CN116070174A - Multi-category target recognition method and system - Google Patents

Abstract

The invention discloses a multi-category target identification method and a system, wherein the method comprises the following steps: step S1: pretreatment: preprocessing the obtained original data; step S2: framing: dividing the original data into a long-time period frame and a short-time period frame; step S3: target identification: respectively extracting features of the two frames in the step S2, and inputting the features into a corresponding neural network for target recognition; step S4: fusion: and according to the signal characteristics, fusing the two neural network results, and outputting the fused result as a sensor identification result. The system is used for implementing the method. The invention has the advantages of simple principle, wide application range, high recognition precision, capability of greatly reducing hardware cost and the like.

Description

Multi-category target recognition method and system

Technical Field

The invention mainly relates to the technical field of security and protection or monitoring, in particular to a multi-class target identification method and system, which are particularly suitable for processing long-short period vibration signals in the technical field of security and protection such as unattended operation.

Background

In many industries and fields where unmanned needs are required, unmanned monitoring, such as identifying vibration signals of pedestrians, vehicles, excavation, blasting, etc., is required according to practical application requirements. In a conventional unattended sensor system, in order to sense the surrounding environment of the sensor, a conventional machine learning method or a deep neural network is adopted to identify an environment signal. However, regardless of the classifier, when the signal to be identified has both long-period and short-period signals, it is difficult to achieve both real-time performance and accuracy with a single classifier.

In specific application, in general, signals of pedestrians and vehicles belong to short periodic signals, and a plurality of signals can be obtained only in a few seconds; the time interval period of the opposite digging signal is much longer and varies from 1s to 5 s. A single classifier does not work well when processing signals of two types with widely differing time periods at the same time. The short signal segment is used for extracting features for identification, the system real-time performance is good, but the accuracy of the long-period signal is poor; the long signal segment is used for extracting the characteristics for identification, the accuracy rate is better for long-period signals, and the system delay is large.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: aiming at the technical problems existing in the prior art, the invention provides the multi-category target identification method and the system which are simple in principle, wide in application range, high in identification precision and capable of greatly reducing hardware cost.

In order to solve the technical problems, the invention adopts the following technical scheme:

a multi-category target recognition method, comprising:

step S1: pretreatment: preprocessing the obtained original data;

step S2: framing: dividing the original data into a long-time period frame and a short-time period frame;

step S3: target identification: respectively extracting features of the two frames in the step S2, and inputting the features into a corresponding neural network for target recognition;

step S4: fusion: and according to the signal characteristics, fusing the two neural network results, and outputting the fused result as a sensor identification result.

As a further improvement of the process of the invention: in step S1, the filtering process for the obtained raw data includes:

step S101: according to the fact that the target signal is a low-frequency signal generally, the noise is a high-frequency signal, low-pass filtering is carried out on the signal, the target signal is reserved, and the high-frequency noise is filtered;

step S102: and carrying out notch filtering or comb filtering on the signals to filter out power frequency and harmonic interference thereof.

As a further improvement of the process of the invention: the step S1 further includes a step S103: when low-frequency interference exists, the band-pass filter is used for band-pass filtering, so that the influence of the interference is reduced.

As a further improvement of the process of the invention: the process of framing in step S2 includes:

step S201: framing the filtered data;

step S202: according to the characteristics of the short period signal and the real-time requirement, the short period frame adopts a signal segment of 2-5s or less than 2s as one frame;

step S203: according to the characteristics of the long-period signal and the requirements on accuracy, the long-period frame adopts a signal segment of 10-20s or more than 20s as one frame.

As a further improvement of the process of the invention: the process of extracting the signal characteristics and identifying in the step S3 includes:

step S301: for short period frames, extracting features using details and multiple dimensions; the output of the classifier is also the corresponding class of the short-period signal;

step S302: for long period frames, only main features are extracted; the classifier output is also the long period signal corresponding class.

As a further improvement of the process of the invention: the process of merging and outputting the identification result in the step S4 comprises the following steps:

for short period signals, using the classification result of the short period frames;

for long period signals, using the classification result of the long period frames;

when the two recognition results are inconsistent, the recognition result of the long period frame is used as output.

The present invention further provides a multi-category target recognition system comprising:

the preprocessing unit is used for preprocessing the acquired original data;

the frame dividing unit is used for dividing the original data into a long-time period frame and a short-time period frame;

the target recognition unit is used for extracting the characteristics of the two frames respectively and inputting the extracted characteristics into the corresponding neural network for target recognition;

and the fusion unit is used for fusing the two neural network results according to the signal characteristics and outputting the fusion result as a sensor identification result.

As a further improvement of the system of the invention: the preprocessing unit is used for filtering the signal in a low-pass mode according to the fact that the target signal is a low-frequency signal and the noise is a high-frequency signal, retaining the target signal and filtering the high-frequency noise; and carrying out notch filtering or comb filtering on the signals to filter out power frequency and harmonic interference thereof.

As a further improvement of the system of the invention: the target recognition unit extracts characteristics of short period frames by using details and multiple dimensions; the output of the classifier is also the corresponding class of the short-period signal; for long period frames, only main features are extracted; the classifier output is also the long period signal corresponding class.

As a further improvement of the system of the invention: the fusion unit uses the classification result of the short period frame for the short period signal; for long period signals, using the classification result of the long period frames; when the two recognition results are inconsistent, the recognition result of the long period frame is used as output.

Compared with the prior art, the invention has the advantages that:

the invention discloses a multi-category target identification method and a system, which are used for processing long-short period vibration signals, have the advantages of simple principle, wide application range and high identification precision, and can greatly reduce hardware cost. The invention performs differentiation processing according to the characteristics of long and short periodic signals. A longer time window is used for long period signals and thus a higher accuracy can be obtained. For short periodic signals, better real-time performance can be obtained by using a shorter observation time window. And through the fusion of the results of the two classifiers, the identification of the long and short periodic signals can be considered.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

Fig. 2 is a schematic diagram of training set classification results when 4s is used as a frame in a specific application example of the present invention.

FIG. 3 is a schematic diagram of the test set classification result when 4s is used as one frame in the embodiment of the present invention.

Fig. 4 is a schematic diagram of training set classification results when 10s is used as one frame in a specific application example of the present invention.

FIG. 5 is a schematic diagram of the test set classification result when 10s is used as one frame in the embodiment of the present invention.

Detailed Description

The invention will be described in further detail with reference to the drawings and the specific examples.

As shown in fig. 1, the method for identifying multi-category targets according to the present invention is a method for identifying multi-category targets for processing vibration signals with long and short periods, comprising the steps of:

step S1: pretreatment: preprocessing the obtained original data;

the manner in which raw data is acquired may generally be through a sensor or other monitoring unit, such as an unattended sensor apparatus.

Step S2: framing: dividing the original data into a long-time period frame and a short-time period frame;

in specific application, framing can be performed according to the processing capacity of the microcontroller according to the requirements of practical application.

Step S3: target identification: respectively extracting features of the two frames in the step S2, and inputting the features into a corresponding neural network for target recognition;

step S4: fusion: and according to the signal characteristics, fusing the two neural network results, and outputting the fused result as a sensor identification result.

In a specific application example, in step S1, a specific process of filtering the obtained raw data may include:

step S101: according to the fact that the target signal is a low-frequency signal generally, the noise is a high-frequency signal, low-pass filtering is carried out on the signal, the target signal is reserved as far as possible, and the high-frequency noise is filtered;

step S102: carrying out notch filtering or comb filtering on the signals to filter out power frequency and harmonic interference thereof;

step S103: when low-frequency interference exists, the band-pass filter can be used for band-pass filtering, so that the influence of the interference is reduced as much as possible.

In a specific application example, the specific process of framing in step S2 may include:

step S201: framing the filtered data;

step S202: according to the characteristics of the short period signal and the real-time requirement, the short period frame can adopt a signal segment (less than 2 s) with the length of 2-5s or less as one frame;

step S203: depending on the long period signal characteristics and accuracy requirements, long period frames may employ signal segments (greater than 20 s) of 10-20s or longer as a frame.

In a specific application example, in step S3, the specific process of extracting signal features and identifying may include:

step S301: for a short period frame, as the data quantity of the original signal segment is less and the operation quantity is relatively less, the feature extraction can use richer details and multiple dimensions to extract the features so as to improve the recognition performance of the algorithm; the output of the classifier is also a short period signal corresponding class, such as pedestrians, vehicles.

Step S302: for long period frames, such as mining, because the data volume of the original signal segment is larger, details need to be ignored in order to meet the operation requirement of the embedded system, main features are extracted as far as possible, and the operation volume is reduced; the classifier output is also a long period signal corresponding class, such as mining.

In a specific application example, the specific process of identifying the result fusion output in step S4 may include:

for short period signals, the classification result of the short period frames is preferentially used.

The classification result of the long period frame is preferentially used. In the test process, the recognition result of the long period frame is found to have better performance, so that when the recognition results are inconsistent, the recognition result of the long period frame is used as output.

Referring to fig. 2-5, the training performance of selecting 4s and 10s frame lengths is compared, taking the recognition of pedestrians and excavation signals as an example. It can be seen that for short periodic signals, increasing from 4s to 10s, the test set F1 improves recognition performance by about 1%. For long period signals, the test set F1 improves recognition performance by about 3%. Therefore, considering both real-time performance and recognition performance, short-period signals are suitable for using shorter frames, and long-period signals are suitable for using longer frames. The invention performs differentiation processing according to the characteristics of long and short periodic signals. A longer time window is used for long period signals and thus a higher accuracy can be obtained. For short periodic signals, better real-time performance can be obtained by using a shorter observation time window. And through the fusion of the results of the two classifiers, the identification of the long and short periodic signals can be considered.

The method of the invention is realized on a common singlechip with FPU, has simple, convenient and reliable algorithm and high accuracy, can detect pedestrians, vehicles, excavation and blasting in real time, has been applied and verified in a plurality of industries, and has great practical value.

The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the invention without departing from the principles thereof are intended to be within the scope of the invention as set forth in the following claims.

Claims (10)

1. A method for identifying multiple categories of objects, comprising:

step S1: pretreatment: preprocessing the obtained original data;

step S2: framing: dividing the original data into a long-time period frame and a short-time period frame;

step S3: target identification: respectively extracting features of the two frames in the step S2, and inputting the features into a corresponding neural network for target recognition;

step S4: fusion: and according to the signal characteristics, fusing the two neural network results, and outputting the fused result as a sensor identification result.

2. The multi-class object recognition method according to claim 1, wherein in step S1, the process of filtering the acquired raw data includes:

step S101: according to the fact that the target signal is a low-frequency signal generally, the noise is a high-frequency signal, low-pass filtering is carried out on the signal, the target signal is reserved, and the high-frequency noise is filtered;

step S102: and carrying out notch filtering or comb filtering on the signals to filter out power frequency and harmonic interference thereof.

3. The multi-class object recognition method according to claim 2, wherein the step S1 further comprises a step S103: when low-frequency interference exists, the band-pass filter is used for band-pass filtering, so that the influence of the interference is reduced.

4. A multi-category object recognition method according to any one of claims 1 to 3, wherein the framing procedure in step S2 includes:

step S201: framing the filtered data;

step S202: according to the characteristics of the short period signal and the real-time requirement, the short period frame adopts a signal segment of 2-5s or less than 2s as one frame;

step S203: according to the characteristics of the long-period signal and the requirements on accuracy, the long-period frame adopts a signal segment of 10-20s or more than 20s as one frame.

5. A multi-class object recognition method according to any one of claims 1-3, wherein the step S3 of extracting and recognizing signal features includes:

step S301: for short period frames, extracting features using details and multiple dimensions; the output of the classifier is also the corresponding class of the short-period signal;

step S302: for long period frames, only main features are extracted; the classifier output is also the long period signal corresponding class.

6. A multi-category target recognition method according to any one of claims 1 to 3, wherein the process of merging and outputting the recognition result in step S4 includes:

for short period signals, using the classification result of the short period frames;

for long period signals, using the classification result of the long period frames;

when the two recognition results are inconsistent, the recognition result of the long period frame is used as output.

7. A multi-category target recognition system, comprising:

the preprocessing unit is used for preprocessing the acquired original data;

the frame dividing unit is used for dividing the original data into a long-time period frame and a short-time period frame;

the target recognition unit is used for extracting the characteristics of the two frames respectively and inputting the extracted characteristics into the corresponding neural network for target recognition;

and the fusion unit is used for fusing the two neural network results according to the signal characteristics and outputting the fusion result as a sensor identification result.

8. The multi-class object recognition system of claim 7, wherein the preprocessing unit performs low-pass filtering on the signal according to the fact that the object signal is a low-frequency signal and the noise is a high-frequency signal, retains the object signal, and filters out the high-frequency noise; and carrying out notch filtering or comb filtering on the signals to filter out power frequency and harmonic interference thereof.

9. The multi-class object recognition system according to claim 7 or 8, wherein the object recognition unit extracts features using details and multiple dimensions for short period frames; the output of the classifier is also the corresponding class of the short-period signal; for long period frames, only main features are extracted; the classifier output is also the long period signal corresponding class.

10. The multi-class object recognition system according to claim 7 or 8, wherein the fusion unit uses a classification result of a short period frame for a short period signal; for long period signals, using the classification result of the long period frames; when the two recognition results are inconsistent, the recognition result of the long period frame is used as output.

Images