An electric drive non-steady state sound quality prediction method based on intelligent sensor

Abstract

This invention discloses a method for predicting the unsteady acoustic quality of electric drives based on intelligent sensors. Specifically, it includes: acquiring noise signals and extracting sound pressure level, loudness, sharpness, speech intelligibility, and priority speech interference level to form a feature vector sequence and dataset; constructing a neural network to determine nodes and form network parameter vectors; establishing an information entropy budget allocation model, mapping each feature to an information budget variable vector, setting constraints to construct an objective function containing budget constraint terms and prediction error terms, and embedding the budget variables into the feature representation; constructing a particle swarm, uniformly encoding the information budget vector and network parameter vector into particle state vectors and then initializing them; introducing an information budget modulation factor during iteration to weight the velocity update component, achieving co-evolution of the two; updating the optimal particle based on fitness, outputting the optimal information budget and network parameters to construct a prediction model; and inputting the feature vector sequence to output the prediction result. This invention assigns independent budgets to each acoustic feature, dynamically co-optimizes, and achieves high accuracy and fast response.

Description

Technical Field

[0001] This invention relates to the field of electric drive acoustic quality prediction technology, and in particular to a method for predicting the unsteady acoustic quality of electric drives based on intelligent sensors. Background Technology

[0002] The field of electric drive unsteady-state sound quality prediction technology involves the acquisition and processing of noise signals from electric drive systems. It uses intelligent sensors to obtain sound quality characteristic parameters such as sound pressure level, loudness, and sharpness, and employs a neural network model to predict and evaluate sound quality under unsteady-state conditions.

[0003] Existing methods for predicting the acoustic quality of electric drives mostly employ backpropagation (BP) neural network models, using gradient descent to train network parameters and establish a mapping relationship between acoustic quality characteristics and subjective evaluation results. However, BP neural networks are prone to getting trapped in local optima during training and are highly sensitive to the selection of initial weights and thresholds; different initial parameter settings can lead to significant differences in prediction accuracy. To address these issues, some studies have used particle swarm optimization (PSO) algorithms to globally search for the initial parameters of BP neural networks. However, traditional PSO algorithms are prone to problems such as decreased population diversity and slower convergence speed in the later stages of iteration, affecting the parameter optimization effect.

[0004] Furthermore, existing methods typically employ a uniform weighting approach for features such as sound pressure level, loudness, and sharpness during sound quality feature processing, failing to consider the varying degrees of contribution of different features to sound quality under non-steady-state conditions. This makes it difficult to fully extract the sound quality information contained in each feature dimension. The limited feature fusion method restricts the model's ability to represent complex non-steady-state sound quality, affecting the accuracy of prediction results. Summary of the Invention

[0005] One objective of this invention is to propose an electric-driven unsteady-state acoustic quality prediction method based on intelligent sensors. This invention introduces an information entropy budget allocation mechanism and a dynamic adaptive particle swarm optimization model. Independent information budget variables are assigned to sound pressure level, loudness, sharpness, speech intelligibility, and priority speech interference level features. An objective function containing budget constraint terms and prediction error terms is constructed. The information budget vector and network parameter vector are uniformly encoded into particle state vectors. During particle iteration, an information budget modulation factor is introduced to weight the velocity update components. The co-evolution and dynamic recalculation of the information budget vector and network parameter vector are performed to construct a continuously optimized unsteady-state acoustic quality prediction model, which has the advantages of fast response speed, high feature discrimination accuracy, and strong prediction stability.

[0006] A method for predicting the unsteady acoustic quality of an electric drive based on a smart sensor, according to an embodiment of the present invention, includes the following steps: Noise signals from the electric drive system during operation are collected, and feature vector sequences consisting of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level are extracted to establish a dataset corresponding to the sound quality evaluation results. A feedforward neural network structure is constructed, and input nodes, hidden layer nodes, and output nodes are determined to form a network parameter vector containing connection weights and biases. An information entropy budget allocation model is established, mapping the sound pressure level, loudness, and sharpness to information budget variables to form an information budget vector. Information budget constraints are set, and an objective function containing budget constraint terms and prediction error terms is constructed based on these constraints. The information budget variables are embedded into a representation structure composed of sound pressure level, loudness, and sharpness. A particle swarm search space is constructed, and the information budget vector and network parameter vector are uniformly encoded into particle state vectors. Particle position, velocity, individual optimal position, and swarm optimal position are initialized. An information budget modulation factor is introduced during particle iteration to weight the particle velocity update, forming a differentiated search path driven by the information budget, realizing the co-evolution of the information budget vector and the network parameter vector. The optimal position of the particle is updated based on the fitness results, and the optimal information budget vector and the optimal network parameter vector are output to construct the prediction model. The feature vector sequence is input into the prediction model, and the non-steady-state sound quality prediction results are output.

[0007] Optionally, establishing the dataset corresponding to the sound quality evaluation results specifically includes: Several sample segments are extracted from the noise signal collected under unsteady operating conditions. Subjective listening evaluation is performed on these sample segments using multiple sub-adaptive pairwise comparison methods. Several evaluators score the sound quality of each sample segment, obtaining a subjective score value for each segment. For each sample segment, the sound pressure level, loudness, sharpness, speech intelligibility, and priority speech interference level are extracted, and their time-varying sequences are used to construct a feature vector sequence. The feature vector sequence of each sample segment is then stored in correspondence with its subjective score value, forming a dataset.

[0008] Optionally, the construction of the feedforward neural network structure, determining the input nodes, hidden layer nodes, and output nodes, and forming a network parameter vector containing connection weights and biases includes the following steps: Set the number of input nodes to equal the dimension of the feature vector sequence, and the number of output nodes to equal the dimension of the subjective rating value; determine the number of hidden layer nodes, which is equal to the square root of the sum of the number of input nodes and the number of output nodes plus an adjustment parameter, where the adjustment parameter is an integer from 1 to 10. For each value of the adjustment parameter, calculate the corresponding number of hidden layer nodes, construct a feedforward neural network with the specified number of hidden layer nodes, train it with the feature vector sequence as input and the subjective rating value as output, calculate the mean square error of each network after training, and select the number of hidden layer nodes corresponding to the minimum mean square error. The dimension of the network parameter vector is calculated based on the number of input nodes, hidden layer nodes, and output nodes. The network parameter vector consists of four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. The four parts of parameters are randomly initialized and then concatenated in the order of the first to the fourth part to form the network parameter vector.

[0009] Optionally, establishing the information entropy budget allocation model specifically includes: Information budget variables are defined for sound pressure level (SPL), loudness, sharpness, speech intelligibility, and preferred speech interference level, forming an information budget vector. Information budget constraints are set, with each variable taking a value within a closed interval of 0 to 1, and the sum of the five variables equaling 1. The information budget variables are embedded into a representation structure consisting of SPL, loudness, sharpness, speech intelligibility, and preferred speech interference level. Specifically, the SPL sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted SPL sequence; the loudness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted loudness sequence; the sharpness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted sharpness sequence; the speech intelligibility sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted speech intelligibility sequence; and the preferred speech interference level sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted preferred speech interference level sequence. The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted priority speech interference level sequence are concatenated to form a weighted feature vector sequence. An objective function is constructed, which consists of a budget constraint term and a prediction error term. The budget constraint term is the negative of the entropy of the information budget vector, and the prediction error term is the mean squared error generated when the feedforward neural network is trained with the weighted feature vector sequence as input and subjective ratings as output.

[0010] Optionally, constructing the objective function includes the following steps: Obtain the information budget vector, which includes a first information budget variable corresponding to the sound pressure level, a second information budget variable corresponding to the loudness, a third information budget variable corresponding to the sharpness, a fourth information budget variable corresponding to the speech intelligibility, and a fifth information budget variable corresponding to the preferred speech interference level; Calculate the budget constraint term: multiply the first, second, third, fourth, and fifth information budget variables by their respective logarithms to the base 2, sum the five products, and take the opposite of the sum; Obtain the sound pressure level sequence, loudness sequence, sharpness sequence, speech intelligibility sequence, and preferred speech interference level sequence for each sample segment. Multiply the value of each sampling point in each sequence by the corresponding information budget variable to generate a weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, speech intelligibility sequence, and preferred speech interference level sequence. Concatenate the five weighted sequences according to the sampling point index to form a weighted feature vector sequence. Construct a feedforward neural network with the number of input layer nodes equal to the dimension of the weighted feature vector sequence and the number of output layer nodes equal to the dimension of the subjective rating value. Train the network using the weighted feature vector sequence as input and the subjective rating value of the corresponding sample segment as the target. After training, calculate the mean squared error between the network output value and the target output value as the prediction error term. Add the budget constraint term and the prediction error term to obtain the objective function value.

[0011] Optionally, the step of uniformly encoding the information budget vector and the network parameter vector into a particle state vector specifically includes: Obtain the dimension value of the information budget vector, which is 5. Obtain the dimension value of the network parameter vector, calculated from the number of input nodes, hidden layer nodes, and output nodes. Set the total dimension value of the particle state vector to be equal to the sum of the dimensions of the information budget vector and the network parameter vector. Use the three information budget variables of the information budget vector as the 1st, 2nd, 3rd, 4th, and 5th components of the particle state vector, respectively. Use all parameters of the network parameter vector as the 6th to the last component of the particle state vector, respectively. Set the value range of the 1st to 5th components to be a closed interval of 0 to 1. Set the value range of the 6th to the last component to be a closed interval of -3 to +3. Normalize the 1st to 5th components; after normalization, the sum of the three components equals 1.

[0012] Optionally, introducing an information budget modulation factor during particle iteration includes the following steps: The particle state vector dimensions correspond to the following rules: the first three dimensions are the sound pressure level information budget variable, loudness information budget variable, sharpness information budget variable, speech intelligibility information budget variable, and priority speech interference level information budget variable, respectively. Starting from the 6th dimension, the weights are sequentially: input layer to hidden layer connection weight, hidden layer threshold, hidden layer to output layer connection weight, and output layer threshold. Each input weight records the input feature index it connects to. The inertia weight, individual learning factor, and social learning factor for the current iteration number are obtained. For each particle, the current position vector, current velocity vector, individual optimal position vector, and group optimal position vector are obtained. The first 5 dimensions are extracted from the current position vector as the current information budget vector. A modulation factor vector is constructed: the first 5 dimensions are equal to the 5 components of the current information budget vector. For each dimension starting from the 4th dimension, if this dimension corresponds to the input layer to hidden layer connection weight, then the corresponding component is taken from the current information budget vector based on the input feature index associated with this weight as the modulation factor for this dimension. If this dimension corresponds to the hidden layer threshold, the hidden layer-to-output layer connection weight, and the output layer threshold, then the modulation factor for this dimension is set to 1. For each dimension, calculate the original velocity update: multiply the inertia weight by the component of the current velocity vector in this dimension, add the individual learning factor multiplied by the random number multiplied by the difference between the individual's optimal position vector in this dimension and the current position vector in this dimension, add the social learning factor multiplied by the random number multiplied by the difference between the group's optimal position vector in this dimension and the current position vector in this dimension, to obtain the original velocity update. Multiply the original velocity update by the value of this dimension in the modulation factor vector to obtain the modulated velocity update. Add the value of this dimension of the current position vector to the modulated velocity update to obtain the value of this dimension of the updated position vector. After completing the calculation of all dimensions, the updated position vector of the particle is obtained.

[0013] Optionally, the construction of the prediction model by outputting the optimal information budget vector and the optimal network parameter vector specifically includes: The first five components are extracted from the group's optimal position vector as the optimal information budget vector; the remaining components are extracted from the group's optimal position vector as the optimal network parameter vector; based on the number of input nodes, hidden layer nodes, and output nodes, the optimal network parameter vector is divided into four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. Create a feedforward neural network, setting the number of input nodes, hidden layer nodes, and output nodes to be the same as during optimization. Arrange the first part into a matrix with rows of hidden layer nodes and columns of input nodes, and assign it to the connection weights from the input layer to the hidden layer. Arrange the second part into a vector with rows of hidden layer nodes and 1 column, and assign it to the hidden layer bias. Arrange the third part into a matrix with rows of output nodes and columns of hidden layer nodes, and assign it to the connection weights from the hidden layer to the output layer. Arrange the fourth part into a vector with rows of output nodes and 1 column, and assign it to the output layer bias. Set the input feature weighting rule: For the input sound pressure level sequence, multiply the sound pressure level value of each sampling point in the sequence by the first component in the optimal information budget vector to generate a weighted sound pressure level sequence. For the input loudness sequence, the loudness value of each sampling point in the sequence is multiplied by the second component in the optimal information budget vector to generate a weighted loudness sequence; For the input sharpness sequence, the sharpness value of each sampling point in the sequence is multiplied by the third component in the optimal information budget vector to generate a weighted sharpness sequence; For the input speech proficiency sequence, the speech proficiency value of each sampling point in the sequence is multiplied by the fourth component in the optimal information budget vector to generate a weighted speech proficiency sequence; For the input priority language interference level sequence, the priority language interference level value of each sampling point in the sequence is multiplied by the fifth component in the optimal information budget vector to generate a weighted priority language interference level sequence. The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted priority speech interference level sequence are concatenated according to the sampling point index order to form a weighted feature vector sequence. The input feature weighting rules are combined with the assigned feedforward neural network to form a prediction model. The input of the prediction model is the original sound pressure level sequence, original loudness sequence, original sharpness sequence, original speech intelligibility sequence, and original priority speech interference level sequence. The output of the prediction model is the calculation result of the weighted feature vector sequence by the assigned feedforward neural network.

[0014] The beneficial effects of this invention are: (1) This invention constructs an information entropy budget allocation mechanism to allocate independent information budget variables for sound pressure level, loudness and sharpness, thereby realizing the differentiated representation of different sound quality features under non-steady-state conditions. After embedding the information budget variables into the feature representation structure, an objective function containing budget constraint terms and prediction error terms is constructed, so that the model can automatically learn the contribution weight of each feature to sound quality during the training process, obtain a more refined feature fusion method, and improve the accuracy of non-steady-state sound quality prediction.

[0015] (2) The present invention adopts a collaborative evolution mechanism that combines dynamic adaptive particle swarm optimization algorithm with information budget modulation factor. The information budget vector and network parameter vector are uniformly encoded into particle state vector. During the particle velocity update process, the modulation factor is introduced to perform component weighting according to the correlation between parameters and input features, so as to realize the joint optimization of information budget and network parameters. This overcomes the defect of traditional BP neural network being prone to getting trapped in local optima and has the advantages of fast convergence speed and high parameter optimization accuracy. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for predicting the unsteady acoustic quality of an electric drive based on intelligent sensors, as proposed in this invention. Figure 2 This is a schematic diagram of the information entropy budget allocation model for an electric drive unsteady-state acoustic quality prediction method based on intelligent sensors proposed in this invention. Figure 3 This is a schematic diagram illustrating the cooperative evolution of the information budget modulation factor in an electric drive unsteady-state acoustic quality prediction method based on intelligent sensors proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors includes the following steps: Noise signals from the electric drive system during operation are collected, and feature vector sequences consisting of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level are extracted to establish a dataset corresponding to the sound quality evaluation results. A feedforward neural network structure is constructed, determining the input nodes, hidden layer nodes, and output nodes to form a network parameter vector containing connection weights and biases. An information entropy budget allocation model is established, mapping the sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level to information budget variables, forming an information budget vector. Information budget constraints are set, and an objective function containing budget constraint terms and prediction error terms is constructed based on these constraints. The information budget variables are embedded into a representation structure composed of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. A particle swarm search space is constructed, and the information budget vector and network parameter vector are uniformly encoded into particle state vectors, initializing particle position, velocity, individual optimal position, and swarm optimal position. An information budget modulation factor is introduced during particle iteration to weight the particle velocity updates, forming a differentiated search path driven by the information budget, realizing the co-evolution of the information budget vector and the network parameter vector. The optimal position of the particle is updated based on the fitness results, and the optimal information budget vector and the optimal network parameter vector are output to construct the prediction model. The feature vector sequence is input into the prediction model, and the non-steady-state sound quality prediction results are output.

[0019] In this embodiment, establishing the dataset corresponding to the sound quality evaluation results specifically includes: Several sample segments are extracted from the noise signal collected under unsteady conditions. Subjective listening evaluation of the sample segments is conducted using multiple sub-adaptive pairwise comparison methods. Several evaluators score the sound quality of the sample segments to obtain a subjective score value for each sample segment. The time-varying sequences of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level are extracted for each sample segment to form a feature vector sequence. The feature vector sequence of each sample segment is stored in correspondence with the subjective score value of the sample segment to form a dataset.

[0020] In this embodiment, constructing a feedforward neural network structure, determining input nodes, hidden layer nodes, and output nodes, and forming a network parameter vector containing connection weights and biases includes the following steps: Set the number of input nodes to equal the dimension of the feature vector sequence, and the number of output nodes to equal the dimension of the subjective rating value; determine the number of hidden layer nodes, which is equal to the square root of the sum of the number of input nodes and the number of output nodes plus an adjustment parameter, where the adjustment parameter is an integer from 1 to 10. For each value of the adjustment parameter, calculate the corresponding number of hidden layer nodes, construct a feedforward neural network with the specified number of hidden layer nodes, train it with the feature vector sequence as input and the subjective rating value as output, calculate the mean square error of each network after training, and select the number of hidden layer nodes corresponding to the minimum mean square error. The dimension of the network parameter vector is calculated based on the number of input nodes, hidden layer nodes, and output nodes. The network parameter vector consists of four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. The four parts of parameters are randomly initialized and then concatenated in the order of the first to the fourth part to form the network parameter vector.

[0021] In this embodiment, establishing the information entropy budget allocation model specifically includes: Information budget variables are defined for sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. These five information budget variables constitute an information budget vector. Information budget constraints are set, whereby each information budget variable takes values ​​in a closed interval from 0 to 1, and the sum of the five information budget variables equals 1. The information budget variable is embedded into a representation structure consisting of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. Specifically, the sound pressure level sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted sound pressure level sequence; the loudness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted loudness sequence; the sharpness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted sharpness sequence; the speech intelligibility sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted speech intelligibility sequence; and the preferred speech interference level sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted preferred speech interference level sequence. The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted preferred speech interference level sequence are concatenated to form a weighted feature vector sequence. Construct an objective function, which consists of the sum of a budget constraint term and a prediction error term. The budget constraint term is the negative of the entropy of the information budget vector, and the prediction error term is the mean squared error generated when the feedforward neural network is trained with a weighted feature vector sequence as input and subjective rating as output. In this embodiment, information budget variables are defined for sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level, specifically as follows: An adjustable weighting coefficient is assigned to each acoustic feature. Specifically, information budget variables are defined for sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level, and these five variables together constitute the information budget vector. Each information budget variable takes a value in a closed interval from 0 to 1, and the sum of the five variables equals 1. The larger the value of the budget variable, the more important the feature is to the subjective sound quality in the current noise sample, and the higher the weight will be given to it by the model. The smaller the value of the budget variable, the weaker the influence of the feature.

[0022] In this embodiment, constructing the objective function includes the following steps: Obtain the information budget vector, which includes a first information budget variable corresponding to the sound pressure level, a second information budget variable corresponding to the loudness, a third information budget variable corresponding to the sharpness, a fourth information budget variable corresponding to the speech intelligibility, and a fifth information budget variable corresponding to the preferred speech interference level. Calculate the budget constraint term: multiply the first, second, third, fourth, and fifth information budget variables by their respective base-2 logarithms, sum the three products, and take the opposite. Obtain the sound pressure level sequence, loudness sequence, sharpness sequence, speech intelligibility sequence, and preferred speech interference level sequence for each sample segment. Multiply the value of each sampling point in each sequence by the corresponding information budget variable to generate a weighted sound pressure level sequence, a weighted loudness sequence, a weighted sharpness sequence, a weighted speech intelligibility sequence, and a weighted preferred speech interference level sequence. Concatenate the five weighted sequences according to the sampling point index to form a weighted feature vector sequence. Construct a feedforward neural network with the number of input layer nodes equal to the dimension of the weighted feature vector sequence and the number of output layer nodes equal to the dimension of the subjective rating values. Train the network with the weighted feature vector sequence as input and the subjective rating values ​​of the corresponding sample segments as the target. After training, calculate the mean squared error between the network output value and the target output value as the prediction error term. Add the budget constraint term and the prediction error term to obtain the objective function value.

[0023] In this embodiment, encoding the information budget vector and the network parameter vector into a unified particle state vector specifically includes: Obtain the dimension value of the information budget vector, which is 5. Obtain the dimension value of the network parameter vector, calculated from the number of input nodes, hidden layer nodes, and output nodes. Set the total dimension value of the particle state vector to be equal to the sum of the dimensions of the information budget vector and the network parameter vector. Use the five information budget variables of the information budget vector as the 1st, 2nd, 3rd, 4th, and 5th components of the particle state vector, respectively. Use all parameters of the network parameter vector as the 6th to the last components of the particle state vector, respectively. Set the value range of the 1st to 5th components to be a closed interval of 0 to 1. Set the value range of the 6th to the last components to be a closed interval of -3 to +3. Normalize the 1st to 5th components; after normalization, the sum of the three components equals 1.

[0024] In this embodiment, introducing an information budget modulation factor during particle iteration includes the following steps: Define the dimension correspondence rules for the particle state vector: the first three dimensions are the sound pressure level information budget variable, loudness information budget variable, and sharpness information budget variable, respectively. From the fourth dimension onwards, they are the input layer to hidden layer connection weight, hidden layer threshold, hidden layer to output layer connection weight, and output layer threshold, respectively. Each input weight records the input feature index to which it is connected. Obtain the inertia weight, individual learning factor, and social learning factor for the current iteration number. For each particle, obtain the current position vector, current velocity vector, individual optimal position vector, and group optimal position vector. Extract the first 5 dimensions from the current position vector as the current information budget vector. Construct the modulation factor vector: the first 5 dimensions are equal to the 5 components of the current information budget vector. For each dimension starting from the 6th dimension, if this dimension corresponds to the input layer to hidden layer connection weight, then take the corresponding component from the current information budget vector according to the input feature index associated with this weight as the modulation factor for this dimension. If this dimension corresponds to the hidden layer threshold, the hidden layer to output layer connection weight, and the output layer threshold, then the modulation factor for this dimension is set to 1. For each dimension, calculate the original velocity update: multiply the inertia weight by the component of the current velocity vector in this dimension, add the individual learning factor multiplied by the random number multiplied by the difference between the individual optimal position vector in this dimension and the current position vector in this dimension, add the social learning factor multiplied by the random number multiplied by the difference between the group optimal position vector in this dimension and the current position vector in this dimension, to obtain the original velocity update. Multiply the original velocity update by the value of this dimension in the modulation factor vector to obtain the modulated velocity update. Add the value of this dimension of the current position vector to the modulated velocity update to obtain the value of this dimension of the updated position vector. After completing the calculation of all dimensions, the updated position vector of the particle is obtained.

[0025] In this embodiment, the inertia weight is specifically: The degree to which a particle maintains its original motion tendency, from a physical perspective, represents the proportion of its velocity inherited from the previous moment. When the inertial weight is large, the particle tends to maintain its original flight direction, exploring the entire space and avoiding premature aggregation in a local area; When the inertia weight is small, the particle velocity is less affected by history and is more easily driven by individual and group optima, allowing for fine-grained searches in local areas.

[0026] In this embodiment, the individual learning factor is specifically: It reflects the strength of a particle's learning towards its own best historical position and embodies the particle's self-awareness. After being multiplied by a random number, this factor determines the contribution of the difference between the particle's current position and its individual best position to the velocity update. The larger the individual learning factor, the more the particle tends to return to the best position it has found in the past, thus enhancing the utilization of individual experience.

[0027] In this implementation, the social learning factor is specifically: It reflects the strength of a particle's learning toward the best position in the group and embodies the particle's social cooperation ability. After multiplying by a random number, this factor determines the contribution of the difference between the particle's current position and the global best position to the speed update. The larger the social learning factor, the more the particle tends to move toward the current best particle, accelerating the population convergence.

[0028] In this implementation, the output of the optimal information budget vector and the optimal network parameter vector, and the construction of the prediction model, specifically include: The first five components are extracted from the group's optimal position vector as the optimal information budget vector; the remaining components are extracted from the group's optimal position vector as the optimal network parameter vector; based on the number of input nodes, the number of hidden layer nodes, and the number of output nodes, the optimal network parameter vector is divided into four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. Create a feedforward neural network, setting the number of input nodes, hidden layer nodes, and output nodes to be the same as during optimization. Arrange the first part into a matrix with rows of hidden layer nodes and columns of input nodes, and assign it to the connection weights from the input layer to the hidden layer. Arrange the second part into a vector with rows of hidden layer nodes and 1 column, and assign it to the hidden layer bias. Arrange the third part into a matrix with rows of output nodes and columns of hidden layer nodes, and assign it to the connection weights from the hidden layer to the output layer. Arrange the fourth part into a vector with rows of output nodes and 1 column, and assign it to the output layer bias. Set the input feature weighting rule: For the input sound pressure level sequence, multiply the sound pressure level value of each sampling point in the sequence by the first component in the optimal information budget vector to generate a weighted sound pressure level sequence. For the input loudness sequence, the loudness value of each sampling point in the sequence is multiplied by the second component in the optimal information budget vector to generate a weighted loudness sequence; For the input sharpness sequence, the sharpness value of each sampling point in the sequence is multiplied by the third component of the optimal information budget vector to generate a weighted sharpness sequence. For the input speech proficiency sequence, the speech proficiency value of each sample point in the sequence is multiplied by the fourth component of the optimal information budget vector to generate a weighted speech proficiency sequence. For the input priority language interference level sequence, the priority language interference level value of each sampling point in the sequence is multiplied by the fifth component in the optimal information budget vector to generate a weighted priority language interference level sequence. The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted priority speech interference level sequence are concatenated according to the sampling point index order to form a weighted feature vector sequence. The input feature weighting rules are combined with the assigned feedforward neural network to form a prediction model. The input of the prediction model is the original sound pressure level sequence, the original loudness sequence, and the original sharpness sequence. The output of the prediction model is the calculation result of the weighted feature vector sequence by the assigned feedforward neural network.

[0029] Example 1: During the research and testing phase of a new electric vehicle, engineers needed to evaluate the vehicle's in-vehicle noise quality under different driving conditions. The traditional approach involved inviting multiple professional evaluators to conduct subjective listening tests on actual roads. However, this method was not only time-consuming and labor-intensive but also difficult to cover all conditions, and the scores were easily influenced by the evaluators' subjective opinions. To improve development efficiency and achieve objective quantification, the R&D team decided to adopt the method of this invention, which automatically predicts subjective sound quality scores based on noise signals collected by onboard intelligent sensors.

[0030] The test vehicle was equipped with three high-precision microphones, positioned at the driver's right ear, the rear left passenger's ear, and the trunk area, respectively, to simultaneously collect noise data from the vehicle's electric drive system during acceleration, deceleration, and constant-speed operation. A total of 240 valid noise samples were collected during the entire test, each lasting 5 seconds and covering various non-steady-state conditions from low to high speeds and from steady to rapid acceleration. For each sample, the sound pressure level, loudness, and sharpness were extracted as a sequence over time, with a sampling frequency of 51200Hz, resulting in 500 data points for each feature sequence. Simultaneously, a panel of 15 experienced evaluators independently scored the sound quality of each sample using a 10-point scale. The average score was used as the true subjective score for that sample, ultimately forming a complete dataset containing 240 feature vector sequences and their corresponding scores.

[0031] To verify the effectiveness of this invention, the dataset was randomly divided into 180 training samples and 60 test samples. Three methods were used to build the prediction model: the first was a traditional BP neural network; the second was the standard PSO-BP; and the third was the method of this invention. All methods were run on the same computer using the same training and test sets.

[0032] Table 1: Comparison of prediction results for some samples in the test set

[0033] As shown in Table 1, the predicted values ​​of the method of this invention are very close to the actual scores on the six test samples. For sample T01, the actual score is 7.2, the predicted score of this invention is 7.1, a difference of only 0.1, while the traditional BP prediction is 6.8, a difference of 0.4. For sample T02, the actual score is 5.8, the predicted score of this invention is 5.9, and the traditional BP prediction is 6.3, with an error of 0.1 and a traditional BP error of 0.5. On sample T05, which has a higher score, the actual score is 9.1, the predicted score of this invention is 8.9, with an error of 0.2, while the traditional BP prediction is only 8.3, with an error as high as 0.8. On sample T04, which has a lower score, the actual score is 4.3, the predicted score of this invention is 4.4, with an error of 0.1, while the traditional BP prediction is 4.9, with an error of 0.6. Samples T03 and T06 show a similar pattern, with the error of this invention controlled between 0.1 and 0.2, while the error of the traditional BP is in the range of 0.5 to 0.6.

[0034] Statistical analysis of the entire test set of 60 samples showed that the method of this invention had a mean absolute error of 0.12, a mean square error of 0.03, and a mean absolute percentage error of 2.1%. The traditional BP method had a mean absolute error of 0.53, a mean square error of 0.38, and a mean absolute percentage error of 9.2%. The standard PSO-BP method had a mean absolute error of 0.33, a mean square error of 0.15, and a mean absolute percentage error of 5.7%. This invention significantly outperforms the comparative methods in all indicators, demonstrating its ability to effectively improve the accuracy of sound quality prediction under unsteady conditions in practical applications.

[0035] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors, characterized in that, include: Noise signals from the electric drive system during operation are collected, and feature vector sequences consisting of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level are extracted to establish a dataset corresponding to the sound quality evaluation results. A feedforward neural network structure is constructed, determining input nodes, hidden layer nodes, and output nodes to form a network parameter vector containing connection weights and biases. An information entropy budget allocation model is established, mapping the sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level to information budget variables, forming an information budget vector. Information budget constraints are set, and an objective function containing budget constraint terms and prediction error terms is constructed based on these constraints. The information budget variables are embedded into a representation structure composed of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. A particle swarm search space is constructed, and the information budget vector and network parameter vector are uniformly encoded into particle state vectors, initializing particle position, velocity, individual optimal position, and swarm optimal position. An information budget modulation factor is introduced during particle iteration to weight the particle velocity update, forming a differentiated search path driven by the information budget, realizing the co-evolution of the information budget vector and the network parameter vector. The optimal position of the particle is updated based on the fitness results, and the optimal information budget vector and the optimal network parameter vector are output to construct the prediction model. The feature vector sequence is input into the prediction model and the non-steady-state sound quality prediction results are output.

2. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 1, characterized in that, The specific data set established to correspond with the sound quality evaluation results includes: Several sample segments are extracted from the noise signal collected under unsteady operating conditions. Subjective listening evaluation is conducted on the sample segments using multiple sub-adaptive pairwise comparison methods. Several evaluators score the sound quality of the sample segments to obtain a subjective score value for each sample segment. The time-varying sequences of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level are extracted for each sample segment to form a feature vector sequence. The feature vector sequence of each sample segment is stored in correspondence with the subjective score value of the sample segment to form a dataset.

3. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 2, characterized in that, The construction of the feedforward neural network structure, determining the input nodes, hidden layer nodes, and output nodes, and forming a network parameter vector containing connection weights and biases includes the following steps: Set the number of input nodes to be equal to the dimension of the feature vector sequence, and the number of output nodes to be equal to the dimension of the subjective rating values; The number of hidden layer nodes is determined by taking the square root of the sum of the number of input nodes and the number of output nodes, rounded down, and then adding an adjustment parameter. The adjustment parameter is an integer from 1 to 10. For each value of the adjustment parameter, the corresponding number of hidden layer nodes is calculated. A feedforward neural network with the specified number of hidden layer nodes is constructed. The network is trained using a feature vector sequence as input and subjective ratings as output. The mean squared error (MSE) of each network after training is calculated, and the number of hidden layer nodes corresponding to the minimum MSE is selected. The dimension of the network parameter vector is calculated based on the number of input nodes, the number of hidden layer nodes, and the number of output nodes. The network parameter vector consists of four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. After randomly initializing the four parts of parameters, they are concatenated in the order of the first to the fourth part to form the network parameter vector.

4. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 3, characterized in that, The establishment of the information entropy budget allocation model specifically includes: Information budget variables are defined for sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. These five information budget variables form an information budget vector. Information budget constraints are set, with each information budget variable taking values ​​in a closed interval from 0 to 1, and the sum of the three information budget variables equaling 1. The information budget variable is embedded into a representation structure consisting of sound pressure level, loudness, sharpness, speech intelligibility, and preferred speech interference level. Specifically, the sound pressure level sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted sound pressure level sequence; the loudness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted loudness sequence; the sharpness sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted sharpness sequence; the speech intelligibility sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted speech intelligibility sequence; and the preferred speech interference level sequence of each sample segment is multiplied by the corresponding information budget variable to obtain a weighted preferred speech interference level sequence. The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted preferred speech interference level sequence are concatenated to form a weighted feature vector sequence. Construct an objective function consisting of a budget constraint term and a prediction error term. The budget constraint term is the negative of the entropy of the information budget vector, and the prediction error term is the mean squared error generated when the feedforward neural network is trained with a weighted feature vector sequence as input and subjective ratings as output.

5. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 4, characterized in that, The construction of the objective function includes the following steps: Obtain the information budget vector, which includes a first information budget variable corresponding to the sound pressure level, a second information budget variable corresponding to the loudness, a third information budget variable corresponding to the sharpness, a fourth information budget variable corresponding to the speech intelligibility, and a fifth information budget variable corresponding to the priority speech interference level. Calculate the budget constraint term: multiply the first, second, third, fourth, and fifth information budget variables by their respective base-2 logarithms, sum the five products, and take the opposite. Obtain the sound pressure level sequence, loudness sequence, sharpness sequence, speech intelligibility sequence, and priority speech interference level sequence for each sample segment. Multiply the value of each sampling point in each sequence by the corresponding information budget variable to generate a weighted sound pressure level sequence, a weighted loudness sequence, a weighted sharpness sequence, a weighted speech intelligibility sequence, and a weighted priority speech interference level sequence. Concatenate the five weighted sequences according to the sampling point index to form a weighted feature vector sequence. Construct a feedforward neural network with the number of input layer nodes equal to the dimension of the weighted feature vector sequence and the number of output layer nodes equal to the dimension of the subjective rating values. Train the network with the weighted feature vector sequence as input and the subjective rating values ​​of the corresponding sample segments as the target. After training, calculate the mean squared error between the network output value and the target output value as the prediction error term. Add the budget constraint term and the prediction error term to obtain the objective function value.

6. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 5, characterized in that, The specific steps of uniformly encoding the information budget vector and the network parameter vector into a particle state vector include: Obtain the dimension value of the information budget vector, which is 5. Obtain the dimension value of the network parameter vector, which is calculated from the number of input nodes, hidden layer nodes, and output nodes. Set the total dimension value of the particle state vector to be equal to the sum of the dimensions of the information budget vector and the network parameter vector. Use the five information budget variables of the information budget vector as the 1st, 2nd, 3rd, 4th, and 5th components of the particle state vector, respectively. Use all parameters of the network parameter vector as the 6th to the last component of the particle state vector, respectively. Set the value range of the 1st to 5th components to be a closed interval of 0 to 1. Set the value range of the 6th to the last component to be a closed interval of -3 to +3. Normalize the 1st to 5th components; after normalization, the sum of the three components equals 1.

7. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 6, characterized in that, The introduction of the information budget modulation factor during particle iteration includes the following steps: Define the particle state vector dimension correspondence rules: the first five dimensions are sound pressure level information budget variables, loudness information budget variables, sharpness information budget variables, speech intelligibility information budget variables, and priority speech interference level budget variables, respectively. From the sixth dimension onwards, they are the input layer to hidden layer connection weights, hidden layer thresholds, hidden layer to output layer connection weights, and output layer thresholds, respectively. Each input weight records the input feature index connected to it. Obtain the inertia weight, individual learning factor, and social learning factor for the current iteration number. For each particle, obtain the current position vector, current velocity vector, individual optimal position vector, and group optimal position vector. Extract the first five dimensions from the current position vector as the current information budget vector. Construct the modulation factor vector: the first five dimensions are equal to the five components of the current information budget vector. For each dimension starting from the sixth dimension, if this dimension corresponds to the input layer to hidden layer connection weight, then take the corresponding component from the current information budget vector as the modulation factor for this dimension based on the input feature index associated with this weight. If this dimension corresponds to the hidden layer threshold, the hidden layer to output layer connection weight, and the output layer threshold, then the modulation factor for this dimension is set to 1. For each dimension, calculate the original velocity update: multiply the inertia weight by the component of the current velocity vector in this dimension, add the individual learning factor multiplied by the random number multiplied by the difference between the individual optimal position vector in this dimension and the current position vector in this dimension, add the social learning factor multiplied by the random number multiplied by the difference between the group optimal position vector in this dimension and the current position vector in this dimension, to obtain the original velocity update. Multiply the original velocity update by the value of this dimension in the modulation factor vector to obtain the modulated velocity update. Add the value of this dimension of the current position vector to the modulated velocity update to obtain the value of this dimension of the updated position vector. After completing the calculation of all dimensions, the updated position vector of the particle is obtained.

8. The method for predicting the unsteady acoustic quality of an electric drive system based on intelligent sensors according to claim 7, characterized in that, The construction of the prediction model by outputting the optimal information budget vector and the optimal network parameter vector specifically includes: The first three components are extracted from the group's optimal position vector as the optimal information budget vector; the remaining components are extracted from the group's optimal position vector as the optimal network parameter vector; based on the number of input nodes, the number of hidden layer nodes, and the number of output nodes, the optimal network parameter vector is divided into four parts: the first part is the connection weights from the input layer to the hidden layer, with the number of parameters equal to the number of input nodes multiplied by the number of hidden layer nodes; the second part is the bias of the hidden layer, with the number of parameters equal to the number of hidden layer nodes; the third part is the connection weights from the hidden layer to the output layer, with the number of parameters equal to the number of hidden layer nodes multiplied by the number of output nodes; and the fourth part is the bias of the output layer, with the number of parameters equal to the number of output nodes. Create a feedforward neural network, setting the number of input nodes, hidden layer nodes, and output nodes to be the same as during optimization. Arrange the first part into a matrix with rows of hidden layer nodes and columns of input nodes, and assign it to the connection weights from the input layer to the hidden layer. Arrange the second part into a vector with rows of hidden layer nodes and 1 column, and assign it to the hidden layer bias. Arrange the third part into a matrix with rows of output nodes and columns of hidden layer nodes, and assign it to the connection weights from the hidden layer to the output layer. Arrange the fourth part into a vector with rows of output nodes and 1 column, and assign it to the output layer bias. Set the input feature weighting rule: For the input sound pressure level sequence, multiply the sound pressure level value of each sampling point in the sequence by the first component in the optimal information budget vector to generate a weighted sound pressure level sequence. For the input loudness sequence, the loudness value of each sampling point in the sequence is multiplied by the second component in the optimal information budget vector to generate a weighted loudness sequence; For the input sharpness sequence, the sharpness value of each sampling point in the sequence is multiplied by the third component in the optimal information budget vector to generate a weighted sharpness sequence; For the input speech proficiency sequence, the speech proficiency value of each sampling point in the sequence is multiplied by the fourth component in the optimal information budget vector to generate a weighted speech proficiency sequence; For the input priority language interference level sequence, the priority language interference level value of each sampling point in the sequence is multiplied by the fifth component in the optimal information budget vector to generate a weighted language proficiency sequence; The weighted sound pressure level sequence, weighted loudness sequence, weighted sharpness sequence, weighted speech intelligibility sequence, and weighted priority speech interference level sequence are concatenated according to the sampling point index order to form a weighted feature vector sequence. The input feature weighting rules are combined with the assigned feedforward neural network to form a prediction model. The input of the prediction model is the original sound pressure level sequence, original loudness sequence, original sharpness sequence, original speech intelligibility sequence, and original priority speech interference level sequence. The output of the prediction model is the calculation result of the weighted feature vector sequence by the assigned feedforward neural network.

Images