CN114821025A - A method and system for meter recognition based on deep learning - Google Patents

Abstract

The invention discloses a meter identification method and system based on deep learning, which relate to the technical field of image identification and comprise an image analysis module, an information recording module, an information sorting module and a signal verification module; the image analysis module is used for carrying out digital target recognition on the received meter image by adopting a yolov3 detection model to obtain a digital recognition result and finally obtain a meter reading; the information recording module is used for recording the identification record of the image analysis module; the information arrangement module is used for arranging the identification records, constructing a parameter detection training sample, obtaining a parameter compensation model and feeding the parameter compensation model back to the model construction module to correct the yolov3 detection model, so that the identification accuracy is improved; the signal verification module is used for verifying the communication state of the image analysis module in real time, and after a communication early warning instruction is detected, the image analysis module enters an active standby mode, so that the influence of interference signals is effectively reduced, and the identification efficiency and accuracy are improved.

Description

A method and system for meter recognition based on deep learning

technical field

The invention relates to the technical field of image recognition, in particular to a method and system for meter recognition based on deep learning.

Background technique

In the power grid operation and maintenance inspection, due to its strong anti-electromagnetic interference ability, high precision and low price, the pointer meter is still the main measuring instrument for industrial production for a long time; because the pointer meter cannot output digital signals , At present, the reading of the meter is checked manually, which requires the operator to stay in the substation all the time or often, which increases the operator's workload and cannot achieve the effect of unmanned inspection; and it is inconvenient to install meters in high temperature and high pressure environments. Observed;

With the rapid development of robotics technology, inspection robots can replace manual labor, capture meter reading images through cameras or infrared thermal imagers, and finally perform image processing on the acquired images; however, the process of image recognition is often affected by the external environment. It has a great impact on the recognition accuracy; based on the above shortcomings, the present invention proposes a meter recognition method and system based on deep learning.

SUMMARY OF THE INVENTION

The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method and system for meter identification based on deep learning.

In order to achieve the above object, according to the embodiment of the first aspect of the present invention, a deep learning-based meter identification system is proposed, including an image acquisition module, a model construction module, an information recording module, a parameter compensation module, and a signal verification module;

The image acquisition module is used to collect meter images, and transmit the collected meter images to the image analysis module for learning and identification; the image analysis module is connected to the model building module, and is used to obtain the pair of images constructed by the model building module. yolov3 detection model for image training and recognition;

Described image analysis module is used for adopting the yolov3 detection model to carry out digital target recognition to the received meter image to obtain the digital recognition result, and finally obtain the meter reading;

The information recording module is used for recording the identification records of the image analysis module and transmitting the identification records to the information sorting module; the information sorting module is used for sorting the identification records, constructing a training sample for parameter detection, and the training is based on a machine learning method to obtain A parameter compensation model; the parameter compensation module modifies the yolov3 detection model after receiving the parameter compensation model;

The signal verification module is used to verify the communication status of the image analysis module in real time, and calculate the interference coefficient Cy; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication status is abnormal, and a communication warning command is generated to remind the management personnel to deal with it as soon as possible.

Further, the construction process of the yolov3 detection model is as follows:

Collect meter images to construct the data set required for training; the specific performance is as follows: collect n meter images at different shooting angles and under different lighting conditions, label the meter images, and obtain the data set required for subsequent training of the yolov3 model; It is divided into training set, test set and verification set according to the set ratio; the set ratio includes 2:1:1, 3:1:1 and 4:3:1;

Construct a fusion model: The fusion model is a model constructed by at least two fusion methods among support vector machines, deep convolutional neural networks and RBF neural networks. The fusion methods include linear weighted fusion method, cross fusion method, waterfall fusion method, and feature fusion. method and prediction fusion method;

After the training set, test set and verification set are normalized, the fusion model is trained, tested and verified, and the trained fusion model is marked as the yolov3 detection model.

Further, the specific correction steps of the parameter compensation module are:

SS1: Obtain the identification record of the image analysis module at the current moment, input the currently identified meter reading, actual meter reading and corresponding environmental parameter values into the parameter compensation model to obtain the attribute parameter compensation coefficient;

SS2: Modify the yolov3 detection model according to the attribute parameter compensation coefficient.

Further, the specific verification steps of the signal verification module are:

The signal verification module sends a verification configuration message to the FPGA main control of the image analysis module according to a preset verification period, wherein the verification configuration message includes a first signal quality threshold;

In response to receiving the verification configuration message, the FPGA master sends the second synchronization signal to the signal verification module; the signal verification module determines the signal quality of the second synchronization signal, and compares it with the first signal quality threshold to obtain the corresponding poor quality Value Z1; set the response time to XT;

The signal loss index SH is calculated by using the formula SH=Z1×a1+XT×a2, where a1 and a2 are both coefficient factors; the interference coefficient Cy is evaluated according to the change trend of the signal loss index SH.

Further, the specific evaluation process of the interference coefficient Cy is as follows:

Establish a curve graph of the signal loss index SH versus time, and compare the signal loss index SH with the loss threshold; if SH ≥ the loss threshold, intercept the corresponding curve segment in the corresponding curve and mark it as yellow, and record it as the interference curve part;

In the preset time period, the number of statistical interference curve segments is L1, and all interference curve segments are integrated with time to obtain the interference reference energy L2; the interference coefficient Cy is calculated by using Cy=L1×a3+L2×a4, where a3 , a4 are coefficient factors.

Further, the identification record includes the meter reading, the actual meter reading and the corresponding environmental parameter values obtained by the image analysis module each time; the environmental parameter values include temperature, humidity, wind pressure, wind speed and interfere with the signal.

Further, after the image analysis module detects the communication warning command, it enters the active standby mode, that is, the FPGA main control with abnormal communication is no longer used to identify the meter image; after the signal verification module judges that the communication state is normal, continue to the second step. communication between them.

Further, a method for meter identification based on deep learning, comprising the following steps:

Step 1: Build a yolov3 detection model that can train and identify images through the model building module, and transfer the constructed yolov3 detection model to the image analysis module;

Step 2: use the yolov3 detection model to perform digital target recognition on the meter image collected by the image acquisition module through the image analysis module to obtain a digital recognition result, and finally obtain the meter reading;

Step 3: record the identification records of the image analysis module through the information recording module, call the information sorting module to sort the identification records, and construct a training sample for parameter detection; the training is based on a machine learning method to obtain a parameter compensation model;

Step 4: The parameter compensation model is fed back to the model building module through the parameter compensation module to correct the yolov3 detection model;

Step 5: Verify the communication status of the image analysis module in real time through the signal verification module, and calculate the interference coefficient Cy; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication status is abnormal, and a communication warning command is generated to remind the management personnel to deal with it as soon as possible;

Step 6: After the image analysis module detects the communication warning command, it enters the active standby mode, and after the signal verification module judges that the communication state is normal, the communication between the two is continued.

Compared with the prior art, the beneficial effects of the present invention are:

1. In the present invention, the camera ball in the image acquisition module can be rotated in the housing to collect meter images from different shooting angles, providing solid data support for subsequent image recognition. The image analysis module adopts the yolov3 detection model to collect meter images. The image is analyzed and the reading of the meter is identified, so as to achieve the purpose of unmanned inspection and effectively improve the identification efficiency;

2. In the present invention, the information recording module is used for recording the identification records of the image analysis module; the information sorting module is used for sorting the identification records, constructing a training sample for parameter detection, and the training is based on a machine learning method to obtain a parameter compensation model; the parameter compensation module receives After the parameter compensation model is reached, the yolov3 detection model is revised, so that the yolov3 detection model is continuously revised and improved, and the accuracy of recognition is continuously improved. Sex is guaranteed;

3. The signal verification module in the present invention is used to verify the communication status of the image analysis module in real time. First, a verification configuration message is sent to the FPGA main control of the image analysis module according to the preset verification period, and the corresponding quality difference value Z1 and the response time XT are calculated. Obtain the signal loss index SH; calculate the interference coefficient Cy according to the change trend of the signal loss index SH. If Cy ≥ the interference threshold, it is judged that the signal interference is serious and the communication state is abnormal, and a communication warning command is generated to prompt the management personnel to deal with it as soon as possible; The influence of the interference signal, thereby improving the recognition efficiency and accuracy of the image analysis module.

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative efforts.

FIG. 1 is a system block diagram of a deep learning-based meter identification system of the present invention.

FIG. 2 is a flowchart of a deep learning-based meter identification method of the present invention.

Detailed ways

The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

As shown in Figures 1 to 2, a deep learning-based meter recognition system includes an image acquisition module, an image analysis module, a model building module, a controller, a display module, an information recording module, an information sorting module, and a parameter compensation module , signal verification module and alarm module;

The image acquisition module is used to collect meter images, and transmit the collected meter images to the image analysis module for learning and identification;

The image analysis module is connected with the model building module, and the model building module is used to build a yolov3 detection model that can train and identify images. The yolov3 detection model is used to analyze the collected meter images and identify the meter readings; The construction process of the yolov3 detection model is as follows:

Collect meter images to construct the data set required for training; the specific performance is as follows: collect n meter images at different shooting angles and under different lighting conditions, label the meter images, and obtain the data set required for subsequent training of the yolov3 model; It is divided into training set, test set and verification set according to the set ratio; the set ratio includes 2:1:1, 3:1:1 and 4:3:1;

Construct a fusion model: The fusion model is a model constructed by at least two fusion methods among support vector machines, deep convolutional neural networks and RBF neural networks. The fusion methods include linear weighted fusion method, cross fusion method, waterfall fusion method, and feature fusion. method and prediction fusion method;

After the training set, test set and verification set are normalized, the fusion model is trained, tested and verified, and the trained fusion model is marked as the yolov3 detection model;

Among them, the feature extraction network darknet-53 used by the yolov3 detection model uses a large number of 3*3 and 1*1 convolutional layers connected in turn, and there are 53 convolutional layers;

The image analysis module is used to use the yolov3 detection model to perform digital target recognition on the received meter image to obtain the digital recognition result, and finally obtain the meter reading; the image analysis module is used to transmit the final meter reading to the display module through the controller real-time display;

The image acquisition module includes a housing and a camera ball movably arranged in the housing, a camera port is provided on the camera ball, and a camera is provided on the camera port;

In this solution, the camera ball can be rotated in the shell, and the user can adjust the camera angle by adjusting the camera ball, so that the meter reading image can be better collected, and the meter reading in the substation can be identified through the image, so as to achieve The purpose of unmanned inspection;

The information recording module is connected with the display module, and is used to record the identification record of the image analysis module and transmit the identification record to the information sorting module. Various environmental parameter values; various environmental parameter values include temperature, humidity, wind pressure, wind speed, interference signals, etc.;

The information sorting module receives the identification records and sorts the identification records, constructs a training sample for parameter detection, and the training is based on the machine learning method to obtain a parameter compensation model and transmits the parameter compensation model to the parameter compensation module. The yolov3 detection model is corrected, and the specific compensation steps are:

SS1: Obtain the identification record of the image analysis module at the current moment, input the currently identified meter reading, actual meter reading and corresponding environmental parameter values into the parameter compensation model to obtain the attribute parameter compensation coefficient;

SS2: The yolov3 detection model is modified according to the attribute parameter compensation coefficient, so that the yolov3 detection model is continuously revised and improved, and the accuracy of recognition is improved;

The invention uses the results of each link of the image analysis module as a feedback factor, and in turn further verifies the yolov3 detection model in the previous training and identification process, and compensates the attribute parameters, so that the yolov3 detection model is continuously revised and improved, and the correct rate of recognition is continuously improved. The whole process is automatic, so that the work can be widely carried out, and the continuity and effectiveness of the work are guaranteed;

In this embodiment, in order to reduce the influence of interference signals and ensure the accuracy of image recognition; the signal verification module is used to verify the communication state of the image analysis module in real time, and the specific verification steps are:

The signal verification module sends a verification configuration message to the FPGA main control of the image analysis module according to a preset verification period, wherein the verification configuration message includes a first signal quality threshold; in response to receiving the verification configuration message sent by the signal verification module, the FPGA main control sending the second synchronization signal to the signal verification module;

In response to monitoring the second synchronization signal, the signal verification module determines the signal quality of the second synchronization signal, and compares the signal quality of the second synchronization signal with the first signal quality threshold to obtain a corresponding quality difference value Z1; Those skilled in the art should understand that any metric known in the art can be used to characterize signal quality, such as RSRQ, RSRP, RSSI, etc.; the quality difference here can reflect signal attenuation during transmission;

Calculate the time difference between the moment when the signal verification module sends the verification configuration message and the moment when the signal verification module monitors the second synchronization signal again to obtain the response duration XT; the signal loss index SH is calculated by using the formula SH=Z1×a1+XT×a2, where a1 and a2 are both coefficient factors;

Establish a curve graph of the signal loss index SH versus time, and compare the signal loss index SH with the loss threshold; if SH ≥ the loss threshold, intercept the corresponding curve segment in the corresponding curve and mark it as yellow, and record it as the interference curve part;

In the preset time period, the number of statistical interference curve segments is L1, and all interference curve segments are integrated with time to obtain the interference reference energy L2; the interference coefficient Cy is calculated by using Cy=L1×a3+L2×a4, where a3 , a4 are coefficient factors;

Compare the interference coefficient Cy with the interference threshold; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication state is abnormal, and a communication warning command is generated;

The signal verification module is used to transmit the communication warning command to the controller. After the controller receives the communication warning command, it automatically drives the alarm module to issue an alarm to remind managers to deal with it as soon as possible, thereby improving the recognition efficiency and accuracy of the image analysis module;

Among them, the image analysis module enters the active standby mode after detecting the communication warning command, that is, the FPGA main control with abnormal communication is no longer used to identify the meter image. After the signal verification module judges that the communication state is normal, it will continue between the two. Communication.

A method for meter identification based on deep learning, applied to the above-mentioned meter identification system, includes the following steps:

Step 1: Build a yolov3 detection model that can train and identify images through the model building module, and transfer the constructed yolov3 detection model to the image analysis module;

Step 2: Learn and recognize the meter image collected by the image acquisition module through the image analysis module, analyze the collected meter image by using the yolov3 detection model, and identify the meter reading; The controller transmits to the display module for real-time display;

Step 3: Record the identification records of the image analysis module through the information recording module and transmit the identification records to the information sorting module; the information sorting module receives the identification records and sorts the identification records, constructs a training sample for parameter detection, and the training is based on a machine learning method to obtain parameter compensation model;

Step 4: Use the parameter compensation model to correct the yolov3 detection model through the parameter compensation module, so that the yolov3 detection model is continuously revised and improved, and the accuracy of identification is improved;

Step 5: Verify the communication status of the image analysis module in real time through the signal verification module, and calculate the interference coefficient Cy; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication status is abnormal, and a communication warning command is generated to remind the management personnel to deal with it as soon as possible;

Step 6: After the image analysis module detects the communication warning command, it enters the active standby mode, that is, the FPGA main control with abnormal communication is no longer used to identify the meter image. After the signal verification module judges that the communication state is normal, it will continue the two. communication between.

The above formulas are calculated by removing the dimension and taking its numerical value. The formula is a formula that is closest to the real situation by collecting a large amount of data and performing software simulation. The preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation. Or a large amount of data simulation is obtained.

The working principle of the present invention:

A method and system for meter recognition based on deep learning. When working, first the model building module is used to build a yolov3 detection model that can train and recognize images; the image acquisition module is used to collect meter images, and the collected The meter image is transmitted to the image analysis module for learning and recognition; the image analysis module uses the yolov3 detection model to analyze the collected meter image and identify the meter reading, and transmit the final meter reading through the controller to the display The module is displayed in real time to achieve the purpose of unmanned inspection;

The information recording module is connected with the display module, and is used to record the identification records of the image analysis module and transmit the identification records to the information sorting module; the information sorting module receives the identification records and sorts the identification records, constructs a training sample for parameter detection, and the training is based on the machine The learning method is used to obtain the parameter compensation model; after the parameter compensation module receives the parameter compensation model, it corrects the yolov3 detection model, so that the yolov3 detection model is continuously revised and improved, and the recognition accuracy is continuously improved. Make the work can be widely carried out, and the continuity and effectiveness of the work are guaranteed;

The signal verification module is used to verify the communication status of the image analysis module in real time. First, it sends a verification configuration message to the FPGA master of the image analysis module according to the preset verification cycle. In response to receiving the verification configuration message sent by the signal verification module, the FPGA master The control sends the second synchronization signal to the signal verification module; the signal loss index SH is calculated according to the corresponding quality difference Z1 and the response time XT; the interference coefficient Cy is calculated according to the change trend of the signal loss index SH, if Cy ≥ the interference threshold, then It is determined that the signal interference is serious and the communication state is abnormal, and a communication warning command is generated to prompt the management personnel to deal with it as soon as possible; effectively reduce the influence of the interference signal, thereby improving the recognition efficiency and accuracy of the image analysis module.

In the description of this specification, description with reference to the terms "one embodiment," "example," "specific example," etc. means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one aspect of the present invention. in one embodiment or example. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the particular features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

The above-disclosed preferred embodiments of the present invention are provided only to help illustrate the present invention. The preferred embodiments do not set forth all the details and do not limit the invention to mere embodiments. Obviously, many modifications and variations are possible in light of the content of this specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is to be limited only by the claims and their full scope and equivalents.

Claims (8)

1. a meter recognition system based on deep learning, is characterized in that, comprises image acquisition module, model building module, information recording module, parameter compensation module and signal verification module; The image acquisition module is used to collect meter images, and transmit the collected meter images to the image analysis module for learning and identification; the image analysis module is connected to the model building module, and is used to obtain the pair of images constructed by the model building module. yolov3 detection model for image training and recognition; Described image analysis module is used for adopting the yolov3 detection model to carry out digital target recognition to the received meter image to obtain the digital recognition result, and finally obtain the meter reading; The information recording module is used for recording the identification records of the image analysis module and transmitting the identification records to the information sorting module; the information sorting module is used for sorting the identification records, constructing a training sample for parameter detection, and the training is based on a machine learning method to obtain A parameter compensation model; the parameter compensation module modifies the yolov3 detection model after receiving the parameter compensation model; The signal verification module is used to verify the communication status of the image analysis module in real time, and calculate the interference coefficient Cy; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication status is abnormal, and a communication warning command is generated to remind the management personnel to deal with it as soon as possible. 2. a kind of meter recognition system based on deep learning according to claim 1, is characterized in that, wherein, the construction process of yolov3 detection model is as follows: Collect meter images to construct the data set required for training; the specific performance is as follows: collect n meter images at different shooting angles and under different lighting conditions, label the meter images, and obtain the data set required for subsequent training of the yolov3 model; It is divided into training set, test set and verification set according to the set ratio; the set ratio includes 2:1:1, 3:1:1 and 4:3:1; Construct a fusion model: The fusion model is a model constructed by at least two fusion methods among support vector machines, deep convolutional neural networks and RBF neural networks. The fusion methods include linear weighted fusion method, cross fusion method, waterfall fusion method, and feature fusion. method and prediction fusion method; After the training set, test set and verification set are normalized, the fusion model is trained, tested and verified, and the trained fusion model is marked as the yolov3 detection model. 3. a kind of meter identification system based on deep learning according to claim 1, is characterized in that, the concrete correction step of described parameter compensation module is: SS1: Obtain the identification record of the image analysis module at the current moment, input the currently identified meter reading, actual meter reading and corresponding environmental parameter values into the parameter compensation model to obtain the attribute parameter compensation coefficient; SS2: Modify the yolov3 detection model according to the attribute parameter compensation coefficient. 4. a kind of meter identification system based on deep learning according to claim 1, is characterized in that, the concrete verification step of described signal verification module is: The signal verification module sends a verification configuration message to the FPGA main control of the image analysis module according to a preset verification period, wherein the verification configuration message includes a first signal quality threshold; In response to receiving the verification configuration message, the FPGA master sends the second synchronization signal to the signal verification module; the signal verification module determines the signal quality of the second synchronization signal, and compares it with the first signal quality threshold to obtain the corresponding poor quality Value Z1; set the response time to XT; The signal loss index SH is calculated by using the formula SH=Z1×a1+XT×a2, where a1 and a2 are both coefficient factors; the interference coefficient Cy is evaluated according to the change trend of the signal loss index SH. 5. a kind of meter identification system based on deep learning according to claim 4, is characterized in that, the concrete evaluation process of interference coefficient Cy is as follows: Establish a curve graph of the signal loss index SH versus time, and compare the signal loss index SH with the loss threshold; if SH ≥ the loss threshold, intercept the corresponding curve segment in the corresponding curve and mark it as yellow, and record it as the interference curve part; In the preset time period, the number of statistical interference curve segments is L1, and all interference curve segments are integrated with time to obtain the interference reference energy L2; the interference coefficient Cy is calculated by using Cy=L1×a3+L2×a4, where a3 , a4 are coefficient factors. 6. A kind of meter identification system based on deep learning according to claim 1, is characterized in that, described identification record comprises the meter reading that image analysis module identifies each time, actual meter reading and corresponding items Environmental parameter values; the environmental parameter values include temperature, humidity, wind pressure, wind speed, and interference signals. 7. A kind of meter identification system based on deep learning according to claim 1, is characterized in that, described image analysis module enters active standby mode after detecting communication warning instruction, namely no longer uses the FPGA of abnormal communication The main control is used to identify the meter image; after the signal verification module judges that the communication state is normal, the communication between the two is continued. 8. a kind of meter identification method based on deep learning, be applied to a kind of meter identification system based on deep learning as described in any one of claim 1-7, it is characterized in that, comprise the steps: Step 1: Build a yolov3 detection model that can train and identify images through the model building module, and transfer the constructed yolov3 detection model to the image analysis module; Step 2: use the yolov3 detection model to perform digital target recognition on the meter image collected by the image acquisition module through the image analysis module to obtain a digital recognition result, and finally obtain the meter reading; Step 3: record the identification records of the image analysis module through the information recording module, call the information sorting module to sort the identification records, and construct a training sample for parameter detection; the training is based on a machine learning method to obtain a parameter compensation model; Step 4: The parameter compensation model is fed back to the model building module through the parameter compensation module to correct the yolov3 detection model; Step 5: Verify the communication status of the image analysis module in real time through the signal verification module, and calculate the interference coefficient Cy; if Cy ≥ the interference threshold, it is determined that the signal interference is serious and the communication status is abnormal, and a communication warning command is generated to remind the management personnel to deal with it as soon as possible; Step 6: After the image analysis module detects the communication warning command, it enters the active standby mode, and after the signal verification module judges that the communication state is normal, the communication between the two is continued.

Images