CN-122020093-A - Offshore environment monitoring and early warning method based on adaptive BP neural network

CN122020093ACN 122020093 ACN122020093 ACN 122020093ACN-122020093-A

Abstract

The invention relates to the technical field of environment monitoring, in particular to an offshore environment monitoring and early warning method based on a self-adaptive BP neural network, which comprises the following steps of S1, dividing a target offshore area into N differential subareas, constructing an area-space-time-sensitivity three-dimensional initial weight operator to generate an initial weight matrix of the BP neural network, S2, introducing a two-channel error curved surface structure in the training process of the BP neural network, self-adaptively adjusting the learning rate and gradient gain of the network, introducing a movable threshold window early stopping mechanism in the verification stage, S3, inputting real-time buoy observation data and remote sensing data streams into the trained BP neural network to output an environment risk probability sequence, sending early warning information when the risk probability of any subarea continuously meets a dynamic threshold condition, and feeding back an early warning result to a cloud model updating module. The early warning accuracy and robustness of the network in the early stage of the environmental mutation are improved, and the method is suitable for application scenes of offshore ecological monitoring.

Inventors

TAN XIAONAN
WANG DAZHI
LIU NA
QIN GANG

Assignees

青岛黄海学院

Dates

Publication Date: 20260512
Application Date: 20260211

Claims (10)

1. An offshore environment monitoring and early warning method based on an adaptive BP neural network is characterized by being cooperatively executed by a buoy array, a shore-based edge computing node and a satellite remote sensing ground receiving station of N regional sea areas distributed on the coast, and comprising the following steps: the method comprises the steps of S1, dividing a target offshore area into N differential subareas according to priori marine knowledge, for each subarea, fusing remote sensing inversion long-time sequence statistical characteristics of the subareas and buoy short-period high-resolution fluctuation characteristics, constructing a region-space-time-sensitivity three-dimensional initial weight operator, and generating a BP neural network initial weight matrix of the corresponding subarea according to the remote sensing inversion long-time sequence statistical characteristics of the subareas, so that the neural network carries regional differential environment priori response capability at the beginning stage of training; s2, introducing a double-channel error curved surface structure in the BP neural network training process: synchronously establishing a first error channel aiming at a normal environment state; And a second error path for low amplitude anomalies and disaster precursor states; Meanwhile, a movable threshold window early stopping mechanism is introduced in a verification stage, time window statistics is carried out on the node activation distribution of an output layer, and when abnormal silencing characteristics and an error rebound trend are detected to occur simultaneously, the current training is automatically stopped and rolled back to an optimal weight state, so that the false convergence problem in a disaster precursor stage is avoided; S3, inputting real-time buoy observation data and remote sensing data streams into a BP neural network after training according to the subareas, outputting an environment risk probability sequence corresponding to each subarea, when the risk probability of any subarea continuously meets a dynamic threshold condition, sending early warning information comprising subarea numbers, risk grades and confidence degrees by corresponding shore-based edge computing nodes, and simultaneously feeding false alarm or missing alarm marks in the early warning result back to a cloud model updating module for online correction of subarea weight vectors and dynamic threshold parameters.
2. The offshore environment monitoring and early warning method based on the adaptive BP neural network according to claim 1, wherein the S1 specifically comprises: S11, dividing the target offshore area into a plurality of subareas with differentiation in environmental dynamics characteristics and risk sensitivity according to priori marine knowledge including water depth topography, main ocean current paths, tidal characteristics, historical pollution event distribution and ecological sensitivity subareas of the target offshore area; And S12, generating a corresponding BP neural network initial weight matrix for each partition.
3. The method for monitoring and early warning an offshore environment based on an adaptive BP neural network according to claim 2, wherein S12 specifically comprises: s121, data fusion is carried out, namely, the long-time-sequence remote sensing inversion parameter statistical characteristics of the subareas and the buoy array observation fluctuation characteristics with short period and high resolution are integrated; s122, constructing a three-dimensional initial weight operator integrating the region, the space-time and the risk sensitivity information; s123, converting the three-dimensional initial weight operator into an initial connection weight matrix from the neural network input layer to the first hidden layer through a preset micro-mappable function.
4. The offshore environment monitoring and early warning method based on the adaptive BP neural network according to claim 3, wherein the three-dimensional initial weight operator is encoded from three dimensions, and specifically comprises: the regional dimension reflects the inherent geographic and environmental attributes of the partition; the space-time dimension is responsible for aligning and fusing observation features of different time scales to form a unified expression; the sensitivity dimension gives different risk sensitivity weights to different environment parameters according to the historical disaster records.
5. The offshore environment monitoring and early warning method based on the adaptive BP neural network of claim 3, wherein the step S123 specifically comprises: flattening the three-dimensional initial weight operator into a one-dimensional feature vector, and carrying out element-by-element weighting on the feature vector and the risk sensitivity weight vector to obtain a weighted feature vector; the weighted feature vector is input into a micro full connection mapping function, and a vector with the length equal to the connection number of an input layer and a first hidden layer is output after the transformation of an activation function; and reconstructing the vector to generate an initial connection weight matrix.
6. The offshore environment monitoring and early warning method based on the adaptive BP neural network according to claim 1, wherein the S2 specifically comprises: s21, constructing and synchronously calculating a double-channel error, wherein the double-channel error comprises a first error channel and a second error channel, the first error channel forms a main loss function by means of the mean square error between forward propagation output and a normal state label, and the second error channel forms a sensitive loss function by means of the error between forward propagation output and a smoothed enhanced label comprising historical low-amplitude abnormality and disaster precursor characteristics; s22, in each training iteration, calculating the difference between the main loss function and the sensitive loss function, dynamically adjusting the learning rate and the gradient gain coefficient of the current iteration according to the difference, wherein the adjustment rule meets the requirement that when the difference is increased, the sensitivity of the network to the potential abnormal mode in the input and the update step length are improved.
7. The method for monitoring and pre-warning the offshore environment based on the adaptive BP neural network according to claim 6, wherein the step S2 further comprises: In the verification stage, maintaining a detection time window which can move along the verification time sequence; in the detection time window, counting the node proportion of each node activation value of the output layer lower than a preset silencing threshold, and judging that abnormal silencing characteristics occur when the node proportion continuously exceeds a set upper limit proportion threshold; synchronously monitoring a loss curve on the verification set, and judging that an error rebound trend occurs when the loss value continuous rebound exceeds a preset continuous rebound trigger frequency threshold value in the detection time window; When the abnormal silencing feature and the error rebound trend are detected simultaneously in the same detection time window, the network is judged to enter a false convergence state of a disaster precursor stage, training is automatically stopped, and the network weight is rolled back to a historical archiving state with the lowest verification loss.
8. The offshore environment monitoring and early warning method based on the adaptive BP neural network according to claim 1, wherein the step S3 specifically comprises: S31, for each partition, synchronously receiving a short-period high-resolution observation data stream uploaded by a partition buoy array and a satellite remote sensing inversion parameter stream corresponding to a space in each synchronous period, carrying out time alignment and space matching calibration on two types of data, carrying out fusion and standardization processing according to preset characteristic dimensions of the partition to form a real-time input vector, and inputting the real-time input vector into a trained BP neural network model corresponding to the partition; S32, outputting the environment risk probability of the subarea in real time by the BP neural network model, continuously tracking the formed environment risk probability time sequence, comparing the formed environment risk probability time sequence with a dynamic threshold defined based on historical data, and judging to trigger early warning once the environment risk probability time sequence meets early warning conditions in a continuous period; And S33, after triggering the early warning, automatically generating structural early warning information comprising risk level and confidence level, taking the structural early warning information as a feedback sample, and performing on-line fine adjustment on the neural network model parameters of the subareas and optimizing a dynamic threshold value so as to continuously improve the monitoring accuracy.
9. The method for monitoring and early warning an offshore environment based on an adaptive BP neural network according to claim 8, wherein S32 specifically comprises: s321, performing forward propagation processing on real-time input vectors of the partitions through a trained partition BP neural network model, and outputting the environment risk probability at the current moment; S322, continuously recording risk probability values at a plurality of moments, and constructing an environment risk probability time sequence of the subarea; s323, setting a dynamic threshold function based on the historical risk probability distribution, and defining a corresponding early warning judgment rule; S324, when the risk probability time sequence continuously meets the early warning judgment rule in a specified time window, determining that the partition enters an early warning state.
10. The method for monitoring and early warning an offshore environment based on an adaptive BP neural network according to claim 8, wherein S33 specifically comprises: s331, when the partition triggers early warning, a shore-based edge computing node responsible for the partition generates structured early warning information, wherein the structured early warning information comprises partition numbers, risk grades and confidence degrees calculated based on the BP neural network output layer activation entropy; s332, marking the early warning event and the environment verification result actually observed subsequently as feedback samples, and uploading the feedback samples to a cloud model updating module; s333, the cloud model updating module performs fine adjustment and correction on the weight vector of the last layer or layers of the BP neural network model of the partition by using a feedback sample through an online learning algorithm, and synchronously updates parameters of a dynamic threshold function of the partition so as to optimize the accuracy and adaptability of subsequent monitoring.

Description

Offshore environment monitoring and early warning method based on adaptive BP neural network Technical Field The invention relates to the technical field of environmental monitoring, in particular to an offshore environmental monitoring and early warning method based on a self-adaptive BP neural network. Background The offshore area is used as a composite environment with intensive ecological sensitivity and human activity, the monitoring and early warning requirements are increasingly enhanced, the traditional marine environment monitoring means depend on single type observation data sources, such as remote sensing images or buoy monitoring data, spatial coverage and time sequence continuity are difficult to achieve, meanwhile, the rule triggering type early warning method based on a fixed threshold value is limited in identification capability of disaster precursors, particularly when facing low-amplitude and progressive change signals, the problem of response lag or high false alarm rate is easy to occur, and in addition, the existing neural network model has certain capability in terms of pattern identification, but has limitations in terms of regional difference, unreasonable initial weight, local optimum training trapping and the like, and lacks a sensitive sensing and dynamic training regulation mechanism for weak abnormal patterns. In recent years, along with the improvement of remote sensing inversion precision and the enhancement of buoy array sampling capability, the construction of a marine environment monitoring model fused with multi-time-space scale information becomes a research hotspot, and particularly in the disaster early warning field, a modeling method capable of combining long time sequence statistical characteristics and high resolution fluctuation response is needed to be matched with an interpretable dynamic early warning strategy so as to adapt to the complexity and the burstiness of the marine environment. Disclosure of Invention The invention provides an offshore environment monitoring and early warning method based on a self-adaptive BP neural network, which is characterized in that a zonal neural network model is constructed, remote sensing and buoy observation data are fused, an early stopping mechanism of a movable time window and a feedback optimization loop are introduced, so that high-precision prediction and dynamic early warning of offshore environment risk are realized, and the response speed, model robustness and environment adaptability of the system are improved. An offshore environment monitoring and early warning method based on an adaptive BP neural network is characterized by being cooperatively executed by a buoy array, a shore-based edge computing node and a satellite remote sensing ground receiving station of N regional sea areas distributed on the coast, and comprising the following steps: the method comprises the steps of S1, dividing a target offshore area into N differential subareas according to priori marine knowledge, for each subarea, fusing remote sensing inversion long-time sequence statistical characteristics of the subareas and buoy short-period high-resolution fluctuation characteristics, constructing a region-space-time-sensitivity three-dimensional initial weight operator, and generating a BP neural network initial weight matrix of the corresponding subarea according to the remote sensing inversion long-time sequence statistical characteristics of the subareas, so that the neural network carries regional differential environment priori response capability at the beginning stage of training; s2, introducing a double-channel error curved surface structure in the BP neural network training process: synchronously establishing a first error channel aiming at a normal environment state; And a second error path for low amplitude anomalies and disaster precursor states; Meanwhile, a movable threshold window early stopping mechanism is introduced in a verification stage, time window statistics is carried out on the node activation distribution of an output layer, and when abnormal silencing characteristics and an error rebound trend are detected to occur simultaneously, the current training is automatically stopped and rolled back to an optimal weight state, so that the false convergence problem in a disaster precursor stage is avoided; s3, inputting real-time buoy observation data and remote sensing data streams into the BP neural network after training according to the subareas, and outputting an environment risk probability sequence corresponding to each subarea; And simultaneously, feeding back false report or missing report marks in the early warning result to a cloud model updating module for correcting partition weight vectors and dynamic threshold parameters on line. Optionally, the S1 specifically includes: S11, dividing the target offshore area into a plurality of subareas with differentiation in environmental dynamics characteristics and risk