Sound signal noise reduction method for printing equipment

Abstract

The invention discloses a sound signal noise reduction method for printing equipment. The sound signal noise reduction method specifically comprises the following steps: S1, collecting a sound signal of the printing equipment by using a sound sensor; s2, the collected sound signals are preprocessed; s3, constructing a denoising model based on the frequency domain cyclic convolutional neural network; s4, inputting the preprocessed sound signals into a denoising model for denoising; S5, constructing a denoising evaluation mechanism, evaluating the denoised sound signals, and returning the sound signals which do not accord with the evaluation result to S4 for denoising again. The problem that in the prior art, due to the fact that background noise is too much when sound signals of printing equipment are collected, printing and packaging equipment is difficult to achieve fault diagnosis accurately is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of printing and packaging equipment detection, and in particular relates to a sound signal noise reduction method for printing equipment. Background Art

[0002] With the continuous progress of modern science and the improvement of industrial automation level, mechanical equipment is developing towards high precision, high efficiency and high automation. As the core mechanical equipment in the printing and packaging industry, the main function of the printing press is to transfer graphic information to the substrate, thereby realizing the copying and dissemination of information. It is widely used in advertising, packaging and decoration and other fields. Due to the complex structure and high integration of the printing press, the probability of mechanical failure is greatly increased. Once the equipment fails, it will inevitably affect production efficiency, resulting in huge economic losses. Therefore, how to monitor the operating status of printing equipment, discover and diagnose mechanical failures, and further ensure the efficient and high-quality output of printing production has become an important task of industrial maintenance management.

[0003] Vibration analysis is a common method for fault diagnosis of printing equipment. It can detect abnormal conditions of equipment by extracting vibration signal characteristics and comparing different fault types, and accurately identify the fault type. However, this method still has certain limitations: First, to collect vibration data, the acceleration sensor needs to be in close contact with the monitored components such as bearings and gears, and it is necessary to punch holes or paste on the surface of the printing equipment to fix the acceleration sensor, which not only affects production efficiency, but may also damage the acceleration sensor due to high-speed operation. Therefore, it cannot be deployed under many working conditions; second, due to the limited internal space of printing machinery and equipment, the installation position of the acceleration sensor is limited, making it difficult to effectively monitor bearings, gears, etc. at key locations such as rollers and spindles, further exacerbating the difficulty of signal acquisition.

[0004] As another important information during the operation of the equipment, the sound signal can accurately identify the fault type by analyzing the specific frequency in the signal, so as to achieve early warning and rapid processing of the fault. Compared with vibration signals, sound signals have advantages that vibration signals do not have: first, sound signals have the characteristics of non-contact acquisition, which can make up for the shortcomings of vibration signals; second, sound signals are highly sensitive and can capture the sound components generated by mechanical failures, so they are widely used in fault diagnosis. However, due to the high sensitivity of sound signals, they will be interfered by various noises in the project, resulting in invalid diagnosis results. Since there are many mechanical parts in printing equipment and the printing production speed is high, the mechanical structure coupling is serious, and there are many contacts and impacts between the mechanisms, which will generate serious noise and cause the failure of sound monitoring diagnosis. Therefore, it is very important to filter and denoise the collected signals to obtain pure and valuable fault characteristic signals. Summary of the invention

[0005] The purpose of the present invention is to provide a sound signal noise reduction method for printing equipment, which solves the problem in the prior art that when collecting sound signals of printing equipment, the printing and packaging equipment is difficult to accurately perform fault diagnosis due to excessive background noise.

[0006] The technical solution adopted by the present invention is a method for reducing noise of sound signals of printing equipment, which specifically comprises the following steps: S1, using a sound sensor to collect sound signals of printing equipment; S2, preprocessing the collected sound signal; S3, builds a denoising model based on frequency domain recurrent convolutional neural network; S4, inputting the preprocessed sound signal into a denoising model for denoising; S5, build a denoising evaluation mechanism to evaluate the denoised sound signal, and the sound signal that does not meet the evaluation results is returned to S4 for further denoising.

[0007] The present invention is also characterized in that The sound signal of the printing equipment collected by the sound sensor in S1 is a pure signal. Different noises are added to the pure signal to form a filtered signal. The noise includes human speech noise, Gaussian white noise, factory noise, vehicle horn noise, and engine noise.

[0008] The preprocessing in S2 includes pre-emphasis, framing and windowing of the filtered signal.

[0009] The denoising model in S3 specifically includes a denoising input layer, a pre-processing module, an encoder module, a decoder module and a denoising output layer. Two layers of feedforward sequential memory network modules for capturing time series characteristics are inserted between the encoder module and the decoder module. A mask generation module based on the tanh activation function is set after the decoder module.

[0010] The encoder module consists of three layers of convolution-residual modules. Each convolution-residual module consists of a convolution layer and a residual connection, which is used to preserve the original information when extracting local features. The decoder module also contains three convolution-residual modules, which are used to reconstruct the denoised frequency domain signal. Each convolution-residual module of the encoder module is connected to an SE attention module, and the three convolution-residual modules in the encoder module are respectively transmitted to the three convolution-residual modules corresponding to the decoder module through the SE attention module.

[0011] The feedforward sequential memory network module includes an input layer, a hidden layer, a memory module, and an output layer connected in sequence; Eigenvector Input layer, hidden layer is used to receive input feature vector And through linear transformation and activation function processing to obtain , the memory module is used to store historical information, which is passed to the next feedforward sequential memory network layer module or as the input of the decoder through the horizontal connection to the output layer output of the current layer; The calculation formula is as follows ; In the formula, W is the weight matrix, b is the bias term, f is the tanh activation function; The activation function calculation formula is as follows: ; In the formula, e is the base of natural logarithms.

[0012] S4 specifically includes the following steps: S41, dividing the sound signal preprocessed by S2 into a training set and a test set, and inputting the filtered signal containing noise in the training set into the denoising input layer; S42, using the short-time Fourier transform in the pre-processing module to convert the time domain representation of the sound signal into the frequency domain representation and generate a complex spectrum diagram ; The specific formula is as follows: ; In the formula, is the complex spectrum, t represents the time frame, f Indicates frequency, is the window function; S43, after the sound signal is processed by the encoder module and the decoder module, the result is passed to the mask generation module to generate two masks M1 and M2 , and combine the element-wise addition into the final mask M ; The specific formula is as follows: ; S44, the final mask after merging M Applied to the original complex spectrogram , forming the frequency domain signal after filtering and denoising, as shown in the following formula: ; In the formula, represents the frequency domain signal after filtering and denoising, Represents element-wise multiplication; S45, using inverse short-time Fourier transform to convert the filtered and denoised frequency domain signal back to the time domain signal, as shown in the following formula: ; In the formula, It is the time domain representation of the sound signal after filtering and denoising; S46, after multiple iterations of training, the trained optimal model is saved, and then the sample data of the test set is input into the trained optimal model for filtering and denoising to obtain a filtered signal after filtering and denoising.

[0013] S5 specifically includes the following steps: S51, transmitting the filtered signal and the clean signal output by S46 to the Mel filter, and performing Mel cepstral coefficient feature extraction respectively, as shown in the following formula: ; Where DCT stands for discrete cosine transform; Then extract the first three order Mel cepstrum matrices of the filtered signal and the clean signal respectively; S52, respectively calculating the mean matrices of the first-order to third-order Mel-frequency cepstral coefficients of the clean signal and the filtered signal, and calculating the logarithmic power spectrum; S53, the clean signal and the filtered signal are respectively merged by the mean matrix features, and the merged matrix is ​​transposed so that the time step is the first dimension, and the result is normalized; S54, for the two results in S53, a denoising evaluation mechanism is established using the cosine similarity method to evaluate the quality of the sound denoising effect and set a standard value. When the evaluation result does not meet the standard value, return to S4 for denoising again, as shown in the following formula: ; Where A and B are the characteristic matrices of the signal before and after filtering and denoising, respectively. represents the dot product, Represents the norm.

[0014] The calculation formula of the mean matrix in S52 is as follows: ; Where N is the number of time frames, It is i Mel cepstral coefficients of order; The calculation formula of logarithmic power spectrum amplifier is as follows: .

[0015] The calculation formula for merging feature matrices in S53 is as follows: ; In the formula, Indicates vertical splicing; The calculation formulas for transposition and normalization are as follows: .

[0016] The beneficial effects of the present invention are: (1) The method for reducing the noise of sound signals in printing equipment of the present invention introduces a frequency domain recurrent convolutional neural network, converts the signal to the frequency domain for analysis through Fourier transform, extracts features based on multiple convolution-residual modules in the encoder module, and further processes these features through the feedforward sequential memory network and SE attention module. Then, the convolution-residual module in the decoder module decodes the features, and finally, the processed frequency domain signal is converted back to the time domain through the inverse short-time Fourier transform to obtain the denoised acoustic signal. This method can effectively separate noise and signal in the frequency domain, while using the memory capacity of the recurrent neural network to capture the temporal dependency, and relying on the automatic feature extraction capability of the convolutional neural network to directly extract the optimal features from the original audio signal. Effective filtering of complex background noise and accurate extraction of sound features are achieved.

[0017] (2) The sound signal noise reduction method for printing equipment of the present invention adopts a non-contact sound signal acquisition method and utilizes a high-sensitivity sound sensor to collect data. The installation position of the sensor can be flexibly selected, and effective fault monitoring can be achieved without changing the structure of the existing equipment. The need to punch holes or stick sensors on the equipment for vibration signal acquisition is avoided, and the normal operation of the printing equipment is not affected. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow chart of a method for reducing noise of sound signals of printing equipment according to the present invention; Figure 2 It is a structural diagram of a denoising model in a method for reducing noise of a sound signal of printing equipment according to the present invention; Figure 3 It is a structural diagram of a feed-forward sequential memory network module in a sound signal noise reduction method for printing equipment according to the present invention; Figure 4 It is a schematic diagram of the process of S5 in the method for reducing noise of sound signals for printing equipment of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0020] Sound filtering is the process of removing or attenuating unwanted sound components through mathematical algorithms so that the target sound can be heard more clearly. Common sound filtering methods include frequency domain filtering, time domain filtering, and other methods based on statistical models. The working environment of printing presses is complex, and the sources of noise are diverse, including mechanical vibration, motor noise, paper friction, etc. These noises have a wide frequency range and time-varying characteristics. Although common methods can improve signal quality, it is difficult to effectively separate fault characteristic signals.

[0021] In recent years, the development of deep learning technology has provided new ideas for the processing and fault diagnosis of sound signals. The deep learning model has the ability to automatically extract features, which can effectively extract the spatial features and noise features in the signal and capture the time dependency in the signal. At the same time, relying on its good generalization ability, it can be applied to the processing of printing press noise with time-varying characteristics. By training and learning the complex and non-stationary noise signals of the printing press, effective suppression and enhancement of the target signal can be achieved. Therefore, the use of deep learning models for filtering and denoising can more effectively separate fault feature signals, improve the operating efficiency of the printing press and the accuracy of fault diagnosis.

[0022] The present invention aims to provide a method for reducing the noise of sound signals for printing equipment, which effectively filters complex background noise and accurately extracts sound features by introducing a frequency domain recurrent convolutional neural network. The specific scheme is shown in the following embodiments.

[0023] Example 1 The present invention is used for a method for reducing noise of sound signals of printing equipment, such as Figure 1 As shown, the specific steps include: S1, using a sound sensor to collect sound signals of printing equipment; Furthermore, the sound signal of the printing device collected by the sound sensor in step S1 is a pure signal, and different noises are added to the pure signal to form a filtered signal.

[0024] Specifically, the noise includes human speech noise, Gaussian white noise, factory noise, vehicle horn noise, and engine noise.

[0025] S2, preprocessing the collected sound signal; Furthermore, the preprocessing includes pre-emphasis, framing and windowing of the filtered signal.

[0026] S3, builds a denoising model based on frequency domain recurrent convolutional neural network; S4, inputting the preprocessed sound signal into a denoising model for denoising; S5, construct a denoising evaluation mechanism to evaluate the denoised sound signal, and the sound signal that does not meet the evaluation results is returned to step S4 for further denoising.

[0027] Example 2 This embodiment is based on embodiment 1. Figure 2 As shown, the denoising model in step S3 of the sound signal denoising method for printing equipment of the present invention specifically includes a denoising input layer, a pre-processing module, an encoder module, a decoder module and a denoising output layer, two layers of feedforward sequential memory network modules for capturing time series characteristics are inserted between the encoder module and the decoder module, and a mask generation module based on a tanh activation function is arranged after the decoder module.

[0028] Furthermore, the encoder module consists of three layers of convolution-residual modules, each of which consists of a convolution layer and a residual connection, which is used to preserve the original information when extracting local features; the decoder module also contains three convolution-residual modules, which are used to reconstruct the denoised frequency domain signal; Furthermore, a Squeeze-and-Excitation attention module is used after each convolution-residual module of the encoder module to enhance the focus on important features and reduce the influence of irrelevant features. The three convolution-residual modules in the encoder module are respectively transmitted to the three convolution-residual modules corresponding to the decoder module through the SE attention module, forming a one-to-one correspondence.

[0029] Example 3 This embodiment is based on the embodiment 2. Figure 3 As shown, the feedforward sequential memory network module in the sound signal denoising method for printing equipment of the present invention comprises an input layer, a hidden layer, a memory module and an output layer connected in sequence; Eigenvector Input layer, hidden layer is used to receive input feature vector And through linear transformation and activation function processing to obtain , the memory module is used to store historical information, which is passed to the next feedforward sequential memory network layer module or as the input of the decoder through the horizontal connection to the output layer output of the current layer; The calculation formula is as follows ; In the formula, W is the weight matrix, b is the bias term, f is the tanh activation function; The activation function calculation formula is as follows: ; In the formula, e is the base of natural logarithms.

[0030] Example 4 In this embodiment, based on the first embodiment, step S4 of the method for reducing noise of sound signals of printing equipment of the present invention specifically includes the following steps: S41, dividing the sound signal preprocessed by S2 into a training set and a test set, and inputting the filtered signal containing noise in the training set into the denoising input layer; S42, using the short-time Fourier transform in the pre-processing module to convert the time domain representation of the sound signal into the frequency domain representation and generate a complex spectrum diagram ; The specific formula is as follows: ; In the formula, is the complex spectrum, t represents the time frame, f Indicates frequency, is the window function; S43, after the sound signal is processed by the encoder module and the decoder module, the result is passed to the mask generation module to generate two masks M1 and M2 , and combine the element-wise addition into the final mask M ; The specific formula is as follows: ; S44, the final mask after merging M Applied to the original complex spectrogram , forming the frequency domain signal after filtering and denoising, as shown in the following formula: ; In the formula, represents the frequency domain signal after filtering and denoising, Represents element-wise multiplication; S45, using inverse short-time Fourier transform to convert the filtered and denoised frequency domain signal back to the time domain signal, as shown in the following formula: ; In the formula, It is the time domain representation of the sound signal after filtering and denoising.

[0031] S46, after multiple iterations of training, the trained optimal model is saved, and then the sample data of the test set is input into the trained optimal model for filtering and denoising to obtain a filtered signal after filtering and denoising.

[0032] Example 5 This embodiment is based on embodiment 1. Figure 4 As shown, step S5 in the method for reducing noise of sound signals of printing equipment of the present invention specifically comprises the following steps: S51, transmitting the filtered signal and the clean signal outputted in step S46 to the Mel filter, and extracting Mel cepstral coefficient features respectively, as shown in the following formula: ; Where DCT stands for discrete cosine transform; Then extract the first three order Mel cepstrum matrices of the filtered signal and the clean signal respectively; S52, respectively calculating the mean matrices of the first-order to third-order Mel-frequency cepstral coefficients of the clean signal and the filtered signal, and calculating the logarithmic power spectrum; Furthermore, the calculation formula of the mean matrix is ​​as follows: ; Where N is the number of time frames, It is i Mel cepstral coefficients of order; The calculation formula of logarithmic power spectrum amplifier is as follows: .

[0033] S53, the clean signal and the filtered signal are respectively merged by the mean matrix features, and the merged matrix is ​​transposed so that the time step is the first dimension, and the result is normalized; Furthermore, the calculation formula for feature matrix merging is as follows: ; In the formula, Indicates vertical splicing; The calculation formulas for transposition and normalization are as follows: .

[0034] S54. For the two results in step S53, i.e., the result of the pure signal processed in step S53 and the result of the filtered signal processed in step S53, a denoising evaluation mechanism is established using the cosine similarity method, and the value range of cosine similarity is between -1 and 1. The closer the value is to 1, the more similar the two vectors are. The quality of the sound denoising effect is evaluated, and a standard value is set. When the evaluation result does not meet the standard value, return to S4 for denoising again, as shown in the following formula: ; Where A and B are the characteristic matrices of the signal before and after filtering and denoising, respectively. represents the dot product, Represents the norm.

[0035] Example 6 The sound signal noise reduction method for printing equipment in this embodiment specifically includes the following steps: S1, using a sound sensor to collect 500 fault sound signals of printing equipment as pure signals, and adding different noises to the pure signals to form filtered signals; S2, preprocessing the collected sound signal; Specifically, first, the filtered signal is pre-emphasized, and the pre-emphasis coefficient is selected as 0.92 to enhance some high-frequency signals that are attenuated during the acquisition process due to factors such as the transmission channel. Then, the non-stationary continuous sound signal is divided into multiple short-term stationary signals of 40 milliseconds per frame. Finally, each frame is windowed using a Hamming window to reduce signal discontinuity at the frame boundary.

[0036] S3, builds a denoising model based on frequency domain recurrent convolutional neural network; S4, inputting the preprocessed sound signal into a denoising model for denoising; Specifically, S41, divide the sound signal (i.e., filtered signal) preprocessed in step S2 into a training set and a test set, with the ratio of the training set to the test set being 8:2. Input the filtered signal containing noise in the training set to the input layer, set the step size to 1, and the number of training rounds to 150; S42, using the short-time Fourier transform in the pre-processing module to convert the time domain representation of the sound signal into the frequency domain representation and generate a complex spectrum diagram ; S43, after the sound signal is processed by the encoder module and the decoder module, the result is passed to the mask generation module to generate two masks M1 and M2 , and combine the element-wise addition into the final mask M ; S44, the final mask after merging M Applied to the original complex spectrogram , forming a frequency domain signal after filtering and denoising; S45, converting the filtered and denoised frequency domain signal back into a time domain signal using an inverse short-time Fourier transform; S46, after multiple iterations of training, the trained optimal model is saved. Then the sample data of the test set is input into the trained optimal model for filtering and denoising to obtain a sound signal after filtering and denoising, which is used as the filtered signal in step S5.

[0037] S5, constructing a denoising evaluation mechanism to evaluate the denoised sound signal, and returning the sound signal that does not meet the evaluation results to step S4 for further denoising; Specifically, S51, transmitting the filtered signal after short-time Fourier transformation (i.e., the sound signal output in step S46) and the clean signal to a Mel filter, and performing Mel cepstral coefficient feature extraction on each of them; Select a 40 millisecond sound signal after framing, and the first three-order Mel cepstrum matrices of the extracted filtered signal are as follows: ; The first three-order Mel cepstrum matrices of the extracted clean signal are as follows: ; S52, calculating the mean matrix of the first-order to third-order Mel-frequency cepstral coefficients for the clean signal and the filtered signal respectively, and calculating the logarithmic power spectrum; Since the length of the selected signal is one frame, the mean matrix is ​​actually the data of this frame itself. The mean matrices of the filtered signal and the pure signal are shown below.

[0038] Filtered signal mean matrix: ; Clean signal mean matrix: ; The logarithmic power spectrum feature is calculated by taking the logarithm of the spectrum. Here, the spectrum data of the filtered signal and the pure signal are: , .

[0039] The logarithmic power spectrum characteristics of the two signals are calculated as follows: Logarithmic power spectrum characteristics of the filtered signal: ; Pure signal logarithmic power spectrum characteristics: .

[0040] S53, the clean signal and the filtered signal are respectively combined by the mean matrix features; The matrix after filtering signal features are merged is: ; The matrix after the pure signal features are merged is: ; The combined matrix is ​​transposed so that the time step is the first dimension, and the result is normalized; the maximum-minimum normalization is used here, that is, each element is subtracted from the minimum value of the column and then divided by the difference between the maximum and minimum values ​​of the column.

[0041] The normalized filtered signal matrix A: ; Normalized clean signal matrix B: .

[0042] S54: For the two results in step S53, a denoising evaluation mechanism is established using the cosine similarity method to evaluate the quality of the sound denoising effect, and the cosine similarity result is calculated to be 0.9958. The result is greater than 0.94, indicating that the filtering denoising effect is good.

[0043] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0044] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reducing noise of sound signals for printing equipment, characterized in that: The specific steps include: S1, using a sound sensor to collect sound signals of printing equipment; S2, preprocessing the collected sound signal; S3, builds a denoising model based on frequency domain recurrent convolutional neural network; S4, inputting the preprocessed sound signal into a denoising model for denoising; S5, build a denoising evaluation mechanism to evaluate the denoised sound signal, and the sound signal that does not meet the evaluation results is returned to S4 for further denoising.

2. The method for reducing noise of sound signals for printing equipment according to claim 1, characterized in that: The sound signal of the printing equipment collected by the sound sensor in S1 is a pure signal. Different noises are added to the pure signal to form a filtered signal. Different noises are added to the pure signal to form a filtered signal. The noise includes human speech noise, Gaussian white noise, factory noise, vehicle horn noise, and engine noise.

3. The method for reducing noise of sound signals for printing equipment according to claim 2, characterized in that: The preprocessing in S2 includes pre-emphasis, frame division and windowing of the filtered signal.

4. The method for reducing noise of sound signals for printing equipment according to claim 1, characterized in that: The denoising model in S3 specifically includes a denoising input layer, a pre-processing module, an encoder module, a decoder module and a denoising output layer. Two layers of feedforward sequential memory network modules for capturing time series characteristics are inserted between the encoder module and the decoder module. A mask generation module based on a tanh activation function is arranged after the decoder module.

5. The method for reducing noise of sound signals for printing equipment according to claim 4, characterized in that: The encoder module is composed of three layers of convolution-residual modules, each of which is composed of a convolution layer and a residual connection, and is used to maintain the original information when extracting local features; The decoder module also contains three convolution-residual modules to reconstruct the denoised frequency domain signal; Each convolution-residual module of the encoder module is connected to an SE attention module, and the three convolution-residual modules in the encoder module are respectively transmitted to the three convolution-residual modules corresponding to the decoder module through the SE attention module.

6. The method for reducing noise of sound signals for printing equipment according to claim 4, characterized in that: The feedforward sequential memory network module comprises an input layer, a hidden layer, a memory module and an output layer connected in sequence; The feature vector Input layer, hidden layer is used to receive input feature vector And through linear transformation and activation function processing to obtain , the memory module is used to store historical information, which is passed to the next feedforward sequential memory network layer module or as the input of the decoder through the horizontal connection to the output layer output of the current layer; Said The calculation formula is as follows: ； In the formula, W is the weight matrix, b is the bias term, f is the tanh activation function; The activation function calculation formula is as follows: ； In the formula, e is the base of natural logarithms.

7. The method for reducing noise of sound signals for printing equipment according to claim 2, characterized in that: The S4 specifically comprises the following steps: S41, dividing the sound signal preprocessed by S2 into a training set and a test set, and inputting the filtered signal containing noise in the training set into the denoising input layer; S42, using the short-time Fourier transform in the pre-processing module to convert the time domain representation of the sound signal into the frequency domain representation and generate a complex spectrum diagram ; The specific formula is as follows: ； In the formula, is the complex spectrum, t represents the time frame, f Indicates frequency, is the window function; S43, after the sound signal is processed by the encoder module and the decoder module, the result is passed to the mask generation module to generate two masks M1 and M2 , and combine the element-wise addition into the final mask M ; The specific formula is as follows: ； S44, the final mask after merging M Applied to the original complex spectrogram , forming the frequency domain signal after filtering and denoising, as shown in the following formula: ； In the formula, represents the frequency domain signal after filtering and denoising, Represents element-wise multiplication; S45, using inverse short-time Fourier transform to convert the filtered and denoised frequency domain signal back to the time domain signal, as shown in the following formula: ； In the formula, It is the time domain representation of the sound signal after filtering and denoising; S46, after multiple iterations of training, the trained optimal model is saved, and then the sample data of the test set is input into the trained optimal model for filtering and denoising to obtain a filtered signal after filtering and denoising.

8. The method for reducing noise of sound signals for printing equipment according to claim 7, characterized in that: The S5 specifically includes the following steps: S51, transmitting the filtered signal and the clean signal output by S46 to the Mel filter, and performing Mel cepstral coefficient feature extraction respectively, as shown in the following formula: ； Where DCT stands for discrete cosine transform; Then extract the first three order Mel cepstrum matrices of the filtered signal and the clean signal respectively; S52, respectively calculating the mean matrices of the first-order to third-order Mel-frequency cepstral coefficients of the clean signal and the filtered signal, and calculating the logarithmic power spectrum; S53, the clean signal and the filtered signal are respectively merged by the mean matrix features, and the merged matrix is ​​transposed so that the time step is the first dimension, and the result is normalized; S54, for the two results in S53, a denoising evaluation mechanism is established using the cosine similarity method to evaluate the quality of the sound denoising effect and set a standard value. When the evaluation result does not meet the standard value, return to S4 for denoising again, as shown in the following formula: ； Where A and B are the characteristic matrices of the signal before and after filtering and denoising, respectively. represents the dot product, Represents the norm.

9. The method for reducing noise of sound signals for printing equipment according to claim 8, characterized in that: The calculation formula of the mean matrix in S52 is as follows: ； Where N is the number of time frames, It is i Mel cepstral coefficients of order; The calculation formula of logarithmic power spectrum amplifier is as follows: 。 10. The method for reducing noise of sound signals for printing equipment according to claim 8, characterized in that: The calculation formula for merging the feature matrices in S53 is as follows: ； In the formula, Indicates vertical splicing; The calculation formulas for transposition and normalization are as follows: 。

Images