technical effects.
Claims
1. A monitoring system for the condition of key steel structures in dock loading and unloading equipment, characterized in that, include: Sensor arrays deployed at key locations of the steel structure; The data acquisition unit is connected to the sensor array; and an edge computing gateway, wherein the software layer running on the edge computing gateway includes: The feature extraction module for working condition coupling is used to respond to external preset work event signals, extract and process high signal-to-noise ratio event features that are strongly correlated with the preset work event from the continuous vibration data stream of the data acquisition unit; The unsupervised modeling module for dynamic normality is used to train and maintain a dynamic normality model that can reconstruct the features of normal events based on a set of historical normal event features. The dynamic normality model is updated as the structural state evolves normally. An anomaly decoupling and quantization module is used to input the current event features output by the feature extraction module into the dynamic normal model, calculate the residual signal that only represents the anomaly information, and compare the energy of the residual signal and the form of the residual signal in the transform domain with a preset damage mode template to quantize and obtain a comprehensive anomaly score. The decision release module is used to compare the anomaly score output by the anomaly decoupling and quantification module with an alarm threshold dynamically generated based on the statistical distribution of anomaly scores corresponding to historical normal event characteristics, so as to output a health status conclusion or early warning signal of the steel structure.
2. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 1, characterized in that, The preset operation event signal is a container landing confirmation signal from the port equipment control system.
3. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 1, characterized in that, The dynamic normal state model used in the unsupervised modeling module is a deep autoencoder neural network model.
4. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 1, characterized in that, The anomaly decoupling and quantization module is specifically used for: The residual signal is obtained by calculating the vector difference between the current event features and the reconstructed output of the dynamic normal model; The residual signal is transformed to obtain a frequency domain distortion vector that characterizes the distribution of its energy in different frequency domains or time-frequency domain components; Load a predefined damage mode template that characterizes the damage energy distribution of a preset structure; Calculate the morphological similarity between the frequency domain distortion vector and the damage pattern template; The comprehensive anomaly score is calculated based on the norm of the residual signal and the morphological similarity.
5. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 4, characterized in that, The comprehensive anomaly score is calculated using the following functional relationship: The comprehensive anomaly score is determined based on the magnitude of the residual signal and the morphological similarity; and the value of the comprehensive anomaly score is positively correlated with both the magnitude of the residual signal and the morphological similarity.
6. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 5, characterized in that, The comprehensive anomaly score is calculated using the following formula: Where X represents the current event feature, and M(X) represents the reconstructed output of the dynamic normal model. Let be the Euclidean norm of the residual signal. Let be the frequency domain distortion vector. The damage pattern template is defined as follows: cos() is the cosine similarity calculation function, and α and γ are preset non-negative exponential parameters.
7. The monitoring system for the condition of key steel structures of dock loading and unloading equipment as described in any one of claims 4 to 6, characterized in that, The frequency domain distortion vector is obtained by performing a wavelet packet transform on the residual signal.
8. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 1, characterized in that, The dynamically generated alarm threshold used by the decision-making release module is a preset quantile of the statistical distribution of abnormal scores corresponding to the characteristics of historical normal events.
9. The steel structure condition monitoring system for key components of dock loading and unloading equipment as described in claim 1, characterized in that, The sensor array includes at least one broadband piezoelectric vibration sensor configured at the mid-span of the main beam, the weld area at the root of the boom, or the tie rod connecting lug.