AI-driven electric motorcycle intelligent sound synthesis system

Abstract

The present application relates to the technical field of audio terminal, specifically to an intelligent sound wave synthesis system of electric motorcycle based on AI driving, which comprises a vehicle state feature extraction module, a sound wave reasoning and constraint module, a sound field rendering separation module and a power gain compensation module.The vehicle state feature extraction module acquires real-time dynamic parameters to generate multi-dimensional vehicle state features.The sound wave reasoning and constraint module inputs the multi-dimensional features into a generation model for reasoning mapping and superimposes harmonic correction to output an original sound wave tensor.The sound field rendering separation module reconstructs the sound field or limits the frequency band of the tensor to generate immersive ambient sound and safety warning sound.The power gain compensation module generates an adjustment coefficient in combination with the throttle opening and the power ripple amplitude, performs anti-bottom noise filtering and outputs high-fidelity sound waves.The present application provides a simulated sound source through AI architecture and dynamic power linkage, effectively avoiding sudden acceleration bottom noise and sound jamming.

Description

Technical Field

[0001] This invention relates to the field of audio terminal technology, specifically to an AI-driven intelligent sound synthesis system for electric motorcycles. Background Technology

[0002] In electric vehicle engineering (such as electric motorcycles and other scenarios emphasizing driving experience), sound simulation and intelligent audio synthesis based on real-time power parameters are important technological applications for realizing human-machine interaction and state perception in vehicles. Existing technologies generally employ a lookup table mapping method based on preset audio samples. This involves using a microcontroller to collect the vehicle's rotational speed or vehicle speed parameters and mapping them to trigger specific audio segments stored in memory, establishing a sound playback and output model for specific operating conditions. This is currently the mainstream technology for simulating the sound of electric vehicles.

[0003] However, existing sound synthesis schemes have significant shortcomings in terms of audio dynamic tracking capabilities and output stability under complex operating conditions. Their system models for sound generation suffer from unidirectional computation and weak responsiveness: existing sound simulations often rely solely on low-end control chips and single pre-made audio samples for rigid triggering playback. Their audio generation boundaries cannot respond in real-time to subtle changes in the driver's throttle opening, making it difficult to accurately synthesize continuous sound spectra or undefined transitional acoustic states when vehicle power undergoes subtle changes. This results in noticeable stuttering and lag in the output of the sound audio. In summary, their audio rendering and output results are fragmented, only capable of single-channel amplification of basic audio. They fail to establish a complete audio synthesis chain from underlying power supply ripple disturbance resistance to upper-level three-dimensional sound field separation and reconstruction. This makes them highly susceptible to system resets or harsh audio background noise caused by drastic fluctuations in battery voltage during rapid acceleration of electric motorcycles. Furthermore, the single sound output mechanism cannot simultaneously meet the immersive listening experience of the driver inside the vehicle and the frequency requirements of external warning sounds.

[0004] To address this, an AI-driven intelligent sound synthesis system for electric motorcycles is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-driven intelligent sound synthesis system for electric motorcycles to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-driven intelligent sound synthesis system for electric motorcycles, comprising:

[0007] Vehicle state feature extraction module: acquires real-time dynamic parameter signals of the vehicle under the current operating conditions, the real-time dynamic parameter signals include at least rotational speed, throttle opening, vehicle speed and acceleration data; extracts basic dynamic feature matrix and jerk feature from the real-time dynamic parameter signals respectively, and concatenates them to generate a multi-dimensional vehicle state feature vector; Sound wave inference and constraint module: Construct a lightweight sound wave generation model, use multi-dimensional vehicle state feature vectors as input to the model for inference mapping to obtain the initial audio tensor; correct the initial audio tensor through real engine harmonic constraint rules to output the original sound wave tensor; Sound field rendering separation module: Obtain the current application mode command of the vehicle, reconstruct the spatial sound field of the original sound wave tensor according to the application mode command, and generate an immersive ambient sound signal for the driver inside the vehicle and a safety warning sound signal for the outside of the vehicle, which are used as the target sound wave signal. Power gain compensation module: calculates the difference between the throttle opening and the preset sudden change threshold, generates a dynamic gain adjustment coefficient by detecting the amplitude of the system power ripple fluctuation, and uses the dynamic gain adjustment coefficient to perform low-level volume scaling and noise reduction filtering on the target sound wave signal to output a high-fidelity synthesized sound wave.

[0008] Preferably, the specific generation process of the multidimensional vehicle state feature vector includes: acquiring the real-time dynamic parameter signal through the vehicle controller local area network bus; normalizing the speed, throttle opening, vehicle speed and acceleration data within the real-time dynamic parameter signal to eliminate dimensional differences and outputting a basic dynamic feature matrix; performing time difference differentiation on the acceleration data to calculate the acceleration change rate at the current time node and outputting the jerk feature; and concatenating and fusing the basic dynamic feature matrix with the jerk feature to generate a multidimensional vehicle state feature vector.

[0009] Preferably, the specific generation process of the original sound wave tensor includes: inputting the multi-dimensional vehicle state feature vector into the generator network inside the lightweight sound wave generation model, and outputting an initial audio tensor with a target sampling rate; extracting the harmonic spectrum structure of the real internal combustion engine at the specified speed in physical space; comparing the spectral distribution of the initial audio tensor with the harmonic spectrum structure of the real engine; using the energy peak nodes of the real engine harmonic spectrum structure to perform frequency domain energy constraints on the initial audio tensor, completing or weakening the energy spectrum distribution weights of specific frequency bands, and outputting the original sound wave tensor after physical constraint correction.

[0010] Preferably, the specific construction process of the lightweight sound generation model includes: constructing a generative adversarial network model containing the generator network and the discriminator network; acquiring multiple sets of pre-collected locomotive sound data as positive real benchmark samples; inputting random noise and set vehicle dynamic parameters into the generator network to output fake sound audio; inputting the fake sound audio and the positive real benchmark samples into the discriminator network for discriminative feature extraction and binary classification identification, and outputting true and false probability values; calculating the classification loss function of the true and false probability values, adjusting the node connection weight values ​​of the generator network and the discriminator network using the backpropagation mechanism until the classification loss function meets the predetermined convergence condition, compressing and converting the model format, and outputting a lightweight sound generation model with locked parameters.

[0011] Preferably, the specific generation process of the immersive ambient sound signal and the safety warning sound signal includes: extracting the current application mode command of the vehicle; when the mode command is determined to be in-vehicle mode, calling the head-related transfer function, inputting the original sound wave tensor into the head-related transfer function for spatial orientation simulation calculation, and superimposing a low-frequency enhancement algorithm to output an immersive ambient sound signal with three-dimensional spatial directivity; when the mode command is determined to be in-vehicle mode, establishing a specific frequency bandpass filter according to the vehicle external warning sound regulations and standards, filtering out low-frequency vibration components from the original sound wave tensor, and superimposing a directional beamforming algorithm to limit the sound wave radiation angle to output a safety warning sound signal that meets the set decibel and frequency conditions.

[0012] Preferably, the specific generation process of the dynamic gain adjustment coefficient includes: extracting the throttle opening; determining whether the throttle opening exceeds the preset mutation threshold; in response to the state exceeding the preset mutation threshold, generating a power intervention signal to activate the boost mode of the power management unit; in boost mode, acquiring the power ripple fluctuation amplitude of the system in real time; obtaining the reference supply voltage of the digital-to-analog converter, calculating the deviation ratio of the power ripple fluctuation amplitude relative to the reference supply voltage of the digital-to-analog converter using a preset mapping algorithm, establishing a bottom-level anti-distortion dynamic mapping function, and calculating and outputting the dynamic gain adjustment coefficient based on the bottom-level anti-distortion dynamic mapping function.

[0013] Preferably, in response to a state where the preset mutation threshold is not exceeded, the dynamic gain adjustment coefficient remains at the default static gain constant.

[0014] Preferably, the specific generation process of the high-fidelity synthesized sound wave includes: extracting the amplitude sequence of the target sound wave signal; obtaining the dynamic gain adjustment coefficient; converting the dynamic gain adjustment coefficient into the corresponding digital audio amplification factor; applying the digital audio amplification factor frame by frame to the amplitude sequence, and performing smoothing and limiting processing on the abrupt energy range of abnormal amplitude fluctuations; using a digital noise reduction algorithm to perform acoustic noise filtering on the limited audio tensor, eliminating noise data points caused by the sudden drop in power supply voltage during rapid acceleration, and outputting a high-fidelity synthesized sound wave with a pure distribution.

[0015] Preferably, it also includes a hardware encryption / decryption control step, the specific control process of which is as follows: calling an independent security encryption chip to obtain a lightweight sound wave generation model file in an encrypted state; using a preset hardware decryption engine and key system to verify and decrypt the lightweight sound wave generation model file; in response to a successful decryption instruction, loading the decrypted model parameters into the core system memory; in response to a decryption failure or detection of an abnormal signal from an illegal firmware, automatically cutting off the output link of the target sound wave signal and forcibly calling the reserved security prompt audio for playback.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By acquiring real-time dynamic parameter signals of the vehicle under current operating conditions, the basic dynamic feature matrix and jerk characteristics are extracted and fused together to generate a multi-dimensional vehicle state feature vector. Based on this, the multi-dimensional vehicle state feature vector is input into a lightweight sound generation model for inference mapping. In response to the output initial audio tensor, it is input into constraint rules established using the harmonic structure of a real engine for spectrum comparison and energy correction, outputting the original sound tensor. This connects the AI ​​parameter mapping with real physical acoustic constraints, directly establishing a complete response chain from the perception of subtle dynamic changes at the bottom layer to the high-fidelity audio synthesis at the upper layer.

[0017] 2. In the dynamic power supply gain compensation processing, a power intervention signal is generated by calculating the difference between the throttle opening and the preset sudden change threshold. This signal is then combined with the real-time power supply ripple fluctuation amplitude extracted from the digital-to-analog converter's reference supply voltage to perform deviation ratio calculations, outputting a dynamic gain adjustment coefficient. Addressing the severe battery voltage fluctuations unique to electric motorcycles during rapid acceleration, a low-level anti-distortion dynamic mapping function and amplitude limiting calculation are introduced. This smooths and eliminates the non-stationary trend term in the abrupt energy range of abnormal fluctuations while preserving the original sound wave amplitude dynamics characteristic of the vehicle's power burst to the maximum extent. This avoids the slight audio detail attenuation and stuttering caused by conventional rigid truncation or system reset, providing a noise-free and stable synthetic data foundation for subsequent pure audio output.

[0018] 3. In the sound field rendering separation module, application mode commands are used to locate the output scenes inside and outside the vehicle to extract specific separation strategies. Spatial orientation simulation or bandpass filtering is then performed on these scenes to obtain sound source signals with different directional characteristics. These signals are then combined with low-frequency enhancement algorithms or beamforming algorithms to dynamically generate immersive ambient sounds and safety warning sounds, which are then adapted for output through specific hardware channels. By focusing on specific decibel and frequency ranges of vehicle external warning sounds according to regulations, the audio output can adaptively adjust to the acoustic perception needs arising from the switching between the physical spaces inside and outside the vehicle. This dynamic scene separation mechanism based on multi-channel heterogeneous rendering ensures sensitivity to creating a subjective immersive listening experience for the driver inside the vehicle, while limiting disordered noise pollution caused by low-frequency vibrations when switching modes outside the vehicle, providing a high-fidelity separated signal source for subsequent independent power amplifier output.

[0019] 4. By deeply coupling the multi-dimensional vehicle state feature extraction step, the inference correction step based on GAN network and physical harmonic constraints, the sound field rendering separation step based on pattern instructions, and the dynamic gain compensation step combined with power supply ripple feedback, a strong coupling mechanism is achieved. In the overall sound synthesis chain, the high-frequency parameter acquisition and AI inference mapping at the front end provide the audio tensor input for the sound field reconstruction at the back end, eliminating mechanical response hysteresis. Meanwhile, the dynamic gain adjustment coefficient generated based on power fluctuation calculation serves as the underlying scaling threshold, filtering out a large amount of harsh background noise and invalid noise triggering sources caused by sudden power outages for the sound-generating terminal. This strong coupling mechanism effectively blocks the cascading auditory tearing error caused by the defects of a single pre-made audio scheme, enabling the system to achieve synergistic optimization in simulation sensitivity and sound purity under high-voltage rapid acceleration when facing complex and ever-changing extreme driving conditions of electric motorcycles. Attached Figure Description

[0020] Figure 1 This is a system principle block diagram of the AI-driven intelligent sound synthesis system for electric motorcycles proposed in an embodiment of this invention application; Figure 2 This is a flowchart of the algorithm proposed in an embodiment of the present invention. Detailed Implementation

[0021] 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.

[0022] Please see Figures 1-2 The AI-driven intelligent sound synthesis system for electric motorcycles provided by this invention has the following specific modules: Vehicle state feature extraction module: acquires real-time dynamic parameter signals of the vehicle under the current operating conditions, the real-time dynamic parameter signals include at least rotational speed, throttle opening, vehicle speed and acceleration data; extracts basic dynamic feature matrix and jerk feature from the real-time dynamic parameter signals respectively, and concatenates them to generate a multi-dimensional vehicle state feature vector; Sound wave inference and constraint module: Construct a lightweight sound wave generation model, use multi-dimensional vehicle state feature vectors as input to the model for inference mapping to obtain the initial audio tensor; correct the initial audio tensor through real engine harmonic constraint rules to output the original sound wave tensor; Sound field rendering separation module: Obtain the current application mode command of the vehicle, reconstruct the spatial sound field of the original sound wave tensor according to the application mode command, and generate an immersive ambient sound signal for the driver inside the vehicle and a safety warning sound signal for the outside of the vehicle, which are used as the target sound wave signal. Power gain compensation module: calculates the difference between the throttle opening and the preset sudden change threshold, generates a dynamic gain adjustment coefficient by detecting the amplitude of the system power ripple fluctuation, and uses the dynamic gain adjustment coefficient to perform low-level volume scaling and noise reduction filtering on the target sound wave signal to output a high-fidelity synthesized sound wave.

[0023] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0024] Example 1 This application discloses an AI-driven intelligent sound synthesis system for electric motorcycles, see embodiments below. Figure 1 and Figure 2 The specific modules proposed in this invention include: a vehicle state feature extraction module: acquiring real-time dynamic parameter signals of the vehicle under the current operating conditions, wherein the real-time dynamic parameter signals include at least speed, throttle opening, vehicle speed, and acceleration data; extracting the basic dynamic feature matrix and jerk features from the real-time dynamic parameter signals, and concatenating them to generate a multi-dimensional vehicle state feature vector; and a sound wave inference and constraint module: constructing a lightweight sound wave generation model, using the multi-dimensional vehicle state feature vector as input to the model for inference mapping to obtain an initial audio tensor; and applying the initial audio tensor to the real engine harmonic constraint rules. The system then corrects the output and outputs the original sound wave tensor; the sound field rendering separation module obtains the vehicle's current application mode command, reconstructs the spatial sound field of the original sound wave tensor according to the application mode command, and generates an immersive ambient sound signal for the driver inside the vehicle and a safety warning sound signal for the outside of the vehicle, which are used as the target sound wave signal; the power gain compensation module calculates the difference between the throttle opening and the preset sudden change threshold, generates a dynamic gain adjustment coefficient by detecting the amplitude of the system power ripple fluctuation; and uses the dynamic gain adjustment coefficient to perform low-level volume scaling and noise reduction filtering on the target sound wave signal to output a high-fidelity synthesized sound wave.

[0025] Further, real-time dynamic parameter signals of the vehicle under the current operating conditions are acquired, wherein the real-time dynamic parameter signals include at least engine speed, throttle opening, vehicle speed, and acceleration data; basic dynamic feature matrix and jerk feature are extracted from the real-time dynamic parameter signals, and concatenated to generate a multi-dimensional vehicle state feature vector; corresponding to the vehicle state feature extraction module; the specific implementation process includes: The real-time dynamic parameter signals are acquired through the vehicle controller local area network bus; the speed, throttle opening, vehicle speed and acceleration data within the real-time dynamic parameter signals are normalized to eliminate dimensional differences and output a basic dynamic feature matrix; the acceleration data are differentiated by time difference to calculate the rate of change of acceleration at the current time point and output the jerk feature; the basic dynamic feature matrix and the jerk feature are concatenated and fused to generate a multi-dimensional vehicle state feature vector.

[0026] Specifically, the process of generating the multidimensional vehicle state feature vector is as follows: The input is a continuous multi-channel timing data stream based on the vehicle controller area network bus. Taking a high-performance electric motorcycle as an example, its CAN bus physical baud rate is typically set to 500kbps, and the broadcast period of key power frames is 10 milliseconds (i.e., a sampling rate of 100 Hz). The system extracts a 0.5-second time slice during operation, meaning the input contains a timing array with 50 discrete time nodes. This data packet parses at least four core physical dimensions: motor speed, absolute throttle opening (percentage), rear wheel linear speed (km / h), and longitudinal acceleration. On the time axis, the system applies a sliding time window with a length of 20 sampling points, and sets the overlap step of adjacent time windows to 10 sampling points to ensure the continuity of state sampling and prevent the underreporting of power change features. The length of the sliding time window is dynamically calculated based on the system's response requirements and data link layer throughput, ensuring that the window span can at least completely cover one minor throttle twist. The overlap step is typically set to half the length of the sliding time window to balance the timeliness assumptions of feature extraction with the computational resource consumption of the main control chip's NPU. Real-time dynamic parameter signals are low-level telemetry signals collected when the vehicle experiences typical physical conditions such as starting, rapid acceleration, cruising, or coasting. The specific operating mode corresponding to these signals is labeled through message commands from the vehicle control unit. The overlap step is selected by dynamically locking the value based on the least common multiple of the current main control chip's NPU's single inference time and the bus message cycle; it is necessary to ensure that the completion time of AI inference in the previous time window is strictly earlier than the completion time of data aggregation in the next time window.

[0027] Within each defined time window, the system performs multi-scale normalization on the captured RPM, throttle opening, vehicle speed, and acceleration data. Because the dimensions of data transmitted from different sensors vary significantly (for example, RPM can range from 0 to 12000 RPM, while throttle opening is only 0 to 100%), the system uses a combination of Min-Max linear scaling and Z-Score normalization to eliminate these dimensional differences. For throttle opening and engine speed with clear physical extremes, a linear mapping based on the extreme values ​​is performed, strictly limiting them to a numerical range of 0 to 1. For vehicle speed and acceleration, which fluctuate greatly due to road conditions, the arithmetic mean of all vehicle speed or acceleration sampling data within the sliding time window is calculated as the local mean, and the standard deviation of the data within the window is calculated. The vehicle speed or acceleration value at the current time node is subtracted from the local mean, and the difference is divided by the standard deviation to obtain the dimensionless vehicle speed and acceleration feature values. Irregular sporadic sensor pulse information is discarded, and finally a two-dimensional basic dynamic feature matrix is ​​output. The horizontal axis of this matrix represents the time frame dimension, and the vertical axis represents the normalized dynamic channel dimension.

[0028] In this embodiment, a Kalman filter smoothing mechanism based on dynamic covariance is introduced for acceleration data. Acceleration sensor data of the vehicle when it is completely stationary and the motor is enabled is extracted to construct a static drift reference vector. Combined with a trust factor corresponding to the dynamic vehicle speed, the acceleration sequence of the current sliding time window is optimally estimated frame by frame, supplemented by adaptive dead-zone limiting, ultimately outputting a clean longitudinal dynamic acceleration sequence. Specifically, after acquiring acceleration data, the system extracts 50 consecutive time frames in the vehicle's powered-on, stationary state as a reference segment. This frame number is set according to the law of large numbers, sufficient to construct a Gaussian statistical model that approximates the true static distribution of gravity. The mean variance of this reference segment in the stationary state is calculated to generate a one-dimensional steady-state drift reference vector. To suppress road vibration interference under driving conditions, the system updates the current frame's measured acceleration with this reference vector using optimal estimation. A trust weighting factor based on dynamic vehicle speed adjustment is introduced: if the current vehicle speed is greater than 80 km / h, the trust factor is set to 0.3 (reducing the trust in real-time sensor measurements that are easily affected by road bumps); it is set to 0.6 between 20 and 80 km / h; and it is set to 0.9 below 20 km / h. The measured value is multiplied by the corresponding trust factor before being updated from the original dynamic estimation state. Subsequently, an adaptive dead-zone limiting operation is performed, traversing the obtained smoothed sequence and forcibly setting all jump data points with an absolute value change rate less than a preset minimum threshold to the values ​​of the previous frame, preventing subtle high-frequency jitter from interfering with subsequent derivative calculations. The preset minimum threshold is 0.05g, a rejection constant dynamically calculated by the system using the equivalent spring stiffness coefficient of the vehicle suspension system and the vehicle's curb weight to construct a chassis resonance baseline model. This mechanism eliminates the masking effect of mechanical chassis resonance and environmental road noise, improving the signal-to-noise ratio for weak dynamic state changes.

[0029] The peak torque response of modern electric motorcycle motors is typically within tens of milliseconds, resulting in nonlinear step acceleration characteristics. The system performs forward time difference differentiation on the input pure longitudinal acceleration sequence at adjacent time points. The calculation formula is: current frame acceleration minus previous frame acceleration divided by the single-frame time step (set to 20 milliseconds in this embodiment under the vehicle CAN bus sampling frequency). This calculates the rate of change of acceleration at the current time point, thereby extracting and retaining the jerk characteristics containing the transient push-back feeling abrupt change. For the extracted jerk characteristics (whose physical dimension is meters per second squared), the system performs a nonlinear exponential mapping operation using an exponential function with the natural constant as the base. To ensure the mathematical and physical homogeneity of the transcendental function operation, the system first extracts a pre-calibrated physical limit jerk constant for this model as a benchmark reference value. The jerk difference value at the current time point is divided by this benchmark reference value to calculate the dimensionless jerk ratio. Subsequently, the system's specific mathematical mapping logic is as follows: using the natural constant as the base, the product of the dimensionless jerk ratio and the preset sensitivity adjustment coefficient is used as the exponent, and the result is the mapped output feature. After this baseline division and mapping operation, a dimensionless jerk mapping feature vector with physical units removed is output. Then, following a unified data flow format, the system flattens the pre-processed basic dynamic feature matrix along the time axis. This matrix contains the basic physical quantities of the vehicle's steady-state operation. The dimensionless basic dynamic feature matrix is ​​then concatenated and fused with the dimensionless higher-order jerk feature at the memory feature level. This concatenation aligns the dimensions of the data matrix containing steady-state operating condition information and transient change information, thereby outputting a final multi-dimensional vehicle state feature vector containing the dynamic states of the entire scenario.

[0030] By acquiring real-time dynamic parameters through the vehicle controller local area network and performing normalization and difference operations according to different feature dimensions, steady-state and transient feature information containing different vehicle physical states can be extracted, providing a high-quality inference data source for subsequent sound wave tensor generation and effectively reducing the risk of output audio stuttering caused by original parameter jitter.

[0031] Furthermore, a lightweight sound wave generation model is constructed, and a multi-dimensional vehicle state feature vector is input into the model for inference mapping to obtain an initial audio tensor. The initial audio tensor is then corrected using real engine harmonic constraint rules to output the original sound wave tensor. This corresponds to the sound wave inference and constraint module. The specific implementation process includes: A generative adversarial network (GAN) model comprising a generator network and a discriminator network is constructed. Multiple sets of pre-collected locomotive sound data are acquired as positive real-world benchmark samples. Random noise and set vehicle dynamic parameters are input into the generator network to output a fake sound audio. The fake sound audio and the positive real-world benchmark samples are input into the discriminator network for discriminative feature extraction and binary classification to identify the authenticity, outputting a true / false probability value. The classification loss function of the true / false probability value is calculated, and the node connection weights of the generator network and discriminator network are adjusted using a backpropagation mechanism until the classification loss function meets a predetermined convergence condition. The model format is compressed and converted, and a lightweight sound generation model with locked parameters is output.

[0032] Specifically, the construction process of the lightweight sound wave generation model is as follows: The Generative Adversarial Network (GAN) employs a deep acoustic mapping architecture based on one-dimensional convolution. The input layer of its generator network directly receives a multi-dimensional vehicle state feature vector generated by the vehicle state feature extraction module, along with a random noise vector following a Gaussian distribution. Taking the sound simulation of a high-performance electric motorcycle as an example, the input layer of the generator network receives a cascaded feature matrix containing features of specific operating modes, engine speed, throttle opening, and agility. The generator network contains several one-dimensional transposed convolutional layers with 512, 256, 128, and 64 feature channels respectively, performing layer-by-layer audio tensor upsampling expansion. It should be noted that the selection principle for the number of hidden layer channels should ensure a step-like decrease layer by layer, while the temporal dimension feature length expands by doubling layer by layer. This selection is based on constructing an inverse expansion funnel of the audio waveform sequence, forcing the network to accurately map low-dimensional abstract dynamic physical features into high-frequency acoustic waveform details during the upsampling reconstruction process. In each computational layer, the input feature tensor is first convolved with the weights of the transposed convolution kernel of that layer in one dimension. After adding the bias vector, the entire result is fed into a leaky modified linear unit activation function for nonlinear mapping. The processing logic of this activation function is as follows: if the input value is greater than 0, the original value is directly output; if it is less than or equal to 0, a preset minimum positive slope is applied to scale the output to prevent the gradient vanishing problem in the long sequence audio feature mapping process. After layer-by-layer upsampling mapping, the high-dimensional data finally flows into the output layer using the hyperbolic tangent activation function, forcibly normalizing its numerical range to between -1 and 1, and outputting a fake sound wave audio tensor with the target sampling rate. Subsequently, the fake sound audio and the pre-collected positive real locomotive sound benchmark samples are alternately input into the discriminator network. The discriminator network performs layer-by-layer dimensionality reduction convolution extraction and non-linear activation operations, which are the reverse of the generation process. That is, the discriminator feature extraction operation. Finally, the features are compressed to the 0 to 1 range through the Sigmoid activation function in the output layer. The output represents the true or false probability value of the input audio as a real physical sound probability distribution.

[0033] In the adversarial training process combining unsupervised and supervised methods, the true / false probability values ​​output by the discriminator network for different data sources are extracted alternately. For each positive true benchmark sample in the current batch, the true label value (value is 1) is extracted, and multiplied by the natural logarithm of the corresponding output probability value to obtain the first product. For each fake sound wave audio sample in the current batch, the difference between the value 1 and the fake label value (value is 0) is calculated, and the difference between the value 1 and the corresponding output probability value is calculated. The natural logarithm of the latter is extracted, and the two are multiplied to obtain the second product. The first product and the second product corresponding to all samples in the same batch are summed, the negative of the sum is taken, and then averaged by the total number of samples in the current batch. Thus, the binary cross-entropy classification loss function, which represents the zero-sum game state between the generator and the discriminator, is calculated. The backpropagation mechanism works as follows: The system uses the classification loss function as the global objective evaluation standard, calculates the gradient distribution inversely using the chain rule of multivariable calculus, and employs an adaptive moment estimation optimization algorithm with a pre-set initial learning rate parameter of 0.0002. It alternately fine-tunes and updates the connection weights of the convolutional kernel nodes within the generator and discriminator networks along the steepest inverse direction of gradient descent. When the classification loss functions of the discriminator and generator exhibit alternating oscillations over multiple training epochs and the mean deviation is below a specific threshold, the system determines that the classification loss function meets the predetermined convergence condition (i.e., reaches Nash equilibrium). In this embodiment, the specific threshold is set to 0.01. The selection of this specific threshold is based on multiple real-vehicle bench verification data. When the deviation of the loss function fluctuation is below 0.01, the sound audio generated by the system using model inference no longer produces perceptible digital background noise and abrupt changes in subsequent digital-to-analog conversion and power amplification stages. This value represents the optimal empirical critical point that balances model training convergence efficiency and high-fidelity audio output. At this point, the system immediately terminates the backpropagation process, strips and discards the discriminator network branch, and fully solidifies the current optimal generator node connection weight values ​​and bias parameters. Subsequently, the system forcibly calls the model post-quantization algorithm, scaling and truncating the original 32-bit floating-point weight matrix into an 8-bit integer data format, completing the compression and conversion of the model format, thus obtaining a lightweight sound wave generation model with minimal storage space and locked parameters.

[0034] Through an adversarial cyclical game of generator upsampling mapping and discriminator authenticity identification, the network is forced to learn and memorize the potential distribution patterns of physical locomotive voiceprint features under various extreme conditions. Cross-entropy numerical values ​​and alternating backpropagation mechanisms are used to solidify weights. Post-model quantization techniques are employed to eliminate computational redundancy, establishing rigorous edge-side hardware inference specifications.

[0035] The multidimensional vehicle state feature vector is input into the generator network inside the lightweight sound generation model, and an initial audio tensor with a target sampling rate is output. The harmonic spectrum structure of the real internal combustion engine at the specified speed is extracted in physical space. The spectral distribution of the initial audio tensor is compared with the harmonic spectrum structure of the real engine. The energy peak nodes of the real engine harmonic spectrum structure are used to perform frequency domain energy constraints on the initial audio tensor, and the energy spectrum distribution weights of specific frequency bands are supplemented or weakened, and the original sound wave tensor after physical constraint correction is output.

[0036] Specifically, the generation process of the original sound wave tensor is as follows: The system first inputs a multi-dimensional vehicle state feature vector containing steady-state and transient information into the generator network within the lightweight sound generation model. After nonlinear upsampling mapping via transposed convolution within the model, a one-dimensional initial audio tensor with a target sampling rate is output at the output layer. Since purely data-driven generative adversarial networks may produce digital artifacts deviating from mechanoacoustic laws when inferring extreme transient features, the system introduces a constraint mechanism based on the physical laws of real internal combustion engines. The system parses the real-time motor speed (equivalent to the internal combustion engine RPM) from the currently input multi-dimensional vehicle state feature vector and uses this as an index to retrieve and extract the harmonic spectrum structure corresponding to a real high-performance internal combustion engine at that specific speed from a pre-programmed physical acoustic database. This physical spectrum structure clearly defines the fundamental frequency and the center frequencies and standard amplitude ratios of each order of harmonics (such as the 2nd, 3rd, and 4th harmonics). Subsequently, the system performs a short-time Fourier transform on the initial audio tensor in the time domain, converting it to the frequency domain to obtain the spectral distribution matrix. Since there may be slight misalignment between the spectral resolution of the AI ​​output and the preset physical harmonic template at discrete frequency points, the system uses the frequency resolution of the physical template as a unified benchmark and employs a one-dimensional cubic spline interpolation algorithm to perform frequency band energy resampling and padding operations on the spectral matrix output by the AI. This results in a standard AI spectral distribution matrix that is strictly aligned in the frequency dimension, thereby establishing a data benchmark for subsequent accurate comparison.

[0037] After frequency alignment is completed, the system's constraint algorithm module independently calculates the energy difference of each frequency node. The system performs a band-by-band comparison operation between the standard AI spectrum distribution matrix and the real engine harmonic spectrum structure: at the energy peak nodes (i.e., the fundamental frequency and each harmonic band) marked in the real engine harmonic spectrum structure, the physical standard energy weight value corresponding to the band is extracted; if the energy weight value of the current AI spectrum is lower than the physical standard energy weight value, the system uses the physical standard energy weight value as the divisor and the current AI spectrum energy weight value as the divisor to calculate the quotient of the two, and uses the quotient directly as the gain compensation coefficient of the band to make up for the missing energy of the specific frequency band; conversely, in the background band defined by the physical template as the non-harmonic interval (often the invalid digital noise band generated by AI), if the energy weight value of the current AI spectrum is higher than the preset background noise energy standard value, the system uses the preset background noise energy standard value as the divisor and the current AI spectrum energy weight value as the divisor to calculate the quotient of the two and generate an attenuation coefficient that is necessarily less than 1 to weaken its energy spectrum distribution weight. The preset background noise floor energy standard value is typically set as a linear amplitude corresponding to the -80dB to -90dB range relative to the full-scale amplitude. The background noise floor energy standard value is selected as follows: the inherent static signal-to-noise ratio parameter in the current audio DAC chip datasheet is extracted, converted into a frequency domain theoretical quantized noise amplitude, and a safety margin of 3 to 5 dB is added to this amplitude to obtain the final noise floor standard value.

[0038] The system combines all the aforementioned gain and attenuation coefficients in order of frequency occurrence to construct a frequency-domain energy constraint mask matrix that strictly matches the current spectral dimension. The system then performs element-wise multiplication of the standard AI spectral distribution matrix with this mask matrix, forcing the completion of the physically constrained frequency-domain energy correction. Finally, an inverse short-time Fourier transform is performed on the corrected spectral matrix, folding it back to restore it to a time-domain data stream, thereby outputting the physically constrained corrected original acoustic wave tensor.

[0039] By using the harmonic structure of the real engine as a constraint mask to compare and correct the frequency domain energy of the initial audio, high-frequency digital glitches and digital noise that deviate from the physical mechanical sound production law can be eliminated while preserving the extremely high dynamic response flexibility of the AI ​​generation model in the time dimension.

[0040] Furthermore, the system obtains the vehicle's current application mode command, reconstructs the spatial sound field of the original sound wave tensor based on the command, and generates an immersive ambient sound signal for the driver inside the vehicle and a safety warning sound signal for the outside, which are then used as the target sound wave signal; this corresponds to the sound field rendering separation module; the specific implementation process includes: The system extracts the current application mode command of the vehicle. When the mode command is determined to be in-vehicle mode, it calls the head-related transfer function, inputs the original sound wave tensor into the head-related transfer function for spatial orientation simulation calculation, and superimposes a low-frequency enhancement algorithm to output an immersive ambient sound signal with three-dimensional spatial directivity. When the mode command is determined to be in-vehicle mode, it establishes a specific frequency bandpass filter according to the vehicle external warning sound regulations and standards, filters out low-frequency vibration components from the original sound wave tensor, and superimposes a directional beamforming algorithm to limit the sound wave radiation angle, outputting a safety warning sound signal that meets the set decibel and frequency conditions.

[0041] Specifically, the generation process of immersive ambient sound signals and safety warning sound signals is as follows: The system monitors and extracts the current application mode command issued by the vehicle's control unit in real time to determine the target physical space of the audio output. When the mode command is determined to be in-vehicle mode (i.e., the target service object is the driver), the system immediately calls the head-related transfer function database pre-installed in the digital signal processor's memory. The system inputs the original sound wave tensor from the front end into the head-related transfer function algorithm module to perform spatial orientation simulation calculations. During this calculation, the system performs a fast convolution operation on the original one-dimensional audio tensor and the impulse response sequence representing a specific orientation (such as the spatial coordinates of the virtual V8 engine exhaust pipe located below and behind) in the head-related transfer function database. Utilizing the nonlinear acoustic filtering characteristics of binaural time difference and binaural sound level difference, the system reconstructs the originally non-directional audio tensor into a two-channel virtual surround sound data stream with precise three-dimensional spatial directionality. Subsequently, in order to compensate for the low-frequency sound defects of helmet headphones or near-ear miniature speakers in terms of physical size, the system superimposes a low-frequency enhancement algorithm (such as a psychoacoustic harmonic generator or low-frequency filter) onto the surround sound data, performs dynamic gain amplification operation on the frequency band below 150Hz, and finally outputs an immersive ambient sound signal with a deep mechanical roar and three-dimensional spatial directionality without causing hardware clipping distortion.

[0042] When the mode command is determined to be an external vehicle mode (i.e., the target service object is pedestrians or non-motorized vehicles around the vehicle), the system automatically switches to a rendering link that complies with the vehicle external warning sound regulations. First, based on the strict requirements of external warning sound regulations for specific frequency components (such as the requirement to emit sufficient sound within a specific frequency band), the system internally establishes a first- or multi-order specific frequency bandpass filter. The system feeds the original sound wave tensor into this bandpass filter, forcibly filtering out low-frequency vibration components below 200Hz and ultra-high-frequency redundant components that easily cause harsh hearing, preventing low-frequency sound waves from generating omnidirectional noise pollution and chassis resonance interference in street environments. After completing the frequency band limitation, the system extracts the physical topology parameters of the vehicle's external speaker array and superimposes a directional beamforming algorithm onto the filtered audio tensor. This algorithm precisely calculates and injects microsecond-level phase delay and amplitude weights into the multi-channel output, causing constructive interference of the sound waves radiated by the external speaker array in the warning sectors directly in front of and to the sides of the vehicle. This limits and focuses the radiation angle of the sound waves, ultimately outputting a safety warning sound signal that fully meets the set decibel lower limit and frequency distribution conditions, providing compliant acoustic boundaries for vehicle operation.

[0043] By acquiring mode instructions to perform directional rendering separation and calling the HRTF spatial algorithm and bandpass beamforming algorithm respectively, it is possible to achieve high-fidelity distribution of the same original sound wave tensor in different application scenarios without increasing the computing power burden of the front-end AI generation model.

[0044] Further, the difference between the throttle opening and the preset sudden change threshold is calculated, and a dynamic gain adjustment coefficient is generated by detecting the amplitude of the system power supply ripple fluctuation; the dynamic gain adjustment coefficient is used to perform low-level volume scaling and noise reduction filtering on the target sound wave signal to output a high-fidelity synthesized sound wave; corresponding to the power supply gain compensation module; the specific implementation process includes: The throttle opening is extracted; it is determined whether the throttle opening exceeds the preset mutation threshold; in response to exceeding the preset mutation threshold, a power intervention signal is generated to activate the boost mode of the power management unit; in boost mode, the power ripple fluctuation amplitude of the system is collected in real time; the reference supply voltage of the digital-to-analog converter is obtained, and the deviation ratio of the power ripple fluctuation amplitude relative to the reference supply voltage of the digital-to-analog converter is calculated using a preset mapping algorithm, a low-level anti-distortion dynamic mapping function is established, and a dynamic gain adjustment coefficient is calculated and output based on the low-level anti-distortion dynamic mapping function. In response to not exceeding the preset mutation threshold, the dynamic gain adjustment coefficient remains at the default static gain constant.

[0045] Specifically, the generation process of the dynamic gain adjustment coefficient is as follows: The input to this process is the throttle opening data output in real time by the vehicle state feature extraction module, and the power supply ripple parameters in the hardware circuit. Taking a high-performance electric motorcycle as an example, the standard reference supply voltage of its core audio digital-to-analog converter is usually set to 3.3V or 5V. The system first extracts the current real-time absolute throttle opening and determines whether it exceeds a preset mutation threshold. In this embodiment, the preset mutation threshold is strictly set to a throttle opening greater than 90%. This threshold is the power transition boundary point calculated based on the dynamic intersection of the rated output power curve of the motor controller and the current battery state of charge. In response to normal cruising or slow acceleration states that do not exceed the preset mutation threshold, the system determines that the current battery discharge power is stable, and the dynamic gain adjustment coefficient remains at the system's default static gain constant (e.g., 1.0), so that the output gain of the target sound signal is not affected by any interference from the underlying power supply ripple. Conversely, in response to the extreme speed burst state where the throttle opening exceeds 90%, the system immediately generates a microsecond-level power intervention signal, activating the boost mode of the power management unit through the I2C or SPI bus to cope with the sudden drop in battery voltage caused by high current discharge.

[0046] In this embodiment, in boost mode, a dynamic mapping compensation mechanism to prevent clipping is introduced to address the real-time power supply ripple of the system. The real-time power supply ripple fluctuation amplitude is extracted based on the reference supply voltage of the digital-to-analog converter (DAC), and dynamic gain adjustment is performed in conjunction with the deviation ratio to establish a low-level dynamic mapping function for preventing clipping, outputting a dynamic gain adjustment coefficient. Specifically, after activating boost mode, the system uses a high-precision DAC to collect the power supply ripple fluctuation amplitude of the audio motherboard system in real time at a preset ultra-high sampling rate (e.g., 100kHz). The sampling rate is selected to fully cover and capture the high-frequency ripple spikes caused by the PWM switching frequency of the motor driver. The system obtains the inherent reference supply voltage of the DAC, divides the measured power supply ripple fluctuation amplitude by this reference supply voltage, and calculates the deviation ratio. Based on this, the system constructs a low-level dynamic mapping function for preventing clipping: the deviation ratio is substituted into a preset inverse exponential decay function, assigning a lower audio gain weight to the range with larger voltage fluctuations. The specific mapping process is as follows: the mapping variable is calculated with the natural constant e as the base and a negative constant multiple of the deviation ratio as the exponent, and this is directly used as the real-time gain scaling factor. Simultaneously, a hard lower limit penalty mechanism to prevent audio dropouts is introduced: if the calculated deviation ratio of the current frame spikes to more than 15% of the reference voltage (i.e., facing an extremely dangerous power failure edge), 15% is a safety protection constant derived based on the physical critical value of clipping distortion of the operational amplifier inside the corresponding digital-to-analog converter chip. In this case, the dynamic gain adjustment coefficient is forcibly locked and attenuated to 0.5 (i.e., a 50% maximum attenuation limit). The system extracts the calculated real-time gain scaling factor and performs frame-by-frame numerical mapping output. The 15% multiplier is selected based on the physical critical point of digital clipping distortion in the hardware DAC chip, used to identify extreme power supply deterioration; while the 50% forced attenuation is to retain the basic mechanical roar representing the vehicle's power burst while absolutely suppressing audio clipping distortion, preventing the driver from losing power perception due to completely silencing the acceleration sound. The above process suppresses audio distortion caused by unstable power supply under extreme acceleration, improving the hardware reliability of the audio rendering system.

[0047] To eliminate the "volume flickering" phenomenon caused by frequent gain coefficient jumps due to transient voltage glitches, the system performs a local smoothing filter on the output gain sequence. A local sliding window spanning five consecutive time frames is set on the time axis. The window length can be flexibly set between three and seven time frames based on the actual PMIC boost response delay. The arithmetic mean of all gain adjustment coefficients within the window is calculated, and this arithmetic mean value is used to replace the original value at the center node of the window, generating a smoothed dynamic gain adjustment coefficient sequence. Finally, the system buffers and synchronizes this smoothed coefficient sequence, establishing a one-dimensional gain control data stream aligned with the timestamp of the target sound wave signal, in preparation for subsequent multiplication modulation operations.

[0048] By determining whether the throttle opening exceeds a preset sudden change threshold, extreme driving conditions triggering abnormal voltage fluctuations are captured. Dynamic deviation ratio calculations and limiting mapping are performed on the power supply ripple in boost mode, filtering out audio clipping distortion spikes caused by instantaneous power supply deterioration. Finally, a stable dynamic gain adjustment coefficient is generated through local smoothing filtering, condensing the underlying physical power supply fluctuations into an upper-level audio safety scaling ratio. This process establishes a dynamic anti-distortion boundary for subsequent high-fidelity audio output without sacrificing the driving dynamics.

[0049] Extract the amplitude sequence of the target sound wave signal; obtain the dynamic gain adjustment coefficient; convert the dynamic gain adjustment coefficient into the corresponding digital audio amplification factor; apply the digital audio amplification factor frame by frame to the amplitude sequence, and perform smoothing and limiting processing on the abrupt energy range of abnormal amplitude fluctuations; use a digital noise reduction algorithm to perform acoustic noise filtering on the limited audio tensor, eliminate noise data points caused by the sudden drop in power supply voltage during rapid acceleration, and output a high-fidelity synthesized sound wave with a pure distribution.

[0050] Specifically, the process of generating high-fidelity synthesized sound waves is as follows: The system extracts the target sound wave signal output from the sound field rendering and separation module, performs short-time energy calculation on it, and transforms the time-domain tensor into an amplitude sequence characterizing wave dynamics. Simultaneously, the system acquires the dynamic gain adjustment coefficient output in real-time by the power supply gain compensation module. To ensure the linear adjustment coefficient accurately matches the nonlinear characteristics of human ear perception of volume, the system performs a base-10 logarithmic transformation on both the independent variable time node and the dependent variable amplitude value in the amplitude sequence, mapping them to the decibel domain. Subsequently, the system forcibly truncates the top 10% of the data points with the highest energy peaks in the amplitude sequence, constructing a multiple mapping equation based on this dynamic gain adjustment coefficient. The percentage of this data point truncation is adaptively adjusted according to the dynamic range of the power amplifier chip's supply voltage. In low-voltage, high-load conditions (such as rapid acceleration), where the distortion-free thrust is concentrated in a narrower voltage range, a smaller truncation percentage (e.g., 5%) is selected. The system uses envelope fitting calculations to calculate the difference between the logarithmic amplitude and the ambient noise floor energy, dividing by the system safety margin, to find the optimal scaling parameter that minimizes the probability of audio clipping. Based on this, the absolute value of the parameter is extracted, and the system calculates the digital audio amplification factor: the dynamic gain adjustment coefficient is multiplied by the preset audio reference level, and the resulting product is then weighted and summed with the above-mentioned optimal scaling parameter; the inherent ambient noise constant is subtracted from the weighted sum, and the digital audio amplification factor strictly aligned with the current time frame is calculated and output.

[0051] After obtaining the digital audio amplification factor, the system uses this as a parameter to perform frame-by-frame smoothing and amplitude limiting processing on the amplitude sequence of the target sound wave signal. It then uses an audio dynamic range control algorithm to perform nonlinear compression and expansion on the signal stream containing abrupt energy ranges with abnormal amplitude fluctuations. After determining the logarithmic amplitude of the current frame, the system allocates its limiting gain weight coefficient according to the following three piecewise mathematical condition models: Condition 1: When the logarithmic amplitude of the current frame is less than or equal to the safety threshold, for this normal amplitude data, its weight coefficient is fixedly allocated to the digital audio amplification factor calculated above; Condition 2: When the logarithmic amplitude of the current frame is greater than the safety threshold and less than or equal to the soft knee limiting inflection point, for this first-order burst peak data, its weight coefficient is allocated to the digital audio amplification factor minus the over-limit difference penalty term, wherein the over-limit difference penalty term is the product of the over-limit ratio coefficient and the amplitude over-limit difference (i.e., the current amplitude minus the safety threshold); Condition 3: When the logarithmic amplitude of the current frame is greater than the soft knee limiting inflection point, for this second-order high-energy abnormal data, its weight coefficient is allocated to the digital audio amplification factor multiplied by the square of the over-limit difference penalty term, and then the product result is divided by a specific acoustic damping constant. The safety threshold and soft-knee limiting inflection point are determined as follows: The system reads the dynamic supply voltage of the current audio amplifier chip in real time and calculates the theoretical maximum undistorted voltage swing by combining it with the current speaker load impedance. 80% of the energy point of this maximum undistorted voltage swing is mapped to the safety threshold, and 95% of its energy point is mapped to the soft-knee limiting inflection point. This allows the system to adaptively tighten the limiting threshold as the battery voltage decreases.

[0052] Higher-order compression coefficients follow a nonlinear quadratic decreasing pattern. The system extracts the data and its envelope data for the current time node in the amplitude sequence, performs a term-by-term multiplication operation between each amplitude data point and the corresponding calculated compression weight coefficients of each order, and then superimposes all the product results back into the original time-domain audio stream. As the calculation window moves along the time axis, this frame-by-frame multiplication and clipping operation ultimately outputs a clipped audio sequence that eliminates abrupt energy distortion and abnormal peak interference. This sequence presents a stable output state without clipping distortion.

[0053] Furthermore, the system performs digital noise reduction filtering and finite truncation approximation operations on the limited audio tensor: the system uses a digital noise reduction algorithm to extract isolated noise data points generated by the sudden drop in power supply voltage near the current time point in the audio sequence and participates in the filtering calculation. The selection principle for truncating and discarding T spike data points is to reduce the absolute value of the disordered shot noise of the residual signal floor to below -96dB (i.e., the theoretical quantization noise floor lower limit of 16-bit audio). At this point, the residual weak noise floor has a negligible impact on the high-fidelity listening experience of the subsequent power amplifier. Taking the transient power-down spikes caused by rapid acceleration as an example, the system sets a time window of about 200 to 300 sampling points to meet the accuracy requirements of eliminating sudden drop noise without causing perceptible system delay, and the output presents a pure and distributed high-fidelity synthesized sound wave.

[0054] Furthermore, to ensure audio output stuttering is avoided under complex operating conditions, the system introduces asynchronous pipelines and timeout degradation fault tolerance mechanisms in the underlying transmission link for generating high-fidelity synthesized sound waves. The system uses a direct memory access double buffer to handle the high-fidelity synthesized audio tensor after amplitude limiting and noise reduction. On the timeline, the system specifies a worst-case execution time limit for a single sound wave synthesis task, which is strictly less than the overlap step of adjacent sliding time windows (e.g., 100 milliseconds).

[0055] The system's end-to-end total latency consists of CAN bus data acquisition buffer latency, feature extraction preprocessing latency, NPU network inference latency, and audio post-processing and output latency. During normal pipeline operation, the heterogeneous cores of the main control chip asynchronously and in parallel execute the above steps, ensuring that the synthesized audio tensor of the next frame is filled into the spare buffer before the current DMA buffer is exhausted. In response to extreme conditions where NPU inference is blocked or power fluctuations cause the total latency to exceed the worst-case execution time limit, the system triggers a timeout degradation fault tolerance mechanism: it immediately locks and stops waiting for the AI ​​inference data of the current timeout frame, forcibly calls and reuses the valid high-fidelity synthesized audio tensor generated in the previous time window, applies Hanning window functions to its beginning and end for fade-in / fade-out smooth interpolation processing, and directly sends it to the currently idle DMA buffer for playback. This fault tolerance mechanism, under extreme conditions of hardware computing power fluctuations, relies on microsecond-level seamless splicing to replace the digital interruption caused by forced waiting, thus effectively avoiding sound stuttering and tearing in the physical listening experience.

[0056] By solving the scaling parameter through a multiplier mapping equation, a digital audio amplification factor matching the current power supply state is derived. This factor, combined with a dynamic range control algorithm, is used to perform nonlinear smoothing filtering on the amplitude sequence, effectively reducing abnormal noise caused by sudden power outages. This processing method smooths out interference from underlying physical voltage fluctuations and reduces audio distortion and background noise triggering rates caused by sudden power drops.

[0057] The system calls upon an independent security encryption chip to obtain a lightweight sound wave generation model file that is in an encrypted state; it then uses a preset hardware decryption engine and key system to verify and decrypt the lightweight sound wave generation model file; in response to a successful decryption command, it loads the decrypted model parameters into the core system memory; in response to a decryption failure or the detection of abnormal signals from illegal firmware, it automatically cuts off the output link of the target sound wave signal and forcibly plays a reserved security prompt audio.

[0058] Specifically, the hardware encryption / decryption control steps and their specific control process are as follows: This process extracts the root key and unique device identifier that are already hardware-embedded within an independent security encryption chip (such as the ATECC608A integrated IC), and introduces a pre-defined elliptic curve cryptography signature authentication range. The system uses the verification mathematical formula of an asymmetric encryption algorithm for textual calculation: it extracts the percentage digest of the digital certificate from the encrypted lightweight sound wave generation model file in external memory, performs a nonlinear power function decryption operation on the digest with the public key coefficient as the exponent, and determines whether the verification parameters match: if the verification parameters are valid, the encrypted model file is divided into fixed-length data blocks, and then the AES-256 symmetric decryption algorithm is called to perform round function XOR and matrix column obfuscation inverse operations; if the verification parameters are invalid or extremely close to format corruption, the digital signature is directly determined to be invalid, and a very low probability pass threshold is output. Finally, the calculated plaintext result is added to the expected model tensor dimension value for verification. The absolute value of this verification output is the structural integrity threshold, representing the upper limit of dynamic validity. Subsequently, the system extracts the decrypted model format features and subtracts them from the integrity threshold. If the difference is strictly equal to 0, the system responds to the comparison result, determines that the authorization decryption is successful, and immediately outputs a high-level decryption success command to the system bus.

[0059] During the encryption authentication phase, the input is a high-dimensional tensor matrix of a legally authorized sound model and its inherent digital copyright multi-dimensional one-hot encoded tag. The digital copyright tag covers the typical high-frequency mechanical roar and low-frequency sci-fi electronic sound style of electric motorcycles, including, but not limited to, specific physical sound effects such as: V8 muscle engine sound, V12 supercar exhaust sound, interstellar sci-fi pulse sound, and retro twin-cylinder internal combustion engine sound.

[0060] The system constructs a hardware protection architecture that includes a secure boot trust layer and a memory isolation access layer. Specifically, the independent security encryption chip receives the downloaded and updated encryption sound wave model sequence and inputs it frame by frame into the isolation network layer (e.g., an encryption unit containing double obfuscation logic) to prevent physical side-channel attacks, in order to extract the temporal logic dependency relationship between different verification nodes of the encryption model file. Subsequently, the plaintext state vector output by the encryption layer in the last verification step is extracted and sent to the system bus layer (e.g., containing peripheral bus and high-speed AXI bus in sequence) with progressively shrinking memory addresses to perform nonlinear decryption mapping, and a multi-core isolated memory access function is configured in the output layer. The system calculates the XOR check function value between the plaintext features output by the hardware decryption engine and the expected identifier in system memory. The specific calculation process is as follows: For the true officially authorized dimension with a value of 1 in the one-hot encoded tag, the predicted hash value of the verification vector output by the decryption engine in that corresponding dimension is extracted; the hexadecimal format comparison difference of this hash value is calculated, and its bitwise AND operation result is extracted as the integrity assessment value for a single decryption sample; subsequently, the arithmetic mean of the assessment values ​​of all data blocks in the batch is calculated as the global security loss; and a register mechanism with the hardware verification clock set to nanosecond level is selected, using an interrupt return mechanism to update memory loading permissions. This continues until the fluctuation of the XOR check function in multiple consecutive loading rounds meets a predetermined convergence limit (e.g., strictly maintained at a minimum error of 0), at which point the system solidifies the decryption parameter channel and loads the decrypted model parameters into the core system memory.

[0061] The number of key slot nodes in the security encryption chip strictly corresponds to the total number N of authorized sound types contained in the digital rights label. N is pre-configured by the user during system initialization or OTA upgrade phases based on the number of genuine sound skins that the target vehicle may use. Each known authorized sound type corresponds to an N-dimensional verification encoding vector, where the memory dimension value corresponding to that sound type is active, and the other dimensions are locked.

[0062] In actual vehicle operation, the blocking link of the security isolation mechanism remains in a dormant state to conserve computing resources; it is activated only in response to decryption failure or the detection of abnormal signals from illegal firmware. The system directly intercepts illegal output commands containing illegal pirated sound effects, automatically cutting off the output link of the sound field rendering separation module (e.g., forcibly pulling down the I2S communication bus from the DSP to the audio amplifier) ​​via the underlying hardware bus cutting-off logic, outputting a zero-value vector representing the probability of cutting off various sound effects, where the sum of all dimensions of the audio output signal is strictly equal to zero. The system extracts the physically fixed backup tags within the security isolation zone, forcibly calls the reserved security prompt audio for playback, and formally generates a precise security compliance action.

[0063] The system employs a cascaded hardware protection mechanism, activating core memory to load model parameters and render AI sound waves only after capturing a confirmed and legitimate decryption success command. It inputs the current legitimate data into the model for mapping and outputs the frequency distribution of specific sound waves. This progressive processing logic connects the entire process from underlying physical anti-counterfeiting verification to upper-layer digital asset invocation, effectively overcoming the problem of audio models being easily pirated, copied, and maliciously tampered with.

[0064] This invention provides an AI-driven intelligent sound synthesis system for electric motorcycles. A vehicle state feature extraction module acquires basic power characteristics and dynamism characteristics to construct a multi-dimensional vehicle state feature vector, providing a complete data foundation for subsequent analysis. A sound inference and constraint module performs deep mapping on the feature vector, effectively outputting high-fidelity initial audio that incorporates real mechanical harmonics. A sound field rendering and separation module acquires vehicle mode commands and performs spatial orientation reconstruction and bandpass filtering to remove the influence of omnidirectional low-frequency noise. A power gain compensation module uses an anti-clipping dynamic mapping function to limit the voltage ripple at the underlying level, generating an adaptive digital audio amplification factor; further, hardware encryption and decryption control steps are combined to protect core digital assets. The collaborative work of these modules constructs a complete synthesis chain from capturing underlying power parameters to high-fidelity anti-clipping audio output, suppressing voltage drop noise and preventing firmware piracy while improving the immersiveness of human-computer interaction and the accuracy of sound responsiveness.

[0065] Example 2 This embodiment utilizes an AI-driven intelligent sound synthesis system for electric motorcycles, using a high-performance electric motorcycle with a peak power of 25 kW and a top speed of 120 km / h as the operating platform. During the pre-interaction phase, the user connects to the BLE 5.2 low-power Bluetooth module integrated on the PCBA motherboard via a smartphone app. By selecting a specific sound effect in the interface, the app wirelessly over-the-air (OTA) burns the corresponding INT8 quantized and compressed .tflite lightweight model file to the system core's eMMC 5.1 storage medium. The system's main control SoC has an NPU acceleration core computing power of 1 TOPS, a core audio digital-to-analog converter physical sampling rate set to 192 kHz, and a quantization bit depth of 32 bits.

[0066] In the specific implementation of the vehicle state feature extraction module, the system, while the vehicle is running, captures the dynamic timing data continuously transmitted via the CANFD bus (physical baud rate 500kbps, broadcast period 10ms). A sliding time window with a length of 20 sampling points (i.e., a time span of 0.2 seconds) is applied to the time axis, and the overlap step of adjacent time windows is set to 10 sampling points. The system parses and extracts the motor speed, absolute throttle opening, vehicle speed, and acceleration data within this window. Min-Max linear scaling based on extreme values ​​is performed on the throttle opening and speed; simultaneously, a Gaussian steady-state drift reference vector is constructed by capturing 50 consecutive time frames in the vehicle's enabled stationary state. Combined with a confidence factor set by the dynamic vehicle speed (e.g., 0.3 when vehicle speed > 80km / h), dynamic Kalman filtering optimal estimation and 0.05g dead-zone limiting operation are performed on the acceleration data. Subsequently, the system calculates the jerkiness by performing forward time difference differentiation on the smoothed acceleration sequence at a single frame step of 0.01 seconds. After nonlinear exponential mapping, it is flattened along the time axis at the memory level with the basic dynamic features and subjected to cascaded splicing and fusion operations, ultimately outputting a high-dimensional multidimensional vehicle state feature vector containing the dynamic states of the entire scenario. The trust factor is not a fixed constant, but rather an inverse proportional continuous mapping function with the current measured vehicle speed as the independent variable, weighted in conjunction with the tire dynamic friction coefficient. When the vehicle speed increases and the proportion of high-frequency road noise in the total vibration energy exceeds a critical point, the system automatically and smoothly adjusts the trust level of the acceleration sensor according to the preset inverse proportional function curve, thereby dynamically calculating the current trust factor.

[0067] In the sound inference and constraint module, after detecting the ignition signal, the system automatically loads the "V8 muscle engine sound" model flashed by the user over-the-air (OTA) from the eMMC. The generator network of this lightweight model contains a series of one-dimensional transposed convolutional layers, with the number of feature channel nodes decreasing sequentially in a stepwise manner to 512, 256, 128, and 64. After the multi-dimensional feature vector is input into the network, it performs a one-dimensional convolution operation with the weights of the transposed convolutional kernels of each layer and superimposes the bias vector. The entire result is then fed into a linear unit activation function with a preset positive slope of 0.2 and leakage correction, and finally flows into the hyperbolic tangent output layer to generate an initial audio tensor of 16kHz. Subsequently, the system extracts the harmonic spectrum structure of the real V8 internal combustion engine physical acoustic database at the corresponding speed, performs a short-time Fourier transform on the initial audio tensor, and uses a one-dimensional cubic spline interpolation algorithm to perform frequency band energy resampling and alignment. The system compares the peak energy nodes of the physical benchmark, constructs a frequency domain energy constraint mask matrix that is strictly consistent with the current spectral dimension, and performs a forced correction operation by multiplying it element-wise with the AI ​​spectral distribution matrix. Finally, it performs an inverse short-time Fourier transform to output the original acoustic wave tensor restored to the time domain.

[0068] In the power gain compensation module, the system extracts the real-time absolute throttle opening at a frequency of 100Hz. When the throttle opening spikes instantaneously in the current time frame and exceeds the preset abrupt change threshold of 90%, the system immediately activates the boost mode of the intelligent power management unit via the I2C bus. The system activates its internal high-precision ADC, acquiring the power ripple fluctuation amplitude of the audio motherboard system at an ultra-high sampling rate of 100kHz, and dividing it by the DAC's inherent 5V reference supply voltage to calculate the real-time deviation ratio. The system uses an inverse exponential decay mapping function with the natural constant e as the base to solve for the gain scaling factor; when the deviation ratio exceeds 15% of the reference voltage, a hard lower limit penalty mechanism is immediately triggered, forcibly locking and attenuating the dynamic gain adjustment coefficient to 0.5. To eliminate the "volume flickering" interference caused by frequent jumps in the gain coefficient due to transient voltage glitches, the system sets a local sliding window spanning 5 consecutive time frames on the time axis, calculates the arithmetic mean to complete local smoothing filtering, and generates a smooth dynamic gain adjustment coefficient stream with strictly aligned timestamps.

[0069] In the sound field rendering separation module, the system extracts the current application mode command issued by the vehicle control unit for split rendering. When the mode command is determined to be in-vehicle mode, a preset head-related transfer function database is called, and a fast convolution operation is performed on the original sound wave tensor and the impulse response sequence representing the spatial coordinates of the exhaust pipe below the virtual V8 engine. Low-frequency dynamic gain amplification below 150Hz is superimposed to output an immersive ambient sound with precise three-dimensional directivity. When the mode command is determined to be out-of-vehicle mode, the system constructs a first-order specific frequency bandpass filter to forcibly filter out omnidirectional low-frequency vibration components below 200Hz, and superimposes a directional beamforming algorithm to limit the radiation sector, outputting a safety warning sound that meets the decibel lower limit. Finally, the system applies the converted digital audio amplification factor frame by frame to the amplitude sequence of the target sound wave, triggering a soft knee limiting mechanism to perform nonlinear audio dynamic range control operation; a time window of 200 to 300 sampling points is selected to perform finite truncation approximation and digital noise reduction filtering, so that the residual disordered shot noise is attenuated to below -96dB. The processed high-fidelity V8 synthesized sound waves are then distributed to the speakers at the sound output terminal via a Class D dual-channel car amplifier, effectively avoiding noise tearing and audio distortion under rapid acceleration conditions.

[0070] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-driven intelligent sound synthesis system for electric motorcycles, characterized in that, include: Vehicle state feature extraction module: acquires real-time dynamic parameter signals of the vehicle under the current operating conditions, the real-time dynamic parameter signals include at least rotational speed, throttle opening, vehicle speed and acceleration data; extracts basic dynamic feature matrix and jerk feature from the real-time dynamic parameter signals respectively, and concatenates them to generate a multi-dimensional vehicle state feature vector; Sound wave inference and constraint module: Construct a lightweight sound wave generation model, and use multi-dimensional vehicle state feature vectors as input to the model for inference mapping to obtain the initial audio tensor; The initial audio tensor is corrected using the harmonic constraint rules of the real engine to output the original sound wave tensor; Sound field rendering separation module: Obtain the current application mode command of the vehicle, reconstruct the spatial sound field of the original sound wave tensor according to the application mode command, and generate an immersive ambient sound signal for the driver inside the vehicle and a safety warning sound signal for the outside of the vehicle, which are used as the target sound wave signal. Power gain compensation module: calculates the difference between the throttle opening and the preset sudden change threshold, generates a dynamic gain adjustment coefficient by detecting the amplitude of the system power ripple fluctuation, and uses the dynamic gain adjustment coefficient to perform low-level volume scaling and noise reduction filtering on the target sound wave signal to output a high-fidelity synthesized sound wave.

2. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, The specific generation process of the multidimensional vehicle state feature vector includes: acquiring the real-time dynamic parameter signal through the vehicle controller local area network bus; normalizing the speed, throttle opening, vehicle speed, and acceleration data within the real-time dynamic parameter signal to eliminate dimensional differences and outputting a basic dynamic feature matrix; performing time difference differentiation on the acceleration data to calculate the acceleration change rate at the current time node and outputting the jerk feature; and concatenating and fusing the basic dynamic feature matrix with the jerk feature to generate a multidimensional vehicle state feature vector.

3. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, The specific generation process of the original sound wave tensor includes: inputting the multi-dimensional vehicle state feature vector into the generator network inside the lightweight sound wave generation model, and outputting an initial audio tensor with a target sampling rate; extracting the harmonic spectrum structure of the real internal combustion engine at the specified speed in physical space; comparing the spectral distribution of the initial audio tensor with the harmonic spectrum structure of the real engine; using the energy peak nodes of the real engine harmonic spectrum structure to perform frequency domain energy constraints on the initial audio tensor, completing or weakening the energy spectrum distribution weights of specific frequency bands, and outputting the original sound wave tensor after physical constraint correction.

4. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 3, characterized in that, The specific construction process of the lightweight sound generation model includes: constructing a generative adversarial network model containing the generator network and the discriminator network; acquiring multiple sets of pre-collected locomotive sound data as positive real benchmark samples; inputting random noise and set vehicle dynamic parameters into the generator network to output fake sound audio; inputting the fake sound audio and positive real benchmark samples into the discriminator network for discriminative feature extraction and binary classification identification, and outputting true and false probability values; calculating the classification loss function of the true and false probability values, adjusting the node connection weight values ​​of the generator network and the discriminator network using the backpropagation mechanism until the classification loss function meets the predetermined convergence condition, compressing and converting the model format, and outputting a lightweight sound generation model with locked parameters.

5. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, The specific generation process of the immersive ambient sound signal and the safety warning sound signal includes: extracting the current application mode command of the vehicle; when the mode command is determined to be in-vehicle mode, calling the head-related transfer function, inputting the original sound wave tensor into the head-related transfer function for spatial orientation simulation calculation, and superimposing a low-frequency enhancement algorithm to output an immersive ambient sound signal with three-dimensional spatial directivity; when the mode command is determined to be in-vehicle mode, establishing a specific frequency bandpass filter according to the vehicle external warning sound regulations and standards, filtering out low-frequency vibration components from the original sound wave tensor, and superimposing a directional beamforming algorithm to limit the sound wave radiation angle to output a safety warning sound signal that meets the set decibel and frequency conditions.

6. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, The specific generation process of the dynamic gain adjustment coefficient includes: extracting the throttle opening; determining whether the throttle opening exceeds the preset mutation threshold; in response to the state exceeding the preset mutation threshold, generating a power intervention signal to activate the boost mode of the power management unit; in boost mode, acquiring the power ripple fluctuation amplitude of the system in real time; obtaining the reference supply voltage of the digital-to-analog converter, calculating the deviation ratio of the power ripple fluctuation amplitude relative to the reference supply voltage of the digital-to-analog converter using a preset mapping algorithm, establishing a bottom-level anti-distortion dynamic mapping function, and calculating and outputting the dynamic gain adjustment coefficient based on the bottom-level anti-distortion dynamic mapping function.

7. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 6, characterized in that, In response to a state where the preset mutation threshold is not exceeded, the dynamic gain adjustment coefficient remains at the default static gain constant.

8. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, The specific generation process of the high-fidelity synthesized sound wave includes: extracting the amplitude sequence of the target sound wave signal; obtaining the dynamic gain adjustment coefficient; converting the dynamic gain adjustment coefficient into the corresponding digital audio amplification factor; applying the digital audio amplification factor frame by frame to the amplitude sequence, and performing smoothing and limiting processing on the abrupt energy range of abnormal amplitude fluctuations; using a digital noise reduction algorithm to perform acoustic noise filtering on the limited audio tensor, eliminating noise data points caused by the sudden drop in power supply voltage during rapid acceleration, and outputting a high-fidelity synthesized sound wave with a pure distribution.

9. The AI-driven intelligent sound synthesis system for electric motorcycles according to claim 1, characterized in that, It also includes hardware encryption and decryption control steps. The specific control process includes: calling an independent security encryption chip to obtain a lightweight sound wave generation model file in an encrypted state; using a preset hardware decryption engine and key system to verify and decrypt the lightweight sound wave generation model file; in response to a successful decryption command, loading the decrypted model parameters into the core system memory; in response to a decryption failure or detection of abnormal signals from illegal firmware, automatically cutting off the output link of the target sound wave signal and forcibly playing a reserved security prompt audio.

Images