Coke rating method and rating system based on multi-scale residual neural network

Abstract

The invention provides a coke rating method and rating system based on a multi-scale residual neural network. The method comprises the following steps: S1, establishing terminal equipment and a coke quality management system; s2, acquiring coke audio information, and performing audio feature extraction and analysis; s3, forming a function by using the coke components, and marking a quality label on the coke sample in combination with experience; s4, constructing a coke quality rating model; s5, obtaining a quality evaluation result of the current coke sample, and generating a corresponding quality grade; and S6, displaying the quality grade of the coke sample on a terminal equipment interface, and completing the final quality grade of the coke quality. The invention provides a set of complete system from terminal information collection to result display, can unify and standardize the evaluation of the coke quality grade, is beneficial to timely and accurate judgment of the quality level of the coke, promotes judgment and management of the coke, improves the efficiency of coke quality rating management, and enables the coke rating process to be more scientific and normative.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of coke quality rating, more particularly to a coke rating method and system based on a multi-scale residual neural network. BACKGROUND

[0002] As one of the raw materials for blast furnace smelting, the role of coke in the blast furnace is becoming more and more important. How to efficiently rate the quality of coke and help staff quickly identify coke of different grades is the key to improving the utilization efficiency of coke of different qualities and reducing costs.

[0003] The current coke quality rating requires the expert knowledge of experienced technical personnel on site combined with the composition detection results of coke in the laboratory to make a judgment. This method not only puts forward high requirements for technical personnel, but also is time-consuming and laborious, and cannot quickly identify the quality of coke and characterize the rating of coke quality. SUMMARY

[0004] The technical problem to be solved by the present application is to overcome the above-mentioned deficiencies in the prior art, and to provide a coke rating method and system based on a multi-scale residual neural network. A multi-scale residual network structure is designed to extract deep features from the sound of coke falling to the ground and improve the classification performance of coke quality.

[0005] The technical problem to be solved can be implemented by the following technical solutions.

[0006] A coke rating method based on a multi-scale residual neural network comprises the following steps:

[0007] Step S1, a server, a terminal device and a quality management system are provided, the terminal device comprises a hardware device, and the terminal device connected to the server is used to obtain coke audio information;

[0008] Step S2, the terminal device is used to obtain and transmit coke falling audio information, the audio data is transmitted to a cloud server through cloud transmission, cloud management, cloud storage and cloud analysis are performed on the cloud, each coke sample is deeply analyzed, log mel spectrogram features and gamma pass spectrogram features are extracted, and unique feature information of coke of different grades is determined;

[0009] Step S3, a coke component composition function is used in combination with expert experience to label the coke;

[0010] Step S4, the spectrogram features obtained in step S3 are fused, a coke quality rating model is constructed using the fused features, all coke sample features are input into the rating model, and model parameters conforming to the coke quality rating are trained;

[0011] Step S5, input the fusion features of the coke audio into the coke quality rating model constructed in step S4, obtain the quality evaluation result of the current coke sample, and generate the quality grade of the coke;

[0012] Step S6, transmit the evaluation result of step S5 to the display interface of the terminal device through the cloud, display the final quality evaluation result through the display interface, and determine the quality grade of the specific coke.

[0013] Further, the step S1 specifically comprises the following steps:

[0014] The terminal device comprises a coke sound generating device to generate coke audio information, a pickup to collect coke collision audio, a computer to realize cloud service functions, and a mobile phone to realize deployment and result display of the coke management system. The coke audio information includes audio information generated by the collision of coke and iron plate at a set height (for example, a fixed height of 0.7 m).

[0015] Further, the step S2 specifically comprises the following steps:

[0016] Step S21: According to the coke audio data set obtained in step S1, the audio sample is preprocessed, first the sample is judged for effective audio segment, and the audio sample is cut out from the original audio paragraph and unified to an audio sample with a length of 1000 sampling points, and then the audio sample quality is improved through pre-emphasis, etc.

[0017] Step S22: For the length of the preprocessed audio sample, the output signal of the amplitude information is obtained by fast Fourier transform.

[0018] Step S23: The output signal of step S22 is subjected to Mel filter extraction to obtain the Mel spectrogram feature, and the Mel filter bank is composed of s triangular filters. The feature extraction formula of this step is:

[0019]

[0020] Wherein, k represents the kth spectrum line in the spectrum, H m (k) represents the frequency corresponding to the mth Mel filter, f(m) represents the center frequency of the mth filter; wherein, m refers to the number of any filter.

[0021] The spectrum obtained after the Mel filter bank is subjected to logarithmic operation, and the output is a logarithmic Mel spectrogram feature of 222x128.

[0022] Step S24: The output signal of step S22 is subjected to gamma pass filter extraction to obtain the gamma pass spectrogram feature, and the gamma pass filter bank is composed of n equal-height filters. The feature extraction formula of this step is:

[0023]

[0024] wherein a is the amplitude of the Fourier transformed signal, γ represents the order of the filter, f c is the center frequency of the filter, is the phase information of the audio signal, e is the exponential constant, e = 2.718;

[0025] wherein ERB(f c ) is the gamma cut-off frequency, which determines the decay speed of the impulse response and is related to the critical bandwidth of human hearing, and the conversion formula with the center frequency is:

[0026]

[0027] The output of the gamma cut-off filter bank is taken as the absolute value of the cubic root to obtain the gamma cut-off spectrogram feature with a size of 222x128.

[0028] Step S25: The output log-mel spectrogram feature of step S23 and the output gamma cut-off spectrogram feature of step S24 are spliced and fused to form a high-level fusion feature with a size of 222x256, which is used as the input of the subsequent network model.

[0029] The network structure of the coke quality rating model is based on a basic CNN neural network model, multiple residual modules are embedded in the network, and different scale convolution kernels are set, so as to construct a multi-scale residual neural network through the parallel neural network of the multi-scale residual modules. The high-level fusion feature is used as the input of the network, and the feature extraction is performed through channels with different convolution kernel sizes.

[0030] Further, the step S4 specifically includes the following steps:

[0031] Step S41: The original feature map of the output high-level fusion feature 222x256 of step S25 is taken as the input, and is subjected to dimension change and parameter reduction through a 1x1 residual convolution module, so as to realize cross-channel information combination and enhance nonlinear features.

[0032] Step S42: The output of step S41 is input into three channels of 1x1, 3x3 and 5x5 in parallel, and each channel contains two residual convolution modules with corresponding convolution kernel sizes. After passing through the convolution block, each residual convolution module is sequentially input into a batch normalization layer and a RELU layer, so as to relieve the problems of gradient disappearance and gradient explosion, improve the training stability, and improve the convergence speed.

[0033] Step S43: output of three parallel channels in step S42, each channel learns different features through different size convolution kernels, and the multi-feature output is feature spliced to increase the learning ability of the model to the input features. The spliced output is input to the Dropout layer to make the network learn more robust feature representation.

[0034] Step S44: output of step S43 to a fully connected layer, linear combination of features of the previous layer, integration of high-level abstract features.

[0035] Step S45: the output of the fully connected layer is converted into a vector representing a probability distribution by a softmax activation function, which outputs a probability estimate for each element of the input vector, and is used with a cross-entropy loss function to measure the gap between the model output and the actual label, so as to perform gradient descent optimization. The cross-entropy loss function is:

[0036]

[0037] Where p(x) is the label of the training set, specifically the number 1 or 0, representing the coke audio sample of good quality and poor quality, respectively, q(x) is the result of the model predicting the quality of the coke, also represented by the number 1 or 0.

[0038] Further, the step S5 specifically comprises the following steps:

[0039] Step S51: real-time coke audio information obtained from the coke sound device is transmitted to the cloud server;

[0040] Step S52: pre-processing and standardization of the audio information;

[0041] Step S53: feature analysis of the pre-processed audio and samples and input into the coke quality rating model constructed in step S4 to obtain the quality grade of the current coke sample.

[0042] Further, the step S6 specifically comprises the following steps:

[0043] Step S61: complete the deployment of the network model on the server, and transmit the output quality grade of S5 to the mobile phone end through the server interface.

[0044] Step S62: the mobile phone acquires the output information sent by the server, performs field conversion processing, and displays the results on the mobile phone through a pop-up window to complete the coke sample grade evaluation.

[0045] The present application can quickly grade the coke quality through the above steps, and can characterize the coke evaluation features, which is beneficial to reduce the dependence on professional knowledge and realize end-to-end coke quality grade evaluation. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flow chart of the coke rating technology based on the multi-scale residual neural network of the present application.

[0047] Figure 2 is a whole structure diagram of the coke audio generation and collection device.

[0048] Figure 3 is a flow chart of the audio feature acquisition method.

[0049] Figure 4 is a deep multi-scale residual network structure block diagram.

[0050] Figure 5 is a system start and audio recording interface, Figure 5 a, Figure 5 b and Figure 5 c respectively illustrate different interfaces.

[0051] Figure 6 is an audio storage and deletion interface, Figure 6 a, Figure 6 b and Figure 6 c respectively illustrate different interfaces.

[0052] Figure 7 is an audio analysis interface, Figure 7 a and Figure 7 b respectively illustrate different interfaces. DETAILED DESCRIPTION

[0053] The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0054] As shown in Figure 1 , the coke rating system based on the multi-scale residual neural network of the present application comprises the following steps:

[0055] Step S1: Provide a server, an audio transmitter, an audio receiver, a terminal device, a coke audio generation and collection device, and a quality management system, and acquire coke audio information by using the terminal device connected to the server.

[0056] As shown in Figure 2 , when collecting audio, the audio transmitter is placed on the clamp plate of the coke audio generation and collection device to receive the audio signal of the falling coke. The audio receiver is connected to the terminal device through a data line, and the collected audio signal is stored to the server through the connection of the terminal device and the server. When collecting audio, the buttons of the audio transmitter and the audio receiver need to be turned on at the same time to collect audio. The audio collection start button needs to be turned on in advance on the terminal device, and after the coke audio is generated, the audio collection is ended on the terminal device.

[0057] Step S2: using the coke landing audio information obtained on the terminal device 5, the coke landing audio information is transmitted through the upload button, the audio data is transmitted to the cloud server through the cloud transmission mode, the uploaded coke landing audio information is managed, stored and analyzed on the cloud server, the audio information is further processed, the most informative audio information segment is retained, the commonality and difference between the audio information are highlighted, the log mel spectrogram features and gamma pass spectrogram features are extracted, the feature information of the same grade coke is as close as possible in the feature space, and the feature information of different grade coke is as far as possible in the feature space, and the coke audio information is determined.

[0058] Step 2 specifically includes:

[0059] Step S21: in order to extract more coke landing audio effective frequency band, the obtained coke audio data set is preprocessed. First, the sample is judged for effective audio segment, and is cut from the original audio segment. The signal value corresponding to the audio information when the coke lands is certainly greater than the signal value corresponding to the audio information collected when the coke does not land. The time point of the sharp rise of the audio signal value is found by setting the threshold value, the upper boundary of the audio cutting and the lower boundary of the audio cutting are set, the total time length of the cut audio is fixed by the upper and lower boundaries of the audio cutting, and the total time length is determined. The cut audio signal retains the sound information at the coke landing time to the greatest extent, and the redundant information in the audio is discarded. The above operation can realize the audio cutting.

[0060] Step S22: in order to further extract the frequency domain information of the coke landing audio signal, the preprocessed audio sample length is subjected to fast Fourier transform to obtain the output signal of the amplitude information. The purpose of Fourier transform is to convert the signal in time domain into the signal in frequency domain. The fast Fourier transform formula is as follows:

[0061]

[0062] Wherein, x(n) represents the input signal of the audio after windowing, F represents the Fourier transform sampling point number, i represents the current sampling point; e is an exponential constant, e = 2.718; j is a basic imaginary unit, which is also a constant, equivalent to the square root of-1.

[0063] Step S23: the output signal of step S22 is subjected to mel filter to extract mel spectrogram features. The mel spectrogram feature takes the data on the spectrogram as the feature of the signal, the horizontal axis x of the spectrogram is time, the vertical axis y is frequency, and the value corresponding to (x, y) represents the amplitude of frequency y at time x. The frequency of the vertical axis of the mel spectrogram and the original frequency are exchanged through the following formula:

[0064]

[0065]

[0066] wherein f origin represents the original frequency, f mel represents the converted mel frequency, obviously, when f origin is very large, the change of f mel tends to be gentle.

[0067] The mel filter bank is composed of s triangular filters, and the feature extraction formula of this step is:

[0068]

[0069] wherein k represents the kth spectrum line in the spectrum, H m (k) represents the frequency response of the mth mel filter, and f(m) represents the center frequency of the mth filter.

[0070] The mel spectrum feature correlation function is used to calculate the corresponding log mel spectrum feature result of the audio signal. Some important parameters of the function are set, including the sampling rate, the frame length, the pre-emphasis filter coefficient, the number of filters in the filter bank, the mel filter lowest band edge, and the mel filter highest band edge.

[0071] Using the above function, the spectrum obtained after the mel filter bank is logarithmically operated, and the original input audio signal is uniformly output as a 222x128 log mel spectrum feature.

[0072] Step S24:

[0073] The output signal of step S22 is subjected to a gamma pass filter to extract a gamma pass spectrum feature, and the gamma pass filter bank is composed of n equal-height filters. The feature extraction formula of this step is:

[0074]

[0075] wherein a is the amplitude of the Fourier transformed signal, γ represents the order of the filter, f c is the center frequency of the filter, φ is the phase information of the audio signal, e is an exponential constant, e=2.718; wherein ERB(f c ) is the gamma pass frequency, which determines the corresponding decay rate of the pulse and is related to the critical bandwidth of human auditory perception. The conversion formula of the center frequency is:

[0076]

[0077] The log-mel spectrogram feature corresponding to the audio signal is calculated by using the computed gamma-tunneled spectrogram feature correlation function. Some important parameters of the function are set, including the sampling rate, the frame length, the pre-emphasis filter coefficient, the number of filters in the filter bank, the lowest and highest band edges of the mel filter, etc.

[0078] The preprocessed audio signal is converted into the corresponding gamma-tunneled spectrogram feature by using the above function. The output of the gamma-tunneled filter bank is taken as the absolute value of the cubic root, and the gamma-tunneled spectrogram feature with a size of 222x128 is obtained.

[0079] Wherein, n or s in steps S23 and S24 represents the meaning of several, only a letter code; the numerical size corresponds to m≤n or s in the formula.

[0080] Step S25:

[0081] The output log-mel spectrogram feature of step S23 and the output gamma-tunneled spectrogram feature of step S24 are spliced and fused to extract multi-dimensional features of the audio information, and a high-level fusion feature with a size of 222x256 is formed as the input of the subsequent network model.

[0082] Step S3 uses the coke component composition to combine expert experience to label the coke; wherein the good quality coke label is "A grade" and the bad quality coke label is "B grade".

[0083] Step S4 constructs a multi-scale residual convolutional neural network based on Pytorch, inputs the high-level fusion feature obtained in S2 into the network, and performs supervised constraint training based on the label in S3. The model parameters with the highest accuracy on the validation set are used as the final model parameters. The network structure is as shown in Figure 4 Step S4 specifically includes:

[0084] Step S41: The model is trained in batch training mode, and the input size is a four-dimensional tensor X∈R B ×1×222×256 Wherein, B represents the model training batch size. X is changed in dimension by two residual convolution modules with a kernel size of 1x1, and the feature dimension is reduced by using the Pooling layer. After this step, the output feature F∈R B×C×W×H Wherein, B represents the model training batch size, C represents the model channel number, W represents the feature map width, H represents the feature map height, and R is a commonly agreed real number space.

[0085] Step S42: the output feature F obtained in S41 is input into three channels with convolution kernel size of 1*1, 3*3 and 5*5 in parallel, each channel contains two residual convolution modules with corresponding convolution kernel size, and each residual convolution module is sequentially input into a batch normalization layer and a Relu layer after passing through a convolution block. Taking the 3*3 channel as an example, each residual convolution module (RCM) is composed of two convolution blocks (CB) with convolution kernel 1*1 and 3*3:

[0086] CB_1(x) = Relu(BN(Conv_1(x)))

[0087] CB_3(x) = Relu(BN(Conv_3(x)))

[0088] RCM(x) = x + CB_3(CB_1(x))

[0089] Wherein, x represents the model input, CB_1(x) represents the output feature of the 1*1 convolution block, CB_3(x) represents the output feature of the 3*3 convolution block, BN represents the batch normalization function, and Relu() represents the activation function layer.

[0090] The feature F passes through two residual convolution modules to obtain the feature output H of the channel 3×3 . Similarly, the feature F passes through two other 1*1 and 5*5 channels to obtain the feature outputs H 1×1 , H 5×5 of the respective channels.

[0091] Step S43: the outputs of the three parallel channels in step S42 are spliced to obtain the feature H. H is input into a Dropout layer to force the network to learn more robust feature representation.

[0092] H = cat(H 1×1 , H 3×3 , H 5×5 )

[0093] Wherein, cat() represents a feature splicing function.

[0094] Step S44: the output of step S43 is input into a fully connected layer to linearly combine the features of the previous layer and integrate high-level abstract features.

[0095] Step S45: The output of the step S45 fully connected layer is converted into a vector representing a probability distribution by a softmax activation function, which outputs a probability estimate for each element of the input vector, and is used with a cross-entropy loss function to measure the difference between the model output and the actual label, thereby performing gradient descent optimization. Specifically, the softmax function and the cross-entropy loss function are:

[0096]

[0097]

[0098] where x i is the output of the i-th neuron of the softmax layer, k is the total number of neurons; p(x) is the label of the training set, specifically the number 1 or 0, representing a high-quality or low-quality coke audio sample, respectively, and q(x) is the result of the model predicting the quality of the coke, also represented by the number 1 or 0. Step S5 tests the online coke audio sample based on the model parameters obtained by step S45 to identify the coke grade. Step S5 specifically includes:

[0099] Step S51: Real-time coke landing audio information is obtained from the coke audio generation equipment and transmitted to the cloud server;

[0100] Step S52: The audio information is preprocessed according to step S2, and the Mel spectrogram and gamma pass spectrogram features are extracted to obtain the final input model fusion features.

[0101] Step S53: The fusion features are input into the coke quality rating model trained in step S4 to obtain the predicted quality grade of the current coke sample. Good quality coke corresponds to "coke quality A grade", and bad quality coke corresponds to "coke quality B grade".

[0102] Step S6 displays the coke quality grade prediction result output by the model in step S5 on the terminal device. Step S6 specifically includes:

[0103] Step S61: The trained network parameter model is deployed to the cloud service. When the test sample is sent to the cloud server through the terminal after uploading the instruction, the server stores the audio sample in the cloud server and obtains the training result of step S5 in real time, i.e., the corresponding grade of the current coke sample, which is immediately returned to the terminal device in json format through the output interface.

[0104] Step S62: the terminal device, i.e., the mobile phone, performs field conversion processing on the json result sent by the server, so as to restore the result from the encrypted state to the actual field of the result. After the terminal device processes the result, the terminal device displays the result in a pop-up window, provides the user with an intuitive coke sample quality evaluation result, and allows the user to confirm the result, so as to complete the real-time coke sample grade evaluation.

[0105] The coke audio generation and collection device involved in the present application is described in further detail below.

[0106] Referring to Figure 2 , the device includes a door that can be opened along the side end, and is mainly used for facilitating the collection of coke and the cleaning of dust after the coke falls. The device has a box-shaped frame structure, and is provided with a handle for facilitating carrying. An opening for conveniently throwing coke is arranged on the top of the frame, and the opening is also convenient for observing the position and situation of the coke falling.

[0107] The opening is arranged on the plate body of the acrylic plate constituting the top of the box, so as to conveniently observe the position and situation of the coke falling. The service life of the acrylic plate can be prolonged by at least 3 years compared with other materials; the light transmittance of the acrylic plate can reach more than 92%, and the impact resistance is 16 times that of ordinary glass, which is very suitable for installation in an unsafe environment; and the acrylic plate has strong plasticity, high recyclability, and convenient maintenance.

[0108] A circular hole (i.e., the aforementioned opening) is arranged in the middle of the acrylic plate, and the hole diameter is 80-150 mm, so that the coke to be tested can fall through the hole and accurately fall in the center of the iron plate. The hole diameter cannot be too large or too small. If the hole diameter is too small, the coke with a large volume will be limited to fall by the too small hole diameter, and the coke will be knocked and the experimental effect will be affected; if the hole diameter is too large, the range of the coke falling will be greatly increased, which will affect the judgment of the coke quality.

[0109] Among them,

[0110] (1) Two small holes are arranged at a distance of 10 cm from the ground on the left and right sides of the cabinet / box (the side with the openable door is regarded as the front), so as to facilitate the wiring of the data line.

[0111] (2) A slot is arranged on the top of the cabinet / box, and the width of the slot is 0.8 cm. When the acrylic plate is inserted (the acrylic plate itself has a certain thickness, and the thickness of the customized acrylic plate is 0.6 cm, and the length and width of the acrylic plate are consistent with the length and width of the top of the cabinet, and are both 35 cm), 0.6 cm is slightly smaller than 0.8 cm, so as to reserve some space for the insertion of the acrylic plate.

[0112] (3) The two sides of the small hole of the cabinet are equipped with a pickup holder, which is a cuboid with only three faces (the upper and lower faces and the face close to the cabinet are removed). Since the purchased pickup has a holder, it can be clamped on the pickup holder. The distance of 2 cm between the pickup holder and the cabinet wall ensures that the pickup will not be in contact with the cabinet wall when clamped on the pickup holder, thereby affecting the recording effect.

[0113] (4) An openable door is installed on the front of the cabinet, and a door handle is arranged on the door to facilitate opening and closing of the door. The door is installed to facilitate the collection of coke falling on the iron plate and the cleaning of coke debris by the operator.

[0114] Table 1 below is a parameter table of the pickup.

[0115] Table 1:

[0116]

[0117] The frequency response range of the pickup is 30 Hz-18000 Hz, so the frequency range of the audio that can be collected is also relatively wide. The pickup can collect the sound of coke falling, whether it is crisp or dull. The noise reduction depth and signal-to-noise ratio also indicate that this pickup can avoid noise as much as possible. Generally, the larger the signal-to-noise ratio, the smaller the noise mixed in the signal, and the higher the sound quality of the sound playback, otherwise the opposite. Therefore, the larger signal-to-noise ratio of the pickup indicates that the noise mixed in the collected audio signal is smaller.

[0118] The sampling frequency of the pickup is 48KHZ, and the higher sampling rate can also ensure that the audio quality collected by the pickup and the difference between the audio and the actual original sound will be smaller. The endurance time of the pickup is 8 hours, which indicates that the pickup still has a relatively sufficient endurance time. The charging time ≤2.5H indicates that the charging efficiency of the pickup is relatively high. The longer endurance time and shorter charging time ensure the normal and efficient use of the pickup. The parameter table of the charging compartment is shown in Table 2 below.

[0119] Table 2:

[0120]

[0121] The audio transmitter and the audio receiver of the pickup can be charged as long as they are placed in the corresponding storage position. When collecting coke audio, the audio transmitter is placed on the pickup holder of the coke quality prediction device, and the audio receiver is connected to the mobile phone through the recording line (Android mobile phone uses Android recording line, and Apple mobile phone uses Apple recording line) to collect audio signals. The collected audio signals can be uploaded to the computer through the upload button of the WeChat applet for backup, and the applet also retains the collected audio signals.

[0122] Method for using coke quality prediction system

[0123] Record audio and classify as actual coke landing audio data, which can be applied to actual coke quality management.

[0124] When the user enters the coke quality management system interface, the interface will automatically render the interface, display the audio data storage box, the clear file button and the record audio button, and the interface is as shown in Figure 5 (a).

[0125] Then, when the user confirms the connection of the pickup, click the audio recording button, a start recording popup window will be displayed, and the popup window interface is as shown in Figure 5 (b). When the coke landing action is completed, click the recording button again, a recording success popup window will appear, as shown in Figure 5 (c), indicating that the recording audio operation is complete.

[0126] When the audio recording is successful, the audio storage box of the page will display the storage path, storage time and audio size of the stored audio in time sequence, as shown in Figure 6 (a). When the delete button of an audio file is clicked, the corresponding audio file will be deleted, and the corresponding audio data in the audio storage box has been deleted, as shown in Figure 6 (b). When the clear file button is clicked, all files in the audio storage box will be cleared and become blank, as shown in Figure 6 (c).

[0127] When the user clicks the audio analysis button, the interface will display a result popup window, if the test coke is A-class coke, the interface is as shown in Figure 7 (a) ; if the test coke is B-class coke, the interface is as shown in Figure 7 (b), when the user clicks the confirm button, the interface will return to the storage interface, indicating that the coke quality prediction analysis is complete.

[0128] The above is only a preferred embodiment of the present application, not to limit the scope of the present application, the above embodiment of the present application can be changed. That is, any simple, equivalent changes and modifications made according to the content of the claims and description of the present application, all fall within the scope of the claims of the present application. The present application is not described in detail, which is a conventional technical content.

Claims

1. A coke rating method based on a multi-scale residual neural network, characterized by, The method comprises the following steps: Step S1, providing a server, a terminal device and a quality management system, the terminal device comprising a hardware device, obtaining coke audio information by using the terminal device connected with the server; Step S2, obtaining and transmitting coke landing audio information by using the terminal device, transmitting the audio data to the server, analyzing each coke sample, extracting its log-mel spectrogram feature and gammatone spectrogram feature, and determining the unique feature information of coke of different grades; Step S3, using a coke component composition function and combining experience to label the coke with a quality grade label; Step S4, fusing the spectrogram features obtained in step S3, using the fused features to construct a coke quality rating model, inputting the features of all coke samples into the rating model, and training model parameters conforming to coke quality rating; Step S5, inputting the fused features of the coke audio into the coke quality rating model constructed in step S4, obtaining the quality evaluation result of the current coke sample, and generating the quality grade of the coke; Step S6, transmitting the evaluation result of step S5 to the display interface of the terminal device, displaying the final quality evaluation result through the display interface, and determining the quality grade of the specific coke.

2. The coke rating method based on a multi-scale residual neural network according to claim 1, wherein, In step S1, The terminal device comprises a coke sound generating device for generating coke audio information, a pickup for collecting coke collision audio, and a human-computer interaction terminal device for realizing deployment and result display of the coke management system; The coke audio information comprises audio information generated when the coke falls from a set height and collides with an iron plate.

3. The coke rating method based on a multi-scale residual neural network according to claim 1, wherein, The specific method of step S2 is: Step S21: according to the coke audio data set obtained in step S1, audio sample preprocessing is performed, first, the sample is subjected to effective audio segment judgment, and is cut out from the original audio paragraph to be an audio sample with a set number of sampling points, and then the audio sample quality is improved through a processing mode comprising pre-emphasis; Step S22: the length of the preprocessed audio sample is subjected to fast Fourier transform to obtain an output signal of amplitude information; Step S23: the output signal of step S22 is subjected to mel filter extraction to extract a mel spectrogram feature, and the mel filter bank is composed of s triangular filters, and the feature extraction formula of this step is: where k represents the kth dimension of the spectrogram, H m (k) represents the frequency response of the mth mel filter, and f(m) represents the center frequency of the mth filter. The frequency spectrum obtained after the mel filter bank is subjected to logarithmic operation, and the output is a log-mel spectrogram feature with a size of 222×128; Step S24: the output signal of step S22 is subjected to gammatone filter extraction to extract a gammatone spectrogram feature, and the gammatone filter bank is composed of n equal-height filters, and the feature extraction formula of this step is: where a is the amplitude of the Fourier-transformed signal, γ represents the order of the filter, f c is the center frequency of the filter, is the phase information of the audio signal, e is the exponential constant, e = 2.718; where ERB(f c ) is the gamma band frequency, which is an equivalent rectangular bandwidth frequency that determines the decay rate of the impulse and is related to the critical bandwidth of human hearing. The conversion formula between ERB and center frequency is After the output of the gammatone filter bank, the absolute value of the third cubic root is taken to obtain a gammatone spectrogram feature with a size of 222×128; Step S25: the output log-mel spectrogram feature of step S23 and the output gammatone spectrogram feature of step S24 are spliced and fused to form a high-level fusion feature with a size of 222×256, which is used as the input of the subsequent network model.

4. The coke rating method based on a multi-scale residual neural network according to claim 3, wherein, The specific method of step S4 is: Step S41: taking the original feature map of the output feature high-level fusion feature 222x256 in step S25 as input, performing dimension change and parameter reduction through a 1x1 residual convolution module, and realizing cross-channel information combination to enhance nonlinear features; Step S42: the output of step S41 is input into three parallel channels of 1x1, 3x3 and 5x5, and each channel contains two residual convolution modules with corresponding convolution kernel sizes, respectively. After passing through the convolution block, the batch normalization layer and the RELU layer are input in sequence to alleviate the problems of gradient disappearance and gradient explosion, improve the training stability, and improve the convergence speed; Step S43: the outputs of the three parallel channels in step S42 are input into the Dropout layer to make the network learn more robust feature representations; Step S44: the output of step S43 is input into the fully connected layer to linearly combine the features of the previous layer and integrate high-level abstract features; Step S45: the output of the fully connected layer is converted into a vector representing a probability distribution through the softmax activation function, which estimates the probability of each element of the input vector, and is used together with the cross-entropy loss function to measure the gap between the model output and the actual label, thereby performing gradient descent optimization.

5. The coke rating method based on a multi-scale residual neural network according to claim 1, wherein, The specific method of step S5 is: Step S51: transmitting the real-time coke audio information obtained from the coke sound generating device to the server; Step S52: preprocessing and standardizing the audio information; Step S53: performing feature analysis on the preprocessed audio and samples and inputting them into the coke quality rating model constructed in step S4 to obtain the quality grade of the current coke sample.

6. The coke rating method based on a multi-scale residual neural network according to claim 1, wherein, The specific method of step S6 is: Step S61: deploying the network model on the server and transmitting the output quality grade of step S5 to the human-computer interaction terminal device through the server interface; Step S62: the human-computer interaction terminal device acquires the output information sent by the server, performs field conversion processing, and displays the results through a pop-up window on the human-computer interaction terminal device to complete the coke sample grade evaluation.

7. The coke rating method based on a multi-scale residual neural network according to claim 1, wherein, The audio data is transmitted to the cloud server through cloud transmission, and cloud management, cloud storage and cloud analysis are performed on the cloud.

8. The coke rating method based on a multi-scale residual neural network according to claim 7, wherein, In step S1, the terminal device further includes a computer that implements cloud service functions, and the human-computer interaction terminal device is a handheld terminal device.

9. The coke rating method based on a multi-scale residual neural network according to claim 3, wherein, In step S21, the audio samples are unified to have a length of 1000 sampling points.

10. A coke rating system for implementing the coke rating method of any one of claims 1-9.

Images