CN-121982664-A - Anomaly monitoring method and system based on traffic related data mining

CN121982664ACN 121982664 ACN121982664 ACN 121982664ACN-121982664-A

Abstract

The invention provides an anomaly monitoring method and system based on traffic related data mining, and relates to the technical field of data processing. In the method, firstly, multidimensional traffic related data is collected, wherein the multidimensional traffic related data comprises at least one of traffic flow data, traffic speed data and traffic climate data and does not comprise traffic image data, secondly, deep semantic mining is conducted on the traffic related data of each dimension in the multidimensional traffic related data to obtain traffic related semantic vectors of each dimension, the deep semantic mining comprises semantic mining of at least two domains and semantic enhancement of semantic mining results, then cross-dimensional fusion is conducted on the traffic related semantic vectors of the dimensions to form traffic fusion semantic vectors, and finally, semantic restoration is conducted on the traffic fusion semantic vectors to form target traffic abnormal data. Based on the method, the problem that the reliability of anomaly monitoring is relatively low in the prior art can be solved.

Inventors

DENG XIANG
YANG DA
LUO TAO
LUO LIN
GUO YING
LIU NA
WANG KANG

Assignees

巴中数据集团有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. An anomaly monitoring method based on traffic-related data mining, comprising the steps of: Collecting multi-dimensional traffic related data of a target area, wherein the multi-dimensional traffic related data comprises at least one of traffic flow data, traffic speed data and traffic climate data, and the multi-dimensional traffic related data does not comprise traffic image data; Deep semantic mining is carried out on the traffic related data of each dimension in the multi-dimensional traffic related data to obtain traffic related semantic vectors of each dimension, wherein the deep semantic mining comprises semantic mining of at least two domains and semantic enhancement of semantic mining results, the semantic enhancement comprises multiple local semantic information based on the semantic mining results of a first domain, and multiple cross-domain semantic enhancement is carried out on the semantic mining results of a second domain; performing cross-dimension fusion on the traffic related semantic vectors with multiple dimensions to form traffic fusion semantic vectors; and carrying out semantic reduction on the traffic fusion semantic vector to form target traffic abnormal data, wherein the target traffic abnormal data is used for representing whether traffic in the target area is abnormal or not, and the abnormal condition at least comprises traffic accidents.
2. The anomaly monitoring method based on traffic-related data mining according to claim 1, wherein the step of performing deep semantic mining on traffic-related data of each dimension in the multi-dimensional traffic-related data to obtain traffic-related semantic vectors of each dimension comprises: Performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic related data to form a traffic flow spectrogram, and performing deep semantic mining on the traffic flow data and the traffic flow spectrogram to obtain traffic related semantic vectors of flow dimensions, wherein the traffic flow data belongs to the time domain and comprises flows of a plurality of time points; performing time domain and frequency domain conversion on traffic speed data included in the multidimensional traffic related data to form a traffic speed spectrogram, and performing deep semantic mining on the traffic speed data and the traffic speed spectrogram to obtain traffic related semantic vectors of a vehicle speed dimension, wherein the traffic speed data belongs to the time domain and comprises vehicle speeds of a plurality of time points; And carrying out word embedding processing on traffic climate data included in the multidimensional traffic related data, and carrying out deep semantic mining on the obtained word embedding characteristics to form traffic related semantic vectors of climate dimensions, wherein the traffic climate data belongs to text data.
3. The anomaly monitoring method based on traffic-related data mining according to claim 2, wherein the step of performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic-related data to form a traffic flow spectrogram, and performing deep semantic mining on the traffic flow data and the traffic flow spectrogram to obtain traffic-related semantic vectors of a flow dimension comprises: Performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic related data to form a traffic flow spectrogram, and performing convolution processing on the traffic flow data and the traffic flow spectrogram to form a traffic flow time domain vector and a traffic flow frequency domain vector; mapping the traffic flow frequency domain vector into a plurality of traffic flow local vectors; Performing multiple cross-domain semantic enhancement based on the multiple traffic flow local vectors and the traffic flow time domain vector to form a flow cross-domain enhancement vector; And obtaining traffic related semantic vectors of the traffic dimension based on the traffic cross-domain enhancement vector.
4. The anomaly monitoring method based on traffic-related data mining of claim 3, wherein the step of mapping the traffic flow frequency domain vector into a plurality of traffic flow local vectors comprises: dividing the traffic flow frequency domain vector to form a plurality of flow frequency domain divided vectors, and performing transposition convolution or anti-pooling processing on each flow frequency domain divided vector to obtain each corresponding flow frequency domain expansion vector, wherein the size of each flow frequency domain expansion vector is the same as the size of the traffic flow frequency domain vector; Acquiring a plurality of reference mapping matrixes formed by learning and training, and determining a target mapping matrix from the plurality of reference mapping matrixes based on peak positions in the traffic flow spectrogram; And based on the target mapping matrix, respectively carrying out linear mapping on each flow frequency domain expansion vector to form a plurality of traffic flow local vectors.
5. The anomaly monitoring method based on traffic-related data mining of claim 3, wherein the step of mapping the traffic flow frequency domain vector into a plurality of traffic flow local vectors comprises: based on a plurality of predetermined parameter combination logics, respectively combining the distribution relation of vector parameters in the traffic flow frequency domain vector to form a plurality of traffic flow combination vectors; Aiming at each traffic flow combination vector, carrying out related semantic mining on the traffic flow frequency domain vector based on the traffic flow combination vector to obtain a corresponding traffic flow related vector, and determining a corresponding traffic flow local vector based on the traffic flow related vector; And carrying out related semantic mining on the traffic flow frequency domain vector based on the traffic flow frequency domain vector to obtain a corresponding traffic flow related vector, and determining a corresponding traffic flow local vector based on the traffic flow related vector.
6. The traffic-related data mining-based anomaly monitoring method of claim 3, wherein the cross-domain semantic enhancement is implemented by a traffic anomaly analysis network, the traffic anomaly analysis network comprising a deep semantic mining unit and a cross-domain semantic enhancement unit, the deep semantic mining unit comprising a plurality of deep semantic mining layers, each of the deep semantic mining layers corresponding to one of the traffic local vectors, the step of forming a traffic cross-domain enhancement vector by performing multiple cross-domain semantic enhancement based on the plurality of traffic local vectors and the traffic time domain vector, comprising: Determining an mth traffic flow depth vector through an mth depth semantic mining layer in the depth semantic mining unit, wherein when m=1, the corresponding traffic flow depth vector is the traffic flow time domain vector, and when 1<m is less than or equal to M, performing depth semantic mining on an mth-1 traffic stage enhancement vector to obtain the mth traffic flow depth vector; Through the cross-domain semantic enhancement unit, the mth traffic flow depth vector is subjected to cross-domain semantic enhancement based on the mth traffic flow local vector corresponding to the mth depth semantic mining layer to form an mth traffic phase enhancement vector; and obtaining a traffic cross-domain enhancement vector based on the M-th traffic stage enhancement vector.
7. The anomaly monitoring method based on traffic-related data mining according to claim 6, wherein the step of performing cross-domain semantic enhancement on the mth traffic flow depth vector based on the mth traffic flow local vector corresponding to the mth depth semantic mining layer by the cross-domain semantic enhancement unit to form an mth traffic phase enhancement vector comprises: Projecting an mth traffic flow local vector corresponding to the mth depth semantic mining layer through a plurality of linear projection matrixes in the cross-domain semantic enhancement unit to form a plurality of traffic flow projection vectors, and performing activation processing on each traffic flow projection vector to form a plurality of traffic flow activation distribution, wherein each parameter in each traffic flow activation distribution belongs to 0-1 and is used for representing the importance of a corresponding vector position; determining semantic enhancement capability indexes corresponding to the traffic flow activation distribution according to the distribution condition of each activation parameter in the traffic flow activation distribution aiming at each traffic flow activation distribution; Determining a target parameter distribution based on the traffic flow activation distributions and the corresponding semantic enhancement capability indexes, and performing parameter adjustment on the mth traffic flow depth vector based on the target parameter distribution to obtain an mth flow stage enhancement vector.
8. The anomaly monitoring method based on traffic-related data mining according to claim 2, wherein the step of performing time domain and frequency domain conversion on traffic speed data included in the multidimensional traffic-related data to form a traffic speed spectrogram, and performing deep semantic mining on the traffic speed data and the traffic speed spectrogram to obtain traffic-related semantic vectors of a vehicle speed dimension comprises: Performing time domain and frequency domain conversion on traffic speed data included in the multidimensional traffic related data to form a traffic speed spectrogram, and performing convolution processing on the traffic speed data and the traffic speed spectrogram respectively to form a traffic speed time domain vector and a traffic speed frequency domain vector; Mapping the traffic speed frequency domain vector into a plurality of traffic speed local vectors; Performing multiple cross-domain semantic enhancement based on the multiple traffic speed local vectors and the traffic speed time domain vector to form a speed cross-domain enhancement vector; and obtaining traffic related semantic vectors of the vehicle speed dimension based on the vehicle speed cross-domain enhancement vector.
9. The anomaly monitoring method based on traffic-related data mining according to any one of claims 1 to 8, wherein the step of cross-dimensionally fusing the traffic-related semantic vectors of multiple dimensions to form a traffic fusion semantic vector comprises: Based on a cross attention mechanism, fusing the traffic related semantic vectors in multiple dimensions in pairs to form multiple traffic attention vectors; And splicing the plurality of traffic attention vectors to form traffic spliced vectors, and carrying out rolling and pooling processing on the traffic spliced vectors to form traffic fusion semantic vectors.
10. An anomaly monitoring system based on traffic-related data mining, comprising a processor and a memory, the memory for storing a computer program, the processor for executing the computer program to implement the anomaly monitoring method based on traffic-related data mining of any one of claims 1-9.

Description

Anomaly monitoring method and system based on traffic related data mining Technical Field The invention relates to the technical field of data processing, in particular to an anomaly monitoring method and system based on traffic related data mining. Background In monitoring traffic anomalies such as traffic accidents, images or videos are generally collected based on traffic cameras installed on roads, and then a rear end performs analysis to determine whether traffic anomalies exist, if so, whether traffic accidents exist. However, in the prior art, the problems of cost, visibility and the like are limited, and the monitoring based on the traffic camera is difficult to realize the whole coverage of the road, so that reliable abnormal monitoring is difficult to perform on the area without the traffic camera or the area where the camera is difficult to effectively image, and thus the processing such as timely intervention or early warning cannot be performed. That is, the prior art has a problem that the reliability of anomaly monitoring is relatively low. Disclosure of Invention In view of the above, the present invention is directed to an anomaly monitoring method and system based on traffic-related data mining, so as to solve the problem of relatively low reliability of anomaly monitoring in the prior art. In order to achieve the above purpose, the embodiment of the present invention adopts the following technical scheme: An anomaly monitoring method based on traffic-related data mining, comprising: Collecting multi-dimensional traffic related data of a target area, wherein the multi-dimensional traffic related data comprises at least one of traffic flow data, traffic speed data and traffic climate data, and the multi-dimensional traffic related data does not comprise traffic image data; Deep semantic mining is carried out on the traffic related data of each dimension in the multi-dimensional traffic related data to obtain traffic related semantic vectors of each dimension, wherein the deep semantic mining comprises semantic mining of at least two domains and semantic enhancement of semantic mining results, the semantic enhancement comprises multiple local semantic information based on the semantic mining results of a first domain, and multiple cross-domain semantic enhancement is carried out on the semantic mining results of a second domain; performing cross-dimension fusion on the traffic related semantic vectors with multiple dimensions to form traffic fusion semantic vectors; and carrying out semantic reduction on the traffic fusion semantic vector to form target traffic abnormal data, wherein the target traffic abnormal data is used for representing whether traffic in the target area is abnormal or not, and the abnormal condition at least comprises traffic accidents. In some preferred embodiments, in the anomaly monitoring method based on traffic-related data mining, the step of performing deep semantic mining on traffic-related data of each dimension in the multi-dimensional traffic-related data to obtain traffic-related semantic vectors of each dimension includes: Performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic related data to form a traffic flow spectrogram, and performing deep semantic mining on the traffic flow data and the traffic flow spectrogram to obtain traffic related semantic vectors of flow dimensions, wherein the traffic flow data belongs to the time domain and comprises flows of a plurality of time points; performing time domain and frequency domain conversion on traffic speed data included in the multidimensional traffic related data to form a traffic speed spectrogram, and performing deep semantic mining on the traffic speed data and the traffic speed spectrogram to obtain traffic related semantic vectors of a vehicle speed dimension, wherein the traffic speed data belongs to the time domain and comprises vehicle speeds of a plurality of time points; And carrying out word embedding processing on traffic climate data included in the multidimensional traffic related data, and carrying out deep semantic mining on the obtained word embedding characteristics to form traffic related semantic vectors of climate dimensions, wherein the traffic climate data belongs to text data. In some preferred embodiments, in the anomaly monitoring method based on traffic-related data mining, the step of performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic-related data to form a traffic flow spectrogram, and performing deep semantic mining on the traffic flow data and the traffic flow spectrogram to obtain traffic-related semantic vectors of a flow dimension includes: Performing time domain and frequency domain conversion on traffic flow data included in the multidimensional traffic related data to form a traffic flow spectrogram, and performing convolution processing on the