Flotation process condition recognition method based on audio signal characteristics

Abstract

The present invention provides a method for identifying working conditions of a flotation process based on audio signal features, including: step 1, data preparation and data preprocessing, specifically including data acquisition, noise reduction and working condition classification; step 2, according to the physical meaning of the frequency represented by the vertical axis of the Mel spectrogram, amplifying the proportion of key frequency bands in the flotation audio signal, reducing the influence of unimportant frequency bands, and constructing a flotation audio Mel spectrogram based on a feature attention mechanism. The present invention observes the influence of different frequency bands in the Mel spectrogram on the recognition result, finds out the key frequency bands, constructs a flotation audio Mel spectrogram based on an attention mechanism, performs preliminary feature extraction, introduces transfer learning when establishing the model, and selects the most likely test set data to be correctly identified through the output characteristics of the model itself and adds pseudo labels, and aligns the marginal distribution and conditional distribution of the data to ensure that a good classification boundary can still be generated during migration, so as to achieve the purpose of improving the generalization performance of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of flotation process data analysis, and in particular to a flotation process working condition identification method based on audio signal characteristics. Background Art

[0002] Flotation is widely used in industries such as non-ferrous metals, non-metallic minerals and coal. It is a mineral separation method that uses the different physical and chemical properties of the mineral surface to separate minerals. It achieves the purpose of separating useful minerals from impurities by selectively attaching useful minerals to air bubbles in the slurry and then floating them to the surface of the slurry. Flotation is a key process in many industrial processes, but it is not easy to control the effect of flotation. This is because there are many factors that affect the flotation effect and adjustable variables in the operation. The flotation process has always been adjusted manually. At present, most on-site workers adjust the production process based on their own experience and the test results of the final product grade. However, this rough adjustment mode first requires the on-site operators to have very strong professionalism; second, the operation is highly subjective. Different workers will have completely different operations under the same conditions, and the working conditions change frequently, resulting in unstable product quality; third, the results of offline tests cannot reflect the current flotation effect in real time, and there is a large time lag.

[0003] Online identification of working conditions is a prerequisite for achieving objective operation optimization. In the past, most of the research on online working condition identification of flotation processes focused on foam image analysis. However, in the actual flotation process, in addition to observing the shape of the foam, on-site operators also need to obtain important information, which is the sound produced when the scraper scrapes the foam and the scraped foam falls into the bottom tank. Because the sound signals corresponding to different working conditions are different, there are currently few studies using audio signals for working condition identification. In some studies using public data sets for audio classification, due to different research objects, it is often necessary to pay attention to the information of all frequency bands in the spectrogram. In contrast, the audio signals generated during the flotation process have their own unique properties. The differences in audio signals generated by different working conditions may only be concentrated in a certain frequency band. How to better capture the tiny features that produce differences between different categories and amplify and eliminate the interference of irrelevant features is crucial to improving the model recognition performance.

[0004] In addition, actual industrial data sets often have different distributions for the training set and test set of the working condition recognition model due to the small amount of training set data or the time series of the data, differences in environment, noise, operation, etc. This causes a model that performs well on the training set to perform poorly on the test set, resulting in overfitting. For the problems described above, transfer learning can be used to improve the generalization performance of the model. The usual transfer learning method only focuses on the marginal distribution between the two domain data, but only considering the marginal distribution does not necessarily produce a favorable classification boundary. How to reasonably incorporate the conditional distribution of the data is the key to further improving the performance of the model. Summary of the invention

[0005] The present invention provides a flotation process condition recognition method based on audio signal features, which aims to solve the problem that traditional flotation process condition recognition methods cannot capture the tiny features that produce differences between different categories and there is a data set drift between the training set and the test set.

[0006] In order to achieve the above object, an embodiment of the present invention provides a method for identifying working conditions of a flotation process based on audio signal characteristics, comprising:

[0007] Step 1: Data preparation and data preprocessing, including data acquisition, noise reduction, working condition classification, and data division into training set and test set;

[0008] Step 2: Based on the physical meaning of the frequency represented by the vertical axis of the Mel spectrogram, the proportion of the key frequency band in the flotation audio signal is amplified, the influence of the unimportant frequency band is reduced, and the flotation audio Mel spectrogram based on the feature attention mechanism is constructed;

[0009] Step 3: Build a deep convolutional network model, use the flotation audio Mel spectrogram based on the feature attention mechanism as the input of the deep convolutional network model, use the classification loss of the training set and the maximum mean difference between the training set and the test set as the loss, train the deep convolutional network model to automatically learn and extract deep audio signal features, and finally input the extracted deep features into the classifier for classification;

[0010] Step 4: Input the unlabeled test set data into the deep convolutional network model, select some data in the recognition results and label them with pseudo labels, and retrain the deep convolutional network model using the test set data and training set data with pseudo labels;

[0011] Step 5: Repeat step 4 until the specified number of times is reached, and use the deep convolutional network model obtained last time as the flotation process working condition identification model.

[0012] Wherein, the step 1 specifically includes:

[0013] Step 11, using a microphone to collect audio signals generated when the scraper scrapes the foam and the scraped foam falls into the bottom tank during the flotation process, collect two minutes of audio every hour, and record the foam concentration and grade at the corresponding time as the label of the audio data;

[0014] Step 12, designing a high-pass filter with a preset cutoff frequency to achieve smooth signals above the cutoff frequency of the high-pass filter and rapid attenuation of signals below the cutoff frequency of the high-pass filter;

[0015] Step 13, segment each filtered audio signal. Multiple short audio segments obtained by segmenting the same audio segment correspond to the same data label. According to the labels corresponding to all audios, they are divided into N categories. Corresponding to N different working conditions in the production process, all the obtained data are divided into training sets and test sets.

[0016] Wherein, the step 12 specifically includes:

[0017] Step 121, performing FFT transformation on the original signal;

[0018] Step 122, filtering is performed using the following formula:

[0019]

[0020] Among them, |Y(ω)| * represents the amplitude of the signal after filtering, |Y(ω)| represents the amplitude of the signal before filtering, and ω represents the frequency;

[0021] Step 123, perform inverse FFT transform to obtain the filtered signal.

[0022] Wherein, the step 2 specifically includes:

[0023] Step 21, convert each flotation audio signal into a Mel spectrogram and establish a deep convolutional network model, use the obtained Mel spectrogram as the input of the deep convolutional network model, use the training set to train the deep convolutional network model and perform working condition recognition on the test set data, and select the deep convolutional network model with the highest recognition accuracy of the test set from different deep convolutional network models obtained by multiple sets of hyperparameter training as the benchmark model;

[0024] Step 22, masks are added to different frequency bands in the Mel spectrogram to shield the features of some frequency bands, and the masked Mel spectrogram is used as input to train a new deep convolutional network model. The accuracy of the new deep convolutional network model on the test set is compared with the results of the baseline model to determine the impact of shielding different frequency bands on the accuracy.

[0025] Wherein, the step 2 further comprises:

[0026] Step 23, the frequency band with the greatest impact on accuracy is used as the key feature, and the frequency band with less impact is used as the secondary feature;

[0027] Step 24, in the process of converting the audio signal into a mel spectrogram, weighting the key features, enhancing the proportion of the key features and weakening the influence of the secondary features.

[0028] Wherein, the step 3 specifically includes:

[0029] Step 31, construct a deep convolutional network model, whose output is an N-dimensional vector normalized by a Softmax function containing a temperature coefficient, and each element of the N-dimensional vector represents the prediction probability of the deep convolutional network model for each class; wherein the expression of the Softmax function containing a temperature coefficient is Among them, z i represents the i-th element of the vector before Softmax normalization, N represents the number of categories, and T represents the temperature coefficient;

[0030] Step 32: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n is the number of training samples, represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample.

[0031] Wherein, the step 3 specifically includes:

[0032] Step 33: Input the training set data and the test set data into the constructed deep convolutional network model and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector;

[0033] Step 34, calculate the total loss Among them, λ represents the weight hyperparameter;

[0034] Step 35, train the constructed deep convolutional network model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the constructed deep convolutional network model.

[0035] Wherein, the step 4 specifically includes:

[0036] Step 41, setting a probability threshold θ. When the maximum Softmax output of a certain test set data in the current deep convolutional network model is greater than θ, it is determined that the recognition result of the current deep convolutional network model for the test set data is correct, and the recognition result output by the current deep convolutional network model is used as a pseudo label for the data.

[0037] Step 42: Input the pseudo-labeled test set data into the current deep convolutional network model and calculate the cross entropy loss function of the test set data. Among them, k represents the number of test samples with pseudo labels, represents the pseudo label of the i-th test sample, Represents the recognition result of the current model for the i-th test sample.

[0038] Wherein, the step 4 further comprises:

[0039] Step 43: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n represents the number of training samples. represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample;

[0040] Step 44: Input the training set data and the test set data into the constructed deep convolutional network model, and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector.

[0041] Wherein, the step 4 further comprises:

[0042] Step 45, recalculate the total loss Among them, γ and λ represent weight hyperparameters;

[0043] Step 46, training the model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the model;

[0044] Step 47, jump to step 41, repeat step 41 to step 46 until a specified number of times is reached, and a set of working condition identification models under hyperparameters including T, θ, γ and λ are obtained;

[0045] Step 48, resetting multiple sets of hyperparameters, repeating step 47 until a specified number of times is reached, obtaining different models based on multiple sets of hyperparameter training, and selecting the model with the highest recognition accuracy of the test set data as the flotation process condition recognition model based on the obtained different models.

[0046] The above scheme of the present invention has the following beneficial effects:

[0047] The flotation process condition identification method based on audio signal characteristics described in the above embodiment of the present invention,

[0048] (1) The present invention combines the on-site experience of operators in actual production and uses flotation audio signals, which have rarely been studied in previous studies, to identify working conditions. This method does not interfere with the industrial process, and the information is easy to obtain at a low cost, thus broadening the channels for obtaining information in the industry. The identification results also provide important information for the subsequent optimization of industrial production operations.

[0049] (2) The present invention does not require an accurate process mechanism model. It only needs to collect enough audio signal data and corresponding labels to use the powerful fitting ability of deep learning for learning, avoiding the tedious manual design of features and reducing the difficulty of modeling.

[0050] (3) The present invention focuses on the key frequency band features that are beneficial to recognition accuracy by constructing a feature attention mechanism, and reduces the proportion of non-key features. This can not only reduce the feature dimension and greatly reduce the amount of calculation, but also improve the generalization ability of the model and the recognition accuracy of the model;

[0051] (4) The present invention uses transfer learning to solve the problem of data set offset, which is essentially a direct learning method. The model will be updated as the test set changes. Compared with the inductive learning method that only uses the training set to train the model, it has better generalization performance and is more in line with industrial practice, which can significantly improve the performance of the model on the test set;

[0052] (5) The present invention utilizes the output characteristics of the test data in the model itself, and selects samples with similar distribution to the training data through the output of Softmax and adds pseudo-labels. The conditional distribution of the test data is taken into account during the migration. As the model continues to iterate and learn, more and more test data will be accurately pseudo-labeled, and better recognition accuracy can be obtained than conventional transfer learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the present invention;

[0054] Figure 2 It is a schematic diagram of the flotation process of the present invention;

[0055] Figure 3 The audio waveform diagram before and after filtering of the present invention;

[0056] Figure 4 A schematic diagram of the deep convolutional neural network structure of the present invention;

[0057] Figure 5 This is the Mel spectrogram before and after adding the mask according to the present invention. DETAILED DESCRIPTION

[0058] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0059] Aiming at the problem that the existing flotation process condition recognition method cannot capture the tiny features that produce the differences between different categories and there is a data set drift between the training set and the test set, the present invention provides a flotation process condition recognition method based on audio signal features.

[0060] like Figures 1 to 5 As shown, an embodiment of the present invention provides a method for identifying working conditions of a flotation process based on audio signal features, including: step 1, data preparation and data preprocessing, specifically including data acquisition, noise reduction, working condition classification, and data division of a training set and a test set; step 2, based on the physical meaning of the frequency represented by the vertical axis of the Mel spectrogram, amplifying the proportion of key frequency bands in the flotation audio signal, reducing the influence of unimportant frequency bands, and constructing a flotation audio Mel spectrogram based on a feature attention mechanism; step 3, constructing a deep convolutional network model, using the flotation audio Mel spectrogram based on the feature attention mechanism as the deep convolutional network model. The input of the model is taken as the classification loss of the training set and the maximum mean difference between the training set and the test set as the loss, the deep convolutional network model is trained to automatically learn to extract deep audio signal features, and finally the extracted deep features are input into the classifier for classification; step 4, the unlabeled test set data is input into the deep convolutional network model, part of the data in the recognition result is selected to be pseudo-labeled, and the deep convolutional network model is retrained using the test set data and the training set data with pseudo-labels; step 5, step 4 is repeated until the specified number of times is reached, and the deep convolutional network model obtained for the last time is used as the flotation process working condition recognition model.

[0061] The flotation process condition identification method based on audio signal characteristics described in the above embodiment of the present invention is very important to the quality of the final product in the potash fertilizer production process of the enterprise. Figure 1It can be seen that the raw ore is flotated after being decomposed in the decomposition tank. The flotation process mainly includes primary roughing, secondary roughing and fine selection. Its purpose is to select the useful potassium ions in the minerals through roughing and fine selection, and avoid mixing useless sodium ions into the product. According to the actual on-site investigation, a key flotation tank with obvious sound changes in the roughing is selected as the recording point. In order to achieve objective, real-time and accurate judgment of the current flotation process conditions, provide conditions for optimizing the flotation process.

[0062] Wherein, the step 1 specifically includes:

[0063] Step 11, using a microphone to collect audio signals generated when the scraper scrapes the foam and the scraped foam falls into the bottom tank during the flotation process, collect two minutes of audio every hour, and record the foam concentration and grade at the corresponding time as the label of the audio data;

[0064] Step 12, designing a high-pass filter with a preset cutoff frequency to achieve smooth signals above the cutoff frequency of the high-pass filter and rapid attenuation of signals below the cutoff frequency of the high-pass filter;

[0065] Step 13, segment each filtered audio signal. Multiple short audio segments obtained by segmenting the same audio segment correspond to the same data label. According to the labels corresponding to all audios, they are divided into N categories. Corresponding to N different working conditions in the production process, all the obtained data are divided into training sets and test sets.

[0066] Wherein, the step 12 specifically includes:

[0067] Step 121, performing FFT transformation on the original signal;

[0068] Step 122, filtering is performed using the following formula:

[0069]

[0070] Among them, |Y(ω)| * represents the amplitude of the signal after filtering, |Y(ω)| represents the amplitude of the signal before filtering, and ω represents the frequency;

[0071] Step 123, perform inverse FFT transform to obtain the filtered signal.

[0072] Wherein, the step 2 specifically includes:

[0073] Step 21, convert each flotation audio signal into a Mel spectrogram and establish a deep convolutional network model, use the obtained Mel spectrogram as the input of the deep convolutional network model, use the training set to train the deep convolutional network model and perform working condition recognition on the test set data, and select the deep convolutional network model with the highest recognition accuracy of the test set from different deep convolutional network models obtained by multiple sets of hyperparameter training as the benchmark model;

[0074] Step 22, masks are added to different frequency bands in the Mel spectrogram to shield the features of some frequency bands, and the masked Mel spectrogram is used as input to train a new deep convolutional network model. The accuracy of the new deep convolutional network model on the test set is compared with the results of the baseline model to determine the impact of shielding different frequency bands on the accuracy.

[0075] Wherein, the step 2 further comprises:

[0076] Step 23, the frequency band with the greatest impact on accuracy is used as the key feature, and the frequency band with less impact is used as the secondary feature;

[0077] Step 24, in the process of converting the audio signal into a mel spectrogram, weighting the key features, enhancing the proportion of the key features and weakening the influence of the secondary features.

[0078] The flotation process condition identification method based on audio signal characteristics described in the above embodiment of the present invention converts each flotation audio signal into a Mel spectrogram, wherein the conversion frequency range is 0 to 22.05kHz, and the specific conversion process includes: inputting the denoised speech signal, performing pre-emphasis, framing and windowing; performing FFT on each frame signal to obtain its frequency spectrum, and then obtaining its amplitude spectrum; filtering the amplitude spectrum through a Mel filter group; taking the logarithm of the output of the filter group to obtain the final spectrogram.

[0079] Wherein, the step 3 specifically includes:

[0080] Step 31, construct a deep convolutional network model, whose output is an N-dimensional vector normalized by a Softmax function containing a temperature coefficient, and each element of the N-dimensional vector represents the prediction probability of the deep convolutional network model for each class; wherein the expression of the Softmax function containing a temperature coefficient is Among them, z i represents the i-th element of the vector before Softmax normalization, N represents the number of categories, and T represents the temperature coefficient;

[0081] Step 32: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n is the number of training samples, represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample.

[0082] Wherein, the step 3 specifically includes:

[0083] Step 33: Input the training set data and the test set data into the constructed deep convolutional network model and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector;

[0084] Step 34, calculate the total loss Among them, λ represents the weight hyperparameter;

[0085] Step 35, train the constructed deep convolutional network model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the constructed deep convolutional network model.

[0086] Wherein, the step 4 specifically includes:

[0087] Step 41, setting a probability threshold θ. When the maximum Softmax output of a certain test set data in the current deep convolutional network model is greater than θ, it is determined that the recognition result of the current deep convolutional network model for the test set data is correct, and the recognition result output by the current deep convolutional network model is used as a pseudo label for the data.

[0088] Step 42: Input the pseudo-labeled test set data into the current deep convolutional network model and calculate the cross entropy loss function of the test set data. Among them, k represents the number of test samples with pseudo labels, represents the pseudo label of the i-th test sample, Represents the recognition result of the current model for the i-th test sample.

[0089] Wherein, the step 4 further comprises:

[0090] Step 43: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n represents the number of training samples. represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample;

[0091] Step 44: Input the training set data and the test set data into the constructed deep convolutional network model, and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector.

[0092] Wherein, the step 4 further comprises:

[0093] Step 45, recalculate the total loss Among them, γ and λ represent weight hyperparameters;

[0094] Step 46, training the model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the model;

[0095] Step 47, jump to step 41, repeat step 41 to step 46 until a specified number of times is reached, and a set of working condition identification models under hyperparameters including T, θ, γ and λ are obtained;

[0096] Step 48, resetting multiple sets of hyperparameters, repeating step 47 until a specified number of times is reached, obtaining different models based on multiple sets of hyperparameter training, and selecting the model with the highest recognition accuracy of the test set data as the flotation process condition recognition model based on the obtained different models.

[0097] Table 1 shows the condition recognition accuracy of the same model under multiple training strategies for the flotation process condition recognition method based on audio signal features described in the above-mentioned embodiment of the present invention. It can be seen that when only ordinary Mel spectrogram is used as input and transfer learning is not used, the model generalization performance is poor and the accuracy is low; when the Mel spectrogram is reconstructed as input based on the feature attention mechanism, the model accuracy is greatly improved; on the other hand, when only the maximum mean difference loss is added to align the marginal distribution between the two data sets, the model accuracy is also significantly improved; when the test set data is pseudo-labeled using the flotation process condition recognition method based on audio signal features and the marginal distribution and conditional distribution of the data are aligned, the model accuracy is further improved; finally, the highest accuracy is obtained when the feature attention mechanism is combined with the proposed transfer learning method to model, which verifies the effect of the proposed method.

[0098] Table 1 Comparison of working condition identification results

[0099]

[0100]

[0101] In summary, the flotation process condition identification method based on audio signal characteristics has the following advantages:

[0102] (1) The present invention combines the on-site experience of operators in actual production and uses flotation audio signals, which have rarely been studied in previous studies, to identify working conditions. This method does not interfere with the industrial process, and the information is easy to obtain at a low cost, thus broadening the channels for obtaining information in the industry. The identification results also provide important information for the subsequent optimization of industrial production operations.

[0103] (2) The present invention does not require an accurate process mechanism model. It only needs to collect enough audio signal data and corresponding labels to use the powerful fitting ability of deep learning for learning, avoiding the tedious manual design of features and reducing the difficulty of modeling.

[0104] (3) The present invention focuses on the key frequency band features that are beneficial to recognition accuracy by constructing a feature attention mechanism, and reduces the proportion of non-key features. This can not only reduce the feature dimension and greatly reduce the amount of calculation, but also improve the generalization ability of the model and the recognition accuracy of the model;

[0105] (4) The present invention uses transfer learning to solve the problem of data set offset, which is essentially a direct learning method. The model will be updated as the test set changes. Compared with the inductive learning method that only uses the training set to train the model, it has better generalization performance and is more in line with industrial practice, which can significantly improve the performance of the model on the test set;

[0106] (5) The present invention utilizes the output characteristics of the test data in the model itself, and selects samples with similar distribution to the training data through the output of Softmax and adds pseudo-labels. The conditional distribution of the test data is taken into account during the migration. As the model continues to iterate and learn, more and more test data will be accurately pseudo-labeled, and better recognition accuracy can be obtained than conventional transfer learning.

[0107] The flotation process operating condition identification method based on audio signal features described in the above-mentioned embodiment of the present invention performs operating condition identification based on the audio signal generated during the flotation process. It is an innovative attempt at flotation process data analysis, which not only broadens the information acquisition channels in industry, but also provides guidance for subsequent multi-view learning research that integrates multiple aspects of information such as images. The operating condition identification results can provide important status information for operational optimization of the flotation process.

[0108] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for identifying flotation process conditions based on audio signal characteristics, characterized in that: include: Step 1: Data preparation and data preprocessing, including data acquisition, noise reduction, working condition classification, and data division into training set and test set; Step 2: Based on the physical meaning of the frequency represented by the vertical axis of the Mel spectrogram, the proportion of the key frequency band in the flotation audio signal is amplified, the influence of the unimportant frequency band is reduced, and the flotation audio Mel spectrogram based on the feature attention mechanism is constructed; Step 3, constructing a deep convolutional network model, taking the flotation audio Mel spectrogram based on the feature attention mechanism as the input of the deep convolutional network model, taking the classification loss of the training set and the maximum mean difference between the training set and the test set as the loss, training the deep convolutional network model to automatically learn and extract deep audio signal features, and finally inputting the extracted deep features into the classifier in the deep convolutional network model for classification; Step 4: Input the unlabeled test set data into the deep convolutional network model, select some data in the recognition results and label them with pseudo labels, and retrain the deep convolutional network model using the test set data and training set data with pseudo labels; Step 5, repeat step 4 until the specified number of times is reached, and the deep convolutional network model obtained last time is used as the flotation process working condition identification model; Wherein, the step 1 specifically includes: Step 11, using a microphone to collect audio signals generated when the scraper scrapes the foam and the scraped foam falls into the bottom tank during the flotation process, collect two minutes of audio every hour, and record the foam concentration and grade at the corresponding time as the label of the audio data; The step 2 specifically includes: Step 21, convert each flotation audio signal into a Mel spectrogram and establish a deep convolutional network model, use the obtained Mel spectrogram as the input of the deep convolutional network model, use the training set to train the deep convolutional network model and perform working condition recognition on the test set data, and select the deep convolutional network model with the highest recognition accuracy of the test set from different deep convolutional network models obtained by multiple sets of hyperparameter training as the benchmark model; Step 22, adding masks to different frequency bands in the Mel spectrogram to shield the features of some frequency bands, using the Mel spectrogram with the mask as input to train a new deep convolutional network model, comparing the accuracy of the new deep convolutional network model on the test set with the result of the benchmark model to determine the impact of shielding different frequency bands on the accuracy; The step 2 also includes: Step 23, the frequency band with the greatest impact on accuracy is used as the key feature, and the frequency band with less impact is used as the secondary feature; Step 24, in the process of converting the audio signal into a mel spectrogram, weighting the key features, enhancing the proportion of the key features and weakening the influence of the secondary features.

2. The method for identifying flotation process conditions based on audio signal characteristics according to claim 1, characterized in that: The step 1 specifically includes: Step 12, designing a high-pass filter with a preset cutoff frequency to achieve smooth signals above the cutoff frequency of the high-pass filter and rapid attenuation of signals below the cutoff frequency of the high-pass filter; Step 13, segment each filtered audio signal. Multiple short audio segments obtained by segmenting the same audio segment correspond to the same data label. According to the labels corresponding to all audios, they are divided into N categories. Corresponding to N different working conditions in the production process, all the obtained data are divided into training sets and test sets.

3. The method for identifying flotation process conditions based on audio signal characteristics according to claim 2, characterized in that: The step 12 specifically includes: Step 121, performing FFT transformation on the original signal; Step 122, filtering is performed using the following formula: Among them, |Y(ω)| * represents the amplitude of the signal after filtering, |Y(ω)| represents the amplitude of the signal before filtering, and ω represents the frequency; Step 123, perform inverse FFT transform to obtain the filtered signal.

4. The method for identifying flotation process conditions based on audio signal characteristics according to claim 1, characterized in that: The step 3 specifically includes: Step 31, construct a deep convolutional network model, whose output is an N-dimensional vector normalized by a Softmax function containing a temperature coefficient, and each element of the N-dimensional vector represents the prediction probability of the deep convolutional network model for each class; wherein the expression of the Softmax function containing a temperature coefficient is Among them, z i represents the i-th element of the vector before Softmax normalization, N represents the number of categories, and T represents the temperature coefficient; Step 32: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n is the number of training samples, represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample.

5. The method for identifying flotation process conditions based on audio signal characteristics according to claim 4, characterized in that: The step 3 specifically includes: Step 33: Input the training set data and the test set data into the constructed deep convolutional network model and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector; Step 34, calculate the total loss Among them, λ represents the weight hyperparameter; Step 35, train the constructed deep convolutional network model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the constructed deep convolutional network model.

6. The method for identifying flotation process conditions based on audio signal characteristics according to claim 5, characterized in that: The step 4 specifically includes: Step 41, setting a probability threshold θ. When the maximum Softmax output of a certain test set data in the current deep convolutional network model is greater than θ, it is determined that the recognition result of the current deep convolutional network model for the test set data is correct, and the recognition result output by the current deep convolutional network model is used as a pseudo label for the data. Step 42: Input the pseudo-labeled test set data into the current deep convolutional network model and calculate the cross entropy loss function of the test set data. Among them, k represents the number of test samples with pseudo labels, represents the pseudo label of the i-th test sample, Represents the recognition result of the current model for the i-th test sample.

7. The method for identifying flotation process conditions based on audio signal characteristics according to claim 6, characterized in that: The step 4 also includes: Step 43: Input the training set data into the constructed deep convolutional network model and calculate the cross entropy loss function of the training set data. Where n represents the number of training samples. represents the true label of the i-th training sample, Represents the recognition result of the model for the i-th training sample; Step 44: Input the training set data and the test set data into the constructed deep convolutional network model, and calculate the maximum mean difference loss between the training set and the test set. Where n represents the number of training samples, m represents the number of test set samples, represents the i-th training set sample input, represents the j-th test set sample input, φ(x) represents the most hidden feature vector learned by the deep convolutional network, and ||x||2 represents the 2-norm of the vector.

8. The method for identifying flotation process conditions based on audio signal characteristics according to claim 7, characterized in that: The step 4 also includes: Step 45, recalculate the total loss Among them, γ and λ represent weight hyperparameters; Step 46, training the model with the goal of minimizing the total loss, and after the training is completed, input the test set data to obtain the recognition result on the model; Step 47, jump to step 41, repeat step 41 to step 46 until a specified number of times is reached, and obtain a set of working condition identification models under hyperparameters including T, θ, γ and λ; T represents the temperature coefficient; Step 48, resetting multiple sets of hyperparameters, repeating step 47 until a specified number of times is reached, obtaining different models based on multiple sets of hyperparameter training, and selecting the model with the highest recognition accuracy of the test set data as the flotation process condition recognition model based on the obtained different models.

Images