Noise control method, device, apparatus, vehicle, and program product

Abstract

The embodiment of the application provides a noise control method, device, equipment, vehicle and program product, actual noise data is determined according to current noise data and running state data of the vehicle; the actual noise data is input into a pre-constructed noise recognition model, noise features of the actual noise data are extracted, a noise type probability distribution vector and a noise quantization feature vector of the actual noise data are output according to the noise features; based on the noise type probability distribution vector and the noise quantization feature vector, the actual noise type, the actual confidence level and / or the actual noise value are determined, and a processing instruction for making the actual noise value less than a preset noise threshold matched with the current running condition is generated in combination with the current running condition, and the vehicle is controlled to execute the processing instruction, which not only guarantees the noise recognition accuracy, but also shortens the model reasoning time consumption, can issue accurate noise reduction processing instructions in real time, and improves the real-time performance and pertinence of vehicle noise control.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a noise control method, apparatus, equipment, vehicle, and program product. Background Technology

[0002] In related technologies, residual noise signals can be collected by an error microphone and sent to a digital signal processor. The digital signal processor then executes a preset fixed adaptive filtering algorithm to generate a cancellation signal, which drives the speaker to emit anti-phase sound waves to neutralize the original noise, thereby forming a closed-loop control loop for noise reduction control. Alternatively, the sound pressure level of the audio signal or the energy of a specified frequency band can be calculated, and the calculation result can be compared with a preset fixed threshold to classify the noise level or call a predetermined working mode, thus achieving simple noise determination.

[0003] However, in related technologies, the parameters of active noise reduction algorithms are fixed and cannot be adapted to complex and ever-changing noise conditions. Furthermore, the filter parameters rely on manual experience for tuning, resulting in slow convergence speed and easy system instability. On the other hand, noise detection relies solely on sound pressure level and frequency band energy threshold classification, which has a single perception dimension and lacks effective information. It cannot support refined noise reduction control and is prone to misjudgment due to the convergence of different noise energies. Therefore, improvements are urgently needed. Summary of the Invention

[0004] This application provides a noise control method, device, equipment, vehicle, and program product, aiming to improve the problems in related technologies, such as fixed parameters of active noise reduction algorithms, weak scene adaptability, and slow convergence and system instability caused by manual parameter tuning; noise detection relying solely on energy index thresholds, resulting in coarse information granularity, making it difficult to achieve precise noise reduction control, and the high probability of misjudgment due to noise type confusion.

[0005] The first aspect of this application provides a method for controlling vehicle noise, comprising the following steps: determining the actual noise data of the vehicle based on current noise data and operating status data; inputting the actual noise data into a pre-constructed noise recognition model, extracting noise features from the actual noise data, and outputting a noise type probability distribution vector and a noise quantization feature vector based on the noise features, wherein the noise recognition model is constructed by a pruned and quantized deep learning model, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold; determining the actual noise type, actual confidence level, and / or actual noise value based on the noise type probability distribution vector and the noise quantization feature vector, and generating a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating condition based on the current operating condition of the vehicle, the actual noise type, the actual confidence level, and / or the actual noise value, and controlling the vehicle to execute the processing instruction.

[0006] The above technical solution integrates current vehicle noise data with operational status data to obtain actual noise data, eliminating data interference and ensuring the authenticity and reliability of the noise data source. This actual noise data is then input into a noise recognition model constructed from a pruned and quantized lightweight deep learning model. The noise recognition model simultaneously outputs a noise type probability distribution vector and a noise quantization feature vector, compressing computational load while maintaining recognition accuracy to meet the real-time inference requirements of the vehicle. Based on this, the noise category and noise quantization index are accurately determined. Furthermore, the actual noise type, actual confidence level, and / or actual noise value are determined based on the noise type probability distribution vector and noise quantization feature vector. This generates a processing instruction that ensures the actual noise value is below a preset noise threshold, which is then executed by the vehicle. By specifically generating and issuing processing instructions for the actual noise data, differentiated and refined noise control is achieved.

[0007] Optionally, in one embodiment of this application, the step of generating a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating condition based on the vehicle's current operating condition, the actual noise type, the actual confidence level, and / or the actual noise value includes: acquiring target noise data of the target vehicle, and determining the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data, wherein the similarity between the target noise data and the actual noise data is greater than a preset similarity threshold; determining a target processing instruction matching the target noise data based on the target noise type, the target confidence level, the target noise value, and the target operating condition; determining a mapping relationship between the target noise type, the target confidence level, the target noise value, the target operating condition, and the target processing instruction; and matching the processing instruction from the mapping relationship based on the current operating condition, the actual noise type, the actual confidence level, and / or the actual noise value.

[0008] The above technical solution can determine the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data of the target vehicle. This allows for the generation of target processing instructions that match the target noise data. Furthermore, it establishes a mapping relationship between the target noise type, target confidence level, target noise value, target operating condition, and target processing instructions. Based on this mapping relationship, processing instructions can be matched and customized according to the target noise data. A one-to-one mapping relationship is constructed, and instructions can be directly matched by looking up a table during the operation phase. This ensures that the noise handling strategy is accurately adapted to the scenario, simplifies online calculations, and accelerates control response speed.

[0009] Optionally, in one embodiment of this application, the step of generating a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating conditions based on the vehicle's current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value includes: if the actual confidence level is greater than or equal to a preset confidence level, then determining an initial processing instruction for the actual noise data based on the current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value; and correcting the initial processing instruction based on a correction parameter matching the actual noise data in the mapping relationship to obtain the processing instruction.

[0010] The above technical solution allows for the determination of initial processing instructions based on current operating conditions, actual noise type, actual confidence level, and / or actual noise value, provided the actual confidence level is greater than or equal to a preset confidence level. Correction parameters for these initial processing instructions are then determined through a mapping relationship. This process first filters reliable identification results using a preset confidence level and generates initial processing instructions based on multi-dimensional operating condition information. Finally, the mapping relationship is used to solve for correction parameters and optimize the instruction content. This approach avoids erroneous control caused by low-confidence identification results and optimizes control instructions step-by-step, improving the accuracy and stability of vehicle noise management.

[0011] Optionally, in one embodiment of this application, before inputting the actual noise data into the pre-built noise recognition model, the method further includes: determining the pruning dimensions of the deep learning model and the pruning ratio corresponding to each pruning dimension based on the actual noise data; pruning the deep learning model based on the pruning dimensions and the pruning ratio to obtain a pruned model; quantizing and calibrating the pruned model to determine the quantization parameters of each quantization layer to obtain a quantized model; and loading the quantized model into an NPU (Neural Processing Unit) to construct the noise recognition model.

[0012] The above technical solution allows for the determination of pruning dimensions and pruning ratios for each pruning dimension in the deep learning model before inputting actual noise data into the noise recognition model. This enables the deep learning model to be pruned, resulting in a pruned model. Furthermore, quantization calibration of the pruned model determines the quantization parameters for each quantization layer, yielding a quantized model. This quantized model is then loaded into the NPU to complete the construction of the noise recognition model. By adaptively determining the pruning dimensions and ratios based on actual noise data, the network structure is simplified. Quantization calibration further optimizes the model parameters before deployment to the NPU to build the noise recognition model. This approach effectively reduces computational overhead, improves inference speed, and maintains noise recognition accuracy, thus meeting the real-time computing needs of in-vehicle hardware.

[0013] Optionally, in one embodiment of this application, determining the actual noise data of the vehicle based on the vehicle's current noise data and operating status data includes: collecting multi-source audio data of the vehicle and determining the vehicle's non-noise audio data; determining the vehicle's non-noise characteristics based on the non-noise audio data; and removing audio data that meets the non-noise characteristics from the multi-source audio data to obtain the current noise data.

[0014] The above technical solution allows for the first collection of multi-source audio data from the vehicle, the identification of non-noise audio data, and the extraction of non-noise features from the non-noise audio data. Audio data that meets the non-noise features is then removed to obtain the current noise data. By collecting multi-source audio data from the vehicle and extracting non-noise features as a filtering basis, interference-type audio signals are removed, and the current noise data is purified. This eliminates data interference caused by irrelevant acoustic signals, ensuring the accuracy of data for subsequent noise analysis and control decisions.

[0015] Optionally, in one embodiment of this application, determining the actual noise data of the vehicle based on the vehicle's current noise data and operating status data includes: extracting a first time feature of the current noise data and a second time feature of the operating status data; calculating the time difference between the current noise data and the operating status data based on the first time feature and the second time feature; and if the time difference is less than a preset time threshold, aligning the current noise data and the operating status data based on a preset time reference to obtain the actual noise data.

[0016] The above technical solution allows for the extraction of first-time features from current noise data and second-time features from operating status data. The time difference between the current noise data and operating status data is then calculated based on these first and second time features. If the time difference is less than a preset time threshold, time alignment is performed based on either the first or second time feature to obtain the actual noise data. By extracting the time features of the two types of data and calculating the timing deviation, data time alignment is completed when the time difference meets the requirements. This corrects the asynchronous sampling defects of multiple sensors, resulting in time-matched actual noise data. This provides a precise and synchronized data source for subsequent joint noise analysis and control decisions.

[0017] A second aspect of this application provides a vehicle noise control device, comprising: a first determining module, configured to determine the actual noise data of the vehicle based on current noise data and operating status data; an output module, configured to input the actual noise data into a pre-constructed noise recognition model, extract noise features from the actual noise data, and output a noise type probability distribution vector and a noise quantization feature vector of the actual noise data based on the noise features, wherein the noise recognition model is constructed by a pruned and quantized deep learning model, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold; and a control module, configured to determine the actual noise type, actual confidence level, and / or actual noise value based on the noise type probability distribution vector and the noise quantization feature vector, and generate a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating condition based on the current operating condition of the vehicle, the actual noise type, the actual confidence level, and / or the actual noise value, and control the vehicle to execute the processing instruction.

[0018] The above technical solution integrates current vehicle noise data with operational status data to obtain actual noise data, eliminating data interference and ensuring the authenticity and reliability of the noise data source. This actual noise data is then input into a noise recognition model constructed from a pruned and quantized lightweight deep learning model. The noise recognition model simultaneously outputs a noise type probability distribution vector and a noise quantization feature vector, compressing computational load while maintaining recognition accuracy to meet the real-time inference requirements of the vehicle. Based on this, the noise category and noise quantization index are accurately determined. Furthermore, the actual noise type, actual confidence level, and / or actual noise value are determined based on the noise type probability distribution vector and noise quantization feature vector. This generates a processing instruction that ensures the actual noise value is below a preset noise threshold, which is then executed by the vehicle. By specifically generating and issuing processing instructions for the actual noise data, differentiated and refined noise control is achieved.

[0019] Optionally, in one embodiment of this application, the control module includes: a first extraction unit, configured to acquire target noise data of a target vehicle, and determine a target noise type, a target confidence level, a target noise value, and a target operating condition based on the target noise data, wherein the similarity between the target noise data and the actual noise data is greater than a preset similarity threshold; a first determination unit, configured to determine a target processing instruction matching the target noise data based on the target noise type, the target confidence level, the target noise value, and the target operating condition; a second determination unit, configured to determine a mapping relationship between the target noise type, the target confidence level, the target noise value, the target operating condition, and the target processing instruction; and a first generation unit, configured to obtain the processing instruction from the mapping relationship based on the current operating condition, the actual noise type, the actual confidence level, and / or the actual noise value.

[0020] The above technical solution can determine the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data of the target vehicle. This allows for the generation of target processing instructions that match the target noise data. Furthermore, it establishes a mapping relationship between the target noise type, target confidence level, target noise value, target operating condition, and target processing instructions. Based on this mapping relationship, processing instructions can be matched and customized according to the target noise data. A one-to-one mapping relationship is constructed, and instructions can be directly matched by looking up a table during the operation phase. This ensures that the noise handling strategy is accurately adapted to the scenario, simplifies online calculations, and accelerates control response speed.

[0021] Optionally, in one embodiment of this application, the control module includes: a third determining unit, configured to determine an initial processing instruction for the actual noise data based on the current operating condition, the actual noise type, the actual confidence level, and / or the actual noise value when the actual confidence level is greater than or equal to a preset confidence level; and a second generating unit, configured to modify the initial processing instruction based on a correction parameter matching the actual noise data in the mapping relationship to obtain the processing instruction.

[0022] The above technical solution allows for the determination of initial processing instructions based on current operating conditions, actual noise type, actual confidence level, and / or actual noise value, provided the actual confidence level is greater than or equal to a preset confidence level. Correction parameters for these initial processing instructions are then determined through a mapping relationship. This process first filters reliable identification results using a preset confidence level and generates initial processing instructions based on multi-dimensional operating condition information. Finally, the mapping relationship is used to solve for correction parameters and optimize the instruction content. This approach avoids erroneous control caused by low-confidence identification results and optimizes control instructions step-by-step, improving the accuracy and stability of vehicle noise management.

[0023] Optionally, in one embodiment of this application, it further includes: a second determining module, configured to determine the pruning dimension of the deep learning model and the pruning ratio corresponding to each pruning dimension based on the actual noise data before inputting the actual noise data into the pre-built noise recognition model; a first generating module, configured to prune the deep learning model based on the pruning dimension and the pruning ratio to obtain a pruned model; a second generating module, configured to perform quantization calibration on the pruned model to determine the quantization parameters of each quantization layer to obtain a quantized model; and a building module, configured to load the quantized model into the NPU to build the noise recognition model.

[0024] The above technical solution allows for the determination of pruning dimensions and pruning ratios for each pruning dimension in the deep learning model before inputting actual noise data into the noise recognition model. This enables the deep learning model to be pruned, resulting in a pruned model. Furthermore, quantization calibration of the pruned model determines the quantization parameters for each quantization layer, yielding a quantized model. This quantized model is then loaded into the NPU to complete the construction of the noise recognition model. By adaptively determining the pruning dimensions and ratios based on actual noise data, the network structure is simplified. Quantization calibration further optimizes the model parameters before deployment to the NPU to build the noise recognition model. This approach effectively reduces computational overhead, improves inference speed, and maintains noise recognition accuracy, thus meeting the real-time computing needs of in-vehicle hardware.

[0025] Optionally, in one embodiment of this application, the first determining module includes: a collection unit, configured to collect multi-source audio data of the vehicle and determine non-noise audio data of the vehicle; a fourth determining unit, configured to determine non-noise features of the vehicle based on the non-noise audio data; and a rejection unit, configured to reject audio data that satisfies the non-noise features from the multi-source audio data to obtain the current noise data.

[0026] The above technical solution allows for the first collection of multi-source audio data from the vehicle, the identification of non-noise audio data, and the extraction of non-noise features from the non-noise audio data. Audio data that meets the non-noise features is then removed to obtain the current noise data. By collecting multi-source audio data from the vehicle and extracting non-noise features as a filtering basis, interference-type audio signals are removed, and the current noise data is purified. This eliminates data interference caused by irrelevant acoustic signals, ensuring the accuracy of data for subsequent noise analysis and control decisions.

[0027] Optionally, in one embodiment of this application, the first determining module includes: a second extraction unit, configured to extract a first time feature of the current noise data and a second time feature of the operating status data; a calculation unit, configured to calculate the time difference between the current noise data and the operating status data based on the first time feature and the second time feature; and a third generation unit, configured to align the current noise data and the operating status data based on a preset time reference when the time difference is less than a preset time threshold, thereby obtaining the actual noise data.

[0028] The above technical solution allows for the extraction of first-time features from current noise data and second-time features from operating status data. The time difference between the current noise data and operating status data is then calculated based on these first and second time features. If the time difference is less than a preset time threshold, time alignment is performed based on either the first or second time feature to obtain the actual noise data. By extracting the time features of the two types of data and calculating the timing deviation, data time alignment is completed when the time difference meets the requirements. This corrects the asynchronous sampling defects of multiple sensors, resulting in time-matched actual noise data. This provides a precise and synchronized data source for subsequent joint noise analysis and control decisions.

[0029] A third aspect of this application provides an electronic device, including a processor and a memory, wherein the memory is used to store a computer program; and the processor is used to execute the program stored in the memory to implement the vehicle noise control method as described in the above embodiments.

[0030] A fourth aspect of this application provides a vehicle that includes electronic devices as described in the above embodiments.

[0031] A fifth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle noise control method described above. Attached Figure Description

[0032] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a vehicle noise control method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating the working principle of a vehicle noise control method provided in an embodiment of this application; Figure 3 This is a structural diagram of the vehicle noise control device provided in the embodiments of this application; Figure 4 This is a structural diagram of the electronic device provided in the embodiments of this application.

[0033] Figure label: Among them, 10-vehicle noise control device; 100-first determination module, 200-output module, 300-control module; 401-processor, 402-memory. Detailed Implementation

[0034] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] Example 1 This application provides a method for controlling vehicle noise. Please refer to the following embodiments. Figure 1 This includes the following steps: In step S101, the actual noise data of the vehicle is determined based on the vehicle's current noise data and operating status data.

[0036] It is understood that, in the embodiments of this application, the operating status data may include, but is not limited to, vehicle speed, motor speed, window / sunroof status, wiper gear, air conditioning gear, driving mode, etc. This application does not impose specific limitations. It can be collected through the vehicle CAN (Controller Area Network) bus or through communication interfaces such as Ethernet. The specific settings can be made by those skilled in the art according to the actual situation. This application does not impose specific limitations.

[0037] The current noise data in this application embodiment can be understood as audio data that retains only the vehicle's own wind noise, tire noise, or powertrain noise after removing in-vehicle voices, in-vehicle music, and external environmental noise.

[0038] In actual implementation, the embodiments of this application can determine the actual noise data of the vehicle based on the vehicle's current noise data and operating status data.

[0039] Optionally, in one embodiment of this application, determining the actual noise data of the vehicle based on the vehicle's current noise data and operating status data includes: collecting multi-source audio data of the vehicle and determining the vehicle's non-noise audio data; determining the vehicle's non-noise characteristics based on the non-noise audio data; and removing audio data that meets the non-noise characteristics from the multi-source audio data to obtain the current noise data.

[0040] It is understood that, in the embodiments of this application, multi-source audio data may include, but is not limited to, audio data such as human voices, in-vehicle music, external environmental noise, wind noise, tire noise, and powertrain noise, and this application does not impose specific limitations; while non-noise audio data can be understood as audio data in multi-source audio data that does not belong to wind noise, tire noise, or powertrain noise, such as in-vehicle music, navigation voice, passenger conversations, telephone calls, and prompts.

[0041] Furthermore, embodiments of this application can perform signal analysis on non-noise audio data to extract a fixed set of acoustic features corresponding to the non-noise audio data, such as the fundamental frequency range of human voice, the spectral texture of music, and the short-term energy of speech, etc.

[0042] In some embodiments, this application can simultaneously collect multiple audio sources from various parts of the vehicle to obtain multi-source audio data, and determine non-noise audio data that does not need to be processed, such as in-vehicle music, navigation voice, passenger conversations, telephone calls, and prompts. Non-noise features of the non-noise audio data, such as the fundamental frequency range of human voice, music spectrum texture, and short-time energy of speech, are extracted. The multi-source audio data is traversed, and if the audio features in the multi-source audio data match the non-noise features, the corresponding audio data is removed, and the remaining audio data is used as the current noise data.

[0043] For example, in the embodiments of this application, the microphone array accesses multi-source audio data via an I2S (Inter IC Sound) / TDM (Time Division Multiplexing) digital audio interface. The multi-source audio data is transferred to the SoC (System on Chip) shared memory buffer by a dedicated audio DMA (Direct Memory Access) controller. During the data transmission, a low-power coprocessor performs non-noise feature detection in parallel, discards audio data that meets the non-noise features, and outputs the current noise data.

[0044] In this embodiment, the audio sampling rate of the microphone array acquisition end is set to 48kHz, and the audio bit depth is set to 16bit or 24bit. Multi-source audio data is segmented with an audio frame duration of 20ms, and the audio frame duration simultaneously constrains the basic processing latency and frequency domain analysis resolution. Furthermore, this embodiment can configure a voice activity detection sensitivity threshold to determine non-noise audio data in the multi-source audio data, assisting in the extraction of current noise data. Simultaneously, the audio ring buffer depth, a DMA cache scheduling parameter, is configured to effectively prevent data overflow during audio transmission. The relevant parameters are shown in Table 1, where the audio sampling rate is represented as AUDIO_SAMPLE_RATE, the audio bit depth as AUDIO_BIT_DEPTH, the audio frame duration as AUDIO_FRAME_SIZE_MS, the voice activity detection sensitivity threshold as VAD_SENSITIVITY, and the audio ring buffer depth as AUDIO_BUFFER_DEPTH. Table 1 is a schematic table of audio data parameters provided in one embodiment of this application.

[0045] Table 1

[0046] This application first collects multi-source audio data of the vehicle and determines the non-noise audio data of the vehicle. Then, it extracts non-noise features from the non-noise audio data to remove audio data that meets the non-noise features, thereby obtaining the current noise data. By collecting multi-source audio data of the vehicle and extracting non-noise features as a filtering basis, the current noise data is purified after removing interfering audio signals. This can eliminate data interference caused by irrelevant acoustic signals and ensure the accuracy of data for subsequent noise analysis and control decisions.

[0047] Optionally, in one embodiment of this application, determining the actual noise data of a vehicle based on its current noise data and operating status data includes: extracting a first time feature of the current noise data and a second time feature of the operating status data; calculating the time difference between the current noise data and the operating status data based on the first time feature and the second time feature; and if the time difference is less than a preset time threshold, aligning the current noise data and the operating status data based on a preset time reference to obtain the actual noise data.

[0048] It is understood that, in the embodiments of this application, the first time feature can be understood as the time stamp of the current noise data, such as the timestamp, sampling sequence number, and frame number of each frame of data acquisition time; the second time feature can be understood as the time stamp of the running status data, such as the timestamp and message cycle number attached when the vehicle controller sends CAN messages.

[0049] In some embodiments, this application can first read the first time feature inherent in the current noise data, such as a first timestamp, and the second time feature inherent in the running status data, such as a second timestamp. Based on the first and second timestamps, the time difference between the current noise data and the running status data is calculated, and it is detected whether the time difference is less than a preset time threshold. If it is less, time alignment is performed according to a preset time reference, such as the first timestamp, the second timestamp, or other timestamps, to complete the time synchronization of the current noise data and the running status data, thereby determining the actual noise data. The preset time reference and preset time threshold can be set by those skilled in the art according to actual conditions, and this application does not impose specific limitations.

[0050] For example, in this embodiment of the application, a hardware timestamp unit can accurately record the acquisition time of the current noise data and the operating status data, determine the first timestamp of the current noise data and the second timestamp of the operating status data, calculate the time difference between the first timestamp and the second timestamp, and if the time difference is less than a preset time threshold, then the first timestamp is... As a preset time reference, and based on the preset time reference, the operating status data that best matches the current noise data is found in the operating status data. The preset time reference, the operating status data, and the current noise data are then packaged together into a time-domain corresponding data frame to form the actual noise data. The parameters involved are shown in Table 2, which is a schematic table of data time alignment parameters provided in one embodiment of this application.

[0051] This can be understood as meaning that, in this embodiment of the application, time alignment will only be initiated if the time difference between the first timestamp and the second timestamp is less than a preset time threshold. ,in, Indicates the first timestamp. Indicates the second timestamp. This indicates a preset time threshold, such as ±1000μs.

[0052] Furthermore, in this embodiment, the synchronization time accuracy between the aligned running status data and the current noise data is <1μs, which can be set by those skilled in the art according to the actual situation.

[0053] Table 2

[0054] This application embodiment can first extract the first time feature of the current noise data and the second time feature of the operating status data, and calculate the time difference between the current noise data and the operating status data based on the first time feature and the second time feature. Then, if the time difference is less than a preset time threshold, time alignment is performed based on the first time feature or the second time feature to obtain the actual noise data. By extracting the time features of the two types of data and calculating the timing deviation, data time alignment is completed when the time difference meets the requirements, correcting the defect of asynchronous sampling of multiple sensors, and obtaining the actual noise data with timing matching, providing a precise and synchronized data source for subsequent joint noise analysis and control decision-making.

[0055] Optionally, in one embodiment of this application, before inputting the actual noise data into the pre-built noise recognition model, the method further includes: determining the pruning dimensions of the deep learning model and the pruning ratio corresponding to each pruning dimension based on the actual noise data; pruning the deep learning model based on the pruning dimensions and pruning ratio to obtain a pruned model; quantizing and calibrating the pruned model to determine the quantization parameters of each quantization layer to obtain a quantized model; and loading the quantized model into the NPU to construct the noise recognition model.

[0056] It is understood that, in the embodiments of this application, the deep learning model can be understood as a neural network built for noise multi-classification and feature extraction, such as CNN (Convolutional Neural Network) / LSTM (Long Short-Term Memory). Because neural networks have a large number of parameters and long inference time, they cannot be directly run in real-time with low latency on the in-vehicle NPU. Therefore, the embodiments of this application can prune and quantize the neural network. The NPU, as a dedicated in-vehicle hardware AI acceleration chip, is specifically designed to perform neural network convolution and feature extraction operations in parallel. Compared to CPU (Central Processing Unit) / GPU (Graphics Processing Unit), it has faster inference speed and lower power consumption, making it suitable for real-time noise recognition closed-loop control of vehicles.

[0057] In this embodiment of the application, when pruning a deep model, the pruning dimensions for removing executable parameters in the neural network can be determined first based on actual noise data, such as the convolutional kernel dimension of the convolutional layer, the weight channel dimension of the fully connected layer, the temporal feature dimension, and the network branch dimension. Then, the pruning ratio is determined according to the pruning dimensions, such as the pruning ratio of the convolutional layer being 30%, thereby pruning the deep learning model to obtain a simplified neural network model, and constructing a pruned model.

[0058] Furthermore, in embodiments of this application, when the pruning model is a floating-point model and cannot be directly and efficiently deployed on an NPU, quantization calibration is performed on the pruning model to solve the quantization parameters of each quantization layer, converting the floating-point model into an integer quantization model to obtain the corresponding quantization model.

[0059] In essence, before the actual noise data is input into the pre-built noise recognition model, this application first analyzes the redundancy of each layer of the deep learning model based on the actual noise data, determines the pruning dimension and pruning ratio of each layer, and then prunes the deep learning model according to the pruning dimension and pruning ratio to remove redundant structures, thereby obtaining a pruned model. The pruned model is then quantized and calibrated to solve for the quantization parameters required for each quantization layer. The pruned model is then converted into a quantized model and loaded into the NPU, thereby completing the construction of the corresponding noise recognition model.

[0060] For example, in this application embodiment, a quantization model optimized by joint structured pruning and INT8 static quantization of a deep learning model can be loaded and run, and the quantization model can be preloaded into the tightly coupled memory of the NPU to complete the construction of the noise recognition model.

[0061] Structured pruning fixes the computational graph and computational load of the model, eliminating delay fluctuations caused by conditional branching. INT8 static quantization converts computation into periodically determined integer operations on the NPU. The model parameters involved are shown in Table 3. Table 3 is a schematic table of model parameters provided in one embodiment of this application.

[0062] Table 3

[0063] This embodiment of the application can determine the pruning dimensions and pruning ratios of the deep learning model based on the actual noise data before inputting the actual noise data into the noise recognition model. This allows for the pruning of the deep learning model, resulting in a pruned model. Furthermore, the quantization parameters of each quantization layer are determined through quantization calibration of the pruned model, resulting in a quantized model. This quantized model is then loaded into the NPU to complete the construction of the noise recognition model. By adaptively determining the pruning dimensions and pruning ratios based on the actual noise data, the network structure is simplified. Quantization calibration further optimizes the model parameters, and finally, the model is deployed to the NPU to construct the noise recognition model. This approach effectively reduces model computational overhead, improves inference speed, and maintains noise recognition accuracy, thus meeting the real-time computing needs of in-vehicle hardware.

[0064] In step S102, the actual noise data is input into the pre-built noise recognition model to extract the noise features of the actual noise data, so as to output the noise type probability distribution vector and noise quantization feature vector of the actual noise data according to the noise features. The noise recognition model is constructed by a deep learning model that has been pruned and quantized, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold.

[0065] It is understood that, in the embodiments of this application, the noise type probability distribution vector is used to represent the probability vector of the noise category, which may, but is not limited to, be represented as... ,in, This indicates the probability that the noise type is wind noise. This indicates the probability that the noise type is tire noise. This indicates the probability that the noise type is powertrain noise, and .

[0066] The noise quantization feature vector in this application embodiment can be understood as a quantization feature vector used to characterize the noise level. It may include, but is not limited to, the total sound pressure level inside the vehicle, the energy of each frequency band, the center frequency, the bandwidth, etc., and then the actual noise value can be analyzed based on the noise quantization feature vector.

[0067] In some embodiments, the present application can input actual noise data into a pre-built noise recognition model, and then use the noise recognition model to extract noise features from the actual noise data. Based on the noise features, the noise type probability distribution vector and noise quantization feature vector of the actual noise data are output within a preset time threshold. The preset time threshold can be set by those skilled in the art according to actual conditions, and the present application does not impose specific limitations.

[0068] It should be noted that the noise recognition model in this embodiment is constructed from a pruned and quantized deep learning model. The pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold, such as the worst-case execution time of a single NPU inference. The worst-case execution time of a single NPU inference can be verified through real-world testing under numerous worst-case inputs, thus ensuring the achievable performance ceiling.

[0069] For example, in this embodiment of the application, the noise recognition model deployed in the NPU receives actual noise data, performs a forward propagation, and extracts the corresponding noise features, such as the Mel spectrogram. Due to model pruning and quantization, the fluctuation of the model's single inference time on the NPU is extremely low, less than the worst-case execution time of a single NPU inference. Therefore, within the worst-case execution time of a single NPU inference, the model outputs two types of results in parallel: a noise type probability distribution vector and a noise quantization feature vector.

[0070] In step S103, based on the noise type probability distribution vector and the noise quantization feature vector, the actual noise type, actual confidence level and / or actual noise value are determined, and according to the vehicle's current operating conditions, actual noise type, actual confidence level and / or actual noise value, a processing instruction is generated to make the actual noise value less than a preset noise threshold that matches the current operating conditions, and the vehicle is controlled to execute the processing instruction.

[0071] It is understood that, in the embodiments of this application, the noise type may include, but is not limited to, wind noise, tire noise, powertrain noise, etc. Among them, wind noise can be understood as the airflow vibration noise generated when the vehicle body breaks through the air, the airflow squeezes the gaps in the vehicle body, and the air is stirred up by the protruding parts such as doors, windows, rearview mirrors, and pillars when the vehicle is in motion; tire noise can be understood as the noise generated by the tire rolling and pressing against the road surface, the tire tread slapping the ground, the vibration of the tire's own rubber, and the resonance of the tire cavity; powertrain noise can be understood as the noise generated by the operation of the power system, mechanical meshing, and exhaust vibration.

[0072] The confidence levels in this application embodiment can be divided into high confidence, medium confidence, and low confidence. High confidence can be understood as the maximum probability in the noise type probability distribution vector ≥ 0.8; medium confidence can be understood as 0.5 ≤ maximum probability < 0.8; and low confidence can be understood as the maximum probability in the noise type probability distribution vector < 0.5. The confidence level threshold can be represented as CLASS_CONFIDENCE_THRESHOLD, which is used to filter out low confidence identification results. The specific threshold can be set by those skilled in the art according to the actual situation, and this application does not impose any specific restrictions.

[0073] Furthermore, in the embodiments of this application, the current operating condition may include, but is not limited to, high-speed driving, driving with windows open, rough road surface driving, low-speed bumpy driving, acceleration, high-speed operation, and vehicle idling stationary conditions, etc., and this application does not impose specific limitations. The current operating condition can be determined based on operating status data. Different operating conditions correspond to different noise thresholds; for example, the noise threshold for high-speed driving is 65dB; for rough road surface driving, the noise threshold is 63dB; and for vehicle idling stationary conditions, the noise threshold is 57dB, and this application does not impose specific limitations.

[0074] In some embodiments, the present application embodiments can determine the noise type and confidence level of the actual noise data based on the noise type probability distribution vector, then determine the actual noise value of the actual noise data based on the noise quantization feature vector, and determine the current operating condition of the vehicle and the preset noise threshold matched by the current operating condition based on the operating status data. Then, based on the noise type, confidence level, actual noise value and preset noise threshold, the corresponding processing instruction is determined and the vehicle is controlled to execute the processing instruction.

[0075] For example, in embodiments of this application, the noise type probability distribution vector can be... In this case, the noise type is determined to be tire noise, the confidence level is high confidence, and the actual noise value is determined to be 65.4dB based on the noise quantization feature vector. Based on the operating status data, the current operating condition is determined to be rough road condition, and the preset noise threshold is 63dB. Therefore, it is determined that the noise exceeds the standard, and the corresponding processing command is generated to make the actual noise value less than 63dB, and the vehicle is controlled to execute the corresponding processing command.

[0076] Optionally, in one embodiment of this application, a processing instruction is generated based on the vehicle's current operating condition, actual noise type, actual confidence level, and / or actual noise value to make the actual noise value less than a preset noise threshold matching the current operating condition. This includes: acquiring target noise data of the target vehicle, and determining the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data, wherein the similarity between the target noise data and the actual noise data is greater than a preset similarity threshold; determining a target processing instruction matching the target noise data based on the target noise type, target confidence level, target noise value, and target operating condition; determining a mapping relationship between the target noise type, target confidence level, target noise value, target operating condition, and target processing instruction; and matching the processing instruction from the mapping relationship based on the current operating condition, actual noise type, actual confidence level, and / or actual noise value.

[0077] It is understood that, in the embodiments of this application, the target noise data can be a noise dataset (including the target vehicle's target operating status data) collected and processed on the target vehicle. This dataset is not arbitrary noise, and its similarity to the actual noise data used for online inference is greater than a preset similarity threshold. It is representative and can be used for offline modeling. The preset similarity threshold can be set by those skilled in the art according to the actual situation, and this application does not impose any specific restrictions.

[0078] The processing instructions in the embodiments of this application may include, but are not limited to, a control strategy identifier, used to identify a specific algorithm or mode that the downstream actuator (such as a DSP (Digital Signal Processor)) should enable; and a parameter set, which can be understood as optimized parameters configured for the control strategy, such as center frequency, bandwidth, correction parameters, etc.

[0079] In some embodiments, the present application embodiments may first acquire target noise data of the target vehicle and input the target noise data into a pre-built noise recognition model. Then, the noise recognition model is used to determine the target noise type, target confidence level, and target noise value. Based on the target operating status data, the target operating condition of the target vehicle is determined. Then, using the noise type, confidence level, and operating condition as search indexes and the corresponding noise reduction processing instruction as the output value, the mapping relationship between the target noise type, target confidence level, target noise value, target operating condition, and target processing instruction is determined, and a corresponding instruction pairing rule base is established. Then, based on the current operating condition, actual noise type, actual confidence level, and / or actual noise value, the processing instruction is matched from the mapping relationship.

[0080] For example, in the offline stage, this application collects target noise data of the target vehicle and determines the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data. Then, using the noise type, confidence level, and operating condition as search indexes and the corresponding noise reduction processing instruction as the output value, a complete instruction pairing rule base of noise scene-processing instruction is predefined and compiled. This rule base is implemented in the form of a static lookup table and stored in the processor's cache. Its index keys consist of discrete input conditions, including noise type indexes (e.g., 0 for wind noise, 1 for tire noise, 2 for powertrain noise), confidence level indexes (e.g., high / medium / low), and operating condition indexes (e.g., high-speed driving condition, open-window driving condition, etc.). Each index key corresponds to a predefined processing instruction, including a control strategy identifier and parameter set (e.g., {center frequency: 800Hz, bandwidth: 200Hz, correction parameter: 0.001}). Furthermore, the lookup table has a fixed size and no dynamic growth or modification logic. Furthermore, the search process is a deterministic, non-recursive, and non-looping direct memory access operation, and its execution time depends only on the memory access latency, which is constant time complexity.

[0081] Furthermore, in this embodiment, after determining the current operating conditions, actual noise type, actual confidence level, and actual noise value, the highest-priority decision fusion task in the real-time operating system is triggered. First, the continuous data undergoes rapid deterministic discretization to obtain corresponding discretized data. Then, based on the instruction pairing rule base, the discretized multi-dimensional inputs are combined to calculate a unique lookup index value. This calculation involves only simple arithmetic and bitwise operations, with a fixed execution time. After using the calculated index value, a deterministic memory read operation is performed to directly obtain the corresponding control policy identifier and parameter set from the pre-compiled lookup table. Since the entire process does not involve any branch prediction failures, cache misses, or dynamic memory allocation, the longest execution time of the lookup process (e.g., <500μs) can be precisely defined through static analysis and memory access latency, and verified experimentally. The control policy identifier and parameter set obtained from the lookup table are then encapsulated into a fixed-length control instruction package according to a predefined, simplified binary format to obtain the corresponding processing instructions. This encapsulation process is merely a memory copy, and its time consumption is predictable. This embodiment of the application maps the fusion result to a processing instruction within microseconds through a single deterministic table lookup operation, ensuring stable and controllable processing time across the entire process. Furthermore, this embodiment can directly output the corresponding processing instruction based on a default control strategy when the actual confidence level is less than a preset threshold. The preset threshold can be set by those skilled in the art according to actual conditions, and this application does not impose specific limitations. The parameters involved in this process are shown in Table 4. Table 4 is an illustrative table of parameters involved in generating processing instructions in one embodiment of this application.

[0082] Table 4

[0083] This application embodiment can determine the target noise type, target confidence level, target noise value, and target operating condition based on the target noise data of the target vehicle, thereby obtaining a target processing instruction matched to the target noise data. It also determines the mapping relationship between the target noise type, target confidence level, target noise value, target operating condition, and target processing instruction, and obtains the processing instruction according to the mapping relationship. Adaptive processing instructions are customized based on the target noise data, and a one-to-one mapping relationship is constructed. During the operation phase, the instruction can be directly matched by looking up the table, which not only ensures that the noise handling strategy is accurately adapted to the scenario, but also simplifies online calculation and speeds up the control response.

[0084] Optionally, in one embodiment of this application, a processing instruction is generated based on the vehicle's current operating conditions, actual noise type, actual confidence level, and / or actual noise value to make the actual noise value less than a preset noise threshold that matches the current operating conditions. This includes: if the actual confidence level is greater than or equal to a preset confidence level, determining an initial processing instruction for the actual noise data based on the current operating conditions, actual noise type, actual confidence level, and / or actual noise value; and correcting the initial processing instruction based on a correction parameter that matches the actual noise data in the mapping relationship to obtain the processing instruction.

[0085] In some embodiments, this application can determine the initial processing instruction for the actual noise data based on the current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value, when the actual confidence level is greater than or equal to a preset confidence level. The initial processing instruction is then modified based on correction parameters matching the actual noise data in the mapping relationship to obtain the final processing instruction. The preset confidence level can be set by those skilled in the art according to actual conditions, and this application does not impose specific limitations.

[0086] For example, in order to further reduce the effective delay while ensuring determinism, this application adopts a two-stage output. That is, when the actual confidence level is greater than or equal to the preset confidence level, such as high confidence, the initial processing instruction is generated. After the control strategy identifier and parameter set are determined by the instruction pairing rule base, the correction parameters of the initial processing instruction are determined, and then confirmed or fine-tuned to obtain the processing instruction. In this way, the processing delay is hidden without disrupting the fixed pipeline cycle time, and a near-zero delay response to the controller is achieved.

[0087] Furthermore, the total execution time in this embodiment is determined by the longest execution time of the decision fusion task (FUSION_DECISION_WCET_US) and the worst-case execution time of a single NPU inference (NPU_INFERENCE_WCET_MS). This embodiment is triggered at a globally fixed cycle and, through static priority scheduling of the real-time operating system, ensures that the pipeline completes within each cycle, thereby achieving deterministic end-to-end latency. The relevant parameters are shown in Table 5. Table 5 is a parameter illustration of the initial processing instruction correction process provided in one embodiment of this application.

[0088] Table 5

[0089] This application embodiment can determine the initial processing command based on the current operating conditions, actual noise type, actual confidence level, and / or actual noise value when the actual confidence level is greater than or equal to the preset confidence level. The correction parameters of the initial processing command are determined through mapping relationship, thereby determining the processing command. First, reliable identification results are screened through the preset confidence level, and the initial processing command is generated based on multi-dimensional operating condition information. Then, the correction parameters are solved by using mapping relationship to optimize the command content. This can not only avoid the mis-control caused by low confidence identification results, but also optimize the control command step by step, thereby improving the accuracy and stability of vehicle noise control.

[0090] Example 2 This application also provides an embodiment of a vehicle noise control method. Please refer to the working principle. Figure 2 ,include: Step S201: Collect multi-source audio data of the vehicle and identify non-noise audio data.

[0091] In this embodiment, the microphone array accesses multi-source audio data via an I2S / TDM digital audio interface. The multi-source audio data is transferred to the SoC shared memory buffer by a dedicated audio DMA controller. During the data transmission process, a low-power coprocessor performs non-noise feature detection in parallel, eliminates audio data that meets the non-noise features, and outputs the current noise data.

[0092] Step S202: Obtain running status data.

[0093] In this embodiment, operational status data can be collected through communication interfaces such as the vehicle CAN bus and Ethernet.

[0094] Step S203: Synchronize data frame assembly to determine actual noise data.

[0095] In this embodiment, the current noise data and the operating status data can be time-aligned to determine the corresponding actual noise data.

[0096] Step S204: Prune and quantize the deep learning model, and load it into the NPU to build a noise recognition model.

[0097] The content of constructing the noise recognition model in the actual embodiment of this application is as shown above, and will not be repeated here.

[0098] Step S205: Use the noise recognition model for inference.

[0099] In this embodiment, the noise identification model can be used to output the noise type probability distribution vector and noise quantization feature vector of the actual noise data, thereby determining the actual noise type, actual confidence level and actual noise value.

[0100] Step S206: Check whether the actual confidence level is greater than or equal to the preset confidence level.

[0101] If the result is greater than or equal to the given result, proceed to step S207; otherwise, proceed to step S208.

[0102] Step S207: Output the initial processing instruction and modify it using the instruction matching rule base to obtain the processing instruction.

[0103] Step S208: Determine the processing instruction based on the instruction pairing rule base.

[0104] In this embodiment, the instruction pairing rule base uses noise type, confidence level, and operating condition as search indexes, and the corresponding noise reduction processing instruction as the output value.

[0105] Step S209: Execute the processing instruction to make the actual noise value less than the preset noise threshold.

[0106] It should be noted that the total execution time of this application embodiment is determined by the longest execution time of the decision fusion task, FUSION_DECISION_WCET_US, and the worst-case execution time of a single NPU inference, NPU_INFERENCE_WCET_MS.

[0107] This application also provides a vehicle noise control device 10, please refer to... Figure 3 It includes: a first determining module 100, an output module 200, and a control module 300.

[0108] The first determining module 100 is used to determine the actual noise data of the vehicle based on the vehicle's current noise data and operating status data.

[0109] The output module 200 is used to input actual noise data into a pre-built noise recognition model, extract noise features from the actual noise data, and output the noise type probability distribution vector and noise quantization feature vector of the actual noise data according to the noise features. The noise recognition model is constructed by a pruned and quantized deep learning model, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold.

[0110] The control module 300 is used to determine the actual noise type, actual confidence level and / or actual noise value based on the noise type probability distribution vector and the noise quantization feature vector, and to generate a processing instruction to make the actual noise value less than a preset noise threshold that matches the current operating conditions based on the vehicle's current operating conditions, actual noise type, actual confidence level and / or actual noise value, and to control the vehicle to execute the processing instruction.

[0111] Optionally, in one embodiment of this application, the control module 300 includes: a first extraction unit, a first determination unit, a second determination unit, and a first generation unit.

[0112] The first extraction unit is used to acquire target noise data of the target vehicle and determine the target noise type, target confidence level, target noise value and target operating condition based on the target noise data. The similarity between the target noise data and the actual noise data is greater than a preset similarity threshold.

[0113] The first determining unit is used to determine the target processing instruction that matches the target noise data based on the target noise type, target confidence level, target noise value, and target operating conditions.

[0114] The second determining unit is used to determine the mapping relationship between target noise type, target confidence level, target noise value, target operating condition and target processing instructions.

[0115] The first generation unit is used to obtain processing instructions from the mapping relationship based on the current operating conditions, actual noise type, actual confidence level and / or actual noise value.

[0116] Optionally, in one embodiment of this application, the control module 300 includes: a third determining unit and a second generating unit.

[0117] The third determining unit is used to determine the initial processing instructions for the actual noise data based on the current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value when the actual confidence level is greater than or equal to the preset confidence level.

[0118] The second generation unit is used to modify the initial processing instructions based on the correction parameters that match the actual noise data in the mapping relationship, so as to obtain the processing instructions.

[0119] Optionally, in one embodiment of this application, it further includes: a second determining module, a first generating module, a second generating module, and a building module.

[0120] The second determining module is used to determine the pruning dimensions of the deep learning model and the pruning ratio corresponding to each pruning dimension based on the actual noise data before inputting the actual noise data into the pre-built noise recognition model.

[0121] The first generation module is used to prune the deep learning model based on the pruning dimension and pruning ratio to obtain a pruned model.

[0122] The second generation module is used to perform quantization calibration on the pruning model, determine the quantization parameters of each quantization layer, and obtain the quantization model. The building block is used to load the quantization model into the NPU to build a noise recognition model.

[0123] Optionally, in one embodiment of this application, the first determining module 100 includes: a collection unit, a fourth determining unit, and a rejection unit.

[0124] The acquisition unit is used to acquire multi-source audio data of the vehicle and determine the vehicle's non-noise audio data.

[0125] The fourth determining unit is used to determine the non-noise characteristics of the vehicle based on non-noise audio data.

[0126] The elimination unit is used to remove audio data that meets non-noise characteristics from multi-source audio data to obtain the current noise data.

[0127] Optionally, in one embodiment of this application, the first determining module 100 includes: a second extraction unit, a calculation unit, and a third generation unit.

[0128] The second extraction unit is used to extract the first time feature of the current noise data and the second time feature of the running status data.

[0129] The calculation unit is used to calculate the time difference between the current noise data and the operating status data based on the first time feature and the second time feature.

[0130] The third generation unit is used to align the current noise data and the running status data based on a preset time reference to obtain the actual noise data when the time difference is less than a preset time threshold.

[0131] This application also provides an electronic device 20, please refer to... Figure 4 It includes a processor 410 and a memory 420, wherein the memory 410 is used to store computer programs; the processor 420 is used to execute the programs stored in the memory 410 to implement the vehicle noise control method described in any embodiment of this application.

[0132] This application also provides a vehicle that includes the electronic equipment described in any embodiment of this application.

[0133] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle noise control method described in any embodiment of this application.

[0134] In this application, "multiple" refers to two or more.

[0135] In this application, unless otherwise expressly defined, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0136] The terms “first,” “second,” “third,” “fourth,” etc., in this application (if present) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0137] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0138] Unless otherwise specified, all steps in this application may be performed sequentially or randomly. For example, if a method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if a method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0139] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for controlling vehicle noise, characterized in that, Includes the following steps: The actual noise data of the vehicle is determined based on the vehicle's current noise data and operating status data; The actual noise data is input into a pre-built noise recognition model to extract noise features from the actual noise data. Based on the noise features, the noise type probability distribution vector and noise quantization feature vector of the actual noise data are output. The noise recognition model is constructed by a pruned and quantized deep learning model, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold. Based on the noise type probability distribution vector and the noise quantization feature vector, the actual noise type, actual confidence level and / or actual noise value are determined. Based on the current operating conditions of the vehicle, the actual noise type, actual confidence level and / or actual noise value, a processing instruction is generated to make the actual noise value less than a preset noise threshold that matches the current operating conditions, and the vehicle is controlled to execute the processing instruction.

2. The method according to claim 1, characterized in that, The step of generating a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating conditions, based on the vehicle's current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value, includes: Acquire target noise data of the target vehicle, and determine the target noise type, target confidence level, target noise value and target operating condition based on the target noise data, wherein the similarity between the target noise data and the actual noise data is greater than a preset similarity threshold; Based on the target noise type, the target confidence level, the target noise value, and the target operating condition, a target processing instruction matching the target noise data is determined. Determine the mapping relationship between the target noise type, the target confidence level, the target noise value, the target operating condition, and the target processing command; The processing instruction is obtained by matching the current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value from the mapping relationship.

3. The method according to claim 2, characterized in that, The step of generating a processing instruction to make the actual noise value less than a preset noise threshold matching the current operating conditions, based on the vehicle's current operating conditions, the actual noise type, the actual confidence level, and / or the actual noise value, includes: If the actual confidence level is greater than or equal to the preset confidence level, then based on the current operating conditions, the actual noise type, the actual confidence level and / or the actual noise value, the initial processing instruction for the actual noise data is determined; Based on the correction parameters in the mapping relationship that match the actual noise data, the initial processing instruction is corrected to obtain the processing instruction.

4. The method according to claim 1, characterized in that, Before inputting the actual noise data into the pre-built noise recognition model, the following steps are also included: Based on the actual noise data, the pruning dimensions of the deep learning model and the pruning ratio corresponding to each pruning dimension are determined. Based on the pruning dimension and the pruning ratio, the deep learning model is pruned to obtain a pruned model; The pruning model is quantized and calibrated to determine the quantization parameters of each quantization layer, thereby obtaining the quantization model. The quantization model is loaded into the neural network processing unit (NPU) to construct the noise recognition model.

5. The method according to claim 1, characterized in that, Determining the actual noise data of the vehicle based on its current noise data and operating status data includes: Collect multi-source audio data of the vehicle and determine the non-noise audio data of the vehicle; Based on the non-noise audio data, the non-noise characteristics of the vehicle are determined; The audio data that meets the non-noise characteristics is removed from the multi-source audio data to obtain the current noise data.

6. The method according to claim 5, characterized in that, Determining the actual noise data of the vehicle based on its current noise data and operating status data includes: Extract the first temporal feature of the current noise data and the second temporal feature of the operating status data; Based on the first time feature and the second time feature, calculate the time difference between the current noise data and the operating status data; If the time difference is less than a preset time threshold, the current noise data and the operating status data are aligned based on a preset time reference to obtain the actual noise data.

7. A vehicle noise control device, characterized in that, include: The first determining module is used to determine the actual noise data of the vehicle based on the vehicle's current noise data and operating status data; The output module is used to input the actual noise data into a pre-built noise recognition model, extract the noise features of the actual noise data, and output the noise type probability distribution vector and noise quantization feature vector of the actual noise data according to the noise features. The noise recognition model is constructed by a pruned and quantized deep learning model, and the pruning and quantization are configured to make the single inference time of the noise recognition model less than a preset time threshold. The control module is configured to determine the actual noise type, actual confidence level, and / or actual noise value based on the noise type probability distribution vector and the noise quantization feature vector, and generate a processing instruction to make the actual noise value less than a preset noise threshold that matches the current operating conditions of the vehicle, the actual noise type, the actual confidence level, and / or the actual noise value, and control the vehicle to execute the processing instruction.

8. An electronic device, characterized in that, Including processor and memory, among which, Memory, used to store computer programs; A processor for executing a program stored in a memory to implement the vehicle noise control method according to any one of claims 1-6.

9. A vehicle, characterized in that, It includes the electronic device as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the vehicle noise control method according to any one of claims 1-6.

Images