CN-121982842-A - Forest ecological monitoring and illegal activity identification method based on ecological semantic graph

CN121982842ACN 121982842 ACN121982842 ACN 121982842ACN-121982842-A

Abstract

The application discloses a forest ecological monitoring and illegal activity identification method based on ecological semantic graphs, which belongs to the technical field of ecological protection and comprises the steps of collecting sound signals and GPS coordinates, sending collected data for preprocessing and feature extraction, performing preliminary detection, generating semantic embedded representation of an ecological acoustic event by using a self-supervision learning model based on a preliminary detection result, abstracting the semantic embedded representation into semantic nodes, establishing semantic edges according to space-time association rules to obtain the ecological acoustic semantic graphs, extracting space-time evolution features of short-time scale and long-time scale by using a graph time sequence convolution network based on the ecological acoustic semantic graphs, identifying multi-scale abnormal events by matching with a preset illegal activity combination mode, and performing sound source positioning based on the collected data and triggering a three-level response link to alarm when the abnormal events are identified. The method improves the recognition accuracy, adapts to scenes and response speed, and reduces labor cost.

Inventors

XIN YOUQIANG
DONG FANGJIE
YUAN HAOXUAN

Assignees

西安雀凌飞信息技术有限公司

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (9)

1. A forest ecological monitoring and illegal activity identification method based on ecological semantic graphs is characterized by comprising the following steps: s1, acquiring sound signals by using a microphone array, and recording acquired time stamps and GPS coordinates; S2, sending the collected sound signals to an edge computing node cluster, preprocessing the sound signals, extracting features based on the processed sound signals, and performing primary detection; S3, based on a preliminary detection result, generating a semantic embedded representation of the ecological acoustic event by using a self-supervision learning model; s4, abstracting semantic embedded representation into semantic nodes, and establishing semantic edges according to space-time association rules among the ecological acoustic events to obtain ecological acoustic semantic graphs; S5, extracting time-space evolution features of short-time scale and long-time scale by using a graph time sequence convolution network based on the ecological acoustic semantic graph, and identifying multi-scale abnormal events including short-time combination abnormality and long-time trend abnormality by matching with a preset illegal activity combination mode; And S6, when an abnormal event is identified, sound source positioning is performed based on the collected sound signals and the GPS coordinates, secondary verification is performed through a weighted least square algorithm, and a three-level response link is triggered to alarm according to the level of the abnormal event.
2. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic map according to claim 1, wherein S2 comprises: s201, an edge computing node cluster receives collected sound signals and GPS coordinates, and environmental noise suppression is carried out on the sound signals by utilizing spectral subtraction to obtain enhanced sound signals; S202, screening out sound signals with signal to noise ratio not lower than a preset signal to noise ratio by adopting a short-time energy threshold screening algorithm based on the enhanced sound signals, extracting 24-dimensional Mel frequency spectrum coefficients of the screened sound signals based on light CNN, and matching the extracted 24-dimensional Mel frequency spectrum coefficients with characteristic templates in a voiceprint library to find out the most similar sound templates to obtain sound types; S203, estimating a time delay difference by adopting a GCC-PHAT algorithm, and estimating a sound source position by using a weighted least square algorithm by combining GPS coordinates of a microphone array and the time delay difference; S204, determining an abnormal event according to the sound type and the estimated sound source position, and triggering alarms of different grades according to the grade of the abnormal event.
3. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic map according to claim 1, wherein S3 comprises: S301, utilizing an edge computing node cluster to segment the sound signal determined by the preliminary detection according to time slices; S302, performing time shift, loudness scaling and enhancement processing of a dynamic time-frequency mask on the basis of the segmented sound signals to construct positive and negative sample pairs, wherein the positive sample pairs are different enhancement versions of the same event, and the negative samples are signals of different events; S303, based on the positive and negative samples, carrying out contrast learning training on the teacher model by taking InfoNCE as a loss function; S304, migrating the weight of the trained teacher model to a light student model through a knowledge distillation technology, and deploying the light student model back to the edge computing node cluster to generate a semantic embedded representation.
4. A method for forest ecological monitoring and illegal activity identification based on ecological semantic graphs according to claim 3, wherein the transferring the weight of the trained teacher model to the light-weight student model comprises: and taking the output probability distribution of the teacher model as a soft label, optimizing the student model through KL divergence loss, inserting a low-rank matrix pair into the attention layer of the student model, freezing the weight of the trunk model, and optimizing only low-rank matrix parameters.
5. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic map according to claim 1, wherein S4 comprises: s401, a cloud gateway receives semantic embedded representations uploaded by an edge computing node cluster, and identifies entities, relations and attributes in an acoustic event by adopting a triplet extraction algorithm combining rule matching and BERT models; s402, abstracting an entity into three types of semantic nodes of species sound, environment sound and illegal activity sound, and establishing weighted semantic edges among related nodes based on co-occurrence association rules, causal association rules and rhythm association rules; S403, constructing an initial ecological acoustic semantic map based on semantic nodes and semantic edges, and performing incremental updating and storage on the initial ecological acoustic semantic map according to multi-granularity time scales of a second level, a minute level, a day level and a season level to obtain the ecological acoustic semantic map.
6. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic map according to claim 1, wherein S5 comprises: S501, inputting an ecological acoustic semantic map into a map time sequence convolution network for analysis, wherein the map time sequence convolution network adopts a double-branch structure of short-time branches and long-term branches, the short-time branches use space-time convolution kernels to mainly analyze rapid changes of second-minute ecological acoustic semantic maps and capture local association features among sudden events, and the long-term branches use a time sequence attention mechanism to aggregate evolution information of the day-season ecological acoustic semantic maps and focus long-term trend changes of key nodes and edges; s502, based on local association features and long-term trend changes among events, higher weight is given to high-risk nodes and associated edges in the ecological acoustic semantic graphs by using an attention mechanism; And S503, performing cosine similarity matching on high-risk nodes and associated edges which are given with higher weight in the ecological acoustic semantic graphs and an illegal activity combination mode library prestored in the cloud, judging that illegal activity combination is abnormal if the matching degree exceeds a preset threshold, and generating early warning.
7. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic map according to claim 6, wherein S5 further comprises: And if the new environment is detected or a new sound source is encountered, starting a meta learning algorithm to learn, and fine-tuning the graph time sequence convolution network by utilizing the new environment or the new sound source.
8. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic patterns according to claim 1, wherein the step of triggering a three-level response link to alarm according to the level of the abnormal event comprises the steps of: Triggering alarms of different grades according to the grades of the abnormal events by the edge computing node cluster, and sending positioning information to the unmanned aerial vehicle dispatching system; the unmanned plane scheduling system plans an optimal route through a path planning algorithm based on the received coordinates, controls the unmanned plane to fly to a target area, starts high-definition shooting and recording to carry out on-site rechecking, and transmits a video stream back to a command center in real time, the cloud platform synthesizes real-time feedback data and historical map information, carries out secondary verification on event confidence coefficient by using a Bayesian algorithm, generates a visual report containing event types, positions, confidence coefficient and recommended measures, and pushes the visual report to management staff.
9. The method for identifying forest ecology monitoring and illegal activity based on ecological semantic patterns according to claim 2 or 8, wherein the triggering of alarms of different levels according to the level of abnormal events comprises: The system comprises a first-stage illegal event triggering strong mode audible and visual alarm and a parallel operation unmanned aerial vehicle dispatching system, and a second-stage ecological anomaly triggering weak mode alarm and a parallel operation unmanned aerial vehicle dispatching system.

Description

Forest ecological monitoring and illegal activity identification method based on ecological semantic graph Technical Field The application relates to the technical field of ecological protection, in particular to a forest ecological monitoring and illegal activity identification method based on ecological semantic graphs. Background Forest ecological monitoring and illegal activity prevention and control are core requirements of ecological protection, and acoustic monitoring is a mainstream technical means because the acoustic monitoring is not limited by illumination and terrain. While the prior improved technology can identify single species or mechanical sound, but lacks the analysis capability of a multi-sound event combination mode, the prior improved technology has the defects of difficulty in distinguishing natural disturbance from illegal activities (such as being incapable of identifying the piracy mode of electric saw sound and bird escape) and high false alarm rate. Meanwhile, the prior art relies on a manual labeling training model, the adaptation to a new environment is weak, short-time sudden anomalies can be detected only, and long-term ecological evolution trend can not be captured, so that illegal activity recognition accuracy is low, and early warning is delayed. In addition, the voiceprint recognition monitoring equipment based on supervised learning in the prior art adopts a supervised convolutional neural network to train a voiceprint recognition model, 10 types of acoustic event data such as bird singing, electric saws, gunsounds and the like are required to be marked manually, an acoustic signal is acquired through a single-node acoustic acquisition device, a single target acoustic event is recognized by an input model, only a single event abnormal threshold value is set, no multi-event combined analysis is performed, an abnormal alarm is only notified to a manager through a short message, no positioning and unmanned aerial vehicle linkage function is performed, the technology can only recognize a single event, the combined mode of multiple types of acoustic events cannot be analyzed in an associated mode, the accuracy rate is low, the adaptation cost is high, the position of a sound source cannot be positioned accurately when a new sound source (such as novel electric saw sound and trap touch sound) is encountered, no local acoustic alarm or unmanned aerial vehicle linkage is performed, and the illegal activity interception success rate is low. Disclosure of Invention Aiming at the defects in the prior art, the forest ecological monitoring and illegal activity identification method based on the ecological semantic map solves the problems that the prior art lacks semantic association and combination analysis, the model relies on manual labeling and generalization capability is poor, the monitoring scale is single, the ecological trend cannot be perceived, the response chain is disjointed, the prevention and control timeliness is low and the like. In order to achieve the aim of the application, the application adopts the following technical scheme: The application provides a forest ecological monitoring and illegal activity identification method based on ecological semantic graphs, which comprises the following steps: s1, acquiring sound signals by using a microphone array, and recording acquired time stamps and GPS coordinates; S2, sending the collected sound signals to an edge computing node cluster, preprocessing the sound signals, extracting features based on the processed sound signals, and performing primary detection; S3, based on a preliminary detection result, generating a semantic embedded representation of the ecological acoustic event by using a self-supervision learning model; s4, abstracting semantic embedded representation into semantic nodes, and establishing semantic edges according to space-time association rules among the ecological acoustic events to obtain ecological acoustic semantic graphs; S5, extracting time-space evolution features of short-time scale and long-time scale by using a graph time sequence convolution network based on the ecological acoustic semantic graph, and identifying multi-scale abnormal events including short-time combination abnormality and long-time trend abnormality by matching with a preset illegal activity combination mode; And S6, when an abnormal event is identified, sound source positioning is performed based on the collected sound signals and the GPS coordinates, secondary verification is performed through a weighted least square algorithm, and a three-level response link is triggered to alarm according to the level of the abnormal event. Further, the step S2 includes: s201, an edge computing node cluster receives collected sound signals and GPS coordinates, and environmental noise suppression is carried out on the sound signals by utilizing spectral subtraction to obtain enhanced sound signals; S202, screening out sound signals wit