Search

CN-121980479-A - River course ecological abnormal change identification method based on machine learning

CN121980479ACN 121980479 ACN121980479 ACN 121980479ACN-121980479-A

Abstract

The invention discloses a river channel ecological abnormal change identification method based on machine learning, which relates to the technical field of machine learning and comprises the steps of collecting multi-section ecological monitoring data, preprocessing, constructing a space-time sequence data set, constructing a river channel directional topological structure, generating a topological adjacent matrix, completing association mapping of space-time data and nodes, inputting space-time characteristics and the adjacent matrix, improving DCRNN learning diffusion rules, outputting a dynamic ecological base line sequence, constructing a stable ecological sample set, mapping to generate a stable subspace, correcting a base line and calculating structure deviation, constructing an ecological energy transmission interval matrix, carrying out bidirectional propagation consistency judgment, obtaining a propagation consistency index, calculating a base line residual error and combining the structure deviation, and outputting an abnormal identification result. According to the invention, by constructing a topological constraint dynamic ecological base line and fusing a stable ecological structure with a propagation consistency judgment, accurate identification and early warning of gradual ecological abnormal changes of a river channel are realized.

Inventors

  • LI JIELIN
  • LI YUANYUAN
  • LI FENGMIN
  • LI HUI

Assignees

  • 中国海洋大学

Dates

Publication Date
20260505
Application Date
20260408

Claims (8)

  1. 1. The method for identifying the ecological abnormal change of the river channel based on machine learning is characterized by comprising the following steps of: Collecting ecological monitoring data of a plurality of monitoring sections of a river channel, preprocessing the ecological monitoring data, and constructing a space-time sequence data set; Constructing each monitoring section into a directed topology structure according to the river water flow direction relation, generating a corresponding topology adjacency matrix, and carrying out node association mapping on the space-time sequence data set and the directed topology structure to form a space-time characteristic input sequence; Inputting a space-time characteristic input sequence and a topological adjacency matrix into an improved DCRNN model, learning a space-time diffusion process of the ecological state of the river channel multi-section, outputting an ecological state predicted value of each monitoring section at a target moment, and constructing a dynamic ecological base line sequence conforming to the topological diffusion characteristic of the river channel; Based on the space-time sequence data set, a stable ecological state sample set is constructed, characteristic space mapping is carried out to generate a stable ecological state subspace, a dynamic ecological base line sequence is projected to the stable ecological state subspace to obtain a stable state correction base line, and the ecological structure deviation amount is calculated; constructing an inter-section ecological energy transfer interval matrix based on the river channel distance parameters, the flow velocity parameters and the diffusion parameters among the monitoring sections, and judging bidirectional propagation consistency in the upstream-to-downstream direction and the downstream-to-upstream direction according to the variation of the ecological state of each monitoring section along with time to obtain a propagation consistency index; And calculating a baseline residual error according to the dynamic ecological baseline sequence and the actual ecological monitoring data, and carrying out joint discrimination on the baseline residual error, the ecological structure deviation and the propagation consistency index to output a river ecological abnormal change identification result.
  2. 2. The machine learning-based river course ecological anomaly change identification method of claim 1, wherein the ecological monitoring data comprises water quality index data, hydrodynamic parameter data and environmental auxiliary data, wherein the water quality index data comprises dissolved oxygen, chemical oxygen demand, ammonia nitrogen, total phosphorus, turbidity and chlorophyll concentration, the hydrodynamic parameter data comprises flow, flow rate and water level, and the environmental auxiliary data comprises rainfall and air temperature data.
  3. 3. The machine learning-based river channel ecological anomaly change identification method according to claim 1, wherein the preprocessing of ecological monitoring data comprises unified time alignment processing of multi-source data of different monitoring sections according to time stamps, complementation of missing data by a time sequence interpolation method, sliding window filtering processing of anomaly noise data, standardized normalization processing of each ecological index data and construction of a multi-section multi-index space-time sequence data set.
  4. 4. The machine learning-based river course ecological anomaly change identification method of claim 1, wherein the node-associated mapping of the spatio-temporal sequence dataset and the directed topology structure to form the spatio-temporal feature input sequence comprises: Acquiring space position information, river channel communication relation and water flow direction information of each monitoring section, correspondingly taking each monitoring section as a node, constructing a node set and a directional communication relation according to the relation that actual water flow points from upstream to downstream, and forming a directional topological structure reflecting the actual water power transmission path of the river channel; Establishing a directed edge connection relation for adjacent sections according to the river channel distance, the water flow direction and the hydraulic connectivity among the monitored sections, and performing topological ordering on the nodes according to the sequence from upstream to downstream to obtain a river channel section sequence with direction constraint; Decomposing the space-time sequence data set according to the monitoring section numbers, dividing ecological monitoring data corresponding to each time step into section feature subsequences corresponding to each node one by one, and establishing a unique mapping relation between section identifiers and node identifiers; constructing a topology-aware node association mapping mechanism based on a directed topology structure, sequentially recombining the section feature subsequences according to the river water flow direction, and progressively associating and binding the time sequence feature of each node with the history feature of an upstream node to form a node space-time feature input sequence with a diffusion memory attribute; and carrying out unified coding treatment on the recombined node space-time characteristic input sequence and the directed topological structure to generate topological constraint space-time characteristic data for each monitoring section.
  5. 5. The method for identifying abnormal river channel ecology change based on machine learning according to claim 1, wherein the constructing a dynamic ecology baseline sequence conforming to the topological diffusion characteristics of the river channel comprises: An improved DCRNN model is built, the improved DCRNN model comprises an input coding layer, a topology constraint bidirectional diffusion convolution layer, a section grading gating fusion layer, a time sequence memory updating layer and a dynamic base line output layer, wherein the input coding layer is sequentially connected with the topology constraint bidirectional diffusion convolution layer, the output end of the topology constraint bidirectional diffusion convolution layer is connected with the input end of the section grading gating fusion layer, the output end of the section grading gating fusion layer is connected with the input end of the time sequence memory updating layer, and the output end of the time sequence memory updating layer is connected with the input end of the dynamic base line output layer; the input coding layer receives the topological constraint space-time characteristic data and codes multi-index ecological data in a continuous time window into node time sequence characteristic vectors according to the sequence of monitoring section nodes; In the topological constraint bidirectional diffusion convolution layer, taking a directional topological structure as a diffusion constraint structure, carrying out forward diffusion characteristic transfer along the upstream-to-downstream direction, carrying out reverse feedback characteristic transfer along the downstream-to-upstream direction, and introducing topological attenuation weight based on river channel distance and flow velocity in the diffusion process to obtain node diffusion characteristics with river channel topological diffusion characteristics; Inputting node diffusion characteristics into a section grading gating fusion layer, grading and dividing according to a main river section and a tributary section, setting corresponding gating weights for the diffusion characteristics of different grade sections, carrying out layering fusion on forward diffusion characteristics and reverse feedback characteristics, and outputting topological fusion characteristics simultaneously comprising state information of the section and upstream diffusion accumulation information; Inputting the topology fusion characteristics into a time sequence memory updating layer, and carrying out progressive state updating on the topology fusion characteristics of continuous time steps by using a gating circulating unit to obtain a hidden state sequence reflecting the space-time evolution process of the ecological state of the river channel; in the dynamic baseline output layer, ecological state predicted values of all monitoring sections at target moments are generated according to the hidden state sequences, and dynamic ecological baseline sequences conforming to the topological diffusion characteristics of the river channels are constructed in time sequence.
  6. 6. The method for identifying abnormal river channel ecology change based on machine learning according to claim 1, wherein projecting the dynamic ecology baseline sequence to a steady ecology state subspace to obtain a steady state correction baseline, calculating the ecology structure deviation amount comprises: Based on the space-time sequence data set, continuously calculating the ecological state change amplitude of each monitoring section according to a time window, carrying out stability screening on a continuous time period with the change amplitude lower than the median value in the overall time sequence change, automatically identifying a historical stable operation interval, and extracting multi-section ecological state data in the corresponding time period as candidate stable samples; Carrying out topology consistency verification on candidate stable samples by combining topology constraint space-time characteristic data, and reserving sample data which have synchronous change trend and no mutation transmission characteristics of upstream and downstream nodes in a directed topology structure to construct a stable ecological state sample set; Performing layered characteristic recombination on the stable ecological state sample set according to the monitored section dimension and the ecological index dimension, performing characteristic space mapping processing based on the multi-section multi-index joint distribution relation, projecting original high-vitamin state characteristics to a stable ecological state characteristic space formed by leading the stable samples, and generating a stable ecological state subspace, wherein the characteristic space mapping is obtained by performing aggregation extraction on multi-index collaborative variation directions of the stable samples; inputting the dynamic ecological baseline sequence into a steady ecological state subspace, and carrying out subspace projection correction on the baseline characteristic according to the section node sequence to obtain a steady state correction baseline; And calculating the deviation of the ecological structure based on the overall difference degree between the actual ecological monitoring data at the current moment and the steady state correction base line in multi-section multi-index dimensions, wherein the deviation of the ecological structure is obtained by carrying out weighted aggregation on the difference values of the indexes of each section, and the weight is determined by the connection strength of the corresponding section in the topological structure and the historical stability contribution degree.
  7. 7. The method for identifying a river course ecological anomaly change based on machine learning according to claim 1, wherein the obtaining a propagation consistency index comprises: Calculating ecological transfer influence intensity among the sections according to the connection relation between upstream nodes and downstream nodes based on the directional topological structure, the space distance data, the flow velocity data, the hydraulic retention time data and the historical ecological state change data of each monitoring section, and carrying out interval quantization on the ecological change transmissible range of each node pair by combining the inter-section attenuation coefficient and the flow velocity amplification coefficient to construct an inter-section ecological energy transfer interval matrix; Carrying out time difference processing on ecological state data of each monitoring section in a continuous time window, calculating a multi-index variation sequence of each section between adjacent time steps, and generating a section variation vector set reflecting the evolution of the ecological state along time according to the section node sequence; Establishing a propagation path link along an upstream-to-downstream direction based on a directed topological structure, progressively matching the ecological state variable quantity of a downstream section with the historical variable quantity of a corresponding upstream section in a propagation time delay range, and carrying out joint constraint judgment on the variation amplitude, the variation direction and the variation continuity by combining an ecological energy transmission interval matrix to obtain a propagation consistency judgment result in the upstream-to-downstream direction; Constructing a reverse propagation verification link along the downstream-upstream direction, reversely correlating and comparing the current variation of the upstream section with the historical variation of the downstream section, introducing a reverse attenuation weight based on topological connection strength, and distinguishing the variation which does not accord with a reverse diffusion rule but accords with local abrupt change characteristics to obtain a reverse propagation consistency distinguishing result of the downstream-upstream direction; And comprehensively fusing the propagation consistency results of all sections in the upstream-to-downstream direction and the opposite propagation consistency results in the downstream-to-upstream direction, and carrying out weighted summarization according to the topological path length, the section connection strength and the continuous time step number of the change to generate a propagation consistency index.
  8. 8. The method for identifying a river ecology anomaly change based on machine learning of claim 1, wherein the outputting the river ecology anomaly change identification result comprises: Acquiring a dynamic ecological baseline sequence and actual ecological monitoring data at corresponding time, calculating the difference value of each ecological index of each section one by one according to the node sequence of the monitoring section, and carrying out aggregation treatment on the difference value in index dimension to obtain a baseline residual sequence representing the deviation degree of the actual ecological state relative to the dynamic ecological baseline; The base line residual sequence is weighted and summarized according to the cross section dimension, the weight is determined according to the node connection strength of the directional topological structure and the position of the cross section in the topological path, and a comprehensive base line residual index reflecting the whole deviation degree of multiple cross sections is generated; Acquiring the deviation amount of the ecological structure and the transmission consistency index, and performing reverse conversion treatment on the transmission consistency index to ensure that the transmission consistency index and the deviation index keep consistent in the distinguishing direction, so as to form a combined distinguishing feature set of a base line residual error, the deviation amount of the ecological structure and the transmission consistency deviation amount; the combination judging feature set is subjected to weighted fusion according to a preset weight coefficient, wherein the preset weight coefficient carries out self-adaptive determination on the abnormal recognition contribution degree according to each feature in the historical stable operation stage, and a comprehensive abnormal grading value is obtained; and comparing the comprehensive abnormal grading value with a dynamic discrimination threshold value obtained based on statistics of a historical stable operation stage, judging that the river channel ecology changes when the comprehensive abnormal grading value exceeds the dynamic discrimination threshold value and keeps rising trend in a continuous time window, otherwise judging that the river channel ecology changes normally, and outputting a corresponding ecology abnormal change identification result.

Description

River course ecological abnormal change identification method based on machine learning Technical Field The invention relates to the technical field of machine learning, in particular to a river ecological abnormal change identification method based on machine learning. Background Along with the development of intelligent water affair and ecological environment monitoring technology, the on-line monitoring of the ecological state of the river channel is gradually changed from the traditional manual sampling to a continuous automatic monitoring mode based on multiple sections and multiple indexes, and multi-dimensional ecological data such as dissolved oxygen, ammonia nitrogen, total phosphorus, flow and flow velocity are obtained by arranging a water quality sensor, hydrodynamic monitoring equipment and an environment monitoring terminal, and the ecological change of the river channel is identified by utilizing a time sequence analysis or statistical threshold method. In the prior art, the common method mainly relies on single-section data analysis, fixed threshold judgment or a traditional time sequence prediction model to carry out trend judgment on ecological indexes, and part of the method also introduces a neural network to predict a time sequence, but most of the method only models the ecological state of a single point or an independent section, and lacks systematic description of the upstream and downstream topological relation and the hydrodynamic diffusion process of a river channel. However, since the river channel ecosystem has obvious space-time coupling characteristics, the ecological index change often appears as a gradual diffusion process along the river channel topology structure, and especially under the conditions of upstream disturbance, slow pollution input or internal balance shift of the ecosystem, the ecological abnormal change usually appears as a small-amplitude, continuous and gradual evolution characteristic. The existing method based on the single-point threshold value or the simple time sequence model is difficult to effectively distinguish normal fluctuation from hidden gradual change abnormality, and does not carry out deep correlation modeling on multi-section space-time data and a directional topological structure of a river channel, so that the accuracy of identifying ecological changes conforming to the hydrodynamic force propagation rule is low in a multi-section and multi-index coupling scene, and misjudgment or missed judgment is easy to occur. In the prior art, prediction residual errors or statistical fluctuation is mainly used in the abnormality discrimination process, the combined constraint analysis on the stable structure and the propagation mechanism of an ecological system is lacked, the ecological structure distribution characteristics and the energy transmission and diffusion consistency among sections in the stable running state are not comprehensively considered, and the gradual change abnormal change recognition capability which does not obviously violate physical constraint but deviates from the stable ecological structure is insufficient. In the prior art, the precise identification of ecological abnormal changes which accord with the topological diffusion rule and have concealment and progressive characteristics is difficult to realize in a complex river scene. Therefore, how to provide a method for identifying the ecological abnormal change of a river channel based on machine learning is a problem to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a river channel ecological anomaly change identification method based on machine learning, which is characterized in that by constructing an improved DCRNN dynamic ecological base line which accords with the river channel topology diffusion characteristic and combining a steady ecological state subspace mapping and an ecological energy transmission interval bidirectional propagation consistency discrimination mechanism, multi-section and multi-index space-time ecological data are subjected to joint analysis processing, so that the fine identification of gradual change and hidden ecological anomaly change is realized, the machine learning space-time prediction is fused with a river channel hydrodynamic topological structure and steady ecological structural characteristics, normal working condition fluctuation and real ecological anomaly change can be effectively distinguished, and the method has the advantages of accurate description of the topology diffusion process, strong hidden gradual change anomaly identification capability and high discrimination stability and is suitable for complex river channel multi-section coupling scenes. According to the embodiment of the invention, the method for identifying the ecological abnormal change of the river channel based on machine learning comprises the following steps: Collecting ecological monitoring data of a plurality