PCCP pipeline broken wire acoustic signal recognition method and device based on deep learning

Abstract

The present invention discloses a method and device for identifying broken wire acoustic signals in a PCCP pipeline based on deep learning. The method comprises: building a deep learning model and a corresponding data set to complete model training and obtain a deep learning model for identifying and classifying audio signals; deploying the trained deep learning model in a server, and when the server receives a prediction request, inputting the audio file or data carried by the prediction request into the deep learning model to obtain a prediction result, and returning the prediction result to the requester, wherein the prediction result includes the confidence probability of each category; after receiving the prediction result of the deep learning model, the receiver determines that the signal belongs to the broken wire category according to a set threshold, and then issues a prompt message to warn of the occurrence of a broken wire event. The device comprises: a processor and a memory. The present invention can effectively improve the accuracy of system recognition and the degree of automation of the system.

Description

Technical Field

[0001] The present invention relates to the field of PCCP pipeline safety monitoring, and in particular to a method and device for identifying PCCP pipeline broken wire acoustic signals based on deep learning. Background Art

[0002] Prestressed concrete cylinder pipe (PCCP) is a composite pipe with excellent performance. Due to its excellent pressure resistance and low production and processing costs, PCCP pipes have been widely used in large-scale water supply and water conservancy projects such as urban output and cross-regional water diversion. They are also widely used in the circulating water systems of large industrial facilities and large-scale sewage projects. PCCP pipes have formed a complete industrial chain in China that includes design, production, and construction.

[0003] The excellent compressive performance of PCCP pipelines is largely derived from the circumferentially wound prestressed steel wires. However, over time, the prestressed steel wires will gradually fail and break due to hydrogen embrittlement and corrosion, resulting in a decrease in the strength of the pipeline. When multiple prestressed steel wires break in a pipeline, its compressive performance will be significantly reduced. If the strength of the PCCP pipeline drops to a critical point, the high-pressure liquid inside will cause the pipeline to burst, resulting in internal liquid leakage, causing huge economic losses and even endangering the safety of people around the accident.

[0004] At present, there are many detection or monitoring methods for PCCP pipeline broken wires. Common detection methods include detection robots based on orthogonal electromagnetic method, which detect pipeline broken wires by detecting magnetic field anomalies caused by PCCP broken wires. Common monitoring methods mainly include PCCP broken wire monitoring systems based on distributed optical fiber sensing, which mainly monitor the vibration signals around the sensor optical cables laid along the pipeline to continuously monitor the broken wire events of PCCP pipes.

[0005] Among the existing detection or monitoring methods, most of their recognition algorithms rely on features such as signal amplitude and characteristic frequency. The recognition accuracy of broken wire signals is low, the system has a low degree of automation, and the recognition results still require a lot of manpower to verify and review. Summary of the invention

[0006] The present invention provides a method and device for identifying PCCP pipeline broken wire acoustic signals based on deep learning. The present invention is combined with a PCCP broken wire monitoring system based on distributed optical fiber sensing to determine the category to which the signal of the monitoring system belongs, which can effectively improve the accuracy of system identification and the degree of automation of the system. See the following description for details:

[0007] In a first aspect, a PCCP pipeline broken wire acoustic signal recognition method based on deep learning, the method comprising:

[0008] By building a deep learning model and corresponding data sets, model training is completed to obtain a deep learning model for identifying and classifying audio signals;

[0009] The trained deep learning model is deployed on the server. When the server receives a prediction request, the audio file or data carried in the prediction request is input into the deep learning model to obtain the prediction result, and the prediction result is returned to the requester. The prediction result includes the confidence probability of each category.

[0010] After receiving the prediction result of the deep learning model, the receiver will issue a prompt message if it determines that the signal belongs to the broken wire category according to the set threshold, warning of the occurrence of the broken wire event;

[0011] The model training includes: enhancement before feature extraction, generation of frequency domain coding, generation of time domain coding, data enhancement after feature extraction, packaging and padding;

[0012] The packaging and filling are as follows:

[0013] According to the order of Mel frequency cepstral coefficient logarithmic intensity, frequency domain coding, and time domain coding, the three matrices of size M×N are packed into a tensor of size M×N×3. If M or N is less than 640, padding operations are performed around the matrix, the Mel frequency cepstral coefficients are padded with zeros, the frequency domain coding and time coding are padded with -0.1, and the terminal index is padded when the length is not an integer multiple of 2, and finally a 640×640×3 input tensor is obtained;

[0014] The confidence probabilities of the various categories are: wire breakage, knocking, cutting, and cable contact.

[0015] Among them, the enhancement before feature extraction is low-pass filtering, and the probability of occurrence of low-pass filtering data enhancement is set to 0.1. When it occurs, the cutoff frequency of the low-pass filter is set to meet the probability density function described by the following formula:

[0016]

[0017] Where f is the filter cutoff frequency, is the system sampling frequency.

[0018] Wherein, the generated frequency domain code is:

[0019]

[0020] in,

[0021]

[0022] In the formula, is the frequency corresponding to the current Mel scale, is the system sampling frequency, is the baseline sampling rate, the baseline sampling rate is the system sampling rate of the training set data, and n is the Mel scale.

[0023] Wherein, the generated time domain code is:

[0024]

[0025] in:

[0026]

[0027] in, is the system sampling frequency, m is the matrix abscissa index, and s is the Mel-frequency cepstral coefficient sliding window step size.

[0028] Among them, the data enhancement after feature extraction is:

[0029] When frequency domain truncation occurs, the MFCC logarithmic strength above the frequency domain truncation position is set to 0, and the frequency domain coding and time domain coding above the frequency domain truncation position are set to -0.1. The truncation position is determined by the following probability density function:

[0030]

[0031] When time domain truncation occurs, the MFCC logarithmic strength above the time domain truncation position is set to 0, and the frequency code and time domain code above the time domain truncation position are set to -0.1. The truncation position is determined by the probability density function shown below:

[0032] .

[0033] In the second aspect, a PCCP pipeline broken wire acoustic signal recognition device based on deep learning, the device comprises: a processor and a memory, the memory stores program instructions, and the processor calls the program instructions stored in the memory to enable the device to execute any one of the methods described in the first aspect.

[0034] In a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor executes any one of the methods described in the first aspect.

[0035] The beneficial effects of the technical solution provided by the present invention are:

[0036] 1. The present invention proposes a method for identifying broken wire acoustic signals in PCCP pipelines based on deep learning, which can realize automatic broken wire signal identification with high accuracy;

[0037] 2. The present invention proposes a feature extraction method based on Mel-frequency cepstral coefficients, time domain coding, and frequency domain coding, which can improve the generalization of deep learning models under conditions with different sampling rates and input lengths from training data;

[0038] 3. With respect to the feature extraction method, the present invention proposes three data enhancement methods, namely low-pass filtering, frequency domain truncation, and time domain truncation, to further improve the generalization of high-depth learning models under conditions with different sampling rates and input lengths from training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the data annotation interface;

[0040] Figure 2 A schematic diagram of the data annotation results;

[0041] Figure 3 It is a schematic diagram of the system display interface;

[0042] Figure 4 This is a flowchart of a PCCP pipeline broken wire acoustic signal recognition method based on deep learning;

[0043] Figure 5 Schematic diagram of input features;

[0044] Among them, (a) is the broken wire signal MFCC; (b) is the broken wire signal frequency coding; (c) is the broken wire signal time coding; (d) is the noise signal MFCC; (e) is the noise signal frequency coding; (f) is the noise signal time coding.

[0045] Figure 6 This is a photo of the scene of the steel wire fracture experiment;

[0046] Among them, (a) is the prestressed steel wire cut by a grinding wheel; (b) is the broken steel wire after cutting.

[0047] Figure 7 This is a schematic diagram of the broken wire signal. DETAILED DESCRIPTION

[0048] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0049] Example 1

[0050] A PCCP pipeline broken wire acoustic signal recognition method based on deep learning, see Figure 1-Figure 5 , the method comprises the following steps:

[0051] 101: By building a deep learning model and corresponding data sets, complete model training and obtain a deep learning model that can recognize and classify audio signals;

[0052] 102: Deploy the trained deep learning model in a service. The service may use HTTP, MQTT or other communication methods. When the service receives a prediction request, the service inputs the audio file or data carried in the prediction request into the deep learning model to obtain the prediction result, and returns the prediction result to the requester. The prediction result includes the confidence probability of each category.

[0053] 103: After receiving the prediction result of the deep learning model, the receiver can judge the category to which the signal belongs according to the set threshold. If the confidence probability of the corresponding category in the prediction result is higher than the set threshold, the signal is considered to belong to this category. A signal can belong to multiple categories at the same time. When the signal belongs to the broken wire category, a prompt message is issued to indicate the occurrence of a broken wire event.

[0054] In summary, the embodiment of the present invention can effectively improve the accuracy of system recognition and the degree of automation of the system by determining the category to which the signal of the monitoring system belongs through the above steps 101-103.

[0055] Example 2

[0056] The solution in Example 1 is further introduced below in conjunction with a specific calculation formula, as described below for details:

[0057] The embodiment of the present invention is a deep learning algorithm, which specifically includes three main steps: data set construction, model training, and model deployment. The specific steps of data set construction are as follows:

[0058] 1) Use the acquisition equipment to collect a certain amount of sample data, among which the broken wire signal can refer to Figure 6 The method shown is to collect the wire break signal by cutting the prestressed steel wire with a grinding wheel. Figure 7 As shown, the sampling frequency of the acquisition equipment is required to be at least 5000Hz;

[0059] 2) Use annotation tools to annotate the collected samples, such as Figure 1 As shown, the annotation content is the category to which the audio data at different positions in the audio file belongs;

[0060] 3) After the annotation is completed, export the annotation results in json format, such as Figure 2 As shown, the exported content should include the length of the data file, the start and stop time of each segment of the annotation content, and the category to which it belongs.

[0061] The deep learning model used in the embodiment of the present invention is composed of a convolution layer, a pooling layer, and a fully connected layer. In some convolution layers, the relu activation function and the batchnorm layer are introduced to improve the convergence speed of the model and avoid gradient explosion. The input of the deep learning model used in the embodiment of the present invention is a 640×640×3 tensor, in which the first dimension indicates the time domain, the second dimension indicates the frequency domain, and the three indexes of the third dimension are the Mel frequency cepstral coefficient logarithmic intensity (MFCC), frequency domain coding, and time domain coding. The MFCC, frequency domain coding, and time domain coding of the broken wire signal and the noise signal are shown in Figure 1. Figure 5 The output of the deep learning model used in the embodiment of the present invention is a vector with 4 elements, which respectively represent the confidence probabilities of broken wire, knocking, cutting, and cable touching.

[0062] Model training can be divided into data preprocessing, forward deduction, loss calculation, and back propagation. When performing data preprocessing, the original audio data must first be intercepted, data enhanced, feature extracted, and labels generated. Data interception uses a rectangular window as the window function, and the interception length is set to 1s. Data enhancement includes enhancement methods before feature extraction and enhancement methods during feature extraction. Multiple feature extraction methods have their own probability of occurrence.

[0063] Among them, the feature extraction and data enhancement methods are as follows:

[0064] 1) Enhancement before feature extraction:

[0065] The enhancement method before feature extraction is low-pass filtering, which simulates the actual algorithm application in the case of the same sampling rate as the training set system but with a lower frequency response. The probability of occurrence of low-pass filtering data enhancement is set to 0.1. When it occurs, the cutoff frequency of the low-pass filter is set to meet the probability density function described by the following formula:

[0066]

[0067] 2) Generate the logarithmic intensity of Mel frequency cepstral coefficients:

[0068] The representation is as follows:

[0069]

[0070] in:

[0071]

[0072] In the formula, is the logarithmic strength of the Mel-frequency cepstral coefficient, n is the Mel scale, m is the converted point index, L is the input frame length, s is the step size when calculating the Mel-frequency cepstral coefficient, N is the number of Mel scales, and M is the total number of points after conversion.

[0073] The frame length must satisfy the following formula to ensure that the total number of points after conversion is greater than or equal to 320 and less than or equal to 640.

[0074]

[0075] 3) Generate frequency domain code:

[0076] Its expression is shown in the following formula:

[0077]

[0078] in:

[0079]

[0080] In the formula, is the frequency corresponding to the current Mel scale, is the system sampling frequency, is the baseline sampling rate, which is the system sampling rate of the training set data.

[0081] 4) Generate time domain code:

[0082] Its expression is shown in the following formula:

[0083]

[0084] in:

[0085]

[0086] 5) Data enhancement after feature extraction:

[0087] The enhancement methods after feature extraction are frequency domain truncation and time domain truncation, and the probability of occurrence of the two data enhancements is set to 0.2. When frequency domain truncation occurs, the probability described by the following probability density function determines the frequency domain truncation position, sets the MFCC logarithmic strength above the frequency domain truncation position to 0, sets the frequency domain coding and time domain coding above the frequency domain truncation position to -0.1, and the truncation position is determined by the following probability density function:

[0088]

[0089] When time domain interception occurs, the probability described by the following probability density function determines the time domain interception position, sets the MFCC logarithmic strength above the time domain interception position to 0, sets the frequency code and time domain code above the time domain interception position to -0.1, and the interception position is determined by the following probability density function:

[0090]

[0091] 6) Packing and filling:

[0092] According to the order of Mel frequency cepstral coefficient logarithmic strength, frequency domain coding, and time domain coding, the three matrices of size M×N are packed into a tensor of size M×N×3. If M or N is less than 640, padding operations are performed around the matrix, the Mel frequency cepstral coefficients are padded with zeros, the frequency domain coding and time coding are padded with -0.1, and the terminal index is padded first when the length is not an integer multiple of 2. Finally, a 640×640×3 input tensor is obtained.

[0093] The label information is a vector with 4 elements, as shown in the following formula, which represents the confidence probabilities of the four signal types of broken wire, knocking, cutting, and cable touching. When generating label information, the category is judged according to the position of the intercepted original audio data and the data annotation results. If there is an overlapping area between the intercepted original audio file and a category in the data annotation result, the confidence probability of the corresponding category in the label information of the intercepted original audio file is set to 1.

[0094]

[0095] After feature extraction and label information generation, the model forward deduction can be carried out. Model forward deduction requires inputting the 640×640×3 input tensor obtained by feature extraction into the deep learning model, and the output result is a vector with 4 elements, as shown in the following formula.

[0096]

[0097] Depending on the computing power of the computer, multiple sets of tensors can be input into the deep learning model at the same time, and a 640×640×3×B tensor can be input, and the output will change accordingly to a 4×B tensor.

[0098] After the output result is obtained by forward deduction, the loss is calculated based on the difference between the output result and the label information. The loss calculation function selected by the embodiment of the present invention is the root mean square error loss function, which is expressed as follows:

[0099]

[0100] 7) Training and learning

[0101] After the loss calculation is completed, back propagation is performed, and the model weights are optimized using the SGD optimizer. The magnitude of each optimization is controlled by setting the learning rate. The learning rate should be set to a small value to avoid gradient explosion. The embodiment of the present invention uses 0.0001 as the learning rate during the training process.

[0102] Repeating data preprocessing, forward deduction, loss calculation, and back propagation until all data sets are completely traversed is called a training cycle. After each training cycle, data preprocessing, forward deduction, and loss calculation are performed on part of the data set, also known as the validation set. The loss value obtained by this loss calculation is used as the basic reference for the model training effect.

[0103] The embodiment of the present invention performs model training for a total of 200 training cycles, and finally selects the model weight obtained from the training cycle with the smallest validation set loss as the final deep learning model.

[0104] 8) Prediction

[0105] After the deep learning model training is completed, the model deployment can be carried out. The embodiment of the present invention adopts the HTTP protocol as the communication protocol of the deployed model. The deep learning model will run in the background of the http server. Whenever there is a data request entering the corresponding http method, the deep learning model is called to perform a prediction and return the prediction result. A complete prediction includes the following steps:

[0106] 1. Send the original data or the location of the original data to the http server through the http method. The http server parses and reads the original data to obtain the original data, sampling rate and duration.

[0107] 2. If the duration is less than 1 second, extract features from the original data, input the extracted tensor into the deep learning model, obtain the output result, and return the output result to the requester;

[0108] 3. If the duration is longer than 1 second, the original data is segmented with a segmentation step of 0.5 seconds and a segmentation window of 1 second. The segmented data is subjected to feature extraction, and the tensor after feature extraction is input into the deep learning model to obtain the output result, which is then returned to the requester.

[0109] 4. After receiving the returned result, the requester determines the category to which the data belongs based on the set minimum confidence probability threshold. Each category exceeding the threshold is considered to be the category to which the original data belongs. The original data can belong to multiple categories at the same time.

[0110] The display interface of the system equipped with the embodiment of the present invention is as follows: Figure 3 As shown in the figure, the system collects audio data in real time during operation, sends the audio data of the previous second to the http server every second and obtains the prediction result. If the broken wire confidence probability in the prediction result exceeds the set minimum confidence probability threshold, the sent audio data is considered to be a broken wire signal, and the data waveform and the corresponding confidence probability are displayed in the control below.

[0111] Example 3

[0112] A PCCP pipeline broken wire acoustic signal recognition device based on deep learning, the device comprises: a processor and a memory, the memory stores program instructions, the processor calls the program instructions stored in the memory to enable the device to execute the following method steps in embodiment 1:

[0113] By building a deep learning model and corresponding data sets, model training is completed to obtain a deep learning model for identifying and classifying audio signals;

[0114] The trained deep learning model is deployed on the server. When the server receives a prediction request, the audio file or data carried in the prediction request is input into the deep learning model to obtain the prediction result, and the prediction result is returned to the requester. The prediction result includes the confidence probability of each category.

[0115] After receiving the prediction results of the deep learning model, the receiver determines that the signal belongs to the broken wire category based on the set threshold and issues a prompt message to warn of the occurrence of a broken wire event.

[0116] Among them, model training includes: enhancement before feature extraction, generation of frequency domain coding, generation of time domain coding, data enhancement after feature extraction, packaging and padding.

[0117] Among them, the enhancement before feature extraction is low-pass filtering, and the probability of occurrence of low-pass filtering data enhancement is set to 0.1. When it occurs, the cutoff frequency of the low-pass filter is set to meet the probability density function described by the following formula:

[0118]

[0119] Where f is the filter cutoff frequency.

[0120] Among them, the frequency domain code is generated as:

[0121]

[0122] in,

[0123]

[0124] In the formula, is the frequency corresponding to the current Mel scale, is the system sampling frequency, is the baseline sampling rate, which is the system sampling rate of the training set data.

[0125] Among them, the generated time domain code is:

[0126]

[0127] in:

[0128]

[0129] in, is the system sampling frequency, m is the matrix abscissa index, and s is the Mel-frequency cepstral coefficient sliding window step size.

[0130] Among them, the data enhancement after feature extraction is:

[0131] When frequency domain truncation occurs, the MFCC logarithmic strength above the frequency domain truncation position is set to 0, and the frequency domain coding and time domain coding above the frequency domain truncation position are set to -0.1.

[0132]

[0133] When time domain truncation occurs, the MFCC logarithmic strength above the time domain truncation position is set to 0, and the frequency code and time domain code above the time domain truncation position are set to -0.1.

[0134] .

[0135] Among them, packaging and filling are:

[0136] According to the order of Mel-frequency cepstral coefficient logarithmic intensity, frequency domain coding, and time domain coding, the three matrices of size M×N are packed into a tensor of size M×N×3. If M or N is less than 640, padding operations are performed around the matrix, the Mel-frequency cepstral coefficients are padded with zeros, the frequency domain coding and time coding are padded with -0.1, and the terminal index is padded when the length is not an integer multiple of 2, and finally a 640×640×3 input tensor is obtained.

[0137] It should be pointed out here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0138] The execution subjects of the above-mentioned processor and memory can be devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution subjects and are selected according to the needs of actual applications.

[0139] The data signal is transmitted between the memory and the processor via a bus, which is not described in detail in the embodiment of the present invention.

[0140] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, the storage medium includes a stored program, and when the program is running, the device where the storage medium is located is controlled to execute the method steps in the above embodiment.

[0141] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0142] It should be pointed out here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0143] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present invention are generated.

[0144] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions may be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer-readable storage medium may be any available medium that can be accessed by the computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium or a semiconductor medium, etc.

[0145] Unless otherwise specified, the models of the components in the embodiments of the present invention are not limited, and any device that can perform the above functions may be used.

[0146] Those skilled in the art will appreciate that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A PCCP pipeline broken wire acoustic signal recognition method based on deep learning, characterized in that: The method comprises: By building a deep learning model and corresponding data sets, model training is completed to obtain a deep learning model for identifying and classifying audio signals; The trained deep learning model is deployed on the server. When the server receives a prediction request, the audio file or data carried in the prediction request is input into the deep learning model to obtain the prediction result, and the prediction result is returned to the requester. The prediction result includes the confidence probability of each category. After receiving the prediction result of the deep learning model, the receiver will issue a prompt message if it determines that the signal belongs to the broken wire category according to the set threshold, warning of the occurrence of the broken wire event; The model training includes: enhancement before feature extraction, generation of frequency domain coding, generation of time domain coding, data enhancement after feature extraction, packaging and padding; The packaging and filling are as follows: According to the order of Mel frequency cepstral coefficient logarithmic intensity, frequency domain coding, and time domain coding, the three matrices of size M×N are packed into a tensor of size M×N×3. If M or N is less than 640, padding operations are performed around the matrix, the Mel frequency cepstral coefficients are padded with zeros, the frequency domain coding and time coding are padded with -0.1, and the terminal index is padded when the length is not an integer multiple of 2, and finally a 640×640×3 input tensor is obtained; The confidence probabilities of the various categories are: confidence probabilities of wire breakage, knocking, cutting, and cable contact; The enhancement before feature extraction is low-pass filtering, and the probability of occurrence of low-pass filtering data enhancement is set to 0.

1. When it occurs, the cutoff frequency of the low-pass filter is set to conform to the probability density function described by the following formula: Where f is the filter cutoff frequency, f s is the system sampling frequency; The generated frequency domain code is: in, In the formula, f n is the frequency corresponding to the current Mel scale, f s is the system sampling frequency, F s is the base sampling rate, which is the system sampling rate of the training set data, and n is the Mel scale; The generated time domain code is: in: Among them, f s is the system sampling frequency, m is the matrix abscissa index, and s is the Mel frequency cepstral coefficient sliding window step size; The data enhancement after feature extraction is: When frequency domain truncation occurs, the MFCC logarithmic strength above the frequency domain truncation position is set to 0, and the frequency domain coding and time domain coding above the frequency domain truncation position are set to -0.

1. The truncation position is determined by the following probability density function: P(n)=2 / NN / 2≤n≤NWhen time domain interception occurs, the MFCC logarithmic strength above the time domain interception position is set to 0, and the frequency code and time domain code above the time domain interception position are set to -0.

1. The interception position is determined by the probability density function shown below: P(m)=2 / MM / 2≤n≤M Here, n is the Mel scale.

2. A PCCP pipeline broken wire acoustic signal recognition device based on deep learning, characterized in that: The device comprises: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method according to claim 1.

3. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is caused to perform the method of claim 1 .

Images