Intelligent channel depth prediction method and system based on multi-source information collaborative fusion

Abstract

The invention discloses an intelligent channel water depth prediction method and system based on multi-source information collaborative fusion, and the method comprises the steps: carrying out the processing of collected original multi-source observation data, and obtaining the input data of a unified space-time structure; and inputting the fused data into a trained deep learning water depth prediction model to obtain a minute-level dynamic prediction result of the water depth of the channel, the water depth prediction model being a time sequence feature extraction network comprising a self-attention mechanism and a long and short term memory unit, the fusion Transform structure is used for extracting a long-distance dependency relationship, and the LSTM unit is combined to capture a short-term dynamic change trend; and a monotonic water head operator layer is introduced into the model structure so as to ensure that prediction output meets a predetermined physical monotonic constraint relationship. Through collaborative fusion of multi-source observation data, minute-level dynamic water depth prediction can be realized based on the constructed deep learning water depth prediction model, and the method has the advantages of high prediction precision, stability, interpretability and the like.

Description

Technical Field

[0001] This invention relates to waterway engineering and multi-source data processing, and in particular to a method for real-time prediction of inland waterway depth based on a deep learning model. Background Technology

[0002] With the continuous improvement of the modernization level of inland waterway transportation, the demand for refined management and control of waterways is increasing. As a key parameter that determines the safety of ship navigation, the accuracy of waterway depth prediction is directly related to navigation efficiency and operational safety, and also puts forward higher requirements for waterway operation and maintenance.

[0003] Traditional methods of obtaining water depth in waterways mostly rely on manual surveys and underwater sounding instruments. These methods suffer from poor real-time performance, limited coverage, and significant susceptibility to weather and topography, making it difficult to meet the dynamic water depth response requirements of complex river sections, sudden weather changes, or high-frequency navigation scenarios. Especially in tidal estuaries and inland waterways with frequent siltation, water depth changes drastically, and relying on static measurement data can easily lead to delayed water depth information, affecting vessel scheduling and navigation safety. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing methods for obtaining water depth in waterways, the purpose of this invention is to provide an intelligent waterway water depth prediction method and system based on the collaborative fusion of multi-source information. This method and system can achieve dynamic water depth prediction at the minute level and has advantages such as high prediction accuracy, stability, and interpretability.

[0005] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] A smart channel depth prediction method based on multi-source information collaborative fusion includes:

[0007] The collected raw multi-source observation data are processed to obtain input data with a unified spatiotemporal structure;

[0008] The fused data is input into a trained deep learning water depth prediction model to obtain minute-level dynamic prediction results of channel water depth. The water depth prediction model is a temporal feature extraction network that includes a self-attention mechanism and a long short-term memory unit. It integrates a Transformer structure to extract long-distance dependencies and combines an LSTM unit to capture short-term dynamic change trends. A monotonic head operator layer is introduced into the model structure to ensure that the prediction output meets the predetermined physical monotonic constraint relationship.

[0009] Preferably, the methods for acquiring the original multi-source observation data include:

[0010] Multi-source observation equipment is deployed along typical sections of the waterway. The observation equipment includes: fixed radar water level gauges on the shore, floating or ship-borne multibeam echo sounders, hydrological monitoring systems carried by unmanned survey vessels, remote sensing satellite or UAV imaging devices, automatic weather stations, and GNSS high-precision positioning modules.

[0011] Multi-source observation data are synchronously uploaded to the system data center via wireless communication or local storage, and are timestamped, error checked, and preliminarily cleaned in the data processing unit to form a structured raw input dataset.

[0012] As a preferred approach, the processing of multi-source observation data includes time synchronization, Kalman filtering for noise reduction, outlier detection, and missing value imputation. The fused data is then input into a deep learning water depth prediction model, including: weighted fusion based on the historical confidence of each channel and Bayesian calibration combining prior and likelihood; constructing derived features, including differencing, rate of change, and smoothing, by identifying abrupt changes and gate action events, and calculating autocorrelation and cross-correlation to characterize temporal dependence and cross-source coupling; unifying and standardizing dimensions, performing time encoding, slicing into fixed time windows, and applying causal masks to form a spatiotemporal sequence that can be used for modeling; and retaining the upstream and downstream water levels and gate opening and closing states in a direct pass-through manner to the previous layer for directional correction by the monotonic head operator layer.

[0013] Preferably, the input of the deep learning water depth prediction model is fused multi-source spatiotemporal observation data, and the output is a minute-level predicted value of the channel water depth; its training samples include historical multi-source spatiotemporal observation data and corresponding measured values ​​of channel water depth; the deep learning water depth prediction model includes a monotonic head operator layer, which is set at the output of the Transformer-LSTM network, before the final output layer, and is used to apply monotonicity constraints on the influence of key physical variables; the key physical variables include upstream water level and gate opening, as monotonic non-decreasing variables; downstream water level and gate closure degree, as monotonic non-increasing variables.

[0014] Preferably, the training samples are obtained by the following steps:

[0015] Based on continuous multi-source observation data sequences and corresponding measured channel depth values, the data at each observation time are synchronized and aligned, missing data are detected and filled in, and the multi-source data characteristics and corresponding measured channel depth values ​​at each observation time are obtained.

[0016] By dividing the data into preset time windows, the continuous observation data within each time window are constructed into a model input sequence;

[0017] Based on the time period covered by each input sequence, the measured channel depth corresponding to the last moment of the sequence is extracted as the predicted output label of the input sequence, which is used as the training target of the model in supervised learning, to ensure that each input data corresponds to a unique output result, so as to construct a standard input-output sample pair;

[0018] By traversing all observation sequences, complete input data and output label pairs are collected to construct a sample set for training.

[0019] Preferably, the deep learning water depth prediction model adopts a hybrid neural network architecture, which achieves dynamic prediction of channel water depth by stacking Transformer networks and LSTM networks. The Transformer network, from input to output, includes two encoding layers based on a multi-head self-attention mechanism and two feedforward neural network layers, and combines residual connections and regularization units to extract long-range dependency features of the time series. The LSTM network, from input to output, includes an input layer, two hidden layers, and an output layer. The input layer receives the time series features extracted by the Transformer network, and the two hidden layers extract short-term dynamic change features to obtain the preliminary prediction representation of the backbone network. A monotonic head operator layer is connected in series before the final output layer. This monotonic head operator layer simultaneously receives the hidden representation of the backbone network and key physical variables, generates directional residuals, and performs physical consistency adjustments to the preliminary prediction, thereby outputting the final prediction result that satisfies the monotonic constraint. The final output layer yields the predicted channel water depth.

[0020] Preferably, a monotonic head operator layer is set before the final output layer to jointly map the implicit representation of the backbone network with key physical variables, forming a directional residual correction. The rising branch of the monotonic head operator layer is used to reflect the positive impact of the increase in upstream water level and gate opening on water depth, and the falling branch is used to reflect the suppressive impact of the increase in downstream water level and gate closure degree on water depth. Finally, the monotonic head operator layer and the backbone prediction results are residually fused to output a minute-level water depth prediction value that satisfies the physical monotonic relationship.

[0021] Preferably, the training of the deep learning water depth prediction model is optimized using a joint loss function that includes multiple constraints. The joint loss includes an accuracy term to approximate the measured water depth, a monotonic consistency term to constrain the output of the monotonic head operator layer to satisfy a physically monotonic relationship, and a residual magnitude constraint term to suppress overcorrection. The monotonic consistency term penalizes situations where the upstream water level and gate opening differ only in magnitude, while other conditions remain consistent, by constructing sample pairs in the training set. Similarly, it penalizes situations where the downstream water level or gate closure differs only in magnitude, while the predicted depth is greater. The residual magnitude constraint term limits the correction magnitude of the monotonic head operator layer to the main prediction results, avoiding excessive jumps due to single-point anomalies under extreme conditions.

[0022] A smart channel depth prediction system based on multi-source information collaborative fusion includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the smart channel depth prediction method based on multi-source information collaborative fusion.

[0023] A computer program product includes a computer program that, when executed by a processor, implements the steps of the intelligent waterway depth prediction method based on multi-source information collaborative fusion.

[0024] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. This invention utilizes a collaborative fusion mechanism of multi-source observation data, which can fully integrate heterogeneous data sources such as radar water level gauges, GNSS systems, multibeam echo sounders, and remote sensing images. Combined with time synchronization, noise reduction, anomaly removal, and missing data imputation, it can improve data integrity and consistency, enhance the model's adaptability to complex environments, and overcome the limitations of insufficient prediction accuracy of a single data source in navigation-changing environments.

[0026] 2. The deep learning water depth prediction model in this invention integrates the Transformer structure and LSTM unit, and combines global attention mechanism with local dynamic modeling capability, which can simultaneously capture long-distance dependencies and short-term disturbance features in water depth changes. By introducing a monotonic head operator layer before the output layer, monotonic constraints are applied to the upstream / downstream water level and gate opening and closing, and directional residual correction is performed to avoid anti-physical jumps and improve stability and interpretability.

[0027] 3. This invention can adapt to different waterway environments and different ship types, and has good versatility and scalability. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the main steps in an embodiment of the present invention.

[0029] Figure 2 This is a diagram illustrating the process of obtaining input data from raw data processing in an embodiment of the present invention.

[0030] Figure 3 This is a schematic diagram of the deep learning network model structure in an embodiment of the present invention.

[0031] Figure 4 This is a flowchart illustrating the data output and visualization process in an embodiment of the present invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0033] This invention discloses an intelligent channel depth prediction method based on multi-source information collaborative fusion, such as... Figure 1 As shown, it includes:

[0034] Step S1: Collect raw multi-source observation data;

[0035] In this embodiment, the raw data is acquired by deploying multiple types of hydrological monitoring equipment, specifically including:

[0036] Fixed radar level gauges should be installed on both sides of the waterway shoreline, with installation locations avoiding tree obstruction and wind and wave interference to ensure stable echo signals. Floating or shipborne multibeam echo sounders should be deployed at typical cross-sections to collect water depth profile data. The multibeam emission angle should cover the underwater cross-section, and the overlap rate should be controlled within 30% to ensure data integrity and accuracy. Unmanned survey vessels equipped with current meters, conductivity meters, and turbidity sensors should be used to acquire refined cross-sectional hydrodynamic parameters and water quality information. The vessel's trajectory should be controlled via GNSS positioning to ensure that the observation path matches the sampling points required by the model. High-resolution remote sensing images should be acquired by deploying remote sensing satellites or UAV imaging systems in the air to assist in identifying shoreline, river morphology, and vegetation cover. The spatial resolution of the remote sensing images should be better than 0.5m, and the shooting time should avoid cloud interference. Ground-based automatic weather stations should be set up at key locations to collect data on external influencing factors such as wind speed, wind direction, air pressure, and precipitation.

[0037] Data is uploaded in real time using fixed equipment such as radar level gauges and weather stations via 4G / 5G cellular networks or LoRa low-power wide-area communication. Data collected by multibeam echo sounders and unmanned survey vessels is stored in local data modules and uploaded uniformly after each voyage. Remote sensing images are aggregated and transmitted via ground stations or UAV remote control terminals. Upon receiving the data, the system automatically timestamps it. All data is recorded in Coordinated Universal Time (UTC) format to ensure strict alignment of data from different sources in the time dimension and to perform error verification processes, including static drift identification, sensor mutation point removal, and screening of upper and lower limits of physical parameters, thus eliminating invalid data. In the initial cleaning stage, methods such as cubic spline interpolation or sliding window averaging are used to fill short-term data gaps. Spatial locations are transformed by coordinate projection, unified to WGS-84 format, and spatially corrected in conjunction with terrain elevation correction. All types of observation data are processed through the above process to form a structured raw input dataset with consistent temporal granularity, spatial accuracy, and physical attribute definitions, facilitating subsequent model fusion and time-series prediction input.

[0038] Step S2: Process the raw multi-source observation data to obtain input data with a unified spatiotemporal structure, such as... Figure 2 As shown, it includes:

[0039] Step S2-1: Time synchronization processing of multi-source observation data;

[0040] In shipborne multi-sensor systems, due to the different installation locations of each device, the acquired data are originally distributed in their own independent coordinate reference systems. Without coordinate transformation, significant spatial alignment errors in various observation data will occur, affecting the fusion accuracy. Therefore, it is necessary to first unify the outputs of each sensor to the fixed reference system of the ship's hull, and then perform coordinate transformation using a rigid body transformation model. The transformation formula is shown below:

[0041]

[0042] in, for Rotation matrix, for 1. Translation vector This is the position vector in the sensor coordinate system. These are the world coordinates after transformation to the ship's hull reference frame.

[0043] Sonar repositioning is performed for calibration; if the first Road sonar delay is The speed of sound is The calibration error is The actual water depth is:

[0044]

[0045] Linear interpolation alignment of data streams from different frequencies and sensors based on a unified GPS timestamp:

[0046]

[0047] in , These are the original sampling times of two adjacent frames. , The time is obtained by interpolation to correspond to the measured values. alignment value This method ensures that all sensor data correspond one-to-one on the same timeline, eliminating time delay errors.

[0048] Step S2-2: Kalman filtering for noise reduction;

[0049] Raw sensor signals may contain data fluctuations caused by equipment errors, environmental disturbances, or communication noise, such as abnormal radar reflections, sonar echo jumps, and GPS positioning jitter. Therefore, filtering and noise reduction are necessary. Here, a Kalman filter is used to smooth the estimation of key physical quantities. As a recursive algorithm based on Bayesian minimum mean square error estimation, the Kalman filter can calculate the optimal state estimate based on prior estimates of the current observation and the previous state, given the known system dynamic model and observation model. The estimation results are dynamically adjusted using the Kalman gain, thereby significantly reducing random errors in sensor measurements while preserving true dynamic characteristics.

[0050]

[0051]

[0052] in This is a state vector (e.g., boat speed, water level). Let be the state covariance matrix. Here is the state transition matrix. To control the input matrix, To control the input drive, For process noise covariance. Measurement update phase:

[0053]

[0054]

[0055]

[0056] Where H is the observation matrix and R is the observation noise covariance. For measurement vectors, For Kalman gain.

[0057] Step S2-3: Outlier detection;

[0058] The Z-Source method is used to detect outliers in the filtered data:

[0059]

[0060] in , These are the mean and standard deviation of the sequence, respectively. If the absolute value exceeds the threshold, it is marked as an outlier and repaired by linear or spline interpolation of nearby normal samples.

[0061] Step S2-4: Imputing missing values;

[0062] For gaps in data caused by communication packet loss or transient sensor failures, linear interpolation or cubic spline interpolation methods are used to fill them in, so as to ensure the continuity of the time series and the stability of subsequent feature calculations.

[0063] Through the above processing, the original observation data was uniformly calibrated, high-frequency noise and abrupt changes introduced by equipment errors, environmental interference and acquisition delay were filtered out, and the time gaps caused by signal loss were effectively filled, thus providing a dataset foundation for subsequent data prediction.

[0064] Step S3: Input the fused multi-source spatiotemporal observation data into the trained deep learning water depth prediction model to obtain minute-level predictions of channel water depth at the corresponding time. The water depth prediction model is a temporal feature extraction network incorporating a self-attention mechanism and long short-term memory units. It integrates a Transformer structure to extract long-distance dependencies and combines LSTM units to capture short-term dynamic trends. A monotonic head operator layer is introduced into the model structure to ensure that the prediction output meets predetermined physical monotonic constraints. The multi-source data includes spatiotemporally heterogeneous hydrological and environmental information collected from observation equipment such as radar level gauges, depth sounders, remote sensing images, weather stations, GNSS positioning, and ship motion sensors. In practice, the images can first be extracted into a vector embedding, then aligned with high-frequency numerical features such as those from level gauges and depth sounders on the same time axis, and concatenated into a feature vector of one per minute, which is then fed into the model.

[0065] The prediction model takes fused multi-source spatiotemporal observation data as input and outputs minute-level predicted values ​​of channel depth. Its training samples include multi-source observation data and corresponding measured values ​​of channel depth.

[0066] Specifically, the training samples are obtained through the following steps:

[0067] Based on continuous multi-source observation data sequences and corresponding measured channel depth values, the data at each observation time are synchronized and aligned in time, missing data are detected and filled in, and the multi-source observation characteristics and measured channel depth values ​​at each observation time are obtained.

[0068] By dividing the data into preset time windows, the continuous observation data within each time window are constructed into a model input sequence;

[0069] Based on the time period covered by each input sequence, the measured channel depth corresponding to the last moment of the sequence is extracted as the predicted output label of the input sequence. This label is used as the training target of the model in supervised learning, ensuring that each input data corresponds to a unique output result, so as to construct a standard input-output sample pair.

[0070] Based on the above steps, by traversing all observation sequences, complete input data and output label pairs are collected to construct a sample set for training.

[0071] like Figure 3 As shown, the deep learning water depth prediction model in this embodiment adopts a hybrid neural network architecture, which achieves dynamic prediction of channel water depth by stacking Transformer networks and recurrent neural networks. The Transformer network part includes two encoding layers based on multi-head self-attention mechanism and two feedforward neural network layers from input to output, and combines residual connections and regularization units to extract long-range dependency features of time series. The LSTM network part includes one input layer, two hidden layers and one output layer from input to output. The input layer receives the time series features extracted by the Transformer network part and connects a monotonic head operator layer in series before the output layer. This operator layer simultaneously receives the implicit representation of the backbone network and key physical variables, and generates directional residuals with rising and falling branches to perform physical consistency correction on the initial backbone prediction: there is no abnormal decrease as the upstream water level and gate opening increase, and no abnormal increase as the downstream water level and gate closure degree increase, and the output is a final minute-level water depth prediction value that meets the monotonic constraint.

[0072] Data fusion includes: based on synchronization, denoising, and interpolation, weighted fusion is performed according to the historical confidence of each channel, and Bayesian calibration is performed by combining prior and likelihood to obtain more consistent multi-source water depth and hydrodynamic elements; by identifying key events such as sonar abrupt changes, meteorological sudden changes, and gate actions, derived features such as difference, rate of change, and smoothing are constructed, and autocorrelation and cross-correlation are calculated to characterize temporal dependence and cross-source coupling; by unifying the dimensions and standardizing the full set of features, time interval encoding is supplemented and sine and cosine position encoding is added, and the data is sliced ​​according to a fixed time window and causal mask is applied to form a spatiotemporal sequence that can be used for modeling; at the same time, the upstream and downstream water levels and gate opening and closing states are retained in a direct manner to the layer before the output, so that the monotonic head operator layer can perform directional correction, thereby providing physically consistent and fully represented fusion input for the deep learning prediction module. This is mainly achieved through the following steps:

[0073] Assume the first Road sensor at all times Measured water depth Its uncertainty or historical standard deviation is Then the weight can be set to

[0074]

[0075] Then, by weighting and fusing, the instantaneous water depth estimate is obtained:

[0076] ,

[0077] Further introduction of prior distribution (e.g., historical water depth distribution) and sensor likelihood function According to Bayes' theorem:

[0078]

[0079] Where n is the total number of valid measurements combined with weighted fusion.

[0080] The optimal water depth estimate after multi-sensor joint calibration is obtained by solving for the posterior maximum or posterior mean.

[0081] Feature points are extracted from sonar bathymetry sequences and ship trajectories. In sonar bathymetry sequences, feature points typically refer to the moments with the strongest echoes in the sonar pulses or locations where the echo delay between two adjacent pulses changes significantly. In ship trajectories, feature points typically refer to the peak acceleration during roll or rapid acceleration / deceleration, or the moments when the heading angle suddenly deviates from the intended course. In environmental meteorological data, feature points are typically moments of rapid wind speed increases and points of rapid atmospheric pressure decreases within a short period.

[0082] Regarding the water depth after fusion and ship speed Calculate the first-order difference and the rate of change of velocity:

[0083]

[0084] To mitigate the impact of noise, moving averages can be calculated separately on the feature sequences:

[0085]

[0086] And exponentially weighted moving average:

[0087]

[0088] in, This is the window length for the moving average.

[0089] Within a fixed-length time window, basic statistics such as the mean and variance of the data series are extracted, and the autocorrelation function is calculated to characterize its time-series dependency structure.

[0090]

[0091]

[0092]

[0093] in, The statistical window length is used to extract statistics within a fixed-length time window, where n is the actual number of valid samples within that time window.

[0094] To assess the linear correlation between the characteristic sequences of any two data sets, calculate the Pearson correlation coefficient:

[0095]

[0096] in, ), Let i be the scalar time series values ​​of the i-th and j-th data sources at time t. , To include within the selected statistics window ), The sample mean.

[0097] To ensure the hybrid model obtains optimal input, an integrated processing step is required after fusion and extraction:

[0098] ① Feature integration

[0099] For all sequences Perform Min-Max normalization:

[0100]

[0101] Or Z-Score standardization:

[0102]

[0103] Make the distribution value of each feature dimension equal to 0 and the variance to 1, thus eliminating differences in different dimensions and magnitudes.

[0104] ② Sine and Cosine Position Encoding

[0105] To preserve the relative order of each time point in the sequence, at the first... Each time step and dimension index is and Add them separately:

[0106]

[0107]

[0108]

[0109] The input vector.

[0110] ③ Sliding window and attention mask

[0111] The length is sequence { By fixed window Segmented into overlapping subsequences:

[0112]

[0113] And for each A matrix is ​​constructed mask to ensure that only information from the current position and previous time steps can be accessed during self-attention computation.

[0114]

[0115] By cascading a monotonic head operator layer before the final output layer: This operator layer simultaneously receives the implicit representation of the main body and key physical variables (including upstream water level, downstream water level, gate opening degree and gate closing degree), generates directional residuals in the manner of "rising branch / falling branch", and performs physical consistency adjustment on the preliminary prediction of the main body to ensure that the prediction results do not abnormally decrease with the increase of upstream water level and gate opening degree, and do not abnormally increase with the increase of downstream water level and gate closing degree, and outputs the final water depth prediction value that satisfies the monotonic constraint.

[0116] ① Weight nonnegation

[0117] By training parameters The mapping is to a non-negative weight w, and monotonous activation is used to ensure that the branch mapping is monotonically non-decreasing.

[0118]

[0119] ② Monotonic head operator layer

[0120] The positive impact of upstream water level / gate opening and the suppressive impact of downstream water level / gate closure are separated and modeled to form a directional correction quantity.

[0121]

[0122]

[0123]

[0124] in, The water level at the upstream monitoring station of the sluice gate. The water level at the downstream monitoring station of the sluice gate. The degree to which the gate is open. The degree to which the gate is closed. The contextual features output by the Transformer+LSTM backbone , These are nonnegative mappings for ascending and descending branches, respectively. , This is the weight matrix for the ascending and descending branches, and each element is derived from... The formal constraint is nonnegative. , These are the bias parameters trained for the upward and downward branches. It is a monotonically non-decreasing activation function.

[0125] ③ Residual Fusion

[0126] By adding the backbone prediction and the directional residual before the output layer, the final water depth is obtained, preserving the backbone's fit to complex patterns and forcing it to satisfy the physical orientation.

[0127]

[0128] in, In order to be in Predicting water depth at any given time; This represents the predicted water depth before the network undergoes physical correction.

[0129] ④ Empirical monotonic constraints

[0130] By constraining the network output through sample pair order, inverse physical sorting is avoided.

[0131]

[0132]

[0133] in, To ultimately predict the water depth, , This is the input feature vector for the corresponding sample. These represent the monotonic losses for increasing and decreasing constraints, respectively. , For the index set of paired sample sets, , The values ​​of the comparison variables selected for monotonic constraints on samples i and j.

[0134] ⑤ Total Loss

[0135]

[0136]

[0137] in, To approximate the accuracy of the measured water depth, MSE / MAE is commonly used. To constrain the output of the monotonic head operator layer to satisfy the physical monotonic relationship of the monotonic consistency term; As a residual magnitude constraint term used to suppress overcorrection, the directional residual of MHOL is controlled to be not too large, so as to avoid jitter caused by overcorrection; For regular expressions, For all the parameters in the model that need to be trained, , , For parameters.

[0138] like Figure 3 In the multi-head self-attention mechanism, the number of attention heads is 4 or 8, and the dimension of each attention head is a certain proportion of the dimension of the input features. After the self-attention layer, a feedforward neural network layer with ReLU activation is connected to perform nonlinear transformation on the extracted features. In the recurrent neural network part, the number of hidden units in each hidden layer is 100 to 200, and the activation function of the gated unit is used inside the hidden layer to perform nonlinear mapping on the time series information. During the model training process, a Dropout layer is set to prevent overfitting, and the dropout ratio is 0.2 to 0.3. The final output layer of the prediction model is a fully connected layer and does not use an activation function, directly outputting the predicted channel depth value.

[0139] Figure 4 The data output and visualization process is illustrated. In this embodiment, when the storage and communication module sends the processed data to the display and early warning module via a wireless network, key indicators such as link capacity, channel fading, and latency are also considered.

[0140] To ensure the stability and continuity of surveying data during real-time transmission to the display and early warning modules, it is essential to ensure that the transmission link has sufficient Shannon channel capacity, meaning the link bandwidth must be able to handle peak data rates. Otherwise, issues such as queuing delays, data packet loss, and even image frame loss will arise, severely impacting the system's timeliness and reliability.

[0141]

[0142] in, For channel bandwidth, For received power, This represents the noise power spectral density.

[0143] By quantifying the signal attenuation with distance and frequency, the antenna gain and transmit power configuration are guided to ensure that the receiver can obtain sufficient signal strength.

[0144]

[0145] in, Link distance, For carrier frequency, It is the speed of light and approximately equal to .

[0146] Link and buffer optimization are achieved through end-to-end latency analysis to ensure... Real-time transmission.

[0147]

[0148]

[0149]

[0150] in, Encoding the signal Decoding and packet processing latency, For transmission delay, To delay the transmission time, Due to network equipment queuing delays, For the length of the package, For the transmission rate of the link, For the speed of transmission.

[0151] By quantifying the transmission reliability at a specific signal-to-noise ratio, we can guide the design of link budget, modulation scheme, and error correction coding.

[0152]

[0153] in, () represents the complementary error function. This represents the energy-to-noise ratio per bit.

[0154] After receiving the data, the display and early warning module performs further data analysis and processing to generate surveying results.

[0155] The system outputs comprehensive mapping results, including contour maps, time-series water depth variation maps, and low-water-level risk distribution maps, providing precise decision support for lock scheduling and navigation safety.

[0156] This is achieved by performing finite difference calculations on the grid elevation, providing a basis for decision-making in the identification and risk assessment of low-water-level stagnant areas.

[0157]

[0158]

[0159] in, For elevation right The first-order partial derivative of the direction represents the local horizontal and vertical slope components; The slope vector; , For elevation in Second-order partial derivatives in the direction; The curvature of the cross section.

[0160] By generating a continuous water depth distribution surface through unbiased minimum variance interpolation, the problem of sparse and uneven measurement points from multiple sources can be solved, and a high-precision water depth field can be established.

[0161]

[0162]

[0163] in, It is semivariance. For distance falling The number of sample pairs in the interval. For position , The observation depth at that location To estimate the water depth at unsampled locations, The number of neighboring observations participating in the interpolation. For Kriging weights.

[0164] Quantitative assessment of the total volume and affected area below the safety threshold

[0165]

[0166]

[0167] in, To ensure a minimum safe water depth threshold, For grid Predicted depth at that location Area of ​​each grid cell.

[0168] Error propagation and confidence interval

[0169]

[0170] in This is the predicted water depth vector output by the model. Let be the covariance matrix of the variables. For Jacobian matrices, This is the vector of underlying measurement variables.

[0171] This invention also discloses an intelligent waterway depth prediction system based on multi-source information collaborative fusion, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the intelligent waterway depth prediction method based on multi-source information collaborative fusion.

[0172] This invention also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the intelligent waterway depth prediction method based on multi-source information collaborative fusion.

[0173] The program code used to implement the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the steps of the method of the present invention to be performed. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. All aspects not detailed in this invention are well-known to those skilled in the art.

Claims

1. A method for predicting channel depth based on multi-source information synergy and fusion, characterized in that, include: The collected raw multi-source observation data are processed to obtain input data with a unified spatiotemporal structure; The fused data is input into a trained deep learning water depth prediction model to obtain minute-level dynamic prediction results of channel water depth. The water depth prediction model is a temporal feature extraction network that includes a self-attention mechanism and a long short-term memory unit. It integrates a Transformer structure to extract long-distance dependencies and combines an LSTM unit to capture short-term dynamic change trends. A monotonic head operator layer is introduced into the model structure to ensure that the predicted output meets the predetermined physical monotonic constraint relationship.

2. The intelligent channel depth prediction method based on multi-source information collaborative fusion according to claim 1, characterized in that, The methods for acquiring the raw multi-source observation data include: Multi-source observation equipment is deployed along typical sections of the waterway. The observation equipment includes: fixed radar water level gauges on the shore, floating or ship-borne multibeam echo sounders, hydrological monitoring systems carried by unmanned survey vessels, remote sensing satellite or UAV imaging devices, automatic weather stations, and GNSS high-precision positioning modules. Multi-source observation data are synchronously uploaded to the system data center via wireless communication or local storage, and are timestamped, error checked, and preliminarily cleaned in the data processing unit to form a structured raw input dataset.

3. The intelligent waterway depth prediction method based on multi-source information collaborative fusion according to claim 1, characterized in that, The processing of multi-source observation data includes time synchronization, Kalman filtering for noise reduction, outlier detection, and missing value imputation. The fused data is input into a deep learning water depth prediction model, including: performing weighted fusion based on the historical confidence of each channel and performing Bayesian calibration by combining prior and likelihood. By identifying mutations and gate action events, derived features are constructed, including difference, rate of change, and smoothing. Autocorrelation and cross-correlation are calculated to characterize temporal dependence and cross-source coupling. Dimensions are unified and standardized, and time encoding is performed. The data is sliced ​​into fixed time windows and causal masks are applied to form a spatiotemporal sequence that can be used for modeling. The upstream and downstream water levels and gate opening and closing states are retained in a direct manner to the layer before the output, so that the monotonic head operator layer can perform directional correction.

4. The intelligent channel depth prediction method based on multi-source information collaborative fusion according to claim 1, characterized in that, The input to the deep learning water depth prediction model is the fused multi-source spatiotemporal observation data, and the output is the minute-level prediction value of the channel water depth. Its training samples include historical multi-source spatiotemporal observation data and corresponding measured values ​​of channel water depth; the deep learning water depth prediction model includes a monotonic head operator layer, which is set at the output of the Transformer-LSTM network, before the final output layer, and is used to apply monotonic constraints on the influence of key physical variables; the key physical variables include upstream water level and gate opening, as monotonic non-decreasing variables; downstream water level and gate closure degree, as monotonic non-increasing variables.

5. The intelligent channel depth prediction method based on multi-source information collaborative fusion according to claim 4, characterized in that, The training samples are obtained through the following steps: Based on continuous multi-source observation data sequences and corresponding measured channel depth values, the data at each observation time are synchronized and aligned, missing data are detected and filled in, and the multi-source data characteristics and corresponding measured channel depth values ​​at each observation time are obtained. By dividing the data into preset time windows, the continuous observation data within each time window are constructed into a model input sequence; Based on the time period covered by each input sequence, the measured channel depth corresponding to the last moment of the sequence is extracted as the predicted output label of the input sequence, which is used as the training target of the model in supervised learning, to ensure that each input data corresponds to a unique output result, so as to construct a standard input-output sample pair; By traversing all observation sequences, complete input data and output label pairs are collected to construct a sample set for training.

6. The intelligent channel depth prediction method based on multi-source information collaborative fusion according to claim 1, characterized in that, The deep learning water depth prediction model adopts a hybrid neural network architecture, which realizes dynamic prediction of channel water depth by stacking Transformer network and LSTM network. The Transformer network part includes two encoding layers based on multi-head self-attention mechanism and two feedforward neural network layers from input to output, and combines residual connection and regularization unit to extract long-distance dependency features of time series. The LSTM network consists of an input layer, two hidden layers, and an output layer from input to output. The input layer receives the temporal features extracted by the Transformer network. After the two hidden layers extract the short-term dynamic change features, the preliminary prediction representation of the backbone network is obtained. A monotonic head operator layer is connected in series before the final output layer. The monotonic head operator layer receives the hidden representation of the backbone network and key physical variables, generates directional residuals, and performs physical consistency adjustments to the preliminary prediction, thereby outputting the final prediction result that satisfies the monotonic constraint. The final output layer yields the predicted channel depth.

7. The intelligent waterway depth prediction and multi-source information collaborative fusion method according to claim 6, characterized in that, By setting a monotonic head operator layer before the final output layer, the implicit representation of the backbone network is jointly mapped with key physical variables to form a directional residual correction. The rising branch of the monotonic head operator layer is used to reflect the positive impact of the increase in upstream water level and gate opening on water depth, and the falling branch is used to reflect the suppressive impact of the increase in downstream water level and gate closure degree on water depth. Finally, the monotonic head operator layer and the backbone prediction results are residually fused to output a minute-level water depth prediction value that satisfies the physical monotonic relationship.

8. The intelligent channel depth prediction and multi-source information collaborative fusion method according to claim 1, characterized in that, The deep learning water depth prediction model is trained using a joint loss function with multiple constraints. This joint loss includes an accuracy term to approximate the measured water depth, a monotonic consistency term to constrain the output of the monotonic head operator layer to satisfy a physically monotonic relationship, and a residual magnitude constraint term to suppress overcorrection. The monotonic consistency term penalizes situations where the upstream water level and gate opening differ only in magnitude, while other conditions remain consistent, by constructing sample pairs in the training set. Similarly, it penalizes situations where the downstream water level or gate closure differs only in magnitude, while the predicted depth is greater. The residual magnitude constraint term limits the correction magnitude of the monotonic head operator layer to the main prediction results, preventing excessive jumps due to single-point anomalies under extreme conditions.

9. An intelligent channel depth prediction system based on multi-source information collaborative fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent waterway depth prediction method based on multi-source information collaborative fusion as described in any one of claims 1-8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent waterway depth prediction method based on multi-source information collaborative fusion as described in any one of claims 1-8.

Images