Dynamic risk prediction method and system for offshore construction, medium and equipment

Abstract

The invention discloses a risk dynamic prediction method, system, medium and equipment for offshore construction, and belongs to the field of offshore construction safety prediction.The method comprises the steps that laser point cloud time series data of offshore wind power construction components and meteorological environment monitoring data including wind speed, wind direction, wave height, tide level, flow velocity and rainfall are collected; constructing a component error time sequence and a meteorological environment time sequence which are in time alignment; inputting the two sequences into a preset risk prediction model, extracting component error features layer by layer through a multi-stage time sequence convolutional network, performing weighted correction on the features by using meteorological environment sequence fragments in a corresponding time view in each layer, and transmitting the corrected features layer by layer; and predicting an error of the component in a future preset time period based on the final layer feature regression, and outputting a risk prediction result in combination with a preset risk judgment rule. By implementing the method, the problem that risk prediction is lagged and inaccurate in the offshore wind power construction process in the prior art can be solved.

Description

Technical Field

[0001] This invention belongs to the field of offshore construction safety prediction, and relates to a method, system, medium and equipment for dynamic prediction of risks in offshore construction. Background Technology

[0002] Offshore wind power projects are characterized by complex construction environments, long operation cycles, and large spatial spans. Their construction process heavily relies on real-time perception and risk assessment of the structural component status and external sea conditions. In recent years, with the in-depth application of BIM (Building Information Modeling) technology in infrastructure construction, the ability to represent the construction process in three dimensions and integrate information has been continuously improved, providing fundamental support for the visualization and structured management of construction risks.

[0003] Currently, risk identification in offshore wind power construction mainly relies on sensor monitoring, UAV aerial surveys, or expert experience judgment. These methods are usually based on static threshold judgment or geometric deviation analysis of single-phase data, lacking the ability to fuse and model multi-source time-series data, and cannot effectively capture the evolution trend of risks in the time dimension. At the same time, existing systems generally do not deeply associate monitoring data with BIM three-dimensional spatial semantics, making it difficult to achieve dynamic positioning and prediction of component-level risks, resulting in delayed risk response and passive early warning, which cannot support the proactive management needs for the future. Summary of the Invention

[0004] This application provides a method, system, medium, and equipment for dynamic risk prediction in offshore construction, which can solve the problems of lag and inaccuracy in risk prediction during offshore wind power construction in the prior art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for dynamic risk prediction in offshore construction, comprising:

[0006] The system collects time-series laser point cloud data of current offshore wind power construction components, as well as meteorological and environmental monitoring data, and constructs time-aligned component error time series and meteorological and environmental time series. The meteorological and environmental monitoring data includes wind speed, wind direction, wave height, tide level, current velocity, and rainfall data.

[0007] The component error time series and meteorological environment time series are input into a preset risk prediction model. The risk prediction model extracts features from the component error time series layer by layer through a multi-level temporal convolutional network. After each layer of feature extraction, the meteorological environment sequence segment corresponding to the time horizon of the current layer's feature extraction is used to weight and correct the features output by the current layer. The corrected features are then used as the input for the next layer. Based on the corrected features of the last layer, the error prediction data of the component in the future preset time period is regressed and output.

[0008] Based on the error prediction data and combined with the preset risk assessment rules, the risk prediction result is output.

[0009] Compared to existing technologies, the embodiments of this application have the following beneficial effects: They collect time-series laser point cloud data of current offshore wind power construction components, as well as meteorological and environmental monitoring data including wind speed, wind direction, wave height, tide level, current velocity, and rainfall. They then construct time-aligned component error time series and meteorological and environmental time series, achieving precise synchronization of high-dimensional geometric state observations and multi-dimensional environmental disturbance variables on a unified time axis. This provides a physically complete and temporally consistent input foundation for risk prediction, avoiding modeling biases caused by asynchronous data or missing environmental factors. The component error time series and meteorological and environmental time series are input into a pre-set risk prediction model, enabling the model to simultaneously utilize the component's own historical error evolution information and external meteorological disturbance information, breaking the limitations of traditional reliance on a single data source. Furthermore, a multi-level temporal convolutional network is used to extract features from the component error time series layer by layer, gradually abstracting dynamic patterns at different time scales using a deep convolutional structure, effectively capturing… The system obtains the long-term dependence and nonlinear evolution characteristics of errors in the time dimension. After feature extraction at each layer, the meteorological environment sequence fragments corresponding to the time horizon of the current layer feature extraction are used to weight and correct the output features of the current layer. This allows the meteorological environment information to dynamically modulate the feature representation within the time window that matches the error features, realizing the hierarchical and temporally aligned fusion intervention of exogenous variables on the internal state, avoiding semantic inaccuracies caused by global or fixed window fusion. Based on the error prediction data of the feature regression output component after the last layer correction in the future preset time period, the construction deviation trend is forward extrapolated. The above technical features work together to construct an end-to-end prediction mechanism with multi-level temporal convolution as the core, meteorological environment sequence as the dynamic correction basis, and error time series as the main input. Overall, it realizes the transformation of construction risk from "static threshold judgment" to "dynamic trend prediction", effectively solving the problems of risk prediction lag and inaccuracy in the existing technology.

[0010] In some embodiments of the first aspect of this application, the time-aligned component error time series and meteorological environment time series include:

[0011] The laser point cloud time series data is spatially aligned with the theoretical geometric model of the corresponding component in the preset BIM model, and the distance deviation between the aligned point cloud and the theoretical geometric model is calculated to obtain the original error sequence.

[0012] The original error sequence is time-aligned with the meteorological and environmental monitoring data, and resampled with a uniform time step to obtain a time-synchronized component error time series and meteorological and environmental time series.

[0013] Compared with existing technologies, the above embodiments have the following beneficial effects: spatial alignment of laser point cloud time-series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculation of distance deviation to obtain the original error sequence, so that the error measurement has clear BIM semantic correlation and component-level spatial positioning capability, avoiding the ambiguity of error attribution caused by the lack of structural semantic support in traditional point cloud processing; time alignment of the original error sequence with meteorological and environmental monitoring data and resampling with a unified time step ensures strict synchronization of multimodal data in the time dimension, eliminates time-series distortion caused by sampling frequency differences or asynchronous acquisition, and lays a data consistency foundation for the stable training and prediction of subsequent deep learning models.

[0014] In some embodiments of the first aspect of this application, the step of performing layer-by-layer feature extraction on the component error time series includes:

[0015] Perform causal dilated convolution on the input data of the current layer to obtain the convolutional feature sequence;

[0016] Channel attention enhancement processing is performed on the convolutional feature sequence to generate a channel-weighted feature sequence;

[0017] The channel-weighted feature sequence is residually concatenated with the input data of the current layer to obtain the output features of the current layer.

[0018] The input data for the first layer is the component error time series, and the input data for the remaining layers is the output features of the previous layer.

[0019] Compared with existing technologies, the above embodiments have the following beneficial effects: Causal dilation convolution processing is applied to the current layer input data to ensure that predictions do not depend on future information through causal constraints. Simultaneously, the dilation mechanism exponentially expands the receptive field, effectively capturing long-cycle construction error evolution patterns without significantly increasing the number of parameters. Channel attention enhancement processing is performed on the convolutional feature sequences to dynamically highlight feature channels that contribute highly to the current risk state (such as displacement in a specific direction or key meteorological coupling terms), suppress redundant or noisy channels, and improve feature discriminability. Residual connections are made between the channel-weighted feature sequences and the current layer input data to alleviate the gradient vanishing problem in deep networks, ensuring the training stability and representation fidelity of the model under multi-level stacking, while preserving local details of the original error data and avoiding the loss of key transient information during deep feature extraction.

[0020] In some embodiments of the first aspect of this application, the step of using meteorological environment sequence segments in the meteorological environment time series that correspond to the time horizon of the current layer feature extraction to perform weighted correction on the features output by the current layer, and obtaining the corrected features as input to the next layer, includes:

[0021] Based on the structural parameters of the current layer in the multi-level temporal convolutional network, determine the time window covered by the feature extraction of the current layer;

[0022] Extract the meteorological environment sequence segment corresponding to the time window from the meteorological environment time series;

[0023] The meteorological environment sequence fragments are standardized, and the dimension is adjusted to the number of channels of the output features of the current layer;

[0024] Based on the adjusted meteorological environment sequence segments, a channel-wise weighted coefficient vector and intercept vector are generated through linear mapping;

[0025] Using the weighted coefficient vector and intercept vector, linear correction is performed on the features output by the current layer to obtain the corrected features.

[0026] Compared to existing technologies, the above embodiments have the following advantages: The time window covered by feature extraction is determined based on the structural parameters of the current layer, ensuring strict alignment between the meteorological environment segment extraction and the network's internal temporal perspective, thus guaranteeing the spatiotemporal semantic consistency of exogenous variable correction; segments corresponding to the time window are extracted from the meteorological environment time series, achieving layer-adaptive dynamic environmental context extraction and avoiding information mismatch caused by global averaging or fixed windows; meteorological segments are standardized and their dimensions are adjusted to the number of feature channels in the current layer, ensuring their computability with error features in the spatial dimension; channel-wise weighted coefficient vectors and intercept vectors are generated through linear mapping, transforming meteorological variables into interpretable linear modulation parameters for error features, enabling the model to apply differentiated influences to different meteorological factors (such as wind speed and wave height) on different feature channels; linear correction of the current layer's output features is performed using these coefficients and intercepts, achieving low coupling and explicit control of internal representation by exogenous drivers, improving prediction accuracy while preserving physical interpretability, and avoiding overfitting of the model under extreme sea conditions.

[0027] In some embodiments of the first aspect of this application, the step of outputting a risk prediction result based on the error prediction data and in conjunction with a preset risk assessment rule includes:

[0028] The error prediction data is filtered to obtain the error filtered data;

[0029] Based on the error-filtered data, a rolling statistic within a preset time window is calculated; wherein, the rolling statistic includes amplitude, slope of change, rolling mean, and confidence interval;

[0030] The rolling statistics are compared with the preset multi-level risk threshold range, and combined with the minimum duration constraint, the final risk level is determined as the risk prediction result.

[0031] Compared with existing technologies, the above embodiments have the following beneficial effects: filtering the error prediction data effectively suppresses the influence of extreme outliers, improving the physical rationality and temporal smoothness of the prediction sequence; calculating rolling statistics within a preset time window based on the filtered data, including amplitude, slope of change, rolling mean, and confidence interval, characterizes the dynamic behavior of the error from multiple dimensions, avoiding the limitations of single threshold judgment; comparing the rolling statistics with preset multi-level risk threshold intervals, and combining them with minimum duration constraints, prevents erroneous risk level jumps caused by instantaneous fluctuations, enhancing the stability and engineering applicability of the judgment results; the final output risk level not only reflects the current deviation magnitude but also integrates trend, persistence, and uncertainty information, significantly improving the accuracy and operability of risk warning.

[0032] Secondly, the present invention also provides a dynamic risk prediction system for offshore construction, comprising: a data acquisition module, a model prediction module, and an output module;

[0033] The data acquisition module is used to collect time-series laser point cloud data of current offshore wind power construction components and meteorological environmental monitoring data, and to construct time-aligned component error time series and meteorological environmental time series; wherein the meteorological environmental monitoring data includes: wind speed, wind direction, wave height, tide level, current velocity and rainfall data.

[0034] The model prediction module is used to input the component error time series and meteorological environment time series into a preset risk prediction model, so that the risk prediction model extracts features from the component error time series layer by layer through a multi-level temporal convolutional network. After each layer of feature extraction, the meteorological environment sequence segment in the meteorological environment time series that corresponds to the time horizon of the current layer's feature extraction is used to perform weighted correction on the features output by the current layer, and the corrected features are used as the input for the next layer. Based on the corrected features of the last layer, the error prediction data of the component in the future preset time period is regressed and output.

[0035] The output module is used to output risk prediction results based on the error prediction data and in combination with preset risk determination rules.

[0036] Compared to existing technologies, the above embodiments of this application have the following beneficial effects: They collect time-series laser point cloud data of current offshore wind power construction components, as well as meteorological and environmental monitoring data including wind speed, wind direction, wave height, tide level, current velocity, and rainfall. They then construct time-aligned component error time series and meteorological and environmental time series, achieving precise synchronization of high-dimensional geometric state observation and multi-dimensional environmental disturbance variables on a unified time axis. This provides a physically complete and temporally consistent input foundation for risk prediction, avoiding modeling biases caused by asynchronous data or missing environmental factors. By inputting the component error time series and meteorological and environmental time series into a preset risk prediction model, the model can simultaneously utilize the component's own historical error evolution information and external meteorological disturbance information, breaking the limitations of traditional reliance on a single data source. Furthermore, by using a multi-level temporal convolutional network to extract features layer by layer from the component error time series, and by gradually abstracting dynamic patterns at different time scales using a deep convolutional structure, the model effectively… This method captures the long-term dependence and nonlinear evolution characteristics of errors over time. After feature extraction at each layer, the output features of the current layer are weighted and corrected using meteorological environmental sequence segments corresponding to the temporal perspective of the current layer's feature extraction. This allows meteorological environmental information to dynamically modulate feature representation within a time window that matches the error features, achieving hierarchical and temporally aligned fusion intervention of exogenous variables on internal states, avoiding semantic inaccuracies caused by global or fixed-window fusion. Based on the error prediction data of the feature regression output components after the last layer's correction within a preset future time period, the trend of construction deviations is forward extrapolated. The synergistic effect of the above technical features constructs an end-to-end prediction mechanism with multi-level temporal convolution as the core, meteorological environmental sequences as the dynamic correction basis, and error time series as the main input. Overall, it realizes the transformation of construction risk from "static threshold judgment" to "dynamic trend prediction," effectively solving the problems of lag and inaccuracy in risk prediction in existing technologies.

[0037] In some embodiments of the second aspect of this application, the data acquisition module includes: a deviation calculation unit and an alignment unit;

[0038] The deviation calculation unit is used to spatially align the laser point cloud time series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculate the distance deviation between the aligned point cloud and the theoretical geometric model to obtain the original error sequence.

[0039] The alignment unit is used to time-align the original error sequence with the meteorological environment monitoring data and resample it with a uniform time step to obtain a time-synchronized component error time sequence and meteorological environment time sequence.

[0040] Compared with existing technologies, the above embodiments have the following beneficial effects: spatial alignment of laser point cloud time-series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculation of distance deviation to obtain the original error sequence, so that the error measurement has clear BIM semantic correlation and component-level spatial positioning capability, avoiding the ambiguity of error attribution caused by the lack of structural semantic support in traditional point cloud processing; time alignment of the original error sequence with meteorological and environmental monitoring data and resampling with a unified time step ensures strict synchronization of multimodal data in the time dimension, eliminates time-series distortion caused by sampling frequency differences or asynchronous acquisition, and lays a data consistency foundation for the stable training and prediction of subsequent deep learning models.

[0041] In some embodiments of the second aspect of this application, the model prediction module includes: a convolutional unit, an attention weighting unit, and a residual connection unit;

[0042] The convolutional unit is used to perform causal dilation convolution on the input data of the current layer to obtain a convolutional feature sequence.

[0043] The attention weighting unit is used to perform channel attention enhancement processing on the convolutional feature sequence to generate a channel-weighted feature sequence;

[0044] The residual connection unit is used to perform a residual connection between the channel-weighted feature sequence and the input data of the current layer to obtain the output features of the current layer;

[0045] The input data for the first layer is the component error time series, and the input data for the remaining layers is the output features of the previous layer.

[0046] Compared with existing technologies, the above embodiments have the following beneficial effects: Causal dilation convolution processing is applied to the current layer input data to ensure that predictions do not depend on future information through causal constraints. Simultaneously, the dilation mechanism exponentially expands the receptive field, effectively capturing long-cycle construction error evolution patterns without significantly increasing the number of parameters. Channel attention enhancement processing is performed on the convolutional feature sequences to dynamically highlight feature channels that contribute highly to the current risk state (such as displacement in a specific direction or key meteorological coupling terms), suppress redundant or noisy channels, and improve feature discriminability. Residual connections are made between the channel-weighted feature sequences and the current layer input data to alleviate the gradient vanishing problem in deep networks, ensuring the training stability and representation fidelity of the model under multi-level stacking, while preserving local details of the original error data and avoiding the loss of key transient information during deep feature extraction.

[0047] Thirdly, the present invention also provides a risk dynamic prediction device for offshore construction, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is loaded onto the processor, it implements the steps of any of the risk dynamic prediction methods for offshore construction of the present invention.

[0048] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the risk dynamic prediction methods for offshore construction of the present invention. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a method for dynamic risk prediction in offshore construction provided in some embodiments of the present invention.

[0050] Figure 2 This is a schematic diagram of the structure of a dynamic risk prediction system for offshore construction provided in some embodiments of the present invention.

[0051] Figure 3 This is a structural diagram of a risk dynamic prediction device for offshore construction provided in some embodiments of the present invention.

[0052] Figure 4 This is a structural diagram of a risk prediction model provided in some embodiments of the present invention.

[0053] Figure 5 This is a schematic diagram illustrating data standardization and time alignment provided in some embodiments of the present invention.

[0054] Figure 6 This is a schematic diagram of a TCN residual module and a causal convolutional network architecture provided in some embodiments of the present invention.

[0055] Figure 7 This is a diagram of an ECA attention mechanism architecture provided in some embodiments of the present invention.

[0056] Figure 8 This is a diagram of a regression regulation head module architecture provided in some embodiments of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1:

[0059] Please refer to Figure 1 To address the issues of delayed and inaccurate risk prediction during offshore wind power construction in existing technologies, an embodiment of the present invention provides a dynamic risk prediction method for offshore construction, comprising steps S1 to S3:

[0060] Step S1: Collect laser point cloud time-series data of the current offshore wind power construction components and meteorological environment monitoring data, and construct time-aligned component error time series and meteorological environment time series; wherein, the meteorological environment monitoring data includes: wind speed, wind direction, wave height, tide level, current velocity and rainfall data.

[0061] Furthermore, the time series of component errors and meteorological environment data that are constructed and aligned with the construction time can be implemented through the following preferred embodiments, including steps S11-S12, as follows:

[0062] S11: Spatially align the laser point cloud time series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculate the distance deviation between the aligned point cloud and the theoretical geometric model to obtain the original error sequence.

[0063] S12: The original error sequence is time-aligned with the meteorological and environmental monitoring data, and resampled with a uniform time step to obtain a time-synchronized component error time sequence and meteorological and environmental time sequence.

[0064] In this preferred embodiment, the laser point cloud time series data is spatially aligned with the theoretical geometric model of the corresponding component in the preset BIM model, and the distance deviation is calculated to obtain the original error sequence. This enables the error measurement to have clear BIM semantic correlation and component-level spatial positioning capability, avoiding the ambiguity of error attribution caused by the lack of structural semantic support in traditional point cloud processing. The original error sequence is time-aligned with meteorological and environmental monitoring data and resampled with a unified time step, ensuring strict synchronization of multimodal data in the time dimension. This eliminates the time series distortion caused by sampling frequency differences or asynchronous acquisition, laying a data consistency foundation for the stable training and prediction of subsequent deep learning models.

[0065] Step S2: Input the component error time series and meteorological environment time series into the preset risk prediction model, so that the risk prediction model extracts features from the component error time series layer by layer through a multi-level temporal convolutional network. After each layer of feature extraction, the meteorological environment sequence segment in the meteorological environment time series that corresponds to the time horizon of the current layer feature extraction is used to weight and correct the features output by the current layer, and the corrected features are used as the input of the next layer. Based on the corrected features of the last layer, the error prediction data of the component in the future preset time period is regressed and output.

[0066] Furthermore, the stepwise feature extraction of the component error time series can be implemented through the following preferred embodiments, including steps S21-S23, as follows:

[0067] S21: Perform causal dilation convolution on the input data of the current layer to obtain the convolutional feature sequence;

[0068] S22: Perform channel attention enhancement processing on the convolutional feature sequence to generate a channel-weighted feature sequence;

[0069] S23: Perform residual concatenation between the channel-weighted feature sequence and the input data of the current layer to obtain the output features of the current layer;

[0070] The input data for the first layer is the component error time series, and the input data for the remaining layers is the output features of the previous layer.

[0071] In this preferred embodiment, causal dilated convolution processing is performed on the current layer input data. Causality constraints ensure that the prediction does not depend on future information, while the dilation mechanism exponentially expands the receptive field, effectively capturing long-cycle construction error evolution patterns without significantly increasing the number of parameters. Channel attention enhancement processing is performed on the convolutional feature sequence to dynamically highlight feature channels that contribute highly to the current risk state (such as displacement in a specific direction or key meteorological coupling terms), suppress redundant or noisy channels, and improve feature discriminability. The channel-weighted feature sequence is residually connected with the current layer input data to alleviate the gradient vanishing problem in deep networks, ensure the training stability and representation fidelity of the model under multi-level stacking, and retain the local details of the original error data to avoid the loss of key transient information during deep feature extraction.

[0072] Furthermore, the weighted correction in step S2 can be implemented through the following preferred embodiments, including steps S24-S28, as follows:

[0073] S24: Determine the time window covered by the feature extraction of the current layer based on the structural parameters of the current layer in the multi-level temporal convolutional network;

[0074] S25: Extract the meteorological environment sequence segment corresponding to the time window from the meteorological environment time series;

[0075] S26: Standardize the meteorological environment sequence segment and adjust the dimension to the number of channels of the output feature of the current layer;

[0076] S27: Based on the adjusted meteorological environment sequence segments, a channel-wise weighted coefficient vector and intercept vector are generated through linear mapping;

[0077] S28: Using the weighted coefficient vector and intercept vector, perform linear correction on the features output by the current layer to obtain the corrected features.

[0078] In this preferred embodiment, the time window covered by feature extraction is determined according to the structural parameters of the current layer, ensuring that the truncation of meteorological environment segments is strictly aligned with the internal temporal perspective of the network, thus ensuring the spatiotemporal semantic consistency of exogenous variable correction. Segments corresponding to the time window are extracted from the meteorological environment time series to achieve dynamic environmental context extraction adapted to each layer, avoiding information mismatch caused by global averaging or fixed windows. Meteorological segments are standardized and their dimensions are adjusted to the number of feature channels in the current layer, ensuring their computability with error features in the spatial dimension. Channel-wise weighted coefficient vectors and intercept vectors are generated through linear mapping, transforming meteorological variables into interpretable linear modulation parameters for error features, enabling the model to apply differentiated influences to different meteorological factors (such as wind speed and wave height) on different feature channels. Linear correction of the current layer's output features is performed using these coefficients and intercepts, achieving low coupling and explicit control of internal representation by exogenous drivers. This improves prediction accuracy while preserving physical interpretability, avoiding overfitting of the model under extreme sea conditions.

[0079] Step S3: Based on the error prediction data and combined with the preset risk determination rules, output the risk prediction result.

[0080] Furthermore, step S3 can be implemented through the following preferred embodiments, including steps S31-S33, as follows:

[0081] S31: Filter the error prediction data to obtain filtered error data;

[0082] S32: Based on the error-filtered data, calculate the rolling statistics within a preset time window; wherein, the rolling statistics include amplitude, slope of change, rolling mean, and confidence interval;

[0083] S33: Compare the rolling statistics with the preset multi-level risk threshold range, and combine them with the minimum duration constraint to determine the final risk level as the risk prediction result.

[0084] In this preferred embodiment, the error prediction data is filtered to effectively suppress the impact of extreme outliers and improve the physical rationality and temporal smoothness of the prediction sequence. Based on the filtered data, rolling statistics within a preset time window are calculated, including amplitude, slope of change, rolling mean, and confidence interval, to characterize the dynamic behavior of the error from multiple dimensions and avoid the limitations of single threshold judgment. The rolling statistics are compared with preset multi-level risk threshold intervals and combined with minimum duration constraints to prevent false jumps in risk level due to instantaneous fluctuations, thereby enhancing the stability and engineering applicability of the judgment results. The final output risk level not only reflects the current deviation magnitude but also integrates trend, persistence, and uncertainty information, significantly improving the accuracy and operability of risk warning.

[0085] In practical implementation, the risk prediction model used in this application is constructed and trained using the following method:

[0086] I. Training Data Acquisition: This step aims to construct a multimodal time-series dataset at the construction component level for model training, validation, and testing, and to complete standardized mapping under the unified reference of the BIM 3D model. Each set of samples consists of input data and label data: Input data includes laser point cloud data, BIM data, marine environmental and meteorological parameters (wind speed, wind direction, wave height, tide level, current velocity, rainfall), and construction event anchor points (time axis encoding of key processes of engineering components); Label data consists of corresponding construction environment risk level annotations and construction component attribute information, including two types of information: "peak value" and "whether it exceeds the limit," and mapped to risk levels I–IV according to the grade boundary. Specifically, the following steps are included:

[0087] (1) Spatiotemporal reference and alignment: All input data (point cloud, sea state, work / operation log) uses BIM as the spatial coordinate reference, unifying the engineering coordinate system and elevation datum; the time is uniformly adopted as Coordinated Universal Time (UTC). When aligning time, the component record time is given priority as the reference time scale, and the time deviation of other data sources relative to the reference is controlled within ±1–3 seconds through hard trigger synchronization, and the synchronization method and alignment residual are recorded in the metadata.

[0088] (2) Label generation and risk classification: For construction components and their corresponding errors, the peak value is statistically analyzed within the observation period after the sample window as the "numerical peak value". The "whether it exceeds the limit" is determined according to the specification limit configured for the component, and the peak value is mapped to risk level I-IV according to the preset classification boundary. When there are multiple target quantities, the most unfavorable fusion principle is used to determine the comprehensive level.

[0089] (3) Sample Balance and Enhancement: To improve the robustness and generalization ability of the model, oversampling or cost-sensitive weights are used for rare hazardous samples; stratified sampling is carried out according to sea state (stable, fluctuating, extreme) to avoid the model biased towards normal operating conditions; small noise and disturbances are added to the input sequence to enhance the anti-interference ability; oversampling is performed appropriately for rare or extreme sea state samples to improve coverage; and micro-hourly time misalignment (sampling step level) is applied to some data sources to train the robustness of the model.

[0090] (4) Data set partitioning: The constructed multimodal time series data is divided into training set, validation set and test set according to the proportion. The training set is used for model training, the validation set is used to test the accuracy and generalization ability during the training process, and the test set is used to evaluate the performance and stability of the optimal model in actual application scenarios.

[0091] II. Model Design: Reference Figure 4 The diagram illustrates a risk prediction model structure. This invention employs an output correction architecture of "TCN backbone + multi-level regression control head": The component error time series is first encoded by causal multi-scale TCN to obtain a hidden representation, which is then used to form multiple prediction branches on the right side, divided by horizon or index, generating baseline results. The regression control head, connected in parallel with each branch, uses exogenous quantities such as meteorological / operational conditions to map channel-by-channel weights and intercepts, performing linear, loosely coupled, real-time correction on the regressible features of the corresponding branch to obtain the final predicted value. This design makes the relationship between "exogenous drivers—coefficients—correction amounts" explicit, providing interpretable control channels and improving accuracy and stability under non-stationary sea states. The corrected error output can be directly interfaced with BIM components via a post-processing module to achieve risk classification and alarm linkage, as detailed below:

[0092] (1) Data normalization: Reference Figure 5 The diagram illustrates a data standardization and time alignment process. It cleanses two types of raw data—"construction error" and "meteorological"—to the same time axis, outputting time-series samples that can be directly used to train the model. First, a unified time step (e.g., 5 minutes) is determined to generate a common time axis. Then, the units for each channel are standardized (e.g., error / displacement → mm, wind speed → m / s, wave height → m, air pressure → hPa, tide level → m), and conversion factors are recorded. Each channel is limited according to engineering physical upper and lower limits, and values ​​exceeding these limits are marked as anomalies. A two-level anomaly detection system is used to uniformly write anomalies and missing data into a mask. The error sequence is linearly interpolated onto the common time axis, and the meteorological sequence is aggregated onto the same time axis using the window mean and median. Finally, standard model input data is obtained through time resampling, component error calculation, and data normalization.

[0093] To clarify, the data normalization module in the model is not a core function and can be directly integrated into the model. This allows for direct input of raw data into the model, or it can be separated from the model and used as a data preprocessing step before input. For example, in this embodiment, the "time-aligned component error time series and meteorological environment time series" in step S1 are not obtained from the model's data normalization module but are obtained outside the model. Implementers can choose this method according to their needs, and there are no restrictions on it.

[0094] (2) TCN network residual unit: Reference Figure 6 The diagram illustrates a TCN residual module and causal convolutional network architecture. The TCN consists of three modules: causal convolution, dilated convolution, and residual blocks. Essentially, it's a one-dimensional CNN with dilated causal convolution. Causal convolution is a special convolution operation that ensures each element of the output sequence depends only on the current and previous input sequence elements, without introducing future information, thus maintaining causality. In construction risk prediction tasks, the network needs to remember long-term information and expand the receptive field of the convolutional kernels without increasing computation or the number of parameters; therefore, dilated convolution is added to the TCN. In deep neural networks, as the number of layers increases, the model may experience gradient vanishing or exploding during training. Therefore, residual connections are introduced in the TCN output layer. The stacking of residual blocks ensures the predictive power of the deep network and maintains stability as the number of network layers increases. Each residual module includes a dilated causal convolution and a ReLU activation function, combined with weight normalization and regularization layers to enhance the network's generalization ability. Finally, a 1×1 convolution operation ensures that the input and output dimensions are consistent. Deep TCN is composed of multiple such residual blocks stacked together. This structure can effectively capture long-term dependencies in time series data and improve the performance and stability of the model.

[0095] (3) ECA attention mechanism: Reference Figure 7 The diagram illustrates an ECA attention mechanism architecture. Efficient Channel Attention (ECA) is a lightweight mechanism that improves upon SE (Search Engine) by replacing fully connected layers with one-dimensional convolutions to model the interaction between adjacent channels, thereby enhancing feature selection capabilities with minimal parameters. The approach involves first performing global average pooling on the input features in either the spatial or temporal dimensions to obtain the channel descriptor z, and then adaptively determining the convolution kernel length based on the number of channels C. Then, a one-dimensional convolution of length k is applied to z along the channel dimension to obtain the response s, which is then processed by a sigmoid function to obtain the weight w. Finally, w is multiplied back by the channel to form the original feature to form a weighted output. ECA avoids the loss caused by information compression by eliminating the need for dimensionality reduction and restoration, making it suitable for TCN residual blocks to enhance the response to key channels (such as specific error components or meteorological factors).

[0096] (4) Regression Regulation Head Module: Reference Figure 8 The diagram shows an architecture of a regression control head module, which connects "exogenous features → regression coefficients → correction values". The input is exogenous features such as meteorological data. After input, standardization and dimensionality alignment are performed, and then the data is fed into the "multivariate regression weight mapping" module to generate two sets of coefficients: one column is a channel-wise weighted coefficient vector w. t The other column is the intercept vector b. t This is used to perform regression correction on the implicit representation or output of the error side, thereby instantly mapping the current meteorological effects to the direction and magnitude of the error increase or decrease. This head is the output-level correction quantity. Where φ t These are regressible features initially selected from error characteristics.

[0097] (5) Prediction Post-processing Module: The post-processing module is executed in two layers. The first layer is time series shaping and hierarchical determination. The prediction error sequence is subjected to median filtering based on components and time to weaken the impact of extreme outliers; then, interval statistics are calculated, including the current amplitude, slope, rolling mean, and confidence interval. Thresholds are set based on engineering tolerance and safety margin to form multi-level risk intervals (normal, attention, early warning, alarm). To avoid frequent level jumps, dual threshold determination and minimum duration constraints are introduced. Each time step outputs the level, whether the component is limited, and uncertainty score, and retains the peak information of the most recent window for subsequent linkage.

[0098] The second layer involves BIM mapping and visualization linkage, linking predicted component errors and their levels to BIM instances through attributes, writing them into attribute fields and timestamps, including error value, level code, exceedance limit, and credibility. Rendering rules are generated according to component category and key direction, using hierarchical color schemes and transparency coding, while numerical annotations and small bar charts are overlaid on the components. This is aggregated to subsystem and regional levels, outputting heatmaps and risk ranking tables. If the level reaches a warning or alarm threshold, a review work order is automatically generated with a recent window's evidence snapshot, and pushed to the monitoring panel and scheduling interface.

[0099] III. Model Training:

[0100] (1) Based on the multimodal time-series dataset constructed in the preceding steps, the training set and validation set are divided according to time sequence to ensure that there is no time leakage. Let the error sequence X... err With meteorological sequence X met To input the training model, a sliding window is used to construct samples and forward and backward propagation is performed to update the parameters; after each iteration, the loss and accuracy are monitored using the validation set, and the weights with the best validation metrics are saved.

[0101] (2) The total loss of this invention consists of a prediction error term and a regularization term. The prediction error term calculates the loss based on the prediction output. The regularization term includes regression regulation consistency loss, which is used to constrain the consistency between the final prediction and "initial prediction plus each round of regulation"; physical range and monotonic prior loss, which is used to limit the predicted value from going out of bounds and maintain a reasonable response to key meteorological factors; and time smoothing loss, which is used to suppress non-physical oscillations.

[0102] (3) During training, set hyperparameters such as the number of iterations, learning rate, batch size, and optimizer, and use gradient descent-based optimization methods to minimize the total loss. Introduce learning rate decay, gradient clipping, and early stopping strategies to stabilize the training process, and record the index curves of the training set and validation set to ensure that the model does not overfit; decrease the intensity coefficient of regression regulation by round to improve stability.

[0103] (4) After training, evaluate the optimal model on an independent test set. Verify the model's generalization ability and robustness; classify the prediction results through the post-processing module and map them to BIM components, outputting a classification list and visualization results for engineering application verification.

[0104] IV. Model Usage:

[0105] (1) In online or batch processing scenarios, according to the standardization rules, time alignment and missing measurement mask requirements of the training stage, the input error sequence and meteorological sequence are preprocessed, and the data is organized using a sliding window consistent with the training. The data is then input into the trained model to obtain the error prediction results and uncertainty estimates of each predicted horizon.

[0106] (2) Perform post-processing on the model prediction results, including median or second-order difference smoothing, outlier suppression, threshold determination and double-threshold bounce control, and apply physical range and monotonic prior constraints to generate the risk level, error over-limit and lead time for each time step, while outputting key features and cumulative exposure indicators for explanation.

[0107] (3) Write the predicted values ​​and grading results into the attribute fields of the BIM model instance with the component attributes as the unique primary key. The content includes each horizon error, risk level, over-limit margin, lead time, confidence level, timestamp and evidence reference. Complete the visualization mapping according to the preset rendering rules and generate a list of serious component errors and a summary report of the partitions.

[0108] (4) When the classification result reaches the warning or alarm threshold, a review and handling work order is automatically generated, recording the component identification, spatial location, alarm time, evidence snapshot and suggested measures, and pushing it to the monitoring and scheduling interface; the entire process generates audit logs, which can be retrieved, replayed and exported by component and time range for operation and maintenance closed loop and effect evaluation.

[0109] Compared to traditional methods, which often rely on single-phase thresholds and least-squares registration, resulting in delayed and uncontrollable responses, this invention uses causal multi-scale TCN for long-term series prediction, outputting future peak values ​​and risk levels, along with lead times and uncertainty intervals.

[0110] Furthermore, to emphasize interpretability and controllability, this invention first introduces a "regression control head" on the output side. This module uses exogenous variables such as sea state and operations as conditions to map channel-by-channel weights and intercepts, applying linear correction to the regressible time-series representations, thereby establishing an explicit, loosely coupled functional relationship between predicted values ​​and exogenous drivers. Compared to traditional pure black-box end-to-end approaches, the regression control head has three advantages: First, it provides causal explanations for the coefficients, clarifying the direction and magnitude of risk increase or decrease for each meteorological feature, facilitating the engineering side to formulate intervention strategies accordingly; second, it has a small parameter size and stable training, improving the robustness and transferability of the model under small samples and extreme sea states; third, it forms an audit link of "variable-coefficient-correction amount," meeting the traceability and verification requirements of engineering compliance scenarios.

[0111] Furthermore, traditional outputs are mostly limited to two-dimensional numerical values, making it difficult to locate and trace in three-dimensional space, and lacking actionable suggestions. This invention uses BIM component IDs as anchor points to write back the component error risk heatmap and uncertainty zone in a three-dimensional scene, simultaneously providing the contribution of the main factors (sea state / geometry / process) and standard-oriented I-IV levels; and directly drives the adjustment of work windows, retesting instructions, limit issuance, and temporary reinforcement through standardized interfaces, making "prediction-strategy-execution-feedback" a closed loop. The entire process is recorded for easy auditing and reproduction, and the model is lightweight, easy to deploy, and can run stably at the edge or in the cloud.

[0112] In summary, compared with existing technologies, the embodiments of this application have the following beneficial effects: They collect time-series laser point cloud data of current offshore wind power construction components, as well as meteorological and environmental monitoring data including wind speed, wind direction, wave height, tide level, current velocity, and rainfall, and construct time-aligned component error time series and meteorological and environmental time series. This achieves precise synchronization of high-dimensional geometric state observation and multi-dimensional environmental disturbance variables on a unified time axis, providing a physically complete and temporally consistent input foundation for risk prediction, avoiding modeling bias caused by asynchronous data or missing environmental factors. The component error time series and meteorological and environmental time series are input into a preset risk prediction model, enabling the model to simultaneously utilize the component's own historical error evolution information and external meteorological disturbance information, breaking the limitations of traditional reliance on a single data source. Through a multi-level temporal convolutional network, the component error time series is extracted layer by layer, and the dynamic patterns at different time scales are gradually abstracted using a deep convolutional structure. This approach effectively captures the long-term dependence and nonlinear evolution characteristics of errors over time. After feature extraction at each layer, the meteorological environment sequence fragments corresponding to the time horizon of the current layer's feature extraction are used to weight and correct the output features of the current layer. This allows meteorological environment information to dynamically modulate feature representation within a time window that matches the error features, achieving hierarchical and temporally aligned fusion intervention of exogenous variables on internal states, avoiding semantic inaccuracies caused by global or fixed window fusion. Based on the error prediction data of the feature regression output component after the last layer's correction within a preset future time period, the trend of construction deviation is forward extrapolated. The synergistic effect of the above technical features constructs an end-to-end prediction mechanism with multi-level temporal convolution as the core, meteorological environment sequences as the dynamic correction basis, and error time series as the main input. Overall, it realizes the transformation of construction risk from "static threshold judgment" to "dynamic trend prediction," effectively solving the problems of lag and inaccuracy in risk prediction in existing technologies.

[0113] Example 2:

[0114] Please refer to Figure 2 Based on the same inventive concept, the present invention discloses a dynamic risk prediction system for offshore construction, comprising: a data acquisition module M1, a model prediction module M2, and an output module M3;

[0115] The data acquisition module M1 is used to collect the laser point cloud time series data of the current offshore wind power construction components and meteorological environment monitoring data, and to construct a time-aligned component error time series and meteorological environment time series; wherein the meteorological environment monitoring data includes: wind speed, wind direction, wave height, tide level, current velocity and rainfall data.

[0116] Furthermore, the data acquisition module M1 includes: a deviation calculation unit and an alignment unit;

[0117] The deviation calculation unit is used to spatially align the laser point cloud time series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculate the distance deviation between the aligned point cloud and the theoretical geometric model to obtain the original error sequence.

[0118] The alignment unit is used to time-align the original error sequence with the meteorological environment monitoring data and resample it with a uniform time step to obtain a time-synchronized component error time sequence and meteorological environment time sequence.

[0119] In this preferred embodiment, the laser point cloud time series data is spatially aligned with the theoretical geometric model of the corresponding component in the preset BIM model, and the distance deviation is calculated to obtain the original error sequence. This enables the error measurement to have clear BIM semantic correlation and component-level spatial positioning capability, avoiding the ambiguity of error attribution caused by the lack of structural semantic support in traditional point cloud processing. The original error sequence is time-aligned with meteorological and environmental monitoring data and resampled with a unified time step, ensuring strict synchronization of multimodal data in the time dimension. This eliminates the time series distortion caused by sampling frequency differences or asynchronous acquisition, laying a data consistency foundation for the stable training and prediction of subsequent deep learning models.

[0120] The model prediction module M2 is used to input the component error time series and meteorological environment time series into a preset risk prediction model, so that the risk prediction model extracts features from the component error time series layer by layer through a multi-level temporal convolutional network. After each layer of feature extraction, the meteorological environment sequence segment in the meteorological environment time series that corresponds to the time horizon of the current layer feature extraction is used to weight and correct the features output by the current layer, and the corrected features are used as the input of the next layer. Based on the corrected features of the last layer, the error prediction data of the component in the future preset time period is regressed and output.

[0121] Furthermore, the model prediction module M2 includes: a convolutional unit, an attention weighting unit, and a residual connection unit;

[0122] The convolutional unit is used to perform causal dilation convolution on the input data of the current layer to obtain a convolutional feature sequence.

[0123] The attention weighting unit is used to perform channel attention enhancement processing on the convolutional feature sequence to generate a channel-weighted feature sequence;

[0124] The residual connection unit is used to perform a residual connection between the channel-weighted feature sequence and the input data of the current layer to obtain the output features of the current layer;

[0125] The input data for the first layer is the component error time series, and the input data for the remaining layers is the output features of the previous layer.

[0126] In this preferred embodiment, causal dilated convolution processing is performed on the current layer input data. Causality constraints ensure that the prediction does not depend on future information, while the dilation mechanism exponentially expands the receptive field, effectively capturing long-cycle construction error evolution patterns without significantly increasing the number of parameters. Channel attention enhancement processing is performed on the convolutional feature sequence to dynamically highlight feature channels that contribute highly to the current risk state (such as displacement in a specific direction or key meteorological coupling terms), suppress redundant or noisy channels, and improve feature discriminability. The channel-weighted feature sequence is residually connected with the current layer input data to alleviate the gradient vanishing problem in deep networks, ensure the training stability and representation fidelity of the model under multi-level stacking, and retain the local details of the original error data to avoid the loss of key transient information during deep feature extraction.

[0127] Furthermore, the model prediction module M2 also includes: a time window determination unit, a truncation unit, a dimension adjustment unit, a mapping unit, and a correction unit;

[0128] The time window determination unit is used to determine the time window covered by the feature extraction of the current layer based on the structural parameters of the current layer in the multi-level temporal convolutional network.

[0129] The extraction unit is used to extract a meteorological environment sequence segment corresponding to the time window from the meteorological environment time series.

[0130] The dimension adjustment unit is used to standardize the meteorological environment sequence fragment and adjust the dimension to the number of channels of the output feature of the current layer.

[0131] The mapping unit is used to generate a channel-wise weighted coefficient vector and an intercept vector based on the adjusted meteorological environment sequence fragments through linear mapping.

[0132] The correction unit is used to perform linear correction on the features output by the current layer using the weighted coefficient vector and the intercept vector to obtain the corrected features.

[0133] In this preferred embodiment, the time window covered by feature extraction is determined based on the structural parameters of the current layer, ensuring that the truncation of meteorological environmental segments is strictly aligned with the internal temporal perspective of the network, thus guaranteeing the spatiotemporal semantic consistency of exogenous variable correction. Segments corresponding to the time window are extracted from the meteorological environmental time series to achieve layer-adaptive dynamic environmental context extraction, avoiding information mismatch caused by global averaging or fixed windows. Meteorological segments are standardized and their dimensions are adjusted to match the number of feature channels in the current layer, ensuring their computability with error features in the spatial dimension. Channel-wise weighted coefficient vectors and intercept vectors are generated through linear mapping, transforming meteorological variables into interpretable linear modulation parameters for error features. This allows the model to apply differentiated influences to different meteorological factors (such as wind speed and wave height) on different feature channels. Linear correction of the current layer's output features is performed using these coefficients and intercepts, achieving low coupling and explicit control of internal representation by exogenous drivers. This improves prediction accuracy while preserving physical interpretability, avoiding overfitting of the model under extreme sea conditions.

[0134] The output module M3 is used to output risk prediction results based on the error prediction data and in combination with preset risk determination rules.

[0135] Furthermore, the output module M3 includes: a filtering unit, a statistical unit, and a comparison output unit;

[0136] The filtering unit is used to filter the error prediction data to obtain error filtered data.

[0137] The statistical unit is used to calculate rolling statistics within a preset time window based on error-filtered data; wherein the rolling statistics include amplitude, slope of change, rolling mean, and confidence interval.

[0138] The comparison output unit is used to compare the rolling statistics with a preset multi-level risk threshold range, and combine the minimum duration constraint to determine the final risk level as the risk prediction result.

[0139] In this preferred embodiment, the error prediction data is filtered to effectively suppress the impact of extreme outliers and improve the physical rationality and temporal smoothness of the prediction sequence. Based on the filtered data, rolling statistics within a preset time window are calculated, including amplitude, slope of change, rolling mean, and confidence interval, to characterize the dynamic behavior of the error from multiple dimensions and avoid the limitations of single threshold judgment. The rolling statistics are compared with preset multi-level risk threshold intervals and combined with minimum duration constraints to prevent false jumps in risk level due to instantaneous fluctuations, thereby enhancing the stability and engineering applicability of the judgment results. The final output risk level not only reflects the current deviation magnitude but also integrates trend, persistence, and uncertainty information, significantly improving the accuracy and operability of risk warning.

[0140] In summary, compared with existing technologies, the embodiments of this application have the following beneficial effects: They collect time-series laser point cloud data of current offshore wind power construction components, as well as meteorological and environmental monitoring data including wind speed, wind direction, wave height, tide level, current velocity, and rainfall, and construct time-aligned component error time series and meteorological and environmental time series. This achieves precise synchronization of high-dimensional geometric state observation and multi-dimensional environmental disturbance variables on a unified time axis, providing a physically complete and temporally consistent input foundation for risk prediction, avoiding modeling bias caused by asynchronous data or missing environmental factors. Inputting the component error time series and meteorological and environmental time series into a preset risk prediction model allows the model to simultaneously utilize the component's own historical error evolution information and external meteorological disturbance information, breaking the limitations of traditional reliance on a single data source. Through multi-level temporal convolutional networks, layer-by-layer feature extraction of the component error time series is performed, and the dynamic patterns at different time scales are gradually abstracted using deep convolutional structures. This method effectively captures the long-term dependence and nonlinear evolution characteristics of errors over time. After feature extraction at each layer, the meteorological environment sequence fragments corresponding to the time horizon of the current layer's feature extraction are used to weight and correct the output features of the current layer. This allows meteorological environment information to dynamically modulate feature representation within a time window that matches the error features, achieving hierarchical and temporally aligned fusion intervention of exogenous variables on internal states, avoiding semantic inaccuracies caused by global or fixed window fusion. Based on the error prediction data of the feature regression output component after the last layer's correction within a preset time period, the trend of construction deviation is forward extrapolated. The synergistic effect of the above technical features constructs an end-to-end prediction mechanism with multi-level temporal convolution as the core, meteorological environment sequence as the dynamic correction basis, and error time series as the main input. Overall, it realizes the transformation of construction risk from "static threshold judgment" to "dynamic trend prediction," effectively solving the problems of lag and inaccuracy in risk prediction in existing technologies.

[0141] Example 3:

[0142] Figure 3 A structural diagram of a risk dynamic prediction device for offshore construction, as described in this application, is presented. Figure 3 As shown, the risk dynamic prediction device for offshore construction may include: processor N1, memory N2, data interface N3, and communication bus N4.

[0143] Wherein: processor N1, memory N2, and data interface N3 communicate with each other through communication bus N4; data interface N3 is used for data communication with other devices such as input devices or output devices; processor N1 is used to execute program N5, which can specifically execute the relevant steps in any of the above embodiments of the dynamic risk prediction method for offshore construction.

[0144] Specifically, program N5 may include program code, which includes computer-executable instructions.

[0145] The processor N1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The risk dynamic prediction device for offshore construction includes one or more processors, which may be processors of the same type, such as one or more CPUs, or processors of different types, such as one or more CPUs and one or more ASICs.

[0146] Memory N2 is used to store program N5. Memory N2 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0147] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Furthermore, the embodiments in this application are not directed to any particular programming language.

[0148] Example 4:

[0149] This invention also provides a computer-readable storage medium storing at least one executable instruction. When the executable instruction is run on a risk dynamic prediction device / system for offshore construction, it causes the offshore construction risk dynamic prediction device / system to execute one of the offshore construction risk dynamic prediction methods described in any of the above method embodiments.

[0150] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. Similarly, for the purpose of simplification and aiding understanding of one or more aspects of the invention, in the above description of exemplary embodiments of this application, various features of the embodiments are sometimes grouped together in a single embodiment, figure, or description thereof. The claims, which follow the detailed description, are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0151] Those skilled in the art will understand that the modules in the device of the embodiment can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiment can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components, except that at least some of such features and / or processes or units are mutually exclusive.

Claims

1. A method of dynamic risk prediction for offshore construction, characterized in that, The method comprises the following steps: Collecting current laser point cloud time series data of offshore wind power construction components and meteorological environment monitoring data, and constructing time-aligned component error time series and meteorological environment time series; wherein the meteorological environment monitoring data comprises wind speed, wind direction, wave height, tide level, flow rate and rainfall data; Inputting the component error time series and the meteorological environment time series into a preset risk prediction model, so that the risk prediction model extracts features of the component error time series layer by layer through a multi-level time series convolution network, and after each layer of feature extraction, the meteorological environment sequence segment corresponding to the time field of view of the current layer of feature extraction in the meteorological environment time series is used to weight and correct the features output by the current layer, and the corrected features are used as the input of the next layer; based on the last layer of corrected features, the error prediction data of the component in a future preset period is output by regression. Based on the error prediction data, the risk prediction result is output combined with a preset risk determination rule.

2. A method of dynamic risk prediction for offshore operations as claimed in claim 1, wherein, The construction of the time-aligned component error time series and the meteorological environment time series comprises: Spatially aligning the laser point cloud time series data with the theoretical geometric model of the corresponding component in the preset BIM model, calculating the distance deviation between the aligned point cloud and the theoretical geometric model, and obtaining the original error sequence; Time-aligning the original error sequence with the meteorological environment monitoring data, and resampling at a uniform time step to obtain time-synchronized component error time series and meteorological environment time series.

3. A method of dynamic risk prediction for offshore operations as claimed in claim 1, wherein, The layer-by-layer feature extraction of the component error time series comprises: Performing causal dilated convolution processing on the input data of the current layer to obtain a convolution feature sequence; Performing channel attention enhancement processing on the convolution feature sequence to generate a channel-weighted feature sequence; Residual connecting the channel-weighted feature sequence with the input data of the current layer to obtain the output feature of the current layer; Wherein, the input data of the first layer is the component error time series, and the input data of the remaining layers is the output feature of the previous layer.

4. A method of dynamic risk prediction for offshore operations as claimed in claim 1, wherein, The weighting correction of the features output by the current layer using the meteorological environment sequence segment corresponding to the time field of view of the current layer of feature extraction in the meteorological environment time series to obtain the corrected features as the input of the next layer comprises: Determine the time window covered by the feature extraction of the current layer according to the structure parameters of the current layer in the multi-level time series convolution network; From the meteorological environment time series, the meteorological environment sequence segment corresponding to the time window is intercepted; Standardizing the meteorological environment sequence segment and adjusting the dimension to the number of channels of the output feature of the current layer; Based on the meteorological environment sequence segment after adjusting the dimension, a weight coefficient vector and an intercept vector are generated through linear mapping; Using the weight coefficient vector and the intercept vector, linear correction is performed on the features output by the current layer to obtain the corrected features.

5. A method of dynamic risk prediction for offshore operations as claimed in claim 1, wherein, The error prediction data is filtered to obtain error filtered data. ​ Based on the error filtered data, calculate the rolling statistics in a preset time window; wherein, the rolling statistics include amplitude, change slope, rolling mean and confidence interval; Compare the rolling statistics with the preset multi-level risk threshold interval, and determine the final risk level as the risk prediction result in combination with the minimum duration constraint.

6. A dynamic risk prediction system for offshore construction, characterized in that, It comprises: Data acquisition module, model prediction module and output module; Wherein, the data acquisition module is used for collecting the laser point cloud time series data of the current offshore wind power construction component and the meteorological environment monitoring data, and constructing the time-aligned component error time series and meteorological environment time series; wherein, the meteorological environment monitoring data includes wind speed, wind direction, wave height, tide level, flow rate and rainfall data; The model prediction module is used for inputting the component error time series and meteorological environment time series into the preset risk prediction model, so that the risk prediction model extracts features from the component error time series layer by layer through multi-level time series convolution network, and after each layer of feature extraction, the meteorological environment sequence segment corresponding to the time field of view of the current layer of feature extraction in the meteorological environment time series is used to weight and correct the features output by the current layer, and the corrected features are taken as the input of the next layer; based on the last layer of corrected features, the error prediction data of the component in the future preset period is regressed and output; The output module is used for outputting the risk prediction result based on the error prediction data and in combination with the preset risk determination rule.

7. A dynamic risk prediction system for offshore operations as claimed in claim 6, characterised in that, The data acquisition module comprises a deviation calculation unit and an alignment unit; Wherein, the deviation calculation unit is used for spatially aligning the laser point cloud time series data with the theoretical geometric model of the corresponding component in the preset BIM model, and calculating the distance deviation between the aligned point cloud and the theoretical geometric model to obtain the original error sequence; The alignment unit is used for time aligning the original error sequence with the meteorological environment monitoring data, and resampling with a uniform time step to obtain the time-synchronized component error time series and meteorological environment time series.

8. A dynamic risk prediction system for offshore operations as claimed in claim 6, wherein, The model prediction module comprises a convolution unit, an attention weighting unit and a residual connection unit; Wherein, the convolution unit is used for performing causal dilated convolution processing on the input data of the current layer to obtain a convolution feature sequence; The attention weighting unit is used for performing channel attention enhancement processing on the convolution feature sequence to generate a channel weighted feature sequence; The residual connection unit is used for connecting the channel weighted feature sequence with the input data of the current layer in residual to obtain the output feature of the current layer; Wherein, the input data of the first layer is the component error time series, and the input data of the remaining layers is the output feature of the previous layer.

9. A risk dynamic prediction device for offshore construction, comprising a memory, a processor and a computer program stored on the memory and running on the processor, characterized in that, The computer program is loaded into the processor to realize the steps of the offshore construction risk dynamic prediction method according to any one of claims 1-5.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the steps of the offshore construction risk dynamic prediction method according to any one of claims 1-5.

Images