Deep learning method for parkinson's voice data classification based on hyperparameter optimization

Abstract

The application discloses a kind of deep learning Parkinson's sound data classification method based on hyperparameter optimization, it includes: obtaining Parkinson's sound data;Parkinson's sound data is extracted and data processing, and the Parkinson's sound data after processing is obtained;Utilize the hyperparameter optimization after ResNet50 network to the Parkinson's sound data after processing is classified, and the Parkinson's sound data classification result is obtained;Wherein, the hyperparameter of improved ABC algorithm optimization ResNet50 network is optimized.The application can obtain the hyperparameter value suitable for Parkinson's sound data classification in shorter time, and then improve the accuracy of Parkinson's sound data classification.

Description

Technical Field

[0001] This invention relates to a deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization, belonging to the field of Parkinson's disease sound data classification technology. Background Technology

[0002] Parkinson's disease is a common mental illness among middle-aged and elderly people, mainly characterized by tremors, slowed movement and voice, and changes in speech, tone of voice, and rigidity. Early symptoms of Parkinson's disease are difficult to detect, thus treatment is often delayed. In recent years, Parkinson's researchers have discovered that most patients experience speech impairments in the early stages, such as difficulty articulating and dysarthria. Analyzing voice data to determine if it contains Parkinson's disease characteristics can aid in the early diagnosis of the disease.

[0003] Because computers can accurately capture and record minute changes, they have been used for decades to process medical data and aid in the diagnosis of Parkinson's disease. In recent years, machine learning has become a powerful tool for handling various problems and has achieved good results in classifying Parkinson's disease data. Under the same feature extraction conditions, researchers compared the performance of Support Vector Machines (SVMs), Naive Bayes, and k-Nearest Neighbors (kNNs) classifiers on a Parkinson's speech dataset containing continuous vowels, digits, words, and short phrases. The results showed that SVMs had the highest detection performance, with a correct detection rate of 80%, while kNNs and Naive Bayes achieved correct detection rates of 70% and 65%, respectively. To address the impact of improper feature extraction on subsequent detection results, researchers used a genetic algorithm to extract features from the Parkinson's vowel dataset and improved the accuracy to over 90% using an SVM classifier. Later, researchers first used SAS to analyze and process a sound dataset created by Max Little at Oxford University and then used deep learning methods to classify Parkinson's speech data, achieving a high accuracy of 92.9% on the test set.

[0004] In recent years, with the rapid development of convolutional neural networks (CNNs), many studies have emerged using neural networks for Parkinson's disease sound data classification. The performance of a neural network model varies depending on its hyperparameter settings; the more suitable the hyperparameters, the better the model's performance, theoretically improving the accuracy of Parkinson's disease sound data classification. However, finding suitable hyperparameters for a neural network model requires significant time and effort. Therefore, there is an urgent need to find a more suitable method for selecting CNN hyperparameters. Currently, neural network hyperparameter optimization methods are broadly divided into two categories: model-free and agent-based. Model-free hyperparameter search methods include grid search and random search. Model-based hyperparameter search methods include Bayesian optimization algorithms and swarm intelligence algorithms. Compared to other methods, swarm intelligence algorithms can search in parallel across the entire solution space, making them more suitable for high-dimensional searches of complex problems and for CNN hyperparameter optimization. However, in practical research, their algorithm complexity remains too high, while their accuracy needs improvement. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization. It utilizes an improved ABC algorithm to optimize the hyperparameters of the ResNet50 network, obtaining suitable hyperparameter values ​​for Parkinson's disease sound data classification in a shorter time, thereby improving the accuracy of Parkinson's disease sound data classification.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical means:

[0007] This invention proposes a deep learning-based method for classifying Parkinson's disease sound data based on hyperparameter optimization, comprising the following steps:

[0008] Obtain Parkinson's sound data;

[0009] Feature extraction and data processing were performed on the Parkinson's sound data to obtain the processed Parkinson's sound data.

[0010] The processed Parkinson's sound data was classified using a hyperparameter-optimized ResNet50 network to obtain the classification results of the Parkinson's sound data.

[0011] Among them, the improved ABC algorithm is used to optimize the hyperparameters of the ResNet50 network.

[0012] Furthermore, the method for feature extraction and data processing of Parkinson's disease sound data is as follows:

[0013] Feature extraction of Parkinson's sound data is performed using Mel frequency cepstral coefficients, converting the Parkinson's sound data into Mel frequencies.

[0014] Cepstral analysis of the Mel frequencies yields the Mel cepstral plot.

[0015] The Mel cepstral plots were normalized and cropped to obtain the processed Parkinson's sound data.

[0016] Furthermore, the improved ABC algorithm includes a dimension-by-dimensional adjustment strategy and a range pruning strategy.

[0017] Furthermore, let a set of feasible solutions to the hyperparameters to be optimized in the ResNet50 network be considered as an individual. The method for optimizing the hyperparameters of the ResNet50 network using the improved ABC algorithm is as follows:

[0018] A set of historical Parkinson's disease sound data is obtained, and features are extracted and processed from the historical Parkinson's disease sound data. After classification, a training dataset is obtained.

[0019] By using a mixed variable encoding strategy to randomly initialize individuals within a preset search range, the initialized ResNet50 network is obtained.

[0020] The training dataset is input into the ResNet50 network to obtain the predicted Parkinson's disease sound data classification results;

[0021] The fitness value of the current individual is calculated based on the predicted Parkinson's sound data classification results.

[0022] The local optimal solution is obtained and stored based on the fitness value;

[0023] The process involves three phases: hiring bees, observing bees, and scout bees, to search for new individuals.

[0024] Based on the updated individuals, the Parkinson's sound data classification results are re-predicted, the fitness value of the updated individuals is calculated, and then the local optimum is updated.

[0025] When the local optimum remains unchanged, the range pruning strategy and the dimension-by-dimensional adjustment strategy are used to continue searching for new individuals, update the local optimum, until the iteration termination condition is met, and the optimal hyperparameters are output.

[0026] The hyperparameter-optimized ResNet50 network is obtained based on the optimal hyperparameters.

[0027] Furthermore, the hyperparameters include the type of optimizer, the type of activation function, the pooling type, the initial learning rate, momentum, and inactivation rate in the ResNet50 network.

[0028] Furthermore, the range pruning strategy is as follows:

[0029] During the search process of each generation of individuals, the local optimal solution is recorded in the local optimal solution directory matrix of each generation;

[0030] When the fitness value corresponding to a local optimum remains unchanged for multiple generations, the previous local optimum with a fitness value different from the current local optimum is searched in reverse order according to the local optimum directory matrix. The distance between the current local optimum and the historical local optimum is set as the new search range for the individual.

[0031] Furthermore, the dimension-by-dimensional adjustment strategy is as follows:

[0032] The hyperparameters of the ResNet50 network to be optimized are divided into integer hyperparameters and floating-point hyperparameters;

[0033] Integer hyperparameter optimization is performed, keeping the floating-point hyperparameter values ​​unchanged in stagnant local optima. Within the new search range obtained by the range pruning strategy, each feasible solution with integer hyperparameters is traversed. Each feasible solution with integer hyperparameters is combined with the floating-point hyperparameters to form a new individual. Before evaluating the fitness of the new individual, the bee directory matrix is ​​traversed. If an individual with the same position exists in the bee directory matrix, network training is not performed; the fitness value corresponding to that individual is directly retrieved from the bee directory matrix. If no individual with the same position exists in the bee directory matrix, network training is performed based on the new individual, the fitness value corresponding to the new individual is calculated, and the local optima are updated. All feasible solutions with integer hyperparameters within the new search range are traversed to obtain the optimal integer hyperparameters while keeping the floating-point hyperparameter values ​​unchanged.

[0034] Floating-point hyperparameter optimization is performed, while the optimal integer hyperparameter values ​​remain unchanged. Gaussian randomization is used to fine-tune the floating-point hyperparameters within the new search range of individuals obtained by the range pruning strategy, thereby updating the local optimum.

[0035] Furthermore, in the bee-employment phase, the individual update formula is:

[0036]

[0037] in, This represents the value of the j-th hyperparameter in the i-th individual of generation t+1. Let t represent the value of the j-th hyperparameter in the i-th individual. Let t represent the value of the j-th hyperparameter in the k-th individual, i ≠ k, and ψ be a random number, ψ ∈ [-1, 1].

[0038] Furthermore, the classification accuracy of the Parkinson's sound data is used as the fitness value, calculated using the following formula:

[0039]

[0040] Wherein, Accuracy represents the classification accuracy of Parkinson's sound data, TN represents the number of samples that ResNet50 network predicted as healthy and that are actually healthy, TP represents the number of samples that ResNet50 network predicted as Parkinson's disease and that are actually Parkinson's disease, FP represents the number of samples that ResNet50 network predicted as Parkinson's disease but are actually healthy, and FN represents the number of samples that ResNet50 network predicted as healthy but are actually Parkinson's disease.

[0041] Furthermore, the search for new individuals is continued using range pruning and dimension-by-dimensional adjustment strategies to update the local optimum. When the local optimum remains unchanged for two consecutive generations, the iteration terminates. The optimal hyperparameters are obtained based on the local optimum at the time of iteration termination, and the maximum number of evaluations is output.

[0042] The following advantages can be obtained by adopting the above technical means:

[0043] This invention proposes a deep learning-based method for classifying Parkinson's disease sound data based on hyperparameter optimization. It utilizes an improved ABC algorithm to optimize the hyperparameters of the ResNet50 network, and then uses the optimized ResNet50 network to classify Parkinson's disease sound data, obtaining the classification results. This invention can maximize the optimization of ResNet50 hyperparameters within limited time and resources, enabling the ResNet50 network to exhibit better classification performance and improving the accuracy and reliability of Parkinson's disease sound data classification.

[0044] This invention employs a hybrid variable encoding strategy for hyperparameter initialization, encoding discrete hyperparameters into continuous ones. This allows for subsequent optimization using continuous domain update formulas, simplifying the hyperparameter optimization process. Building upon the existing ABC algorithm, this invention adds range pruning and dimension-by-dimensional adjustment strategies. These strategies enable fine-tuning of hyperparameters even after local optima have stalled. Range pruning reduces the individual search space, lowering the algorithm's time complexity, while dimension-by-dimensional adjustment avoids repeated fitness evaluations at the same individual positions, enhancing the local search capability of the ABC algorithm. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the steps of the deep learning-based Parkinson's sound data classification method based on hyperparameter optimization of the present invention.

[0046] Figure 2 This is a logical block diagram of the deep learning-based Parkinson's sound data classification method based on hyperparameter optimization of the present invention.

[0047] Figure 3 This is a flowchart of feature extraction in an embodiment of the present invention;

[0048] Figure 4 This is a structural diagram of the ResNet50 network in an embodiment of the present invention;

[0049] Figure 5 This is a flowchart illustrating the optimization of hyperparameters of the ResNet50 network using the improved ABC algorithm in an embodiment of the present invention.

[0050] Figure 6 These are the classification results of the training dataset and the validation dataset in this embodiment of the invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings:

[0052] This invention proposes a deep learning-based method for classifying Parkinson's disease sound data based on hyperparameter optimization, such as... Figure 1 , 2 As shown, the specific steps include the following:

[0053] Step A: Obtain Parkinson's sound data.

[0054] Step B: Perform feature extraction and data processing on the Parkinson's sound data to obtain the processed Parkinson's sound data.

[0055] Step B01: Extract features from the original .wav file of MDVR-KCL Parkinson's sound data using Mel frequency cepstral coefficients to convert the Parkinson's sound data into Mel frequencies.

[0056] Step B02: Perform cepstral analysis on the Mel frequency to obtain the Mel cepstral diagram corresponding to the MDVR-KCL Parkinson's sound data.

[0057] like Figure 3 As shown, feature extraction is performed using MFCC feature extraction technology. First, the Parkinson's sound data is divided into frames of 20-40ms. Then, a window function is used to smooth the signal, removing harmonics from the spectrum while preserving the estimate of the spectral envelope. Next, a Discrete Fourier Transform (DFT) is performed on each frame to transform the N samples of each frame from the time domain to the frequency domain, obtaining the power spectrum. Then, the power spectrum is filtered using a Mel filter bank, and the energy of each filter is calculated and its logarithm is taken. Finally, a Discrete Cosine Transform (DCT) is used to convert the logarithmic Mel spectrum to the time domain, obtaining the Mel spectrum.

[0058] Step B03: Normalize and crop the Mel cepstral image, setting the image size to 224*224 to obtain the processed Parkinson's sound data.

[0059] Step C: Use the hyperparameter-optimized ResNet50 network to classify the processed Parkinson's disease sound data, obtaining the classification results; the structure of the ResNet50 network is as follows. Figure 4 As shown, the improved ABC algorithm is used to optimize the hyperparameters of the ResNet50 network. There are 6 hyperparameters, namely the type of optimizer, the type of activation function, the pooling type, the initial learning rate, the momentum, and the inactivation rate in the ResNet50 network.

[0060] In this embodiment of the invention, the improved ABC algorithm includes a dimension-by-dimensional adjustment strategy and a range pruning strategy, which can enhance the local search capability of the ABC algorithm and reduce the time complexity of the algorithm.

[0061] Let a set of feasible solutions to the hyperparameters to be optimized in a ResNet50 network be considered as an individual, such as Figure 5 As shown, the method for optimizing the hyperparameters of the ResNet50 network using the improved ABC algorithm is as follows:

[0062] (1) Obtain a set of historical Parkinson's sound data, perform feature extraction and data processing on the historical Parkinson's sound data. The operation of feature extraction and data processing is the same as step B. Classify the processed historical Parkinson's sound data. The classification result is either healthy or Parkinson's disease. Divide the dataset into training dataset, validation dataset and test dataset in a ratio of 8:1:1.

[0063] (2) Using a mixed variable encoding strategy, individuals are randomly initialized within a preset search range. The algorithm finds SN individuals, resulting in the initialized ResNet50 network. The preset search range is the range of hyperparameter values. In the ResNet50 network, the six hyperparameters are of different types, thus the hyperparameters are discrete. The mixed variable encoding strategy can encode discrete hyperparameters into continuous types, allowing subsequent algorithms to apply the same continuous domain update formula to all hyperparameters, simplifying the hyperparameter optimization process.

[0064] (3) Use the training dataset to train the ResNet50 network. Input the training dataset into the ResNet50 network and use the ResNet50 network as a classifier to process the data and obtain the predicted Parkinson's sound data classification results.

[0065] (4) Calculate the fitness value of the current individual based on the predicted Parkinson's sound data classification results. Evaluate on the validation dataset.

[0066] In this embodiment of the invention, the classification accuracy of Parkinson's sound data is used as the fitness value, and the calculation formula is as follows:

[0067]

[0068] Wherein, Accuracy represents the classification accuracy of Parkinson's sound data, TN represents the number of samples that ResNet50 network predicted as healthy and that are actually healthy, TP represents the number of samples that ResNet50 network predicted as Parkinson's disease and that are actually Parkinson's disease, FP represents the number of samples that ResNet50 network predicted as Parkinson's disease but are actually healthy, and FN represents the number of samples that ResNet50 network predicted as healthy but are actually Parkinson's disease.

[0069] (5) Compare the fitness values ​​of different individuals, select the individual with the largest fitness value as the current local optimum and store it in the local optimum directory matrix.

[0070] (6) The process involves three stages: hiring bees, observing bees, and scout bees, to search for new individuals. Hired bees use information about previous individuals to find new individuals and share this information with the observing bees; the observing bees wait in the hive and search for new individuals based on the information shared by the hired bees; and the scout bees randomly search for a new, valuable individual near the hive.

[0071] During the bee-hiring phase, the expression for finding individual information is as follows:

[0072]

[0073] in, This represents the value of the j-th hyperparameter in the i-th individual of generation t+1. Let t represent the value of the j-th hyperparameter in the i-th individual. Let ψ represent the value of the j-th hyperparameter in the k-th individual of generation t, where k is another randomly selected individual other than i in the current generation, i ≠ k, and ψ is a random number, ψ ∈ [-1, 1].

[0074] (7) Based on the feasible solutions (values) of the hyperparameters in the updated individuals, a new ResNet50 network is obtained. Return to step (3), re-predict the classification results of the Parkinson's sound data, calculate the fitness value corresponding to the updated individuals, and then update the local optimal solution.

[0075] (8) When the local optimal solution remains unchanged, use the range pruning strategy and the dimension-by-dimensional adjustment strategy to continue searching for new individuals, update the local optimal solution, until the iteration termination condition is met, and output the optimal hyperparameter.

[0076] The range pruning strategy, centered on local optima, proposes a search space pruning scheme based on a local optimum directory matrix. During each generation of individual searches, the local optimum directory matrix records the local optima for each generation. When the fitness value corresponding to a local optimum remains unchanged for multiple generations, the previous local optimum with a different fitness value is searched backwards from the current local optimum directory matrix. The distance between the current local optimum and the historical local optimum is set as the new search range for the individual, called the pruning range. This range pruning strategy reduces the search space, lowers the algorithm's time complexity, and improves local search capabilities for gbest dimension adjustment.

[0077] In this embodiment of the invention, if the fitness value corresponding to the local optimum does not change for three consecutive generations, a range pruning strategy is performed.

[0078] The dimension-wise adjustment strategy proposes a new local optimum search method based on the range pruning strategy, which divides the hyperparameters of the ResNet50 network to be optimized into integer hyperparameters and floating-point hyperparameters.

[0079] The specific operation of the dimension-by-dimensional adjustment strategy is as follows:

[0080] During the individual search process, the location and corresponding fitness value of each individual are recorded using a bee directory matrix.

[0081] Integer hyperparameter optimization: Keep the floating-point hyperparameter values ​​in the stagnant local optima unchanged, traverse each feasible solution of integer hyperparameters in the new individual search range obtained by the range pruning strategy, and use each feasible solution of integer hyperparameters and floating-point hyperparameters to form a new individual; before evaluating the fitness of the new individual, traverse the bee directory matrix. If there is an individual with the same position in the bee directory matrix, do not perform network training (i.e., steps (3) and (4)), directly call the fitness value corresponding to the individual from the bee directory matrix, and update the local optima; if there is no individual with the same position in the bee directory matrix, perform network training based on the new individual, calculate the fitness value corresponding to the new individual, and update the local optima; traverse all feasible solutions of integer hyperparameters in the new individual search range to obtain the optimal integer hyperparameters with the floating-point hyperparameter values ​​unchanged.

[0082] Perform floating-point hyperparameter optimization: Keep the optimal integer hyperparameter values ​​unchanged, and use Gaussian randomization to fine-tune the floating-point hyperparameters in the new search range of individuals obtained by the range pruning strategy, and update the local optimum.

[0083] The dimensional adjustment strategy avoids repeated fitness evaluations for the same individual position by using the bee catalog matrix, thereby reducing the time complexity of the algorithm. In addition, Gaussian randomization is used to obtain a more accurate solution.

[0084] In this embodiment of the invention, when the local optimal solution updated by the hired bee, the observation bee, and the scout bee remains unchanged for three consecutive generations, the search for new individuals is continued using the range pruning strategy and the dimension-by-dimensional adjustment strategy to update the local optimal solution. When the local optimal solution updated by the range pruning strategy and the dimension-by-dimensional adjustment strategy remains unchanged for two consecutive generations (the fitness value remains unchanged), the iteration terminates. The optimal hyperparameters are obtained based on the local optimal solution at the time of iteration termination, and the maximum number of evaluations is output.

[0085] (9) Obtain the hyperparameter-optimized ResNet50 network based on the optimal hyperparameters.

[0086] After obtaining the hyperparameter-optimized ResNet50 network, the generalization ability of the network model can be verified using a test dataset.

[0087] To verify the classification performance of the method of this invention, it was compared with other swarm intelligence algorithms, including PSO, QPSO, DE, and the basic ABC algorithm, using ResNet50 on the MDVR-KCL dataset. First, the initialized ResNet50 network was used to perform Parkinson's disease sound data classification, obtaining the number of fitness evaluations. Then, under the same number of fitness evaluations, the same population, and the same network training settings, the hyperparameters of ResNet50 were optimized using PSO, QPSO, DE, the basic ABC algorithm, and the improved ABC algorithm of this invention, respectively, and applied to the MDVR-KCL dataset for Parkinson's disease classification. The classification results are shown in the table below:

[0088]

[0089]

[0090] As can be seen from the data in the table, the training accuracy of the method of the present invention can reach 100%, which is higher than the training accuracy of other swarm intelligence algorithms. In addition, the verification accuracy and test accuracy of the method of the present invention are also higher than other methods, which shows the effectiveness and advancement of the method of the present invention.

[0091] The classification results of the training and validation datasets in the method of this invention, as they change with the number of iterations of the network model, are as follows: Figure 6 As shown in the figure, it can be seen that by optimizing the hyperparameters of the ResNet 50 network through the improved ABC algorithm, the accuracy of Parkinson's sound data classification can be effectively improved, and the final classification accuracy approaches 100%, which is higher than the existing technology.

[0092] Furthermore, the method of this invention can be compared with traditional classical parameter adjustment algorithms, including grid search, random search, and TPE algorithm, as well as other works on Parkinson's classification on the MDVR-KCL dataset. These works were then tested, and the experimental results also proved the effectiveness and rationality of this invention.

[0093] In the feature extraction stage, this invention adopts MFCC cepstral coefficient feature extraction technology because the similarity between the mel scale and human speech perception in MFCC cepstral coefficient feature extraction technology can ensure the accuracy of feature extraction.

[0094] In terms of neural network model selection, this invention chooses ResNet 50 as the basic network framework. The ResNet network model can deepen the number of network layers without vanishing gradients, but it cannot guarantee that the model accuracy will not be reduced. Therefore, this invention needs to optimize the hyperparameters of the ResNet network model.

[0095] In the hyperparameter optimization method, this invention tunes the standard ResNet50 hyperparameters using an improved ABC algorithm, resulting in higher classification accuracy on the MDVR-KCL Parkinson's Voice dataset. The basic ABC algorithm has the advantages of simple parameter settings and strong global search capabilities, but its local search capabilities are weak and its convergence speed is relatively slow. Therefore, this invention improves the basic ABC algorithm through a dimensional adjustment strategy and a range pruning strategy. The range pruning strategy reduces the individual search space, lowers the algorithm's time complexity, and improves local search capabilities. The dimensional adjustment strategy avoids repeated fitness evaluations of the same individual positions and utilizes Gaussian randomization to obtain higher-precision solutions. The improved ABC algorithm can find reasonable ResNet50 hyperparameter settings within limited time and resources.

[0096] In summary, the method of the present invention can effectively improve the accuracy and reliability of Parkinson's disease sound data classification.

[0097] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization, characterized in that, Includes the following steps: Obtain Parkinson's sound data; Feature extraction and data processing were performed on the Parkinson's sound data to obtain the processed Parkinson's sound data. The processed Parkinson's sound data was classified using a hyperparameter-optimized ResNet50 network to obtain the classification results of the Parkinson's sound data. Here, we define a set of feasible solutions to the hyperparameters to be optimized in the ResNet50 network as an individual. The improved ABC algorithm is used to optimize the hyperparameters of the ResNet50 network, including: A set of historical Parkinson's disease sound data is obtained, and features are extracted and processed from the historical Parkinson's disease sound data. After classification, a training dataset is obtained. By using a mixed variable encoding strategy to randomly initialize individuals within a preset search range, the initialized ResNet50 network is obtained. The training dataset is input into the ResNet50 network to obtain the predicted Parkinson's disease sound data classification results; The fitness value of the current individual is calculated based on the predicted Parkinson's sound data classification results. The local optimal solution is obtained and stored based on the fitness value; The process involves three phases: hiring bees, observing bees, and scout bees, to search for new individuals. Based on the updated individuals, the Parkinson's sound data classification results are re-predicted, the fitness value of the updated individuals is calculated, and then the local optimum is updated. When the local optimum remains unchanged, the range pruning strategy and the dimension-by-dimensional adjustment strategy are used to continue searching for new individuals, update the local optimum, until the iteration termination condition is met, and the optimal hyperparameters are output. The hyperparameter-optimized ResNet50 network is obtained based on the optimal hyperparameters.

2. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, The method for feature extraction and data processing of Parkinson's disease sound data is as follows: Feature extraction of Parkinson's sound data is performed using Mel frequency cepstral coefficients, converting the Parkinson's sound data into Mel frequencies. Cepstral analysis of the Mel frequencies yields the Mel cepstral plot. The Mel cepstral plots were normalized and cropped to obtain the processed Parkinson's sound data.

3. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, The improved ABC algorithm includes a dimension-by-dimensional adjustment strategy and a range pruning strategy.

4. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, Hyperparameters include the type of optimizer, type of activation function, pooling type, initial learning rate, momentum, and inactivation rate in the ResNet50 network.

5. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, The pruning strategy for the area is as follows: During the search process of each generation of individuals, the local optimal solution is recorded in the local optimal solution directory matrix of each generation; When the fitness value corresponding to a local optimum remains unchanged for multiple generations, the previous local optimum with a fitness value different from the current local optimum is searched in reverse order according to the local optimum directory matrix. The distance between the current local optimum and the historical local optimum is set as the new search range for the individual.

6. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 5, characterized in that, The dimension-by-dimensional adjustment strategy is as follows: The hyperparameters of the ResNet50 network to be optimized are divided into integer hyperparameters and floating-point hyperparameters; During the individual search process, the location and corresponding fitness value of each individual are recorded using a bee directory matrix; Integer hyperparameter optimization is performed, keeping the floating-point hyperparameter values ​​unchanged in stagnant local optima. Within the new search range obtained by the range pruning strategy, each feasible solution with integer hyperparameters is traversed. Each feasible solution with integer hyperparameters is combined with the floating-point hyperparameters to form a new individual. Before evaluating the fitness of the new individual, the bee directory matrix is ​​traversed. If an individual with the same position exists in the bee directory matrix, network training is not performed; the fitness value corresponding to that individual is directly retrieved from the bee directory matrix. If no individual with the same position exists in the bee directory matrix, network training is performed based on the new individual, the fitness value corresponding to the new individual is calculated, and the local optima are updated. All feasible solutions with integer hyperparameters within the new search range are traversed to obtain the optimal integer hyperparameters while keeping the floating-point hyperparameter values ​​unchanged. Floating-point hyperparameter optimization is performed, while the optimal integer hyperparameter values ​​remain unchanged. Gaussian randomization is used to fine-tune the floating-point hyperparameters within the new search range of individuals obtained by the range pruning strategy, thereby updating the local optimum.

7. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, During the bee-employment phase, the individual update formula is: ； in, Indicates the t+1 generation i Among the individuals, the first j The values ​​of the hyperparameters, Indicates t represents the first i Among the individuals, the first j The values ​​of the hyperparameters, Indicates t represents the first k Among the individuals, the first j The values ​​of the hyperparameters, , It is a random number. [-1,1].

8. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, The classification accuracy of Parkinson's sound data is used as the fitness value, and the calculation formula is as follows: ； in, The values ​​represent the classification accuracy of the Parkinson's disease sound data. TN represents the number of samples that the ResNet50 network predicted as healthy and that are actually healthy. TP represents the number of samples that the ResNet50 network predicted as having Parkinson's disease and that are actually having Parkinson's disease. FP represents the number of samples that the ResNet50 network predicted as having Parkinson's disease but that are actually healthy. FN represents the number of samples that the ResNet50 network predicted as healthy but that are actually having Parkinson's disease.

9. The deep learning-based Parkinson's disease sound data classification method based on hyperparameter optimization according to claim 1, characterized in that, The search continues to find new individuals using range pruning and dimension-by-dimensional adjustment strategies, updating the local optimum. When the local optimum remains unchanged for two consecutive generations, the iteration terminates. The optimal hyperparameters are obtained based on the local optimum at the time of iteration termination, and the maximum number of evaluations is output.

Images