Shipyard steel bent frame load condition monitoring and early warning method and system

Abstract

The invention provides a shipyard steel bent frame load condition monitoring and early warning method and system. The method is applied to the technical field of building monitoring, and comprises the following steps: deploying a plurality of sensors on a shipyard steel frame, collecting stress displacement speed data of the shipyard steel frame, preprocessing the data, performing feature extraction, and collecting to obtain a data feature vector; constructing a load state prediction model based on LSTM-CNN and combined with an attention mechanism, processing the data feature vector, and outputting to obtain a shipyard steel bent frame load prediction value and a state probability; and performing shipyard steel bent frame dynamic threshold early warning and online feedback based on the shipyard steel bent frame load prediction value and the state probability, thereby realizing shipyard steel bent frame load real-time monitoring and early warning. In this way, according to the method, the load condition of the steel frame is monitored in real time by fusing the data of the multi-source sensor and constructing the load state prediction model and adopting a dynamic threshold value adjusting mechanism, and the method has automatic and efficient practical value.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of building monitoring, in particular to a shipyard steel frame load working condition monitoring and early warning method and system. BACKGROUND

[0002] In industrial production, steel frame buildings are widely used in shipyard construction. Since shipyards are usually large, they are easily affected by external forces and seawater corrosion, etc. Therefore, it is necessary to monitor the rigidity and load performance of the steel frame structure of the shipyard. However, the current common monitoring method mainly relies on manual monitoring, which has low automation and high delay, and cannot accurately and timely monitor the load working condition, which may lead to unpredictable structural instability of the building in storm or other extreme weather events.

[0003] Therefore, there is an urgent need for a design scheme that can effectively monitor and early warn the load working condition of the shipyard steel frame. SUMMARY

[0004] The present disclosure provides a shipyard steel frame load working condition monitoring and early warning method and system, which fuses multi-source sensor data and constructs a load state prediction model, realizes real-time monitoring of the steel frame load condition through a dynamic threshold adjustment mechanism, and at least solves the technical problems of low efficiency and low automation of the existing steel frame monitoring method.

[0005] According to a first aspect of the present disclosure, a shipyard steel frame load working condition monitoring and early warning method is provided, comprising the following steps: Deploying a plurality of sensors on the shipyard steel frame to collect shipyard steel frame stress displacement velocity data, performing data preprocessing and feature extraction to collect data feature vectors; Constructing a load state prediction model based on LSTM-CNN combined with an attention mechanism, processing the data feature vectors, and outputting shipyard steel frame load prediction values and state probabilities; Based on the shipyard steel frame load prediction values and state probabilities, the shipyard steel frame dynamic threshold early warning and online feedback are performed to realize real-time monitoring and early warning of the shipyard steel frame load.

[0006] As described above, the aspect and any possible implementation manner further provides an implementation manner, wherein the process of deploying a plurality of sensors on the shipyard steel frame to collect shipyard steel frame stress displacement velocity data, performing data preprocessing and feature extraction, and collecting data feature vectors is as follows: Based on the finite element simulation method, the stress concentration area of the shipyard steel frame is determined, and a plurality of sensors are deployed on the shipyard steel frame using a grid layout strategy to collect stress displacement velocity data; The stress displacement velocity data are denoised by using wavelet threshold denoising combined with Kalman filtering method, and the confidence fusion of the multi-sensor data is obtained by using D-S evidence theory to obtain the fused data. The fused data are subjected to feature extraction to obtain a 32-dimensional data feature vector.

[0007] According to the aspect and any possible implementation manner described above, an implementation manner is further provided, wherein the data feature vector comprises time domain features, frequency domain features, time-frequency domain features and fused features. The time domain features comprise mean value, standard deviation, kurtosis, skewness, peak-to-peak value, waveform factor, pulse factor and zero-crossing rate. The frequency domain features comprise main frequency amplitude, frequency band energy proportion, spectral entropy, harmonic distortion and transfer function amplitude. The time-frequency domain features comprise wavelet packet energy entropy and seven adjacent sub-band correlation coefficients. The fused features comprise strain-displacement coupling coefficient and vibration-load transfer function parameter.

[0008] According to the aspect and any possible implementation manner described above, an implementation manner is further provided, wherein the load state prediction model is constructed based on LSTM-CNN and combined with an attention mechanism, the process of processing the data feature vector and outputting a shipyard steel bent frame load prediction value and a state probability is as follows: A bidirectional LSTM structure is constructed to process the data feature vector and output a time step hidden state, wherein the bidirectional LSTM structure is provided with 128 hidden units in each direction. The time step hidden state is reshaped into a three-dimensional tensor, and a 3x1 convolution kernel is used for local feature extraction, and after processing by a ReLU activation function, a global maximum pooling operation is performed on the convolution result to extract channel significant features and output a multi-dimensional feature vector. An attention mechanism is used to impose attention weights on each time step hidden state output by the LSTM, the importance scores of the time steps are calculated through a learnable weight matrix, and after normalization of the importance scores, an attention weighted feature vector is obtained by weighted summation. The LSTM, CNN and attention weighted feature vector are spliced and connected to a fully connected layer, based on the multi-dimensional feature vector and the attention weighted feature vector, the probability distribution of normal, early warning and dangerous states is output by a Softmax function, and a separate fully connected layer is used to map the features to a continuous prediction result of the load value, and a load value prediction result is output.

[0009] According to the aspect and any possible implementation manner as described above, further provided is an implementation manner of generating a dynamic threshold value for a graded early warning based on historical load data statistical analysis and expert experience rules, and the process is as follows: Based on historical load value prediction result data statistical analysis and expert experience rules, a dynamic threshold value for a graded early warning is generated. The real-time predicted load value is compared with the dynamic threshold value to obtain a warning result, and a graded response mechanism is triggered. Based on the deviation of the warning result and the actual load condition, the model and the threshold value parameters are optimized in real time.

[0010] According to the aspect and any possible implementation manner as described above, further provided is an implementation manner of generating a dynamic threshold value for a graded early warning based on historical load data statistical analysis and expert experience rules, and the process is as follows: The mean and standard deviation of historical load value prediction result data are counted, and the static threshold value interval of three levels of normal, early warning, and danger is preliminarily determined in combination with the safety load limit in the steel frame design specification. Real-time environmental parameters are introduced, and the static threshold value interval is dynamically adjusted according to the thermal expansion coefficient and humidity sensitivity characteristics of the material. Historical false alarm rate and missed alarm rate data are collected, and the threshold value sensitivity coefficient is automatically adjusted to adapt the threshold value interval to the actual situation.

[0011] According to the aspect and any possible implementation manner as described above, further provided is an implementation manner of dynamically adjusting the static threshold value interval according to the thermal expansion coefficient and humidity sensitivity characteristics of the material by introducing real-time environmental parameters, and the process is as follows: Wherein, μ is the mean load value, k is the sensitivity coefficient, α T , and α H are environmental correction factors, is the temperature change amount, is the humidity change amount, μ n is the adaptive threshold value.

[0012] According to the aspect and any possible implementation manner as described above, further provided is an implementation manner of comparing the real-time predicted load value with the dynamic threshold value to obtain a warning result and triggering a graded response mechanism, and the process is as follows: When the real-time predicted load value continuously falls below the early warning minimum threshold value, it is determined to be a normal state, only data is recorded, and a periodic health report is generated. When the real-time predicted load value is between the minimum threshold value and the maximum threshold value, it is determined that a warning state exists, an artificial review process is triggered, and a warning notification including the load exceeding position, the prediction deviation value, and the environmental parameters is sent to the operation and maintenance personnel; When the real-time predicted load value exceeds the maximum threshold value or is in the warning interval for three consecutive times, an emergency response is performed, the emergency support system of the steel frame is activated to limit deformation expansion, the surrounding personnel passage is closed, a sound and light alarm is triggered, the fire and emergency management departments are linked through the Internet of Things platform, and finally a structure damage evaluation report is generated, a high-risk area is marked, and a reinforcement scheme is recommended.

[0013] According to a second aspect of the present disclosure, a shipyard steel frame load working condition monitoring and warning system is provided, comprising: a data acquisition and feature construction module, a steel frame load prediction module, and a load monitoring and warning module; The data acquisition and feature construction module is used to deploy a plurality of sensors on the shipyard steel frame, acquire shipyard steel frame stress displacement speed data, and collect data feature vectors through data preprocessing and feature extraction; The steel frame load prediction module is used to construct a load state prediction model based on LSTM-CNN combined with an attention mechanism, process the data feature vectors, and output shipyard steel frame load prediction values and state probabilities; The load monitoring and warning module is used to perform shipyard steel frame dynamic threshold warning and online feedback based on the shipyard steel frame load prediction values and state probabilities, and realize real-time monitoring and warning of the shipyard steel frame load.

[0014] Compared with the prior art, the present application has the following technical effects: (1) The present application has high precision monitoring and prediction capability. By fusing multi-source sensor data and combining a hybrid neural network model, the system can accurately capture the dynamic response characteristics of the steel frame structure. Compared with traditional methods, the present application significantly improves the accuracy of load prediction, effectively reduces errors, and has higher discrimination ability in structure state classification tasks, can reliably identify subtle abnormal signals, and improves the sensitivity to resonance risk through enhanced frequency domain analysis technology.

[0015] (2) The present application has real-time response and low delay characteristics. Lightweight model architecture and edge computing technology are used, the system realizes millisecond-level real-time data processing capability, meets the demand of rapid capture of high-frequency dynamic load. Through parallel computing and hardware acceleration, the end-to-end response time is significantly shortened, ensuring immediate warning of transient impact events.

[0016] (3) The application has environmental self-adaptation and system robustness. By introducing a dynamic threshold adjustment mechanism, the early warning standard is automatically corrected in combination with environmental parameters (temperature and humidity, material performance), effectively avoiding misjudgment caused by environmental interference. When part of the sensors fail or data is abnormal, the system can still maintain stable operation through redundant design and data interpolation mechanism, ensuring monitoring continuity.

[0017] (4) The application realizes whole life cycle cost optimization. By replacing manual inspection with automatic monitoring, the operation and maintenance manpower demand is greatly reduced; in combination with early risk warning capability, the maintenance cost caused by sudden structure damage is significantly reduced. The combination of low-power design and sustainable power supply scheme further reduces the energy consumption of long-term deployment.

[0018] It should be understood that the content described in the summary section is not intended to limit the key or important features of the embodiments of the present disclosure, nor to limit the scope of the present disclosure. Other features of the present disclosure will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0019] The above and other features, advantages, and aspects of embodiments of the present disclosure will become more apparent by describing in detail the embodiments thereof with reference to the attached drawings. The accompanying drawings are used to better understand the present scheme and do not limit the present disclosure. In the drawings, the same or similar reference numerals refer to the same or similar elements, and: Figure 1 A shipyard steel rack load working condition monitoring and early warning method flow chart is shown. Figure 2 A shipyard steel rack load working condition monitoring and early warning system structure diagram is shown. DETAILED DESCRIPTION

[0020] To make the purposes, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present disclosure.

[0021] To make the purposes, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present disclosure.

[0022] Referring to Figure 1 The present embodiment provides a shipyard steel rack load working condition monitoring and early warning method, including the following steps: S101, deploy a plurality of sensors on the shipyard steel frame to collect stress displacement speed data of the shipyard steel frame, perform data preprocessing and feature extraction, and collect data feature vectors.

[0023] In the embodiment, first, stress concentration areas (such as node connections and mid-span sections) of the steel frame structure are determined based on a finite element simulation method, sensors are deployed on the shipyard steel frame using a meshing layout strategy, and the sensors are arranged at intervals of less than 5m 2 At least one strain sensor and one displacement sensor are arranged.

[0024] In the embodiment, the deployed sensors specifically include: a fiber grating strain sensor with an accuracy of ±1με, a temperature compensation range of -20℃-80℃, and being suitable for long-term monitoring; a MEMS acceleration sensor with a range of ±5g and a frequency response of 0.1-200Hz, and being used to capture low-frequency vibration and impact load; and a dual-frequency GPS displacement meter with a positioning accuracy of ±2mm / ±5mm and a sampling rate of 10Hz, and being used to monitor overall deformation.

[0025] Subsequently, wavelet threshold denoising combined with Kalman filtering is used for data denoising processing, and the wavelet threshold denoising method used in the embodiment is specifically as follows: (1) wherein, DWT is a discrete wavelet transform, T λ (x) is a soft threshold function, and λ is a threshold value, and specifically wherein, is a noise standard deviation, and N is a signal length.

[0026] After denoising is completed, Kalman filtering is used to adopt a state space model for GSP displacement data, so as to suppress multi-path effect noise, and specifically as follows: (2) wherein, x k is a state vector, A is a state transition matrix, w k is a process noise covariance, H is an observation matrix, and v k is an observation noise covariance.

[0027] After data denoising processing is performed, D-S evidence theory is used to fuse confidence degrees of multi-sensor data: (3) wherein, m(A) is a basic probability assignment of proposition A, and m1 and m2 are confidence degree distributions of different sensors.

[0028] For load prediction in subsequent processes, the embodiment performs feature extraction of data to generate a 32-dimensional feature vector, specifically including extracted time domain features, frequency domain features, time-frequency domain features, and fused features. Specifically, in the embodiment, the time domain features are 8-dimensional, specifically including mean, standard deviation, kurtosis, skewness, peak-to-peak value, waveform factor, pulse factor, and zero-crossing rate; the frequency domain features are 8-dimensional, specifically including main frequency amplitude, energy proportion of 3 frequency bands, spectral entropy, harmonic distortion, and transfer function amplitude; the time-frequency domain features are 8-dimensional, specifically including wavelet packet energy entropy and 7 adjacent sub-band correlation coefficients; the fused features are 8-dimensional, specifically including strain-displacement coupling coefficient (reflecting the correlation between local deformation and overall displacement) and vibration-load transfer function parameter (reflecting the frequency domain relationship between acceleration and load).

[0029] The strain-displacement coupling coefficient is specifically: (4) is the strain data, and d is the displacement data.

[0030] The vibration-load transfer function is specifically: (5) where a(t) is the acceleration data, and P(t) is the load data. The function parameters are collected by extracting the function amplitude peak value and the corresponding frequency.

[0031] S102, a load state prediction model based on LSTM-CNN combined with an attention mechanism is constructed, the data feature vector is processed, and a shipyard steel bent frame load prediction value and a state probability are output.

[0032] In the embodiment, the load state prediction model based on LSTM-CNN combined with an attention mechanism is composed of an LSTM layer, a CNN layer, and an attention mechanism. Specifically, the LSTM layer of the embodiment adopts a bidirectional LSTM structure, 128 hidden units are set for each direction, and is used to capture the long-term time sequence correlation of the load data. The LSTM dynamically adjusts the information flow through the forget gate, the input gate, and the output gate, where the forget gate determines how much historical information to retain, the input gate controls the update of new information, and the output gate generates the hidden state of the current time step.

[0033] The bidirectional design enables the model to consider the influence of the previous and subsequent time steps at the same time, and finally outputs the hidden state of each time step, with a dimension of 60x256 (bidirectional splicing result).

[0034] The number of hidden layer units of the LSTM layer is 128, and the bidirectional LSTM captures the forward and backward dependencies, and the unit calculation formula is specifically: (6) ​ (7) (8) (9) (10) (11) where f t is the forget gate, i t is the input gate, O t is the output gate, C t is the memory cell, W f , W i , W c , W o are weight matrices, b f , b i , b c , b o are bias terms, is element-wise multiplication.

[0035] The time series hidden state (60x256) output by the LSTM is reshaped into a three-dimensional tensor (60x256x1) as the input of the CNN. A 3x1 convolution kernel is used for local feature extraction, with a total of 64 convolution channels, and a ReLU activation function is used to enhance the non-linear expression ability. Global max pooling operation is performed on the convolution result to extract the most significant features of each channel, outputting a 64-dimensional feature vector.

[0036] At the same time, the embodiment adopts an attention mechanism to impose attention weights on each time step hidden state output by the LSTM. The importance scores of each time step are calculated through a learnable weight matrix, and the attention weighted feature vector is obtained by weighting and summing the normalized importance scores, which is used to highlight the load change information of the key time step.

[0037] The specific enhancement of the key time step weight is: (12) where W a is the attention mechanism weight matrix, v is the weight vector, and b a is the bias.

[0038] The final feature output by the self-attention mechanism is .

[0039] The LSTM, CNN and attention features are concatenated and connected to a fully connected layer, which outputs the probability distribution of normal, warning and dangerous states through the Softmax function, realizing the classification task. At the same time, a separate fully connected layer is used to map the features to continuous prediction results of the load value.

[0040] Specifically, in the present embodiment, the output Softmax output classification probability p e R 3 , the output regression task load value prediction result .

[0041] The loss function of the neural network constructed in the present embodiment is specifically as follows: (13) Wherein, L cls is the cross-entropy loss, L reg is the MSE regression loss, L aux is the regularization loss, a is the classification loss weight, b is the regression loss weight, and g is the regularization loss weight.

[0042] S103, based on the shipyard steel frame load prediction value and the state probability, the shipyard steel frame dynamic threshold early warning and online feedback are carried out, and the shipyard steel frame load real-time monitoring and early warning are realized.

[0043] In the present embodiment, after completing the steel frame load prediction and state prediction based on the neural network, real-time online early warning needs to be carried out based on the prediction result, and the specific process is as follows: First, the dynamic threshold is generated, and the dynamic threshold of the graded early warning is generated based on the historical load data statistical analysis and expert experience rules, and the specific process is as follows: Baseline threshold calculation: the mean and standard deviation of the historical load data are calculated, and the safety load limit in the steel frame design specification is combined to preliminarily determine the static threshold interval of the normal, early warning and dangerous three levels.

[0044] In the present embodiment, the static threshold interval is specifically as follows: (14) Wherein, m is the historical load mean, is the historical standard deviation, k1 and k2 are threshold coefficients, wherein k1 = 1.5 and k2 = 3.0.

[0045] Environmental parameter correction: real-time environmental parameters such as temperature and humidity are introduced, and the threshold is dynamically adjusted according to the material thermal expansion coefficient and humidity sensitivity characteristics. For example: when the temperature rises by 1℃, the upper limit of the load threshold decreases by 0.5%; when the humidity increases by 10%, the lower limit of the threshold increases by 0.2%.

[0046] Specifically: (15) Wherein, m is the baseline load mean, k is the sensitivity coefficient, a T , a H are environmental correction factors, is the temperature change, is the humidity variation amount, μ n is the adaptive threshold.

[0047] Finally, the dynamic threshold interval [μ low , μ high ] is obtained.

[0048] Sensitivity adaptation: Through historical false alarm rate and false negative rate data, automatically adjust the threshold sensitivity coefficient. When the recent false alarm rate is high, appropriately relax the threshold interval; if the false negative rate rises, tighten the threshold trigger condition.

[0049] Then compare the model prediction value with the dynamic threshold to realize multi-level early warning triggering, the specific process is: Multi-level early warning triggering: compare the real-time predicted load value with the dynamic threshold to trigger a graded response mechanism: ‌Normal state (green), trigger only data operation. Specifically: when the predicted load value is continuously below the early warning threshold, only record data and generate periodic health reports, including load trend, key node strain distribution and abnormal event statistics, which are manifested as.

[0050] Warning state (yellow), trigger manual review. Specifically: if the predicted value first breaks through the early warning threshold, the system immediately triggers the following actions, namely: Send an early warning notice to the operation and maintenance personnel containing the load exceeding position, prediction deviation value and environmental parameters; start the manual review process, retrieve the camera screen and historical data curve of the corresponding position, exclude false judgments caused by transient interference; automatically enhance the data acquisition frequency (such as from 10Hz to 50Hz), continuously monitor for 10 minutes to confirm the trend.

[0051] Dangerous state (red), activate the emergency support system. Specifically: when the predicted value exceeds the danger threshold or is in the early warning interval for 3 consecutive times, execute the emergency response: Activate the steel frame emergency support system (such as hydraulic jacks, prestressed cables) to limit the further expansion of deformation; close the surrounding personnel access and trigger sound and light alarms, and through the Internet of Things platform, link up the fire and emergency management departments; generate a structure damage assessment report, mark high-risk areas and suggest reinforcement schemes.

[0052] Finally, online feedback optimization: according to the deviation of the early warning result and the actual load condition, optimize the model and threshold parameters in real time: Specifically, first, false positive / false negative case learning: collect the sensor raw data, environmental parameters and manual review conclusions in false positive cases, and label them as "normal" samples to join the training set; Reverse analysis of false negative events, extract the load characteristics 1 hour before the false negative event as new "dangerous" samples to strengthen the model's ability to identify early signs.

[0053] Second, incremental model update: automatically input new data into the model every week, slightly adjust the parameters based on the original weights, and avoid performance degradation due to changes in data distribution. For example: only update the weights of the fully connected layer, freeze the LSTM and CNN basic feature extraction layer.

[0054] After updating, use cross-validation to evaluate the classification accuracy and regression error of the model on the mixed data set (historical data + new data). If the performance drops by more than 2%, roll back to the previous version and trigger the manual intervention process.

[0055] Finally, build a threshold self-evolution mechanism: based on the Bayesian optimization algorithm, dynamically adjust the threshold coefficient to minimize the weighted sum of false positive rate and false negative rate. For example: when monitoring a certain area due to material aging, the load fluctuation is intensified, the warning threshold of this area is automatically relaxed; if the new sensor shows that the local stiffness is improved, the corresponding threshold is tightened to improve the monitoring sensitivity.

[0056] Generate a threshold adjustment report every quarter to record the parameter change trajectory and its impact on the warning effect for engineers to review and solidify effective adjustment strategies.

[0057] The method of the embodiment realizes the leap from "passive response" to "active warning" of steel frame structure load monitoring through multi-source data fusion, intelligent algorithm driving and dynamic closed-loop feedback, breaking through the limitations of traditional methods in terms of accuracy, real-time performance, reliability and economy, and providing a practical technical paradigm for intelligent operation and maintenance of infrastructure.

[0058] As shown in Figure 2 The embodiment also provides a shipyard steel bent frame load working condition monitoring and warning system, which includes a data acquisition and feature construction module 1, a steel bent frame load prediction module 2, and a load monitoring and warning module 3. The data acquisition and feature construction module 1 is used to deploy a plurality of sensors on the shipyard steel frame, acquire shipyard steel frame stress displacement velocity data, and collect data feature vectors through data preprocessing and feature extraction. The steel bent frame load prediction module 2 is used to construct a load state prediction model based on LSTM-CNN and combined with an attention mechanism, process the data feature vectors, and output shipyard steel bent frame load prediction values and state probabilities. The load monitoring and warning module 3 is used to perform shipyard steel bent frame dynamic threshold warning and online feedback based on the shipyard steel bent frame load prediction values and state probabilities, and realize real-time monitoring and warning of the shipyard steel bent frame load.

[0059] It should be noted that, for the foregoing method embodiments, the purposes of simplicity and brevity of the description are served as a series of sequential actions are used to illustrate the embodiments; however, a person of ordinary skill in the art should understand that the present disclosure is not limited by the order of the actions, as some steps can be performed in other orders or at the same time, according to the present disclosure. Furthermore, it should be understood that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present disclosure.

[0060] It should be understood that the steps shown in the above various forms of flowcharts can be reordered, added, or deleted. For example, the steps described in the present disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, which are not limited herein.

[0061] The above detailed description does not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for monitoring and early warning of load conditions of steel bents in a shipyard, characterized in that: The following steps are involved: Deploy several sensors on the shipyard steel frame to collect stress, displacement and velocity data of the shipyard steel frame. Through data preprocessing and feature extraction, the data feature vectors are collected. A load state prediction model based on LSTM-CNN and attention mechanism was constructed to process the data feature vectors and output the predicted load value and state probability of the shipyard steel bent frame. Based on the predicted value of the shipyard steel bent load and the state probability, the dynamic threshold warning and online feedback of the shipyard steel bent load are carried out to realize the real-time monitoring and early warning of the shipyard steel bent load.

2. The shipyard steel bent load condition monitoring and early warning method according to claim 1 is characterized in that: The process of deploying a number of sensors on the shipyard steel frame to collect stress and displacement velocity data of the shipyard steel frame, and collecting and obtaining data feature vectors through data preprocessing and feature extraction is as follows: Based on the finite element simulation method, the stress concentration area of ​​the shipyard steel bent frame was determined. A grid layout strategy was adopted to deploy several sensors on the shipyard steel bent frame to collect stress-displacement velocity data. The stress displacement velocity data is denoised using a wavelet threshold denoising method combined with a Kalman filter method, and the multi-sensor data is confidence-fused using the DS evidence theory to obtain fused data; Feature extraction is performed on the fused data to generate a 32-dimensional data feature vector.

3. The shipyard steel bent load condition monitoring and early warning method according to claim 1 is characterized in that: The data feature vector includes time domain features, frequency domain features, time-frequency domain features and fusion features; The time domain characteristics include mean, standard deviation, kurtosis, skewness, peak-to-peak value, shape factor, impulse factor and zero-crossing rate; The frequency domain features include main frequency amplitude, frequency band energy ratio, spectrum entropy, harmonic distortion and transfer function amplitude; The time-frequency domain features include wavelet packet energy entropy and correlation coefficients of seven adjacent sub-bands; The fusion features include strain-displacement coupling coefficients and vibration-load transfer function parameters.

4. The shipyard steel bent load condition monitoring and early warning method according to claim 1 is characterized in that: The process of constructing a load state prediction model based on LSTM-CNN and combined with the attention mechanism, processing the data feature vector, and outputting the predicted load value and state probability of the shipyard steel bent is as follows: Constructing a bidirectional LSTM structure, processing the data feature vector, and outputting a time step hidden state, wherein the bidirectional LSTM structure is set with 128 hidden units in each direction; The time step hidden state is reshaped into a three-dimensional tensor, and a 3×1 convolution kernel is used to extract local features. After processing with a ReLU activation function, a global maximum pooling operation is performed on the processed convolution result to extract channel salient features and output a multi-dimensional feature vector; The attention mechanism is used to apply attention weights to the hidden state of each time step output by the LSTM. The importance score of each time step is calculated through a learnable weight matrix. After normalizing the importance score, the weighted summation is performed to obtain the attention weighted feature vector. The LSTM, CNN and attention-weighted feature vectors are spliced ​​and connected to the fully connected layer. Based on the multidimensional feature vector and the attention-weighted feature vector, the probability distribution of the three states of normal, warning and danger is output through the Softmax function. A separate fully connected layer is used to map the features into continuous prediction results of the load value, and the load value prediction result is output.

5. The shipyard steel bent load condition monitoring and early warning method according to claim 1 is characterized in that: The process of performing dynamic threshold warning and online feedback of shipyard steel bent loads based on the predicted value and state probability of the shipyard steel bent loads and realizing real-time monitoring of the shipyard steel bent loads is as follows: Generate dynamic thresholds for graded warnings based on statistical analysis of historical load value prediction results and expert experience rules; Comparing the real-time predicted load value with the dynamic threshold to obtain an early warning result and trigger a graded response mechanism; Based on the deviation between the early warning result and the actual load situation, the model and threshold parameters are optimized in real time.

6. The shipyard steel bent load condition monitoring and early warning method according to claim 5 is characterized in that: The process of generating dynamic thresholds for graded warnings based on statistical analysis of historical load data and expert experience rules is as follows: The mean and standard deviation of the historical load value prediction results were calculated, and combined with the safe load limits in the steel frame design specifications, the static threshold intervals of the three levels of normal, warning, and danger were preliminarily determined; Introducing real-time environmental parameters to dynamically adjust the static threshold range based on the thermal expansion coefficient and humidity sensitivity of the material; Collect historical false alarm rate and missed alarm rate data, automatically adjust the threshold sensitivity coefficient, and adapt the threshold range according to actual conditions.

7. The shipyard steel bent load condition monitoring and early warning method according to claim 6 is characterized in that: The method of introducing real-time environmental parameters and dynamically adjusting the static threshold range according to the thermal expansion coefficient and humidity sensitivity of the material is as follows: Among them, μ is the mean value of the reference load, k is the sensitivity coefficient, α T , α H is the environmental correction factor, is the temperature change, is the humidity change, μ n is the adaptive threshold.

8. The shipyard steel bent load condition monitoring and early warning method according to claim 5 is characterized in that: The process of comparing the real-time predicted load value with the dynamic threshold to obtain the warning result and trigger the hierarchical response mechanism is as follows: When the real-time predicted load value is continuously lower than the minimum warning threshold, it is considered to be in a normal state, and only data is recorded and periodic health reports are generated; When the real-time predicted load value is between the minimum and maximum thresholds, it is considered a warning state, which triggers the manual review process and sends a warning notification to the operation and maintenance personnel, including the load exceeding position, predicted deviation value, and environmental parameters; When the real-time predicted load value exceeds the maximum threshold or is in the warning range for three consecutive times, an emergency response is executed, activating the steel frame emergency support system to limit the expansion of deformation, closing the surrounding personnel passages, and triggering the sound and light alarm. The fire and emergency management departments are linked through the Internet of Things platform, and finally a structural damage assessment report is generated, marking high-risk areas and recommending reinforcement plans.

9. A shipyard steel bent load condition monitoring and early warning system, characterized in that: include: Data acquisition and feature construction module (1), steel bent load prediction module (2) and load monitoring and early warning module (3); The data acquisition and feature construction module (1) is used to deploy a number of sensors on the shipyard steel frame to collect stress, displacement and velocity data of the shipyard steel frame, and collect data feature vectors through data preprocessing and feature extraction; The steel bent load prediction module (2) is used to construct a load state prediction model based on LSTM-CNN and combined with an attention mechanism, process the data feature vector, and output the predicted value of the shipyard steel bent load and the state probability; The load monitoring and early warning module (3) is used to perform dynamic threshold early warning and online feedback of the shipyard steel bent frame based on the shipyard steel bent frame load prediction value and state probability, thereby realizing real-time monitoring and early warning of the shipyard steel bent frame load.

Images