Intelligent control system for monitoring data processing of coal mill based on multimode fusion voiceprint recognition

Abstract

This invention discloses an intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition, belonging to the field of industrial regulatory data processing and control technology. Through a joint noise reduction mechanism of wavelet packet multi-band decomposition, adaptive spectral subtraction, and attention masking, the signal-to-noise ratio of the voiceprint signal can be effectively improved, effectively solving the problem of voiceprint signal submersion under strong noise. Self-supervised pre-training on unlabeled noisy data enables the model to learn essential voiceprint features independent of noise, effectively reducing the noise contamination rate of traditional MFCC features. Optimizing the attention mask through pre-trained feature feedback effectively improves the model's ability to identify early weak fault features, providing highly reliable input for subsequent fault diagnosis models and solving the problem of high misjudgment rate in the early fault stage. By dynamically adjusting the Mel filter bank, fault features are stably extracted, solving the problem of feature distortion in traditional fixed MFCCs when operating conditions change.

Description

Technical Field

[0001] This invention relates to the field of industrial monitoring data processing and control technology, specifically to an intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition. Background Technology

[0002] Intelligent control for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition is essentially an industrial intelligent system that uses sound as its core, integrates multi-source information, and drives closed-loop control, aiming to improve the safety, reliability, and energy efficiency of coal mill operation.

[0003] Because the working environment of the coal mill has multiple strong noise sources such as fans, boilers, and coal conveyor belts, the background noise can reach more than 90dB. Moreover, the spectrum coverage is wide, the signal-to-noise ratio (SNR) of the voiceprint signal is extremely low, and traditional features such as MFCC and spectrogram are severely contaminated. The model has a high misjudgment rate, especially in the early stage of weak faults, there is a problem of voiceprint feature distortion under strong noise interference. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition, which solves the technical problem of voiceprint feature distortion under strong noise interference in existing technical solutions.

[0005] The objective of this invention can be achieved through the following technical solutions: The intelligent control system for coal mill monitoring data processing based on multi-modal fusion voiceprint recognition includes: Coal mill operation monitoring data processing and analysis module: Real-time acquisition of raw acoustic signature signals from coal mills and wavelet packet decomposition, sub-band level noise suppression, and secondary optimization of the denoised signal are performed. A label-free dataset containing the raw noise signal and the jointly denoised signal is constructed and self-supervised pre-training is performed. The robust features obtained from the pre-training are fed back and utilized to dynamically adjust the mask weights and enhance the identification of fault features. Feature extraction and optimization adaptive control module: Real-time acquisition and processing of current operating parameters; dynamic adjustment of the center frequency and bandwidth of the Mel filter bank based on the processed data; application of the adjusted parameters to the original Mel filter bank to obtain an adaptive filter; extraction of 13-dimensional MFCC feature values ​​based on the adaptive filter; data analysis of the extracted MFCC feature values; and dynamic control of the coal mill operation based on the analysis results.

[0006] Furthermore, the raw acoustic signature signal of the coal mill is acquired in real time, and the raw acoustic signature signal is decomposed into 6-layer wavelet packet decomposition, which decomposes the signal into different frequency bands and accurately separates the noise and potential fault characteristic frequency bands.

[0007] Furthermore, for each wavelet packet subband signal decomposed, its power spectrum is calculated, and the noise power spectrum is estimated using adaptive spectral subtraction to achieve subband-level noise suppression.

[0008] Furthermore, the power spectrum after noise reduction is converted into a time-domain signal by inverse Fourier transform to complete subband noise reduction; Based on the subband signal-to-noise ratio, the attention mask weights are calculated, and the denoised signal is optimized in a second step to generate a mask that highlights fault-related features and suppresses residual noise. The mask is applied to the subband noise-reducing signal to obtain the mask-optimized subband signal.

[0009] Furthermore, an unlabeled dataset containing the original noise signal and the jointly denoised signal was constructed, and the Wav2Vec 2.0 industrial version model was used for self-supervised pre-training to learn robust voiceprint feature representation in a strong noise environment. Furthermore, the robust features obtained from Wav2Vec 2.0 pre-training are embedded and fed back into the attention mask module to dynamically adjust the mask weights, thereby further enhancing the recognizability of fault features.

[0010] Furthermore, real-time data collection of current operating parameters, including coal mill load and primary air volume; The collected current operating condition parameters are normalized. Based on the normalized operating condition parameters, the center frequency and bandwidth of the Mel filter bank are dynamically adjusted, including the calculation of center frequency offset and bandwidth scaling factor.

[0011] Furthermore, when extracting 13-dimensional MFCC features, a first-order high-pass filter is applied to the denoised voiceprint signal to compensate for high-frequency signal attenuation, enhance the high-frequency components of the fault impact features, and output the pre-emphasized signal. The pre-emphasized signal is subjected to frame-by-frame windowing processing.

[0012] Furthermore, a 1024-point Fast Fourier Transform is performed on the windowed frame signal to convert it into a frequency domain signal and the power spectrum is calculated. The power spectrum is filtered by an adaptively adjusted Mel filter bank to simulate the characteristics of human hearing and output the Mel spectrum.

[0013] Furthermore, by taking the natural logarithm of the Mel spectrum, the dynamic range is compressed, simulating the logarithmic response of the human ear, and the logarithmic Mel spectrum is output. The log-Mel spectrum is subjected to discrete cosine transform, and features are extracted.

[0014] Furthermore, the first 13 dimensions are taken as the final features and output. The final extracted 13-dimensional MFCC feature values ​​are then input into the coal mill monitoring model for data analysis. Based on the analysis results, the operation of the coal mill is dynamically controlled.

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention effectively improves the signal-to-noise ratio of voiceprint signals by employing a joint denoising mechanism of wavelet packet multi-band decomposition, adaptive spectral subtraction, and attention masking, thus effectively solving the problem of voiceprint signal submersion under strong noise. Through self-supervised pre-training on unlabeled noisy data using the industrial version of Wav2Vec 2.0, the model learns essential voiceprint features independent of noise, effectively reducing the noise contamination rate of traditional MFCC features. Furthermore, by optimizing the attention mask through pre-trained feature feedback, the model's ability to identify early, weak fault features is effectively improved, providing highly reliable input for subsequent fault diagnosis models and addressing the problem of high misjudgment rates in the early fault stage.

[0016] This invention achieves stable fault feature extraction of MFCC features within the range of 30%~100% load and 15000~50000 m³ / h air volume in coal mills by dynamically adjusting the Mel filter bank, thus solving the problem of feature distortion in traditional fixed MFCCs when operating conditions change. By combining robust feature feedback pre-trained with Wav2Vec 2.0, it can effectively improve the anti-interference ability of adaptive MFCC features against strong noise of 80~100dB and the fault identification accuracy of adaptive MFCC features. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the operation of the intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition, as described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, the present invention is an intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition, including a coal mill operation monitoring data processing and analysis module and a feature extraction, optimization, and adaptive control module; The coal mill operation monitoring data processing and analysis module: Real-time acquisition of the raw acoustic signature signal from the coal mill, followed by wavelet packet decomposition, sub-band noise suppression, and secondary optimization of the denoised signal. This constructs an unlabeled dataset containing both the raw noise signal and the jointly denoised signal, and performs self-supervised pre-training. The robust features obtained from the pre-training are then used for feedback, dynamically adjusting the mask weights to enhance the identification of fault features. Specific steps include: The raw acoustic signature signal of the coal mill is acquired in real time, and a 6-level wavelet packet decomposition is performed on the raw acoustic signature signal to decompose the signal into different frequency bands, accurately separating noise and potential fault characteristic frequency bands. The expression involved in the decomposition is as follows: ;in, The signal amplitude of the k-th frequency band sub-band at time t in the j-th level wavelet packet decomposition; j is the wavelet packet decomposition level, which is fixed at 6 in this implementation, j∈[1,6]. 6 levels of decomposition correspond to the generation of... =64 sub-bands, covering the full audio range of 0~20kHz, matching the spectral coverage characteristics of strong noise above 90dB in the working environment of coal mill, namely fan noise 2~8kHz, boiler noise 0.5~2kHz, and coal conveyor belt noise 0~15kHz, ensuring that all noise source frequency bands are accurately separated; k is the sub-band number, which changes dynamically with the number of decomposition layers j, and the range is [1, The range of the 6-layer decomposition is [1, 64]. Each number corresponds to an independent frequency band. For example, the first sub-band of the 6th layer corresponds to 0~312.5Hz, and the 16th sub-band corresponds to 4.6875~5kHz. This enables precise positioning of different noise source frequency bands, which facilitates subsequent enhancement of low-frequency sub-bands with concentrated fault characteristics. These are the time-domain sampled values ​​of the original voiceprint signal. The time integral variable represents the original voiceprint signal at a historical moment. The sampling amplitude is input as the raw acoustic data collected on-site at the coal mill; is the db8 wavelet packet basis function, corresponding to the wavelet analysis basis function of the kth sub-band of the jth layer. The db8 wavelet is selected because it has good temporal locality and frequency resolution, and can accurately capture the low-frequency impact acoustic features generated by early faults in coal mills, such as the periodic impact sound of grinding roller wear, while effectively suppressing high-frequency noise interference; t is the variable at the current time. This is the convolution operator; For each wavelet subband signal of the decomposition Calculate its power spectrum The noise power spectrum is estimated using an adaptive spectral subtraction method. To achieve subband-level noise suppression, the relevant expressions are as follows: ;in, The power spectrum of the subband after denoising represents the frequency domain power distribution of the k-th wavelet packet subband in the j-th layer after spectral subtraction denoising. It is the core input for the subsequent inverse Fourier transform to generate the denoised time domain signal. The original subband power spectrum is obtained by performing a fast Fourier transform on the subband signal obtained from wavelet packet decomposition. For angular frequency variables, , This is the actual frequency; This is an adaptive over-subtraction factor used to control the subtraction amplitude of the noise power spectrum. Its value is dynamically adjusted according to the subband signal-to-noise ratio (SNR): when the subband SNR < 5dB, it is a frequency band dominated by strong noise, such as the high-frequency subband where wind turbine noise is concentrated. A value of 2.0~2.5 enhances noise suppression; when the sub-band SNR ≥ 5dB, it corresponds to the mid-to-low frequency sub-band where fault characteristics are concentrated. The value should be between 1.2 and 1.5 to avoid excessive noise reduction that could distort fault characteristics. For the estimated subband noise power spectrum, , This is the smoothing factor, with a default value of 0.95. The lower limit protection factor is fixed at 0.05 and is used to constrain the minimum power spectrum after noise reduction to prevent the fault features from being completely submerged by noise and the useful signal from being excessively subtracted when the original signal power is lower than the noise power, such as in the early stage of weak faults. The maximum value function is used to ensure that the power spectrum after noise reduction is not lower than the lower limit protection value, thus avoiding the occurrence of a negative power spectrum. The denoised power spectrum is converted into a time-domain signal using inverse Fourier transform. Complete subband noise reduction; Based on the subband signal-to-noise ratio, the attention mask weights are calculated, and the denoised signal is then further optimized to generate a mask that highlights fault-related features and suppresses residual noise. The expression for mask generation is as follows: ;in, The attention mask weights for the k-th subband of the j-th layer are defined, with values ​​ranging from [0,1]. This is the gain coefficient, fixed at 0.5, which controls the mask's sensitivity to changes in signal-to-noise ratio. Let be the signal-to-noise ratio of the k-th subband in the j-th layer; The signal-to-noise ratio threshold is set to 5dB. When the subband SNR is higher than the threshold, the mask approaches 1 to preserve the signal; when it is lower than the threshold, the mask approaches 0 to suppress noise. Applying the mask to the subband noise-reduced signal yields the mask-optimized subband signal. After reconstruction, the final speaker signal after joint noise reduction is obtained. ; Furthermore, an unlabeled dataset containing the original noise signal and the jointly denoised signal was constructed. The Wav2Vec 2.0 industrial version model was used for self-supervised pre-training to learn robust speaker signature representations under strong noise conditions. The pre-training objective function is as follows: ;in, The loss function is for contrastive learning; B is the batch size, which is fixed at 32 in industrial scenarios; i is the sample index, representing the i-th training sample in the batch, i∈[1,B]. Each sample consists of the original noise signal segment (negative sample) and the signal segment after joint noise reduction (positive sample), forming a sample pair for contrastive learning, allowing the model to learn to distinguish between noise and effective voiceprint features. For in-batch sample summation, the loss values ​​of all samples within a training batch are accumulated to calculate the batch average loss, which guides the update of model parameters; To compare feature segments, from the input signal The 100ms duration voiceprint segment randomly extracted from the data corresponds to the typical time scale of coal mill fault characteristics. For example, the periodic impact interval of grinding roller wear is usually 50~200ms. For voiceprint feature embedding, the Wav2Vec2.0 industrial encoder processes the input signal The encoded high-dimensional feature vector has a fixed dimension of 768. The training input samples consist of two classes of unlabeled data: positive samples and negative samples. The positive samples are segments of coal mill acoustic signature signals after joint noise reduction, while the negative samples are segments of pure noise signals collected on-site from the coal mill, such as the independent noise from the fan, boiler, and coal conveyor belt. It should be explained that Wav2Vec 2.0 Industrial Edition is a dedicated voiceprint feature extraction and fault diagnosis model based on the Wav2Vec 2.0 self-supervised speech model open-sourced by Facebook AI Research. It is designed to address the pain points of industrial voiceprint monitoring scenarios (such as coal mills, fans, pumps, etc.) such as strong noise, variable operating conditions, and imbalanced samples. It achieves high-precision and high-robustness industrial equipment status monitoring by self-supervised pre-learning of general features of industrial voiceprints and then fine-tuning with a small number of labeled fault samples.

[0021] In this embodiment of the invention, the model uses a 12-layer Transformer encoder, with an input of 80-dimensional Mel spectrum, a frame length of 25ms, a frame shift of 10ms, 100k pre-training iterations, and a learning rate of 5e-5. The robust features obtained from the pre-trained Wav2Vec 2.0 industrial version model are embedded and fed back into the attention mask module. The mask weights are dynamically adjusted to further enhance the identification of fault features. The optimization formula is as follows: ;in, The optimized subband attention mask weights represent the final mask weights of the k-th wavelet packet subband at the j-th layer after feature feedback adjustment. The initial attention mask weights are preset based on the subband signal-to-noise ratio (SNR), with values ​​ranging from 0 to 1. When the subband SNR < 5dB, the initial weights are set to 0.4; when the subband SNR ≥ 5dB, the initial weights are set to 0.8. The Sigmoid activation function is used to map the input values ​​to the 0~1 range and convert the cosine similarity result into a mask scaling factor to ensure that the weight adjustment range is within a reasonable range. This is the temperature regulation coefficient, dynamically set to 1.5~2.5 in industrial scenarios. It is used to adjust the sensitivity of the cosine similarity. When the ambient noise intensity is ≥95dB, the value is 2.5 to enhance the response to weak fault characteristics; when the noise intensity is ≤85dB, the value is 1.5 to avoid feature distortion caused by over-adjustment. This is the function for calculating cosine similarity. For feature embedding of the masked subband, the signal of the k-th subband at the j-th level after wavelet packet decomposition is... Input the Wav2Vec 2.0 pre-trained model to obtain a 768-dimensional high-dimensional feature vector; To reference fault feature embedding, feature templates were obtained by encoding the acoustic fingerprint data of typical coal mill faults through a Wav2Vec 2.0 pre-trained model; typical coal mill faults include grinding roller wear, liner detachment, and coal block jamming.

[0022] In this embodiment of the invention, a joint denoising mechanism combining wavelet packet multi-band decomposition, adaptive spectral subtraction, and attention masking effectively improves the signal-to-noise ratio of the voiceprint signal, effectively solving the problem of voiceprint signal submersion under strong noise. Through self-supervised pre-training of the Wav2Vec 2.0 industrial version model on unlabeled noisy data, the model learns essential voiceprint features independent of noise, effectively reducing the noise contamination rate of traditional MFCC features. Optimizing the attention mask through pre-trained feature feedback effectively improves the model's ability to identify early, weak fault features, providing highly reliable input for subsequent fault diagnosis models and solving the problem of high misjudgment rates in the early fault stage.

[0023] Feature extraction and optimization adaptive control module: This module collects and processes current operating parameters in real time. Based on the processed data, it dynamically adjusts the center frequency and bandwidth of the Mel filter bank. The adjusted parameters are then applied to the original Mel filter bank to obtain an adaptive filter. Based on this adaptive filter, 13-dimensional MFCC feature values ​​are extracted. Data analysis is performed on the extracted MFCC feature values, and the operation of the coal mill is dynamically controlled based on the analysis results. Specific steps include: The coal mill DCS control system collects current operating parameters in real time, including coal mill load and primary air volume; the coal mill load reflects the actual coal processing capacity of the coal mill; and the primary air volume reflects the ventilation intensity inside the coal mill. The collected operating parameters are mapped to the [0,1] interval to eliminate the difference in dimensions. This is a conventional technical solution. The specific implementation steps will not be elaborated here. When dynamically adjusting the center frequency and bandwidth of the Mel filter bank based on normalized operating condition parameters, the calculation includes center frequency offset and bandwidth scaling factor. The expression for calculating the center frequency offset is as follows: ;in, Let m be the center frequency offset of the m-th Mel filter; m is the Mel filter. This is the influence coefficient of the load on the center frequency, with a default value of 200. This is the influence coefficient of air volume on center frequency, with a default value of 150. These are the normalized load data and air volume data, respectively. The expression for calculating the bandwidth scaling factor is: ;in, Let be the bandwidth scaling factor for the m-th Mel filter; It should be noted that the higher the load, the wider the fault characteristic bandwidth, and the greater the filter bandwidth needs to be; the larger the air volume, the wider the airflow noise bandwidth, and the greater the filter bandwidth needs to be compressed to focus the fault characteristics.

[0024] Applying the adjusted parameters to the original Mel filter bank yields an adaptive filter, with the following expression: ; ;in, This is the center frequency of the original Mel filter; This represents the original Mel filter bandwidth. , The center frequency and bandwidth are adaptively adjusted. Based on the adjusted Mel filter bank, when extracting 13-dimensional MFCC features, a first-order high-pass filter is applied to the denoised acoustic signature signal to compensate for high-frequency signal attenuation, enhance the high-frequency components of the fault impulse features, and output the pre-emphasized signal. The expression involved is: ;in, This is the noise-reduced voiceprint signal; n is the sampling point number, n=0, 1, ..., N-1, N=44100 corresponds to a 1-second signal; settings The first sampling point has no prior data, so the original value is directly retained; It should be explained that the pre-emphasis coefficient is set to 0.97, which is optimized based on the high-frequency attenuation characteristics of the voice signal. The attenuation is 6dB per octave, which increases the gain of the high-frequency components above 5kHz of the coal mill's acoustic signal by about 6dB. This can effectively enhance the impact sound characteristics of faults such as grinding roller wear and liner detachment.

[0025] For the pre-emphasis signal Frame segmentation and windowing processing is performed, with a frame length of L = 0.025 × 44100 = 1102 points, corresponding to a duration of 25ms, covering the typical duration of 10~30ms of coal mill fault impact; Frame shift S = 0.01 × 44100 = 441 points, corresponding to a duration of 10ms, with an inter-frame overlap rate of approximately 56%, thus preventing the loss of fault features across frames; Total frames ; Applying a Hamming window to each frame of the signal reduces spectral leakage; the relevant expression is: ;in, The value of the window function for the Hamming window represents the window function weight corresponding to the nth sampling point, and its value ranges from 0.08 to 1.0. The frame signal after windowing is ;in, The frame signal after windowing represents the first frame. The signal value of the nth sampling point in the frame; The frame index ranges from 0, 1, ..., M−1, where M is the total number of frames, and M represents the sampling point of the original signal after pre-emphasis. Locate the specific sampling point in the original signal, where It is the first The position of the starting sampling point of the frame; Perform a 1024-point Fast Fourier Transform (FFT) on the windowed frame signal to convert it into a frequency domain signal and calculate the power spectrum. The relevant expressions are as follows: ;in, Let be the frequency domain complex signal after FFT transformation, representing the th The frame signal is in the first The frequency domain components of each frequency point consist of a real part and an imaginary part. The real part corresponds to the in-phase component of the signal, and the imaginary part corresponds to the quadrature component. The modulus of the complex number represents the signal amplitude at that frequency point, and the amplitude represents the phase information, which is the basis for subsequent calculation of the power spectrum. This is a frequency index, with values ​​ranging from 1, 2, ..., 1023, corresponding to 1024 frequency points after the FFT transformation. Each index corresponds to an actual physical frequency. , Sampling rate; ;in, The power spectral density value represents the power spectrum value of the first power spectrum. Frame number Signal energy at each frequency point; The power spectrum is filtered by an adaptively adjusted Mel filter bank to simulate the characteristics of human hearing, and the Mel spectrum is output. The relevant expression is: ;in, This is the Mel spectrum, with a dimension of 24, corresponding to 24 Mel filters; For adaptively adjusted Mel filter banks, Each Mel filter is triangular, with parameters satisfying: center frequency The system adaptively adjusts based on current operating conditions, favoring low frequencies under low loads and high frequencies under high loads; bandwidth... Adaptive adjustment based on current operating conditions; high load amplification bandwidth; high air volume compression bandwidth; filter overlap rate, with approximately 50% overlap of the -3dB bandwidth between adjacent filters to ensure characteristic continuity; 24 Mel filters: covering 0~20kHz, with the low-frequency (0~5kHz) filter density being twice that of the high-frequency (10~20kHz) filter density, matching the characteristics of coal mill faults that are mainly concentrated in the low frequency range. Taking the natural logarithm of the Mel spectrum to compress the dynamic range, simulating the logarithmic response of the human ear, and outputting the logarithmic Mel spectrum, the relevant expression is: ;in, This is a correction factor, with a value of [value missing]. To avoid the fact that the logarithm is meaningless when the Mel spectrum is 0; It is a logarithmic function; It should be noted that by converting the linear energy of the Mel spectrum into logarithmic energy to match the logarithmic perception of sound intensity by the human ear, the difference between subtle fault characteristics and strong noise can be made more apparent.

[0026] The discrete cosine transform is performed on the log-Mel spectrum, and features are extracted. The relevant expression is as follows: ;in, The final extracted MFCC feature value represents the th... The d-th dimension MFCC feature of the frame acoustic signature signal is the core input feature of the coal mill fault diagnosis model. The 13-dimensional MFCC feature retains only 1.3% of the original Mel spectrum dimension, but contains more than 90% of the fault information, which greatly reduces the computational load of industrial edge equipment. The orthogonal basis functions of DCT are constructed using the orthogonal transformation matrix built by the cosine function, which transforms the highly correlated Mel spectrum into independent MFCC features. 2f+1 is used to ensure the symmetry of the basis functions, avoid the DC component shift after transformation, and ensure the stability of the MFCC features. It needs to be explained that MFCC features, specifically Mel-Frequency Cepstral Coefficients (MFCC), are a voiceprint feature extraction algorithm that simulates the auditory characteristics of the human ear. By mapping linear frequencies to Mel frequencies that conform to human auditory perception, and then performing cepstral transformation, low-dimensional and robust voiceprint features are obtained. It is widely used in fields such as speech recognition and industrial equipment fault diagnosis (e.g., voiceprint monitoring of coal mills). It should be noted that by using Discrete Cosine Transform (DCT), redundant information in the Mel spectrum can be removed, and highly correlated Mel spectra can be converted into orthogonal MFCC features, reducing the feature dimensionality and solving the problem of feature instability of traditional spectral features when the operating conditions of the coal mill are variable. The output 13-dimensional features reduce the feature dimensionality from 24 to 13 while ensuring the completeness of fault information. In coal mill fault diagnosis, the first 13 dimensions of MFCC features contain more than 96% of the fault feature energy, while the subsequent dimensions are mainly high-frequency noise details, contributing less than 1% to fault identification. Therefore, choosing 13 dimensions can reduce the feature dimensionality from 24 to 13 while ensuring diagnostic accuracy, improving computational efficiency by 50%. The first 13 dimensions are taken as the final feature value output, which contains more than 90% of the feature energy. The final 13-dimensional MFCC feature values ​​are then input into the coal mill monitoring model for data analysis. Based on the analysis results, the operation of the coal mill is dynamically controlled. Among them, the coal mill monitoring model is an existing conventional model, which is constructed by processing and training the historical fault data of the coal mill using machine learning algorithms. The specific machine learning algorithms can be random forest or XGBoost. Dynamic control of the operation of the coal mill, including but not limited to maintaining the operation of the coal mill, automatically adjusting the opening of the cold / hot air damper (temperature control), and linking the start and stop of the coal feeder to start the inerting system or perform emergency tripping.

[0027] It should be noted that the training, construction, and data analysis of the coal mill monitoring model, as well as the dynamic control scheme for coal mill operation, are all existing conventional technical means, and the specific implementation steps will not be elaborated here.

[0028] In this embodiment of the invention, by dynamically adjusting the Mel filter bank, the MFCC features can stably extract fault features within the range of 30%~100% load of the coal mill and an air volume of 15,000~50,000 m³ / h, thus solving the problem of feature distortion of traditional fixed MFCCs when operating conditions change. By combining robust feature feedback pre-trained with Wav2Vec 2.0, the anti-interference ability of adaptive MFCC features against strong noise of 80~100dB and the fault identification accuracy of adaptive MFCC features can be effectively improved.

[0029] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0030] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0031] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0032] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A coal mill monitoring data processing intelligent control system based on multi-mode fusion voiceprint recognition, characterized in that, include: Coal mill operation monitoring data processing and analysis module: Real-time acquisition of raw acoustic signature signals from coal mills and wavelet packet decomposition, sub-band level noise suppression, and secondary optimization of the denoised signal are performed. A label-free dataset containing the raw noise signal and the jointly denoised signal is constructed and self-supervised pre-training is performed. The robust features obtained from the pre-training are fed back and utilized to dynamically adjust the mask weights and enhance the identification of fault features. Feature extraction and optimization adaptive control module: Real-time acquisition and processing of current operating parameters; dynamic adjustment of the center frequency and bandwidth of the Mel filter bank based on the processed data; application of the adjusted parameters to the original Mel filter bank to obtain an adaptive filter; extraction of 13-dimensional MFCC feature values ​​based on the adaptive filter; data analysis of the extracted MFCC feature values; and dynamic control of the coal mill operation based on the analysis results.

2. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 1, characterized in that, The raw acoustic signature signal of the coal mill is acquired in real time, and the raw acoustic signature signal is decomposed into 6-layer wavelet packet decomposition, which decomposes the signal into different frequency bands and accurately separates the noise and potential fault characteristic frequency bands.

3. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 2, characterized in that, For each wavelet packet subband signal decomposed, its power spectrum is calculated, and the noise power spectrum is estimated by adaptive spectral subtraction to achieve subband-level noise suppression.

4. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 3, characterized in that, Subband denoising is achieved by converting the denoised power spectrum into a time-domain signal using inverse Fourier transform. Based on the subband signal-to-noise ratio, the attention mask weights are calculated, and the denoised signal is optimized in a second step to generate a mask that highlights fault-related features and suppresses residual noise. The mask is applied to the subband noise-reducing signal to obtain the mask-optimized subband signal.

5. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 4, characterized in that, An unlabeled dataset containing the original noise signal and the jointly denoised signal was constructed. The Wav2Vec2.0 industrial version model was used for self-supervised pre-training to learn robust voiceprint feature representation in a strong noise environment. Furthermore, the robust features obtained from Wav2Vec 2.0 pre-training are embedded and fed back into the attention mask module to dynamically adjust the mask weights, thereby further enhancing the recognizability of fault features.

6. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 5, characterized in that, Real-time collection of current operating parameters, including coal mill load and primary air volume; The collected current operating condition parameters are normalized. Based on the normalized operating condition parameters, the center frequency and bandwidth of the Mel filter bank are dynamically adjusted, including the calculation of center frequency offset and bandwidth scaling factor.

7. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 6, characterized in that, When extracting 13-dimensional MFCC features, a first-order high-pass filter is applied to the noise-reduced acoustic signature signal to compensate for high-frequency signal attenuation, enhance the high-frequency components of the fault impact features, and output the pre-emphasized signal. The pre-emphasized signal is subjected to frame-by-frame windowing processing.

8. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 7, characterized in that, Perform a 1024-point Fast Fourier Transform on the windowed frame signal to convert it into a frequency domain signal and calculate the power spectrum; The power spectrum is filtered by an adaptively adjusted Mel filter bank to simulate the characteristics of human hearing and output the Mel spectrum.

9. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 8, characterized in that, By taking the natural logarithm of the Mel spectrum, the dynamic range is compressed, simulating the logarithmic response of the human ear, and the logarithmic Mel spectrum is output. The log-Mel spectrum is subjected to discrete cosine transform, and features are extracted.

10. The intelligent control system for coal mill monitoring data processing based on multi-mode fusion voiceprint recognition according to claim 9, characterized in that, The first 13 dimensions are taken as the final features and output. The final 13-dimensional MFCC feature values ​​are then input into the coal mill monitoring model for data analysis. Based on the analysis results, the operation of the coal mill is dynamically controlled.

Images