A data management system and management method for set-top boxes

Abstract

This invention relates to the field of set-top box data management technology, specifically to a set-top box data management system and method. It aims to solve the problem of noise and user interference caused by high disk I / O data processing tasks in quiet environments. This invention collects ambient audio signals, extracts short-time average energy, short-time average zero-crossing rate, and the proportion of low-frequency spectrum energy, and generates an acoustic comprehensive feature vector. A pre-trained multilayer perceptron model outputs a quietness index, which is compared with a preset threshold to generate a binary control signal. When the environment is quiet, high disk I / O background data processing tasks are paused; when the environment is not quiet, all tasks are allowed to start. This invention can adaptively schedule background maintenance tasks, balancing user quiet experience with device data management needs. The lightweight model is adapted to set-top box hardware, improving user experience and operational efficiency.

Description

Technical Field

[0001] This invention relates to the field of set-top box data management technology, and more specifically, to a set-top box data management system and management method. Background Technology

[0002] Set-top box data management is an important technology, specifically applied to the intelligent scheduling of set-top box background data processing tasks. It achieves adaptive task control through acoustic scene recognition, balancing user experience with device maintenance needs. Set-top box background data processing tasks involve high disk I / O load operations with frequent read / write operations. Because the task scheduling mechanism cannot perceive and identify the acoustic quietness of the current usage scenario in real time, data processing tasks continue to execute automatically even in quiet scenarios such as user viewing or nighttime rest, generating storage read / write noise that interferes with normal user operation. Furthermore, it is difficult to efficiently complete data maintenance work during non-quiet periods, making it difficult to balance the set-top box user's quiet experience with the actual needs of device data management and maintenance. To solve this problem, we provide a set-top box data management system and method. Summary of the Invention

[0003] The purpose of this invention is to provide a data management system and method for set-top boxes to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a data management system for a set-top box is provided, comprising: The environmental sensing unit acquires the original environmental audio signal, preprocesses the original environmental audio signal to obtain the time domain signal, calculates the short-time average energy and short-time average zero-crossing rate in parallel on the time domain signal, and performs a fast Fourier transform on the time domain signal to obtain its frequency domain power spectrum. The spectral energy ratio of the preset low-frequency band is extracted from the frequency domain power spectrum. Finally, the short-time average energy, short-time average zero-crossing rate and spectral energy ratio are fused to generate a comprehensive feature vector to characterize the current acoustic environment. The audio scene classification unit inputs the comprehensive feature vector into a pre-trained multilayer perceptron neural network model. The multilayer perceptron neural network model includes an input layer, a hidden layer, and an output layer. The input layer receives the comprehensive feature vector and passes it to the hidden layer. The hidden layer is used to linearly weight and sum the comprehensive feature vector with the pre-trained weight matrix to obtain high-order mode features used to distinguish different acoustic scenes. The high-order mode features are then mapped and transformed by a preset nonlinear activation function to obtain nonlinear acoustic mode features. The output layer processes and calculates the nonlinear acoustic mode features and outputs a scalar value representing the level of quietness of the environment, denoted as the quiet index. The quiet index is compared with a preset quiet threshold, and a binary control signal is generated based on the comparison result. The task scheduling and execution unit organizes and schedules tasks based on binary control signals.

[0005] The second objective of this invention is to provide a management method for implementing a data management system for a set-top box, comprising the following steps: S1. Obtain the original environmental audio signal through the audio acquisition module, preprocess the signal to obtain the time domain signal, calculate the short-time average energy and short-time average zero-crossing rate of the time domain signal in parallel, perform fast Fourier transform on the time domain signal to obtain the frequency domain power spectrum and extract the spectral energy ratio of its preset low frequency band, and fuse the short-time average energy, short-time average zero-crossing rate and spectral energy ratio to generate a comprehensive feature vector. S2. Input the comprehensive feature vector into a pre-trained multilayer perceptron neural network model. Through the hidden layer of the model, perform linear weighted summation and nonlinear activation function mapping on the vector based on pre-trained weights and biases to extract nonlinear acoustic pattern features. Then, process and calculate through the output layer to output a quiet index representing the level of environmental quietness. S3. Compare the calculated quiet index with a preset quiet threshold, and generate a binary control signal based on the comparison result; S4. Execute data processing task scheduling according to the generated binary control signal. When the signal indicates that the environment is quiet, suspend high disk I / O load tasks. When the signal indicates that the environment is not quiet, allow the background data processing task to start.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention can accurately identify quiet and non-quiet usage states by sensing the acoustic environment of the set-top box in real time and extracting multi-dimensional acoustic features. It can adaptively schedule high-load data processing tasks in the background, fundamentally avoiding data read and write noise from interfering with users' use in quiet scenarios such as watching movies and resting, thus improving the user experience. It adopts a lightweight classification model to achieve low computing power and high real-time scene determination, which not only ensures the need for quiet use, but also efficiently completes device data maintenance, perfectly balancing user experience and device management needs. Attached Figure Description

[0007] Figure 1 This is an overall block diagram of the present invention; Figure 2 This is the overall flowchart of the present invention.

[0008] The meanings of the labels in the diagram are as follows: 1. Environmental perception unit; 2. Audio scene classification unit; 3. Task scheduling and execution unit. Detailed Implementation

[0009] 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.

[0010] This invention provides a data management system for set-top boxes. Please refer to [link / reference]. Figure 1 As shown, it includes: The environmental sensing unit acquires the original environmental audio signal, preprocesses the original environmental audio signal to obtain the time domain signal, calculates the short-time average energy and short-time average zero-crossing rate in parallel on the time domain signal, and performs a fast Fourier transform on the time domain signal to obtain its frequency domain power spectrum. The spectral energy ratio of the preset low-frequency band is extracted from the frequency domain power spectrum. Finally, the short-time average energy, short-time average zero-crossing rate and spectral energy ratio are fused to generate a comprehensive feature vector to characterize the current acoustic environment. The audio scene classification unit inputs the comprehensive feature vector into a pre-trained multilayer perceptron neural network model. The multilayer perceptron neural network model includes an input layer, a hidden layer, and an output layer. The input layer receives the comprehensive feature vector and passes it to the hidden layer. The hidden layer is used to linearly weight and sum the comprehensive feature vector with the pre-trained weight matrix to obtain high-order mode features used to distinguish different acoustic scenes. The high-order mode features are then mapped and transformed by a preset nonlinear activation function to obtain nonlinear acoustic mode features. The output layer processes and calculates the nonlinear acoustic mode features and outputs a scalar value representing the level of quietness of the environment, denoted as the quiet index. The quiet index is compared with a preset quiet threshold, and a binary control signal is generated based on the comparison result. Task scheduling and execution unit 3 performs task scheduling based on binary control signals and data processing.

[0011] After completing the initial preprocessing, three calculations are performed synchronously based on the same frame of discrete-time domain signal: short-time average energy, short-time average zero-crossing rate, and frequency domain power spectrum. The specific implementation method is as follows: Short-time average energy is the most fundamental time-domain characteristic of discrete-time signals, used to quantitatively characterize the volume intensity of time-domain signals. It is calculated as follows: The amplitudes of all sample points within a single frame of the time-domain signal are squared and summed. The sum of squares is then divided by the frame length to obtain the average value. This process eliminates interference from abrupt amplitude changes at individual sample points, yielding a stable characteristic value reflecting the instantaneous energy level of the signal. Before calculation, the time-domain signal refers to a discrete digital signal with time as the horizontal axis and signal amplitude as the vertical axis. It consists of sample points evenly arranged in the order of sampling time and is the basic representation of the digitized signal. A frame is the basic unit of short-time signal analysis. To capture the dynamic changes of the signal, continuous time-domain signals are divided into short-time frames of fixed length. The frame length is the total number of discrete sample points contained in each frame. Frame lengths are typically chosen as integer powers of 2, such as 128, 256, 512, or 1024, to meet the accuracy requirements of short-time signal analysis and the subsequent Fast Fourier Transform calculation requirements. It is the unified basic parameter for calculating short-time average energy and short-time average zero-crossing rate. The specific formula for calculating short-time average energy is as follows: ,in The short-time average energy, calculated from the final value of a single-frame time-domain signal, is a dimensionless characteristic value that directly corresponds to the signal's volume intensity. The frame length of the current analysis frame, i.e., the total number of discrete sample points contained in this frame, must be completely consistent with the frame length of all subsequent calculations to ensure the temporal synchronization of features. This is the index of the sample point within the frame, ranging from 0 to N-1, used to sequentially identify each independent sample point within the frame. For the intra-frame sequence number is The sample point amplitude is a discrete value of the time-domain signal at that sampling moment. It can be positive or negative and represents the instantaneous vibration amplitude of the signal. The result of squaring the amplitude of the sample points eliminates the influence of the sign of the amplitude, retaining only the absolute value of the signal energy. The sum of squares of the amplitudes of all sample points within the frame represents the total energy of the time-domain signal in that frame. Finally, the total energy is divided by the frame length. The average value is obtained as the short-time average energy. This calculation process is performed independently for each frame of the time-domain signal, ultimately yielding a short-time average energy sequence that dynamically changes over time. This sequence completely and continuously reflects the real-time variation of the volume intensity of the time-domain signal. Simultaneously with the calculation of the short-time average energy, the system performs parallel calculations of the short-time average zero-crossing rate. The short-time average zero-crossing rate is another time-domain characteristic specifically used to reflect the frequency of zero-crossing in the time domain. Zero-crossing refers to a change in the sign of the amplitude of a discrete sample point, i.e., from positive to negative amplitude, or from negative to positive amplitude, representing the moment when the signal crosses the zero amplitude axis. A higher zero-crossing rate indicates more drastic changes in the time domain and richer high-frequency components, while a lower zero-crossing rate indicates smoother changes in the time domain and a dominance of low-frequency components. The calculation involves counting the total number of sign changes between adjacent sample points within a single frame of the time-domain signal, then rationing this number to the frame length to obtain the standardized short-time average zero-crossing rate. During the calculation process, each group of adjacent sample points within the frame is first... and Perform sign determination; the sign determination function is: When the sign function values ​​of two adjacent sample points are not equal, a zero-crossing event is determined to have occurred. After traversing all adjacent sample points in the frame, the total number of sign changes within a single frame is counted. Then through the formula The short-time average zero-crossing rate is calculated in the formula. This represents the short-time average zero-crossing rate of a single-frame time-domain signal, ranging from 0 to 1. The magnitude of this value directly corresponds to the frequency of zero-crossing. This represents the total number of sign changes between adjacent sample points within a single frame, i.e., the cumulative number of zero-crossing events. To ensure that the frame length is completely consistent with the short-time average energy calculation, both time-domain features are calculated based on the same frame of data, thus achieving synchronous extraction of time-domain features. This calculation does not require additional signal preprocessing and shares sample point data with the short-time average energy calculation, which greatly improves the efficiency of signal time-domain feature extraction. The two time-domain features complement each other and comprehensively characterize the amplitude energy and time-domain variation characteristics of the time-domain signal.

[0012] While simultaneously calculating two time-domain features, to uncover the essential frequency-domain characteristics of the time-domain signal and clarify its constituent frequency components and energy proportions, the system applies a Fast Fourier Transform (FFT) algorithm to the same frame of the time-domain signal to transform it from the time domain to the frequency domain. FFT is an optimized and fast algorithm based on the Discrete Fourier Transform (DFT), which uses a divide-and-conquer strategy to reduce the computational complexity of the traditional DFT from... Significantly reduced to This is an industry-standard algorithm for converting discrete digital signals from the time domain to the frequency domain. It efficiently extracts all frequency components of a signal, achieving a lossless mapping between the time and frequency domains. In practical implementation, a length of... Single-frame discrete-time signal As input to the Fast Fourier Transform, perform Point FFT operation directly outputs a complex sequence in the frequency domain. ,in Here, represents the index of a frequency point in the frequency domain, ranging from 0 to N-1. The actual physical frequency corresponding to each frequency point is . , The sampling rate of the signal, i.e., the number of sample points acquired per unit time, is a fundamental parameter for signal digitization and is the output of the FFT. It is in complex form and contains the real part. With the imaginary part These correspond to the amplitudes of the cosine and sine components of the corresponding frequency component, respectively. The modulus of the complex number represents the magnitude of the frequency component. The frequency domain power spectrum is used to quantitatively reflect the frequency domain characteristics of the energy distribution of a time-domain signal at different frequency components. Its calculation involves taking the square of the modulus of the complex number corresponding to each frequency point after the Fast Fourier Transform. The specific calculation formula is as follows: , in the formula For the first The power spectral density value corresponding to each frequency point directly represents the energy level of that frequency component. The formula for calculating the modulus of the complex number output by the FFT is as follows: ; Let be the real part of the complex number in the frequency domain. The power value is obtained by squaring the modulus of the imaginary part of the complex number in the frequency domain, eliminating interference from complex phase information, and retaining only the correspondence between frequency and energy, thus providing the power spectral values ​​for all frequency points. Arranged sequentially by frequency point number, the frequency domain power spectrum corresponding to the time domain signal of that frame is obtained. This power spectrum fully reveals the frequency energy distribution law of the time domain signal. The high proportion of low frequency power spectrum values ​​indicates that the signal is dominated by low frequency components, while the prominent high frequency power spectrum values ​​indicate that the signal is rich in high frequency components. The overall distribution pattern of the power spectrum directly corresponds to the essential frequency domain characteristics of the signal, providing a frequency domain basis for subsequent signal analysis, identification and classification.

[0013] After calculating the short-time average energy, short-time average zero-crossing rate, and low-frequency spectrum energy proportion of a single frame signal, the environmental perception unit performs statistical smoothing on the instantaneous features of multiple consecutive frames based on a preset time window. It calculates the mean and variance of the feature sequence and concatenates the multi-dimensional statistical features into a comprehensive feature vector. The preset time window is typically set to 1 to 5 seconds. Within this window, the system continuously segments the audio signal into M ordered time-domain frames according to a fixed frame length and frame shift. M is the total number of frames included in the time window, determined by the window duration, sampling rate, frame length, and frame shift. The calculation formula is as follows: ,in For the preset time window duration, For audio sampling rate, For single frame length, The number of frame-shifted samples. The function rounds down to ensure the number of frames within the window is an integer. Frame shift refers to the number of overlapping sample points between two adjacent frames. Frame shift design achieves continuous, gapless signal coverage within the window, avoiding signal loss and reducing redundant computation. Within this time window, the system strictly follows the standard procedure for single-frame feature extraction, processing each continuously acquired time-domain signal frame independently and extracting three sets of instantaneous features for each frame. The first set is the short-time average energy, which is the arithmetic mean of the squared amplitudes of all sample points in a single frame signal, used to characterize the volume intensity of the single-frame signal. The second set is the short-time average... The zero-crossing rate is the ratio of the total number of sign changes between adjacent sample points in a single frame to the frame length. It reflects the frequency of zero crossings in the time domain of a single frame signal. The third group is the spectral energy ratio, which is the ratio of the total energy of a preset low-frequency band (usually 0Hz to 300Hz, which is the distribution range of steady-state background noise and can effectively distinguish between steady-state environment and sudden sound) in the frequency domain power spectrum to the total energy of the entire frequency band. It is used to characterize the low-frequency energy distribution characteristics of the signal. After continuous frame-by-frame extraction, the system obtains three ordered instantaneous feature sequences within the time window, namely the short-time average energy sequence. Short-time average zero-crossing rate sequence Spectral energy proportion sequence Each element in the sequence corresponds to the instantaneous characteristic value of a frame of time-domain signal within a time window, completely recording the continuous change process of acoustic features within the window.

[0014] After acquiring three sets of instantaneous feature sequences, the system calculates the mean and variance for each set of feature sequences within a preset time window. The mean is used to characterize the overall average level of the features within the window, eliminating the influence of instantaneous fluctuations. The variance is used to characterize the degree of dispersion of the features within the window, reflecting the stability of the environment. The combination of these two statistical measures can comprehensively characterize the static attributes and dynamic changes of the acoustic environment. First, the short-time average energy sequence is statistically calculated, and the short-time average energy mean... The average volume intensity of ambient audio within a time window is a fundamental indicator for judging the quietness of the environment. The calculation formula is: Where M is the total number of frames within the preset time window, and k is the frame number (ranging from 1 to M). Let k be the instantaneous value of the short-time average energy in the k-th frame. The sum of the short-time average energy of all frames within the window; the variance of the short-time average energy. The variance reflects the fluctuation range of the ambient volume within a time window. A larger variance indicates that the ambient volume fluctuates wildly and is unstable, while a smaller variance indicates that the ambient volume is stable and without drastic fluctuations. The calculation formula is as follows: ,in The variance is calculated by squaring the squared deviation of the energy from the mean in a single frame, eliminating the influence of positive and negative deviations, and then summing and averaging the results. Simultaneously, the same statistical calculation is performed on the short-time average zero-crossing rate sequence. This reflects the average frequency of signal time-domain changes within the window. A higher mean indicates more high-frequency components in the signal and a more complex ambient sound environment, while a lower mean indicates a smoother signal and a quieter environment. The calculation formula is as follows: ; short-time mean zero-crossing rate variance It reflects the degree of fluctuation in the zero-crossing rate within the window, embodying the stability of the environmental time-domain characteristics. The calculation formula is as follows: At the same time, the mean and variance of the spectral energy proportion sequence were calculated, and the mean of the spectral energy proportion was... The average proportion of low-frequency steady-state noise within the window is a key characteristic for distinguishing between quiet environments, background noise environments, and sudden sound environments. The calculation formula is as follows: ; Spectral energy proportion variance This reflects the fluctuation in the proportion of low-frequency energy within the window. A small variance indicates that low-frequency noise is stably present, while a large variance indicates that low-frequency noise is intermittent and the environmental conditions are unstable. The calculation formula is as follows: The mean and variance calculations of the above three sets of features are all performed synchronously based on the same time window and the same batch of frame data, with data sharing and parallel computation, which is fully adapted to the low power consumption and high real-time requirements of set-top box embedded systems.

[0015] After completing all statistical calculations, the system obtained six dimensions of stable statistical characteristics, namely, the short-time average energy mean. Short-time average energy variance Short-time average zero-crossing rate Short-time mean zero-crossing rate variance , average proportion of spectrum energy Spectral energy proportion variance These six features, from six dimensions—volume level, volume stability, time-domain variation frequency, time-domain variation stability, low-frequency energy proportion, and low-frequency energy stability—describe the essential attributes of the current acoustic environment. The system then arranges these six feature values ​​in a preset, fixed order, forming a one-dimensional array. This preset order is a system-defined rule and remains unchanged throughout the process. The specific order is as follows: The sequence of short-time average energy mean → short-time average energy variance → short-time average zero-crossing rate mean → short-time average zero-crossing rate variance → spectral energy percentage mean → spectral energy percentage variance perfectly matches the input dimension and feature weights of the multilayer perceptron model in the subsequent audio scene classification unit, ensuring that the model correctly interprets the physical meaning of each feature. The concatenated one-dimensional array can be represented as follows: This one-dimensional array is a comprehensive feature vector used to characterize the current acoustic environment. After the comprehensive feature vector is generated, it will be transmitted to the audio scene classification unit in real time as the standard input of the pre-trained multilayer perceptron neural network model, realizing the transformation from the original audio signal to stable acoustic features.

[0016] The comprehensive feature vector is transmitted in real time to the audio scene classification unit and enters the hidden layer feature transformation stage of the pre-trained multilayer perceptron neural network model. The parameters upon which the model's hidden layer performs linear weighted summation have all been optimized offline and locally fixed. The pre-trained weight matrix consists of learnable parameters that enable the multilayer perceptron model to achieve linear mapping of input features. This weight matrix was trained offline using a large amount of labeled sample data from different acoustic scenes, such as quiet and noisy, before model deployment. After iterative optimization to its optimal state using the gradient descent algorithm, it is permanently stored in the embedded storage module of the audio scene classification unit. It does not require retraining or adjustment during online runtime. Its dimensionality follows the matching rules between input features and network structure. The number of rows in the pre-trained weight matrix... Similar to the comprehensive feature vector, the comprehensive feature vector in this system is fixed at a six-dimensional structure. Therefore, the number of rows in the weight matrix is ​​fixed at 6, and the number of columns in the pre-trained weight matrix directly corresponds to the number of neurons in the hidden layer. The number of neurons in the hidden layer is a fixed structural parameter preset according to the performance of the set-top box embedded hardware and the requirements of scene classification accuracy. It is usually set to 8, 16 or 32. Assuming that the number of neurons in the hidden layer is H, the dimension of the pre-trained weight matrix W is 6×H. Each weight value in the weight matrix corresponds to the influence strength of a specific feature in the comprehensive feature vector on a specific neuron in the hidden layer. The magnitude of the value directly determines the contribution ratio of the feature in the neuron activation process and is a linear parameter for realizing acoustic scene pattern differentiation.

[0017] When performing the linear weighted summation operation, the system uses the comprehensive feature vector as a one-dimensional input vector F (dimension 1×6) and performs a standard matrix multiplication operation with the pre-trained weight matrix W (dimension 6×H). This operation aggregates the input acoustic features to each hidden neuron according to the weight ratio. The specific implementation logic is as follows: Each eigenvalue of the composite feature vector is multiplied one by one with all the weight values ​​in the corresponding row of the weight matrix. Then, all the products in the same column (corresponding to the same hidden neuron) are summed along the row direction to generate an intermediate weighted sum vector of dimension 1×H. The matrix multiplication formula is as follows: Each parameter in the formula has a clear definition and physical meaning. The intermediate weighted sum vector is obtained after matrix multiplication. F is a 1×6 dimensional comprehensive feature vector, and W is a 6×H dimensional pre-trained weight matrix. "·" indicates a standard matrix multiplication operation, which projects the six dimensions of the comprehensive feature vector onto all neurons in the hidden layer according to the influence strength defined by the weight matrix, forming the initial weighted input for each neuron. After obtaining the intermediate weighted sum vector, the system performs element-wise addition of the pre-trained bias vector (also stored in the audio scene classification unit) with the intermediate weighted sum vector. The pre-trained bias vector b is a baseline adjustment parameter that was trained offline synchronously with the weight matrix. Its dimension is 1×H, completely consistent with the dimension of the intermediate weighted sum vector. Its function is to provide an independent activation baseline value for each neuron in the hidden layer, avoiding the problem of single pattern representation caused by neuron input relying solely on feature weighting. The formula for element-wise addition is... ,in The final high-order pattern features, Let b be a 1×H dimensional intermediate weighted sum vector, and b be a 1×H dimensional pre-trained bias vector. "+" indicates element-wise addition, that is, the i-th element in the intermediate weighted sum vector is added to the i-th element in the bias vector to obtain the i-th element of the higher-order pattern feature. After the complete operation of linear weighted summation and element-wise addition of bias, the dimension of the final higher-order pattern feature is exactly equal to the number of neurons in the hidden layer. The higher-order pattern feature has completed the linear fusion and abstract expression of the original acoustic features, and has initially acquired the ability to distinguish different acoustic scenes. This lays a solid linear feature foundation for the subsequent nonlinear mapping transformation to extract more discriminative nonlinear acoustic pattern features through preset nonlinear activation functions. The entire linear operation process is logically rigorous and computationally lightweight, and is fully compatible with the hardware resources and real-time operation requirements of the set-top box embedded system.

[0018] After obtaining the high-order mode features of the hidden layer output, the pre-trained weight matrix is ​​explained. The weight matrix serves as a learnable parameter for the linear mapping of acoustic features in the hidden layer. Each weight value has a unique and definite physical correspondence. Each weight value in the weight matrix corresponds to the influence strength of a specific feature in the comprehensive feature vector on a specific neuron in the hidden layer. The specific features in the comprehensive feature vector refer to the six fixed-dimensional acoustic features obtained through statistical concatenation via time windows: mean short-time average energy, mean short-time average energy variance, mean short-time average zero-crossing rate, mean short-time average zero-crossing rate variance, mean spectral energy percentage, and mean spectral energy percentage variance. In a layer, a specific neuron refers to each independent computational unit configured according to a pre-defined structure in the hidden layer. The magnitude and sign of the weight value directly represent the strength and direction of the corresponding input feature's effect on the target neuron. A positive weight value with a larger absolute value indicates a greater positive activation contribution of the acoustic feature to the target neuron; a negative weight value with a larger absolute value indicates a stronger inhibitory effect of the feature on the target neuron; and a weight value close to zero indicates that the feature has almost no effect on the activation of the target neuron. This one-to-one parameter correspondence allows the model to autonomously learn the importance weights of different acoustic features in quiet / noisy scene classification after offline training, which is the foundation for achieving acoustic scene differentiation. Based on a clear understanding of the weight parameters, the matrix multiplication operation is not a simple numerical calculation but an operation with a clear feature aggregation logic. This operation scales all six feature values ​​of the comprehensive feature vector according to their corresponding weight values ​​in the weight matrix, and then systematically aggregates all scaled feature values ​​to each neuron in the hidden layer, ultimately forming the initial weighted input for each neuron in the hidden layer. The specific formula for calculating the initial weighted input of each neuron is as follows: Each parameter in the formula has a clear physical definition. Indicates the hidden layer number 1 The initial weighted input corresponding to the nth neuron, which is the nth intermediate weighted sum vector. One element, The first eigenvector represents the composite eigenvector. Each acoustic feature value Represents the weight matrix of the first element. Line 1 The weight value of the column, i.e., the weight of the composite feature vector. The first feature affects the hidden layer. The influence strength of each neuron This represents the summation of the scaled results of the six input features. Through this matrix multiplication operation, each hidden layer neuron can receive the aggregated information of all input acoustic features after being scaled according to their weights, achieving efficient linear fusion of multi-dimensional acoustic features and providing a foundation for subsequent pattern recognition. Simultaneously, the bias vector, which is trained offline and stored in the audio scene classification unit in sync with the weight matrix, is a key auxiliary parameter for adjusting the activation of hidden layer neurons. Each element in the bias vector uniquely corresponds to a neuron in the hidden layer, providing an independent baseline adjustment for each neuron. This baseline adjustment is a fixed offset value learned by the model through a large number of acoustic scene samples during offline training. Its function is to adjust the activation threshold of each neuron, avoiding the homogenization of activation thresholds and the simplification of pattern representation caused by all neurons relying solely on feature weighting. This allows each neuron to possess independent feature response characteristics, significantly improving the model's generalization ability to recognize different acoustic scenes.

[0019] After obtaining the initial weighted input for each neuron through matrix multiplication, the system performs an element-wise addition operation. This operation adds the baseline adjustment provided by the bias vector for each neuron to the corresponding initial weighted input in a one-to-one manner. The specific calculation formula is as follows: , in the formula The first character representing the higher-order pattern features One element, Represents the corresponding bias vector in the first bias vector. The baseline adjustment of each neuron, when summed, forms the complete computational basis for each element in the higher-order pattern feature. In other words, each element of the higher-order pattern feature is composed of two parts: the initial weighted input obtained by weighted scaling of the input features and the baseline adjustment of the neuron. This not only fully preserves the linear fusion information of the multi-dimensional acoustic features but also incorporates independent neuron activation baseline adjustment, giving the higher-order pattern feature a stronger ability to distinguish acoustic scene patterns. It completes the linear transformation from low-dimensional original acoustic features to high-dimensional abstract higher-order features, laying a solid and high-quality feature foundation for subsequent nonlinear mapping transformation through preset nonlinear activation functions and extraction of more discriminative nonlinear acoustic pattern features.

[0020] After obtaining the high-order pattern feature vectors representing the linear correlation laws of the acoustic scene, in order to overcome the limitation that pure linear operations can only fit simple feature relationships and enable the multilayer perceptron model to learn the complex nonlinear patterns of the acoustic environment in which the set-top box is located, and thus distinguish different scenes such as quiet, noisy, and sudden sound, the system immediately performs an element-wise numerical mapping transformation of the high-order pattern feature vectors using a preset nonlinear activation function. This step is the link that upgrades linear features to nonlinear pattern features, and it is also the key to the deep learning model fitting complex acoustic laws. The linear feature transformation forms a tight connection, and the specific implementation method is clear, complete, and feasible for embedded devices. The detailed description is as follows: First, let's clarify the concepts. The high-order pattern feature vector is the hidden layer output vector obtained by linearly weighted summation and adding the bias. Each element corresponds to the linear computation result of a neuron in the hidden layer, fully carrying the linear fusion information of multi-dimensional acoustic features, but it can only express simple linear relationships between features. The preset nonlinear activation function is the Corrected Linear Unit (ReLU) function selected by this system for the hidden layer of the multilayer perceptron. This function is a classic nonlinear transformation function in the field of deep learning that is adapted to embedded terminals, has high computational efficiency, and has no gradient saturation problem. Its role is to introduce nonlinear fitting ability into the linear network, allowing the model to learn the complex mapping relationship between acoustic features and scene classification, rather than being limited to a simple linear combination. When the system performs the transformation, it will traverse each element in the high-order pattern feature vector obtained by linearly weighted summation and adding the bias, that is, the independent computation result of each neuron in the hidden layer, and apply the preset nonlinear activation function point-to-point numerical mapping in turn. This transformation is an element-by-element independent operation and does not change the overall dimension of the feature vector. It only performs nonlinear correction on the value of each element. The calculation formula is as follows: The parameters in the formula have clear physical definitions and operational rules, among which This represents the single-element value output after the nonlinear activation transformation, and also the corresponding element in the final nonlinear acoustic mode feature vector. This represents the value of a single element to be transformed in the higher-order pattern feature vector, i.e., the linear calculation result of a single neuron in the hidden layer after linear weighted summation and bias addition. The specific execution logic for the maximum value operation is as follows: When the element value in the higher-order pattern feature vector When the value is positive, the preset non-linear activation function directly retains and outputs that value, i.e. The response intensity of effective features is preserved; when the element values ​​in the higher-order mode feature vector... When the value is zero or negative, a pre-defined non-linear activation function forces it to be mapped to zero, i.e. This suppresses interference from invalid features. By applying a pre-defined nonlinear activation function to all elements of the higher-order pattern feature vector, the higher-order pattern features, originally derived entirely by linear calculation, are successfully introduced into a nonlinear transformation. This element-wise nonlinear transformation breaks through the expressive bottleneck of linear operations, allowing hidden layer neurons to autonomously learn and extract complex nonlinear correlation patterns of acoustic features such as short-time average energy, short-time average zero-crossing rate, and spectral energy proportion under different acoustic scenarios. For example, it reveals a complex nonlinear pattern of low mean energy, low zero-crossing rate, and high low-frequency proportion in a quiet environment, rather than a simple linear superposition of features. All elements after the nonlinear mapping transformation remain unchanged in their original order, and the combination... A nonlinear acoustic mode feature vector is formed that is completely consistent with the dimension of the high-order mode feature vector. This vector not only retains the information of the original acoustic features, but also has strong nonlinear mode expression ability and scene discrimination, characterizing the essential differences of different acoustic environments. This nonlinear acoustic mode feature vector will be used as the output of the last hidden layer and directly input to the output layer of the multilayer perceptron, providing nonlinear feature support for subsequent dot product operations, superposition of output bias, and generation of quiet index. The entire nonlinear transformation process has a very small amount of computation and a very fast execution speed, which is fully compatible with the hardware resources and real-time operation requirements of the set-top box embedded system, ensuring the classification and efficiency of acoustic scenes.

[0021] The preset nonlinear activation function is the modified linear unit function. The modified linear unit function calculates the input value for each element. If the input value is positive, it outputs the input value directly; if the input value is negative, it outputs zero.

[0022] After performing a nonlinear activation transformation on the high-order pattern feature vector in the last hidden layer of the audio scene classification unit, resulting in a nonlinear acoustic pattern feature vector capable of fitting complex acoustic scenes and distinguishing between quiet and noisy environments, this vector serves as the sole input data for the output layer of the multilayer perceptron model. In the final scene quantification scoring stage, through linear weighting, bias stacking, and activation function mapping in the output layer, the high-dimensional nonlinear features are transformed into a single scalar form of a quietness index, achieving a quantitative representation of the quietness of the acoustic environment. This step is the final output step from feature extraction to scene determination. The linear transformation and nonlinear activation process in the hidden layers form a complete neural network forward inference loop. The specific implementation method is complete and adapted to the lightweight operation requirements of set-top box embedded terminals. First, the parameters and operation definitions are clarified. The nonlinear acoustic mode feature vector output by the last hidden layer is a high-dimensional feature vector obtained after linear weighting, bias addition, and ReLU nonlinear activation of the hidden layer. Its dimension is exactly the same as the number of neurons in the last hidden layer. The depth carries all the nonlinear mode information of acoustic features such as short-time average energy, short-time average zero-crossing rate, and spectral energy ratio. It is the input basis for the output layer to complete the final calculation. The pre-trained output weight vector is a one-dimensional vector obtained by offline training and optimization of a large number of acoustic scene samples in sync with the overall model weights. Its length is the same as the nonlinear acoustic mode feature vector. The vectors are all of equal dimension. Each weight value in the vector corresponds to the contribution strength of an element in the nonlinear acoustic mode feature vector to the final quiet index. After training, they are stored in the embedded storage module of the audio scene classification unit and can be directly called during the online inference stage without readjustment. The pre-trained output bias scalar value is a single numerical form of the baseline bias parameter, and its optimal value is also determined through offline training. It is used to provide a fixed offset for the linear operation of the output layer, avoiding numerical deviation caused by the output result relying solely on the feature weighted sum. The preset output activation function used in this system is the Sigmoid function, which is a classic normalized activation function in deep learning that can map any real number to a fixed value between 0 and 1. Within the specified range, the quietness index is set as a scalar value between 0 and 1. A value closer to 0 indicates a quieter environment, while a value closer to 1 indicates a noisier environment. In the output layer's computation phase, the input nonlinear acoustic mode feature vector is first multiplied by the pre-trained output weight vector. The dot product, a linear operation specific to one-dimensional vectors, is implemented by multiplying element-wise and then summing the results. This involves sequentially multiplying each element of the nonlinear acoustic mode feature vector with the corresponding weight value in the output weight vector, then summing all the product results to obtain a preliminary weighted sum in a single numerical form. The formula for this dot product operation is as follows: , This is the initial weighted sum obtained after the dot product operation of the output layer. This represents the total number of neurons in the last hidden layer, and is also the dimension of the nonlinear acoustic pattern feature vector and the output weight vector. The eigenvector of the nonlinear acoustic mode The element value at each position. The first element in the pre-trained output weight vector The system assigns weights to each position, efficiently compressing high-dimensional nonlinear features into a single weighted value, thus completing the initial linear mapping from feature to score. After obtaining the initial weighted sum, the system performs a scalar addition operation with the pre-trained output bias scalar value to obtain the original linear output value, calculated using the following formula: ,in This is the original linear output value, which is an arbitrary real number with no fixed range of values ​​and cannot be directly used as a standardized quietness index. The pre-trained output bias scalar value is a single constant parameter that provides a baseline adjustment for the linear output, ensuring the reasonableness and accuracy of the output value. To transform the unconstrained raw linear output value into a standardized quiet index that can be directly used for scene determination, the system inputs the raw linear output value into a preset output activation function, performing nonlinear mapping and numerical range constraints. The Sigmoid output activation function used in this system is calculated using the following formula: , in the formula This is the final quiet index obtained after mapping constraints. It is a natural constant. As the original linear output value, this function can map any real number to a fixed numerical range of 0 to 1, perfectly meeting the quantification requirements of the quietness index. Through this final mapping transformation, the model completes the entire process of inference calculation from multi-dimensional acoustic features to a single quietness index. After the quietness index is generated, it will be immediately transmitted to the comparison module of the audio scene classification unit to be compared with the preset quietness threshold, thereby generating the corresponding binary control signal. This provides a basis for judgment for the set-top box backend data processing task scheduling of the subsequent task scheduling execution unit. The entire output layer operation process is logically rigorous and computationally lightweight, with no complex calculations or high memory consumption. It is fully compatible with the hardware performance and real-time operation requirements of the set-top box, realizing a seamless connection between acoustic scene perception and intelligent task scheduling.

[0023] After the audio scene classification unit completes dot product operations, output bias superposition, and Sigmoid activation function mapping through the output layer to obtain a quiet index that quantifies the quietness of the current acoustic environment, this quiet index is transmitted in real time and without distortion to the comparison module integrated within the audio scene classification unit. This then formally enters the acoustic scene compliance judgment and binary control signal generation stage. This stage is a crucial link between intelligent acoustic perception and set-top box backend task scheduling. Standardized control instructions are generated through numerical comparison at fixed thresholds, providing a clear execution basis for subsequent task scheduling execution units. This forms a complete and rigorous logical loop with the calculation and reasoning process of the quiet index. The comparison module is a hardware logic unit or embedded software function module within the audio scene classification unit specifically responsible for high-speed numerical comparison, instruction logic judgment, and control signal output. It features low latency, high stability, and no calculation errors, and can quickly complete the comparison of floating-point values ​​and output a standard level signal. The quiet threshold is determined during the factory deployment or on-site configuration phase based on typical scenarios such as home viewing, nighttime quiet, and daily use of the set-top box. The actual noise requirements of the environment are pre-calibrated and permanently stored in the set-top box's non-volatile memory as a fixed critical value. This value, along with the quiet index, is in the standardized range of 0 to 1 and serves as the benchmark for determining whether the current environment meets the conditions for silent operation. The first and second control signals are standard binary digital control signals commonly used within the set-top box system. They are distinguished by high and low levels, ensuring stable transmission and rapid response. They are the standard communication carriers for realizing instruction interaction between modules. After receiving the quiet index, the comparison module immediately retrieves the pre-set quiet threshold from local storage and performs a real-time high-precision comparison operation on the two sets of values. The comparison logic strictly follows the preset rules. If the value of the quiet index is less than the value of the quiet threshold, it indicates that the short-time average energy of the current environment is low, the short-time average zero-crossing rate is flat, and the proportion of preset low-frequency spectrum energy is stable. The acoustic environment is in a quiet state without significant noise interference, corresponding to the set-top box user being in a noise-sensitive usage scenario such as watching a movie or resting. At this time, the comparison module officially determines that the current environment's quietness level meets the standard and then generates and outputs the first control signal.If the quietness index is greater than or equal to the quietness threshold, it indicates that the current ambient audio energy is high, the time domain changes drastically, or there is ambient noise masking, and the acoustic environment is in a non-quiet state. However, this will not cause interference to the user due to equipment noise. At this point, the comparison module formally determines that the current ambient quietness level is non-standard, and then generates and outputs a second control signal. Both control signals are transmitted synchronously to the task scheduling execution unit 3 in real-time at standard levels to ensure the real-time performance and reliability of instruction transmission. The task scheduling execution unit 3 is the execution unit in the set-top box data management system responsible for background task start / stop control, instruction parsing and execution, and scheduling status feedback. It establishes a two-way real-time communication connection with the audio scene classification unit and the set-top box's built-in data processing module, possessing the ability to quickly parse instructions, schedule tasks, and link device status. The data processing module is a background function module inside the set-top box responsible for cache cleaning, log archiving, application data optimization, firmware defragmentation, and storage partition maintenance. This module performs high disk I / O operations involving frequent hard drive read / write. Background data organization tasks with (I / O) load refer to maintenance tasks that require continuous, high-frequency read and write operations on the set-top box's built-in mechanical hard drive or flash memory. These tasks generate hard drive read / write noise or circuit noise during operation, which can seriously affect the user experience in a quiet environment. Therefore, it is necessary to achieve intelligent start-stop control through environmental perception. After receiving the binary control signal sent by the comparison module, the task scheduling execution unit 3 will immediately parse the signal type and strictly execute the corresponding background data task scheduling operation according to the preset strategy. When the received signal is identified as the first control signal, the task scheduling execution unit 3 will immediately send a task pause command to the set-top box's data organization module. After the command is issued, the data organization module will immediately terminate all running high disk I / O load background data organization tasks and refuse to respond to any new similar task start requests, thereby eliminating noise interference caused by frequent hard drive read / write from the root and ensuring an interference-free user experience in a quiet environment. When the received signal is identified as the second control signal, the task scheduling execution unit 3 immediately sends a task start permission command to the set-top box's data processing module. After the command is issued, the data processing module can start and run all the preset background data processing tasks normally according to the system's preset task scheduling priority and execution sequence. It makes full use of non-quiet environment periods to complete the set-top box's data optimization, storage cleanup, system maintenance, and other tasks, ensuring the set-top box's operating performance and storage efficiency without interfering with the user experience. The entire process of comparison judgment, signal generation, command transmission, and task scheduling is fully automated and closed-loop, requiring no manual operation, no online parameter adjustment, and no additional hardware overhead. It has fast calculation response speed and high execution stability, perfectly realizing intelligent scheduling of set-top box data management based on acoustic environment perception. It not only takes into account the user's need for noise-free use in quiet scenarios, but also ensures the normal execution of the set-top box's background maintenance tasks, giving the entire set-top box's data management system the application capabilities of environmental adaptation and intelligent collaboration.

[0024] Please see Figure 2 As shown, the second objective of this invention is to provide a management method for implementing a data management system for a set-top box, comprising the following steps: S1. Obtain the original environmental audio signal through the audio acquisition module, preprocess the signal to obtain the time domain signal, calculate the short-time average energy and short-time average zero-crossing rate of the time domain signal in parallel, and perform fast Fourier transform on the time domain signal to obtain the frequency domain power spectrum and extract the spectral energy ratio of the preset low frequency band. Combine the short-time average energy, short-time average zero-crossing rate and spectral energy ratio to generate a comprehensive feature vector. S2. Input the comprehensive feature vector into the pre-trained multilayer perceptron neural network model. The hidden layer of the model performs linear weighted summation based on pre-trained weights and biases and nonlinear activation function mapping on the vector to extract nonlinear acoustic pattern features. Then, the output layer processes and calculates to output a quiet index that represents the quietness of the environment. S3. Compare the calculated quiet index with a preset quiet threshold, and generate a binary control signal based on the comparison result. S4. Execute data processing task scheduling based on the generated binary control signal. When the signal indicates that the environment is quiet, suspend high disk I / O load tasks. When the signal indicates that the environment is not quiet, allow the background data processing task to start.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data management system for a set-top box, characterized by, include: The environmental perception unit (1) acquires the original environmental audio signal, preprocesses the original environmental audio signal to obtain the time domain signal, calculates the short-time average energy and short-time average zero-crossing rate in parallel for the time domain signal, performs a fast Fourier transform on the time domain signal to obtain its frequency domain power spectrum, extracts the spectral energy ratio of the preset low-frequency band from the frequency domain power spectrum, and finally fuses the short-time average energy, short-time average zero-crossing rate and spectral energy ratio to generate a comprehensive feature vector for characterizing the current acoustic environment. The audio scene classification unit (2) inputs the comprehensive feature vector into the pre-trained multilayer perceptron neural network model. The multilayer perceptron neural network model includes an input layer, a hidden layer, and an output layer. The input layer receives the comprehensive feature vector and passes it to the hidden layer. The hidden layer is used to linearly weight and sum the comprehensive feature vector with the pre-trained weight matrix to obtain the high-order mode features used to distinguish different acoustic scenes. The high-order mode features are mapped and transformed by a preset nonlinear activation function to obtain nonlinear acoustic mode features. The output layer processes and calculates the nonlinear acoustic mode features and outputs a scalar value representing the quietness of the environment, which is called the quietness index. The quietness index is compared with the preset quietness threshold, and a binary control signal is generated based on the comparison result. The task scheduling execution unit (3) performs task scheduling based on the binary control signal data.

2. The data management system of a set-top box according to claim 1, wherein: The sum of squares and average of the amplitudes of all sample points of the time-domain signal are calculated to obtain the short-time average energy, which is used to characterize the volume intensity of the time-domain signal. Simultaneously, the ratio of the number of symbol changes between adjacent sample points of the time-domain signal to the frame length is calculated to obtain the short-time average zero-crossing rate. The short-time average zero-crossing rate is used to reflect the frequency of zero crossings of the time-domain signal in the time domain. The fast Fourier transform algorithm is applied to the time-domain signal to transform it from the time domain to the frequency domain. The square of the modulus of the complex number corresponding to each frequency point after the transformation is calculated to obtain the frequency domain power spectrum of the time-domain signal. The frequency domain power spectrum reflects the energy distribution of the time-domain signal at different frequency components.

3. The data management system of a set-top box according to claim 2, wherein, For multiple time-domain signals acquired continuously within a preset time window, the short-time average energy, short-time average zero-crossing rate, and spectral energy ratio corresponding to each frame are extracted to obtain multiple sets of instantaneous features; The mean and variance of the short-time average energy, the short-time average zero-crossing rate, and the spectral energy percentage are calculated within the time window, respectively. Finally, the mean and variance of the short-time average energy, the mean and variance of the short-time average zero-crossing rate, and the mean and variance of the spectral energy percentage are concatenated and combined into a one-dimensional array in a preset order. The one-dimensional array constitutes the comprehensive feature vector used to characterize the current acoustic environment.

4. The data management system for a set-top box according to claim 3, characterized in that: The pre-trained weight matrix is ​​stored in the audio scene classification unit (2). The number of rows of the pre-trained weight matrix is ​​the same as the dimension of the comprehensive feature vector. The number of columns of the pre-trained weight matrix corresponds to the number of neurons contained in the hidden layer. When performing linear weighted summation, the comprehensive feature vector is used as the input vector and matrix multiplication is performed with the weight matrix, specifically as follows: Each feature value of the comprehensive feature vector is multiplied by all weight values ​​in the corresponding row of the weight matrix and summed along the row direction to generate an intermediate weighted sum vector. The pre-trained bias vector is then added element-wise to the intermediate weighted sum vector to obtain the higher-order pattern feature. The dimension of the higher-order pattern feature is equal to the number of neurons in the hidden layer.

5. The data management system for a set-top box according to claim 4, characterized in that: Each weight value in the weight matrix corresponds to the influence strength of a specific feature in the comprehensive feature vector on a specific neuron in the hidden layer. The matrix multiplication operation scales all the feature values ​​of the comprehensive feature vector according to their corresponding weights and summarizes them to each neuron to form the initial weighted input of each neuron. The bias vector provides a baseline adjustment for each neuron. The element-wise addition operation adds the baseline adjustment to the initial weighted input of the corresponding neuron, which together constitute the calculation basis for each element in the higher-order pattern feature.

6. The data management system for a set-top box according to claim 5, characterized in that: For each element in the higher-order pattern feature vector obtained after linear weighted summation and bias addition, i.e. the calculation result of each neuron, the preset nonlinear activation function is applied for numerical mapping. By applying a preset nonlinear activation function to each element of the higher-order mode feature vector, the higher-order mode features originally calculated linearly are introduced into a nonlinear transformation to obtain a nonlinear acoustic mode feature vector.

7. A set-top box data management system according to claim 6, characterized in that: The preset nonlinear activation function specifically adopts the modified linear unit function. The modified linear unit function calculates the input value of each element. If the input value is positive, it directly outputs the input value; if the input value is negative, it outputs zero.

8. The data management system for a set-top box according to claim 7, characterized in that: The nonlinear acoustic pattern feature vector output from the last hidden layer is used as the input to the output layer. In the output layer, the nonlinear acoustic pattern feature vector is multiplied by the pre-trained output weight vector, that is, each element of the nonlinear acoustic pattern feature vector is multiplied by the corresponding weight value in the output weight vector. All product results are summed to obtain a preliminary weighted sum. The preliminary weighted sum is added to the pre-trained output bias scalar value to obtain the original linear output value. Finally, the original linear output value is input to a preset output activation function. The output activation function maps and constrains the original linear output value within a fixed numerical range. The final value obtained after mapping and constraint is the quietness index.

9. A data management system for a set-top box according to claim 8, characterized in that: The comparison module receives the quiet index and compares the value of the quiet index with the value of the pre-set and stored quiet threshold. If the value of the quiet index is less than the value of the quiet threshold, it is determined that the current environment quietness meets the standard, and the first control signal is generated and output. If the value of the quiet index is greater than or equal to the value of the quiet threshold, it is determined that the current environment quietness does not meet the standard, and the second control signal is generated and output. When the task scheduling execution unit (3) receives the first control signal, it sends an instruction to the set-top box's data processing module to suspend all background data processing tasks involving high disk input / output loads that frequently read and write hard disks. When it receives the second control signal, it sends an instruction to the data processing module to allow the start of all preset background data processing tasks.

10. A management method for implementing a data management system for a set-top box according to any one of claims 1-9, characterized in that: Includes the following steps: S1. Obtain the original environmental audio signal through the audio acquisition module, preprocess the signal to obtain the time domain signal, calculate the short-time average energy and short-time average zero-crossing rate of the time domain signal in parallel, perform fast Fourier transform on the time domain signal to obtain the frequency domain power spectrum and extract the spectral energy ratio of its preset low frequency band, and fuse the short-time average energy, short-time average zero-crossing rate and spectral energy ratio to generate a comprehensive feature vector. S2. Input the comprehensive feature vector into a pre-trained multilayer perceptron neural network model. Through the hidden layer of the model, perform linear weighted summation and nonlinear activation function mapping on the vector based on pre-trained weights and biases to extract nonlinear acoustic pattern features. Then, process and calculate through the output layer to output a quiet index representing the level of environmental quietness. S3. Compare the calculated quiet index with a preset quiet threshold, and generate a binary control signal based on the comparison result; S4. Execute data processing task scheduling according to the generated binary control signal. When the signal indicates that the environment is quiet, suspend high disk I / O load tasks. When the signal indicates that the environment is not quiet, allow the background data processing task to start.

Images