A low-power composite multi-channel acoustic imaging method and system

Abstract

This application relates to the technical field of acoustic imaging, and in particular, to a low-power composite multi-channel acoustic imaging method and system, which includes: acquiring sampled acoustic signals and sending them to a three-stage pipeline structure; at the front end, acquiring the signal characteristics of the sampled acoustic signals to select corresponding deserialization paths to obtain several parallel data streams; at the middle end, distributing the several parallel data streams to each processing unit, acquiring acoustic environment information and allocating corresponding target frequency points k and frequency point requirements x1 to each processing unit; performing DFT operations on each parallel data stream to respectively obtain corresponding covariance matrices; acquiring a steering vector matrix and combining several covariance matrices to generate parallel beams; at the back end, acquiring phase compensation coefficients, and converting them into fir filter parameters based on the phase compensation coefficients to correct the phase errors of the parallel beams and then performing parallel-to-serial processing to generate serial outputs. This application has the effect of reducing the computational power consumption during acoustic imaging.

Description

Technical Field

[0001] This application relates to the technical field of acoustic imaging, and particularly to a low-power composite multi-channel acoustic imaging method and system. Background Art

[0002] In the field of acoustic signal processing, the improvement of composite multi-channel beamforming algorithms and low power consumption have become current research hotspots.

[0003] As a typical application device of this technology, an acoustic imager usually consists of a high-density microphone array, a high-precision camera, a multi-channel signal processing unit, and intelligent analysis software. Its working principle is based on beamforming technology: First, a high-density microphone array (such as a circular, linear, or planar array layout) synchronously collects acoustic wave signals according to a preset geometric structure. The array spacing usually satisfies the Nyquist sampling theorem to avoid spatial aliasing. Subsequently, the system calculates the phase difference (IPD) and time difference (TDOA) between each array element, and synthesizes a directional beam based on the phased array principle - by dynamically adjusting the weight coefficients of each array element, the acoustic waves in a specific direction are coherently superimposed to form a directional beam, suppressing sidelobe interference. Finally, through spatial scanning or beam scanning technology, the sound source intensity in different directions is mapped into a pseudo-color map (sonogram), and is fused with the visible light image collected by the camera to achieve visual superposition display of the sound field.

[0004] However, traditional acoustic imaging systems have problems of high computational intensity and excessive power consumption, with low overall resource utilization and an energy efficiency bottleneck in multi-channel signal processing. Summary of the Invention

[0005] In order to reduce the computational power consumption during acoustic imaging, this application provides a low-power composite multi-channel acoustic imaging method and system.

[0006] In a first aspect, a low-power composite multi-channel acoustic imaging method provided by this application adopts the following technical solution:

[0007] A low-power composite multi-channel acoustic imaging method includes the following steps:

[0008] Obtain sampled acoustic signals and send them to a three-level pipeline structure configured with a front end, a middle end, and a back end;

[0009] At the front end,

[0010] Obtain the signal characteristics of the sampled acoustic signals to select corresponding deserialization paths, and perform serial-to-parallel processing on the sampled acoustic signals based on the deserialization paths to obtain several parallel data streams;

[0011] At the middle end,

[0012] Invoke the parallel beamforming array to configure a number of processing units and distribute several paths of the parallel data streams to each of the processing units, obtain the acoustic environment information and allocate the corresponding target frequency point k and the number of frequency point requirements x1 to each of the processing units;

[0013] Perform DFT operations on each path of the parallel data streams based on the target frequency point k and the number of frequency point requirements x to respectively obtain the corresponding covariance matrices;

[0014] Obtain the steering vector matrix corresponding to the parallel beamforming array and combine a number of the covariance matrices to generate parallel beams;

[0015] At the backend,

[0016] Obtain the phase compensation coefficient, and convert it into fir filter parameters based on the phase compensation coefficient to correct the phase error of the parallel beams and then perform parallel-to-serial processing to generate a serial output.

[0017] In some embodiments, obtain the signal characteristics of the sampled acoustic signal to select the corresponding deserialization path, and perform serial-to-parallel processing on the sampled acoustic signal based on the deserialization path to obtain several paths of parallel data streams, including the following steps:

[0018] Obtain the high-speed serial signal corresponding to the sampled acoustic signal;

[0019] Collect the corresponding data rate and processing requirements in the signal characteristics, and match the corresponding deserialization ratio n based on the data rate and the processing requirements to select the deserialization path, where the deserialization ratio includes 1:4, 1:8, 1:16;

[0020] Deserialize the high-speed serial signal into several paths of the parallel data streams based on the deserialization path.

[0021] In some embodiments, after deserializing the high-speed serial signal into several paths of the parallel data streams based on the deserialization path, the following steps are further included:

[0022] Obtain the number of microphone channels x2, and calculate the number of multiplexed signals m required to be configured for each path of the parallel data streams in combination with the deserialization ratio n, where the number of multiplexed signals is inversely proportional to the reciprocal of the deserialization ratio;

[0023] Obtain x2 paths of 1bit pdm signals, double the number of the 1bit pdm signals according to the left and right channels to obtain the sampling processing amount x3;

[0024] Pass the x3 paths of the 1bit pdm signals through a cic filter to convert them into multi-bit pcm signals.

[0025] In some of these embodiments, a parallel beamforming array is invoked to configure a number of processing units and distribute a number of paths of the parallel data streams to each of the processing units, and acoustic environment information is obtained and a corresponding target frequency point k and a frequency point requirement number x1 are assigned to each of the processing units, including the following steps:

[0026] Obtain the number x4 of the processing units included in the parallel beamforming array, and calculate the number of paths x5 that each of the processing units needs to process based on x3 / x4 to distribute a number of multi-bit pcm signals to each of the processing units;

[0027] Determine the environment type based on the acoustic environment information, and determine the target frequency range according to the environment type;

[0028] Determine the environment complexity based on the acoustic environment information, and determine the frequency point requirement number x1 according to the environment complexity;

[0029] In each of the processing units, select a corresponding number of frequency feature points in the target frequency range corresponding to each of the multi-bit pcm signals as the target frequency point k based on the frequency point requirement number x1.

[0030] In some of these embodiments, perform DFT operations on each path of the parallel data streams based on the target frequency point k and the frequency point requirement number x to respectively obtain corresponding covariance matrices, including the following steps:

[0031] Analyze the frequency domain distribution of the multi-bit pcm signal to generate a first coefficient, obtain the imaging accuracy requirement to generate a second coefficient, and obtain the signal quality to generate a third coefficient;

[0032] Dynamically configure the corresponding initial bit width based on the first coefficient, the second coefficient, and the third coefficient;

[0033] Select the corresponding target frequency point k based on the initial bit width to perform DFT operations to obtain the frequency domain data of each path of the multi-bit pcm signal;

[0034] Generate a frequency domain vector based on the frequency domain data of a number of the multi-bit pcm signals, and calculate the upper triangular part covariance of each of the multi-bit pcm signals through conjugate eigenvectors according to the frequency domain vector to generate a covariance matrix.

[0035] In some of these embodiments, obtain the steering vector matrix corresponding to the parallel beamforming array and combine a number of the covariance matrices to generate a parallel beam, including the following steps:

[0036] Determine the steering vectors corresponding to each element in the parallel beamforming array to form the steering vector matrix;

[0037] Reduce the dimension of the covariance matrix to obtain a number of sub - matrices, associate each of the array elements with each of the multi - bit pcm signals for correlation mapping, and multiply each of the sub - matrices of the steering vector matrix respectively to obtain a number of sub - results. Among them, when each of the sub - matrices is multiplied by the steering vector matrix, the same ip core is reused;

[0038] Combine the several sub - results in the form of a block matrix into the complete parallel beam.

[0039] In some embodiments, obtaining the steering vector matrix corresponding to the parallel beamforming array and combining several covariance matrices to generate a parallel beam further includes the following steps:

[0040] Generate a number of neuron clusters of the same size based on the target frequency range, and each neuron cluster is configured with a processing strategy including synaptic weights and spike - time coding;

[0041] Obtain the corresponding frequency based on the target frequency point k, map the frequency to the corresponding neuron cluster to obtain the corresponding processing strategy, and adjust the output of the parallel beam. Specifically;

[0042] Generate the firing probability of the pulse signal based on the synaptic weights in the processing strategy for random weighting. The pulse signal includes 0 or 1. When the pulse signal is 0, the covariance corresponding to the target frequency point k does not participate in the calculation of the covariance matrix;

[0043] Adjust the phase to supplement the delay time based on the spike - time coding in the processing strategy.

[0044] In some embodiments, obtaining the phase compensation coefficient and converting it into fir filter parameters to correct the phase error of the parallel beam further includes the following steps:

[0045] Obtain a phase compensation table containing the results of several tested array elements. The phase compensation table contains the phase deviation amounts of different array elements relative to the reference array element;

[0046] Determine the phase compensation coefficients corresponding to each array element based on the target frequency point k corresponding to each array element and the number of frequency point requirements x1 in the phase compensation table, and integrate them into the target frequency - domain response;

[0047] Perform an inverse Fourier transform on the target frequency - domain response and calculate the fir filter parameters in combination with the filter order;

[0048] Perform a convolution operation on the steering vector corresponding to each array element based on the fir filter parameters and output to complete the phase error correction.

[0049] In some of these embodiments, after correcting the phase error of the parallel beams, serial output is generated through parallel-to-serial processing, and the following steps are further included:

[0050] Set a number of ping-pong buffers, and configure the read-write order for each path based on the number of the parallel data streams;

[0051] Based on the read-write order, write the array elements corresponding to each of the parallel data streams into the ping-pong buffer in each frame for phase error correction of the steering vector;

[0052] When the depth corresponding to one of the ping-pong buffers is fully written, read and output the corresponding parallel data stream, and replace the ping-pong buffer to continue writing the array elements corresponding to each of the parallel data streams for phase error correction of the steering vector;

[0053] Stitch together a number of continuously read parallel data streams into a serial data stream to generate a serial output.

[0054] In a second aspect, a low-power composite multi-channel acoustic imaging system provided by the present application adopts the following technical solution:

[0055] A low-power composite multi-channel acoustic imaging system is used to implement the above method.

[0056] Based on the above technical solution, at least the following technical effects exist:

[0057] Reconstruct the acoustic signal processing flow through a three-stage pipeline of a serial-to-parallel conversion architecture, optimize the calculation bottleneck with matrix features, select different numbers of parallel channels and processing precisions in different signal features of sampled acoustic signals and different acoustic environments, optimize the high full-band calculation load brought by traditional FFT operations through DFT operations, and only select target frequency points for individual operations according to the corresponding environment and features, further reducing the load power during beamforming, and achieving real-time high-precision acoustic positioning with micro-watt-level power consumption. Description of the Drawings

[0058] Figure 1 It is a step schematic diagram of a low-power composite multi-channel acoustic imaging method provided in this embodiment. Detailed Embodiments

[0059] To more clearly understand the purpose, technical solution and advantages of the present application, the present application will be described and explained below with reference to the accompanying drawings and embodiments. However, those of ordinary skill in the art should understand that the present application can be implemented without these details. In some cases, well-known methods, processes, systems, components and / or circuits that have been described at a higher level are not described in detail to avoid obscuring aspects of the present application with unnecessary descriptions. For those of ordinary skill in the art, it is obvious that various changes can be made to the disclosed embodiments of the present application, and without departing from the principles and scope of the present application, the general principles defined in the present application can be applied to other embodiments and application scenarios. Therefore, the present application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope claimed in the present application.

[0060] It should be noted here that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation to the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0061] In the description of the present application, the meaning of "a number of" is one or more, the meaning of "a plurality of" is two or more, "greater than", "less than", "exceeding", etc. are understood as not including the number itself, and "above", "below", "within", etc. are understood as including the number itself. If there is a description of "first" and "second", it is only used to distinguish technical features and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence relationship of the indicated technical features.

[0062] In the description of the present application, the description with reference to terms such as "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a combined manner.

[0063] As Figure 1 shown, the embodiments of the present application disclose a low-power composite multi-channel acoustic imaging method, including the following steps:

[0064] S100, obtain the sampled acoustic signal and send it to the configured three-level pipeline structure including the front end, the middle end and the back end.

[0065] The embodiment of this application constructs a three - level pipeline architecture of "serial input - parallel processing - serial output". After acoustic information is collected by several microphones in the microphone matrix on the acoustic acquisition device, the serial data corresponding to the acoustic signal is sent to the three - level pipeline architecture for corresponding subsequent operations.

[0066] At the front - end:

[0067] S200, obtain the signal characteristics of the sampled acoustic signal to select the corresponding deserialization path, and perform serial - to - parallel processing on the sampled acoustic signal based on the deserialization path to obtain several parallel data streams.

[0068] First, according to the signal characteristics of the sampled acoustic signal, such as the data rate, user requirements, etc., obtain its corresponding acoustic environment, and select different deserialization paths according to the acoustic environment. Different deserialization paths correspond to different deserialization ratios, that is, different numbers of channels.

[0069] Based on the selected deserialization path, convert the serial data into several parallel data streams for subsequent processing.

[0070] At the middle - end:

[0071] S300, call the parallel beamforming array to configure several processing units and distribute several parallel data streams to each processing unit, obtain the acoustic environment information, and allocate the corresponding target frequency point k and the number of frequency point requirements x1 to each processing unit.

[0072] Send the parallel data stream to the parallel beamforming matrix (PEA). There are several processing units (PE) configured in the parallel beamforming matrix. Each processing unit integrates several - bit fixed - point operation cores, and the fixed - point operation cores are used to perform corresponding operation processing on each parallel data stream.

[0073] At the same time, in order to avoid the high computing power burden caused by performing operations on all frequency band feature points in each data stream, based on the acoustic data analysis of the sampled acoustic environment, select the frequency band points with higher correlation in the corresponding environment for subsequent processing according to different acoustic environments, and do not perform operations on other frequency bands. At the same time, select the processing precision based on the complexity and noise of the acoustic environment. Different processing precisions correspond to different numbers of selected frequency points.

[0074] S400, perform DFT operations on each parallel data stream based on the target frequency point k and the number of frequency point requirements x to obtain the corresponding covariance matrix respectively.

[0075] Perform DFT processing (discrete Fourier transform) on the target frequency point k with the corresponding number of frequency point requirements x. After the DFT processing, integrate the calculation results of several parallel data streams to form a covariance matrix.

[0076] S500, obtain the steering vector matrix corresponding to the parallel beamforming array and combine a number of covariance matrices to generate parallel beams.

[0077] The steering vector matrix is a mathematical matrix composed of the steering vectors between a number of parallel data streams, and the steering vector is characterized as a vector representing the direction between two points in space.

[0078] Multiplying the covariance matrix and the steering vector matrix can obtain the positioning result corresponding to beamforming.

[0079] At the backend:

[0080] S600, obtain the phase compensation coefficient, and convert it into the corresponding FIR filter parameters based on the phase compensation coefficient to correct the phase error of the parallel beam, and then generate a serial output through parallel-to-serial processing.

[0081] Since there are inevitably errors in the steering vector, which will lead to the distortion of the positioning result, it is necessary to perform corresponding phase compensation and correction on the steering vector before calculating the steering vector matrix and the covariance matrix to accurately position the result after beamforming.

[0082] When performing phase compensation, first obtain the phase compensation coefficient stored during parameter debugging in the laboratory in advance, and convert this phase compensation coefficient into the corresponding filtering parameters of the FIR filter in the FPGA, and use the FIR filter to perform accurate time delay compensation on the multi-channel signal.

[0083] Finally, calculate the steering vector after phase compensation and the covariance matrix to obtain the corrected parallel beam. At the same time, convert a number of parallel data streams into serial data through ping-pong buffering for output.

[0084] Through the above steps, reconstruct the acoustic signal processing flow through a three-stage pipeline of the serial-parallel conversion architecture, optimize the computing bottleneck through matrix features, select different parallel channel numbers and processing precisions in different signal features of sampled acoustic signals and different acoustic environments, optimize the high full-band computing load brought by traditional FFT operations through DFT operations, only select the target frequency points for separate operations according to the corresponding environment and features, further reduce the load power during beamforming, and achieve real-time high-precision acoustic positioning with micro-watt-level power consumption.

[0085] In some other embodiments, obtaining the signal features of the sampled acoustic signal to select the corresponding deserialization path, and performing serial-to-parallel processing on the sampled acoustic signal based on the deserialization path to obtain a number of parallel data streams, includes the following steps:

[0086] S210, obtain the high-speed serial signal corresponding to the sampled acoustic signal.

[0087] S220. Collect the data rate and processing requirements corresponding to the signal characteristics, and match the corresponding deserialization ratio n based on the data rate and processing requirements to select the deserialization path. The deserialization ratios include 1:4, 1:8, and 1:16.

[0088] Based on the ISERDES technology, first obtain the data rate and processing requirements corresponding to the signal characteristics, and deserialize the serial high-speed signal into several parallel low-speed signals according to the corresponding signal characteristics.

[0089] In the embodiments of the present application, variable magnification conversion of 1:4, 1:8, and 1:16 is supported. First, obtain the data rate. Different data rates correspond to different usage scenarios. For example, when the data frequency is 2.5 Gbps, it represents a standard high-precision imaging environment, and when the data frequency suddenly changes to 5 Gbps, it represents a high-speed motion mode.

[0090] When the data frequency is 2.5 Gbps, the deserialization ratio n can be set to 1:16 to convert the single-channel frequency to 156.25 MHz and improve the phase calculation accuracy. When the data frequency is 5 Gbps, a deserialization ratio of 1:8 can be used to avoid motion blur.

[0091] At the same time, obtain the processing requirements. The processing requirements are characterized on the one hand by the processing capabilities (bit width, processing frequency) that the backend module can support during parallel processing, and on the other hand, the overall processing energy consumption is considered. For example, in the energy-saving mode, in the face of some situations with relatively low data rates, the deserialization ratio can be appropriately reduced to reduce the number of processing channels and thus reduce the energy consumption; while when some high-speed data needs to be processed and the processing capabilities of the system are excessive, the deserialization ratio with a higher number of channels such as 1:16 can be switched to improve the parallel processing ability, thereby improving the flexibility and adaptability of the system.

[0092] Dynamically select the optimal deserialization path according to the data rate and processing requirements. First, ensure that the multiple parallel data stream channels after deserialization can ensure normal parallel processing, and at the same time, select the deserialization ratio with the lowest power consumption on this basis.

[0093] S230. Deserialize the high-speed serial signal into several parallel data streams based on the deserialization path.

[0094] Deserialize the high-speed serial signal into several parallel data according to the ratio of the corresponding deserialization path. For example, at a deserialization ratio of 1:16, the 2.5 Gbps signal is converted into 16 parallel data streams, and the single-channel processing frequency is reduced to 156 MHz, reducing the clock power by 40%.

[0095] In some other embodiments, after deserializing the high-speed serial signal into several parallel data streams based on the deserialization path, the following steps are further included:

[0096] S240. Obtain the number of microphone channels x2, and calculate the number of multiplexed signals m required to be configured for each parallel data stream in combination with the deserialization ratio n. Here, the number of multiplexed signals is inversely proportional to the reciprocal of the deserialization ratio.

[0097] The number of microphone channels is characterized as the number of channels of the acoustic signals collected by the microphone matrix on the corresponding acoustic device. Depending on the microphone matrix, the number of microphone channels x2 may be 8, 16, 32, etc.

[0098] Generate the number of multiplexed signals m based on the selected different deserialization ratios n. After deserialization, each parallel path can multiplex and carry multiple low-speed signals. The number of multiplexed signals m for each channel is determined by the number of paths supported by the system and required by the user.

[0099] For example, if during subsequent DFT operations, in order to maintain accuracy, 64 parallel data are required. At this time, if the deserialization ratio is 1:16, then the number of multiplexed signals m corresponding to each parallel path can be set to 4. In this way, the number of physical parallel channels of 15 can be converted into 64 PDM signals to fully utilize the parallel bandwidth of the ISERDES.

[0100] S250. Obtain 1-bit PDM signals with the number of x2 channels, and double the number of 1-bit PDM signals according to the left and right channels to obtain the sampling processing quantity x3.

[0101] Taking 64 channels as an example, when 64 low-speed 1-bit PDM signals (pulse density modulation) are obtained, they correspond to 64 single-vibrator mono channels. However, in order to improve the spatial positioning accuracy of the acoustic signals, it is also necessary to distinguish between the left and right sound fields during the processing of the acoustic signals. Therefore, it is necessary to generate two signals for the left and right channels by interpolating or time-division multiplexing the 64 PDM signals, that is, 64 PDM signals for each of the left and right channels. In this way, the final sampling processing quantity x3 is 128 channels.

[0102] S260. Pass the 1-bit PDM signals with the number of x3 channels through a CIC filter to convert them into multi-bit PCM signals.

[0103] Finally, in order to facilitate subsequent beamforming calculations, it is also necessary to convert the PDM signal (pulse density modulation) into a PCM signal (pulse code modulation). This is because PDM is a 1-bit digital signal that represents the amplitude of an analog signal through the density (duty cycle) of high-frequency pulses. For example, the higher the amplitude of the analog signal, the higher the proportion of "1" and the lower the proportion of "0" in the PDM signal. When performing beamforming, multi-bit amplitude information is required, and PCM is beneficial for simplifying processing and signal processing. This enables signal processing operations to be completed on the audio stream, such as mixing, filtering, and equalization. Therefore, it is necessary to convert the high-frequency 1-bit PDM signal into a PCM signal with a low sampling rate and high quantization accuracy.

[0104] The specific conversion method is through a CIC filter. Specifically:

[0105] First, the continuous PDM signal is accumulated (integrated) through the first-stage integrator of the CIC filter. N cycles of the PDM signal (1 bit, 1 or 0 per cycle) are input, and after integration, a multi-bit value (such as N bits) is output, representing the total number of "1" within N cycles, thus obtaining the corresponding multi-bit signal.

[0106] Secondly, the high-frequency noise is filtered through the second-stage comb filter of the CIC filter, and at the same time, the sampling rate is reduced through the decimation step. When the integrated multi-bit signal passes through filtering and decimation, a PCM signal with a fixed bit width is finally output. In this application, the PCM signal is 16 bits.

[0107] At the same time, this application uses a CIC filter to achieve a multiplier-free design, which is suitable for multi-channel parallel processing, meets the high-efficiency requirements of the microphone array system, and suppresses quantization noise based on the oversampling and filtering processes, making the signal-to-noise ratio of the 16-bit PCM signal much lower than that of direct 1-bit quantization, laying a foundation for subsequent high-precision positioning algorithms.

[0108] In some other embodiments, a parallel beamforming array is called to configure a number of processing units and distribute a number of parallel data streams to each processing unit, obtain acoustic environment information, and allocate corresponding target frequency points k and frequency point requirements x1 to each processing unit, including the following steps:

[0109] S310, obtain the number x4 of processing units included in the parallel beamforming array, and calculate the number x5 of paths that each processing unit needs to process based on x3 / x4 to distribute a number of multi-bit PCM signals to each processing unit.

[0110] Different parallel beamforming matrix configurations have different numbers of processing units. First, obtain the current number of processing units x4, and divide the number of processing units by the number of paths of the pcm signal to calculate the number of paths that each processing unit needs to process.

[0111] For example, if there are 32 processing units in the parallel beamforming matrix in this application, then when there are 128 paths of pcm signals, each processing unit needs to process 4 paths of pcm signals.

[0112] In this way, the 128 paths of signals are evenly distributed to each processing unit, and the fixed-point operation cores configured in each processing unit perform operation processing on the corresponding pcm paths.

[0113] S320, determine the environmental type based on the acoustic environment information, and determine the target frequency range according to the environmental type.

[0114] In this application, the DFT algorithm is used for processing because the FFT is a butterfly operation, which calculates the entire frequency band of 1024 in one acoustic signal path at a time, while the DFT algorithm can only calculate a certain characteristic frequency band. This can greatly save the calculation amount for 128 channels. At the same time, the DFT does not require caching, and all results are just calculated when the data reading ends, saving resources.

[0115] Therefore, in order to determine which frequency band to select for operation processing in different scenarios, it is necessary to determine the environmental type based on the acoustic environment information and select different target frequency ranges according to the corresponding environmental type.

[0116] For example, in a noisy environment, the noise frequency band is generally between 50 - 200 Hz. The voice frequency band in the office area is generally between 300 - 3400 Hz. In a shopping mall, there may be human voices between 300 - 3400 Hz, music between 5 - 10 kHz, and air conditioner noise between 100 - 200 Hz, etc. Then, different target frequency ranges are selected according to different acoustic environments.

[0117] S330, determine the environmental complexity based on the acoustic environment information, and determine the number of frequency points required x1 according to the environmental complexity.

[0118] Obtain the environmental complexity in the acoustic environment. The complexity mainly depends on whether the frequency range corresponding to various noises is included in the sound frequency range, such as construction, vehicle flow, industrial lathes, etc., and determine how many different frequencies of sounds exist in the parallel data stream, that is, determine whether it is a multi-source sound scene.

[0119] In scenarios with a high environmental complexity, to ensure the accuracy and precision of the subsequent beamforming calculations, it is necessary to collect more frequency points for subsequent operations. Conversely, when the environmental complexity is low, fewer frequency points can be collected to reduce the computational pressure on the system.

[0120] S340. In each processing unit, based on the number of required frequency points x1, select the corresponding number of frequency feature points within the target frequency range corresponding to each multi-bit pcm signal as the target frequency points k.

[0121] Based on the number of required frequency points, select the corresponding number of frequency feature points within the selected target frequency range as the target frequency points k. In this way, when performing DFT operations on each acoustic channel subsequently, only the several target frequency points k collected on each sound data stream need to be operated on, which can skip low-frequency noise and high-frequency irrelevant frequency bands and save computing power as much as possible.

[0122] At the same time, the size of the target frequency range can be further adjusted dynamically. Calculate the entropy value of each current pcm signal. When the entropy value is high due to multiple sound sources in the environment, dynamically expand the range of the target frequency range. When there is only a single sound source in the environment and the entropy value is low, shrink the size of the target frequency range.

[0123] In some other embodiments, before performing the DFT operation, spectrum monitoring is also required to further optimize the specific parameters of the target frequency points k selected by each processing unit. Specifically:

[0124] In a multi-sound-source environment, different calculation frequency ranges are independently configured for each processing unit. For example, processing units 1-8 calculate the speech frequency band (300-3400 Hz), and processing units 9-16 calculate the high-frequency noise frequency band, etc. In this way, when several pcm signals are sent to each processing unit, based on the target frequency range of each pcm signal, the corresponding target frequency points are sent to the corresponding processing unit for operation.

[0125] In some other embodiments, based on the target frequency points k and the number of required frequency points x, perform DFT operations on each parallel data stream to respectively obtain the corresponding covariance matrices, including the following steps:

[0126] S410. Analyze the frequency-domain distribution of the multi-bit pcm signal to generate a first coefficient, obtain the imaging accuracy requirement to generate a second coefficient, and obtain the signal quality to generate a third coefficient.

[0127] S420. Dynamically configure the corresponding initial bit width based on the first coefficient, the second coefficient, and the third coefficient.

[0128] The bit width is characterized by the amount of data that can be transmitted at one time. The larger the bit width, the greater the amount of data that can be transmitted per unit time, which directly determines the data processing efficiency.

[0129] Before performing the DFT operation, the bit width must be dynamically adjusted based on the specific parameter details of the signal and the requirements of imaging accuracy. This application supports dynamic bit width adjustment of 8 / 12 / 16.

[0130] First, analyze the frequency domain distribution of each PCM signal. First, perform a short-time Fourier transform on each PCM signal to obtain the energy distribution of each frequency point. If the energy proportion of the high-frequency band is greater than that of the low-frequency band, it means that the signal is rich in high-frequency details. For such high-frequency signals, the quantization error has a greater impact on the phase accuracy. At this time, it is necessary to reduce the phase error as much as possible by increasing the bit width to avoid main lobe splitting. The higher the first coefficient, the more the signal tends to be high-frequency, so it is more inclined to choose a higher initial bit width.

[0131] At the same time, the imaging accuracy set by the user is obtained. In high-precision scenarios, the maximum quantization error must be less than 1 degree. The accuracy requirement formula based on the maximum quantization error is:

[0132] .

[0133] in, is the phase quantization error, d represents the array element spacing. If the maximum quantization error is less than 1 degree, the phase quantization error needs to be less than 0.1 , the minimum bit width corresponds to 12 bits. (8 bits allow for an error > 5°, 12 bits allow for an error 1°~5°, and 16 bits allow for an error < 1°).

[0134] Then, when the second coefficient is higher, it means that the required precision is greater and the required initial bit width is larger.

[0135] Finally, obtain the signal quality corresponding to the PCM signal, that is, the signal-to-noise ratio. In low signal-to-noise ratio scenarios, quantization noise may drown out signal details. At this time, the bit width needs to be increased to reduce the quantization noise power. The bit width and signal-to-noise ratio satisfy the signal-to-noise ratio = 6.02N + 1.76 dB (N is the bit width). 8 bits correspond to a 49.9 dB signal-to-noise ratio, which is sufficient to process high signal-to-noise ratio signals; low signal-to-noise ratio signals require 16 bits (97.9 dB) to prevent quantization noise from becoming the main error source.

[0136] Then, when the signal-to-noise ratio is smaller, the third coefficient is larger, and a higher initial bit width is more likely to be selected.

[0137] Among them, during the subsequent DFT operation, the initial bit width can also be dynamically adjusted according to different operation steps. For example, when performing operations such as filtering the signal in the preprocessing stage, due to the low precision requirement, an 8-bit width can be selected to reduce the calculation delay and power consumption. When performing the DFT matrix operation, a 12-bit width can be selected to balance the precision and efficiency. Finally, when performing the final beamforming or weight calculation that is sensitive to precision, a 16-bit width can be selected to ensure the pointing precision of the beam.

[0138] S430, based on the initial bit width, select the corresponding target frequency point k to perform the DFT operation to obtain the frequency domain data of each multi-bit PCM signal.

[0139] Perform the DFT operation on the target frequency point k in each PCM signal path according to the selected initial bit width to obtain the frequency domain data corresponding to each PCM signal.

[0140] S440, generate a frequency domain vector based on the frequency domain data of several multi-bit PCM signals, and calculate the upper triangular part covariance of each multi-bit PCM signal through conjugate eigenvectors according to the frequency domain vector to generate a covariance matrix.

[0141] When the frequency data of all PCM signals are calculated, generate the corresponding frequency domain vectors to form a covariance matrix. At the same time, since the covariance matrix is conjugate symmetric, that is, the results of the upper triangular matrix and the lower triangular matrix are conjugate symmetric. Therefore, in order to reduce the calculation amount of the system, the present application directly calculates the covariance matrix for the upper triangular part, and then directly obtains the lower triangular values through conjugate transpose, so that the calculation amount is reduced by 50%.

[0142] In some other embodiments, obtaining the steering vector matrix corresponding to the parallel beamforming array and combining several covariance matrices to generate parallel beams includes the following steps:

[0143] S510, determine the steering vectors corresponding to each element in the parallel beamforming array to form a steering vector matrix.

[0144] First, when performing beamforming, it is necessary to multiply the covariance of each acoustic signal by the steering vector of the element where the signal is located. Therefore, it is necessary to determine the steering vectors corresponding to each element of the parallel beamforming array. Specifically:

[0145] When all elements in the microphone matrix sample simultaneously, it is regarded as a snapshot. When there is a signal from the direction incident on the array, the received signal of each snapshot of the array is (the subscript represents the dimension):

[0146] .

[0147] Among them, characterized as a steering matrix, where M and N represent the row dimension and column dimension respectively, M is the number of array elements, and N is the number of incident signals.

[0148] S520, perform dimensionality reduction on the covariance matrix to obtain several sub-matrices, and perform associated mapping between each array element and each multi-bit PCM signal, and then multiply each sub-matrix of the steering vector matrix respectively to obtain several sub-results. Among them, the same IP core is reused when each sub-matrix is multiplied by the steering vector matrix.

[0149] After calculating the steering vector matrix, in order to further reduce the computational complexity of the system, the present application performs dimensionality reduction on the complete covariance matrix to obtain several sub-matrices. For example, a 128*128 covariance matrix is reduced to several 32*32 matrices, so as to decompose large matrix operations into several small matrix operations, and only one IP core of a 32*32 matrix needs to be reused to perform operations on all sub-matrices, greatly saving computational complexity.

[0150] Specifically, first, in the FPGA development environment, configure the 32×32 matrix operation IP core. This includes setting parameters such as the data bit width of its input and output ports (for example, matching the 16-bit fixed-point operation core of the PE unit), operation mode (such as matrix multiplication, conjugate transpose, etc.), and clock frequency, to ensure that it can work properly and meet the performance requirements of the system.

[0151] Then connect the input and output ports of the IP core to the tree-shaped data path of the PBA system. Specifically, the grouped 32-channel signal data (from 128-channel signal grouping) needs to be correctly transmitted to the input port of the IP core, and at the same time, the 32×32 sub-matrix result calculated by the IP core is output to the subsequent data processing module or storage unit.

[0152] First, calculate the covariance matrix within the group. Taking four sub-matrices within the group as an example, they are respectively , , , . First, calculate and the result of the steering vector matrix, obtain the corresponding data from the cache, and transmit it to the two input ports of the IP core at the same time. The IP core first calculates according to the configured operation mode. After the calculation is completed, is stored in the specified storage unit.

[0153] Calculate , , in turn according to the same method.

[0154] Secondly, perform the inter-group covariance matrix operation. The inter-group covariance sub-matrix includes , , , , , , first calculate The result of the steering vector matrix reads the two covariances x1 and x2 respectively and transmits them to the two input ports of the IP core respectively to calculate And stored in the specified storage unit.

[0155] Using the same method, calculate , , , , .

[0156] S530: Combining a plurality of sub-results into a complete parallel beam in the form of a block matrix.

[0157] After all the sub-results of the sub-matrices are calculated, several sub-results are combined in the form of block matrices to form parallel beams corresponding to the complete 128*128 covariance matrix.

[0158] In some other embodiments, obtaining a steering vector matrix corresponding to a parallel beamforming array and combining a plurality of covariance matrices to generate parallel beams further includes the following steps:

[0159] S540, generating a plurality of neuron clusters of the same size based on the target frequency range, each neuron cluster being configured with a processing strategy including synaptic weights and pulse time coding.

[0160] First, frequency clustering is performed. Several different frequency bands are divided based on the frequency range of the parallel data stream. Each frequency band corresponds to a neuron cluster. Each neuron cluster corresponds to one or more processing units and each neuron cluster corresponds to multiple sensors. At the same time, an IIR or FIR filter is configured at each processing unit to filter the sensor signal to the corresponding frequency band in real time. In this way, the filter of each frequency band only allows the signal within the frequency band to pass, while the sound signals in other frequency bands will be suppressed.

[0161] At the same time, each neuron cluster also corresponds to synaptic weights and pulse time codes. The synaptic weights correspond to the probability of the event that the sensor corresponding to each neuron cluster is stimulated by the pulse signal, while the pulse time code is characterized by the time difference of the pulse arriving at the neuron cluster. The time when the pulse triggers the membrane potential to rise will affect the subsequent phase delay compensation.

[0162] S550, obtain the corresponding frequency based on the target frequency point k, map the frequency to the corresponding neuron cluster to obtain the corresponding processing strategy, and adjust the output of the parallel beam.

[0163] First, configure each channel in the neuron cluster corresponding to the frequency range based on the frequency of the target frequency point corresponding to each channel, and adjust the output of the parallel beam based on the independent processing strategies corresponding to different frequency bands of each neuron cluster. The direction-of-arrival estimation accuracy of signals with different frequencies is different. Low-frequency signals (such as 20 Hz, wavelength 17 m) are difficult to distinguish narrow angles, and high-frequency signals (such as 20 kHz, wavelength 1.7 cm) are sensitive to phase errors.

[0164] For example, there are 16 sensors receiving signals as , the target direction is , then the beam output is:

[0165] .

[0166] is the synaptic weight, is the phase compensation delay (the time difference corresponding to the direction of arrival).

[0167] Then in the embodiment of the present application, the core goal of the neuron cluster is to make the signals in the target direction in-phase superposition (maximize the gain) and cancel the signals in the non-target direction (suppress noise) by adjusting and .

[0168] Each frequency band calculates and independently. The low-frequency band focuses on suppressing noise, and the high-frequency band focuses on precise phase compensation, forming a "frequency-direction" adaptive beam.

[0169] Specifically:

[0170] S560, generate the firing probability of the pulse signal based on the synaptic weight in the processing strategy for random weighting. The pulse signal includes 0 or 1. Among them, when the pulse signal is 0, the covariance corresponding to the target frequency point k does not participate in the calculation of the covariance matrix.

[0171] The neuron cluster outputs a pulse signal of 0 or 1, and controls the pulse firing probability through the synaptic weight . For example, if = 0.8, it means that the probability of outputting 1 is 80% and the probability of outputting 0 is 20%, so as to achieve random weighting in turn. When the neuron does not fire a pulse (output 0), the corresponding data does not participate in the covariance calculation, which is equivalent to dynamically eliminating noise samples and improving the estimation robustness.

[0172] Specifically, the current acoustic scenario is to monitor the periodic impact noise generated by gearbox faults, whose main frequency is in the range of 5 kHz - 7 kHz, while the background noise is click power frequency noise (50 Hz - 1 kHz).

[0173] First, configure its corresponding processing unit in neuron cluster x, whose corresponding frequency range is 5 kHz - 7.5 kHz. Only when the amplitude of the sound signal exceeds the preset value, the corresponding neuron cluster fires pulses with a probability corresponding to the synaptic weight (such as a sharp increase in pulses at the moment of gear impact), and it remains dormant at other times.

[0174] S570, adjust the phase to supplement the delay time based on the pulse time encoding in the processing strategy.

[0175] Using pulse time encoding (PTM), convert the delay into the time difference of pulse arrival, and naturally compensate it during the integration of the neuron membrane potential. For example, for a signal with a delay of 5 ns, the pulse triggers a later rise in the membrane potential.

[0176] Through the above method, activation calculation is only performed when the signal is valid, transforming the acoustic signal processing from "full-cycle synchronous calculation" to "on-demand activated pulse processing", achieving high energy efficiency + high robustness beamforming in a low duty cycle scenario, replacing complex floating-point operations with binary processing of pulse neurons, and implementing adaptive beamforming in FPGA with extremely low resource consumption (<10% logic units), taking both performance and cost into account.

[0177] Due to the inevitable error in the steering vector, which will lead to the distortion of the positioning result, it is necessary to perform phase compensation on the steering vector. Convert the phase compensation parameters of the steering vector obtained in the laboratory into FIR filter parameters in the FPGA, and use the FIR filter to accurately compensate the time delay of the multi-channel signal.

[0178] Specifically,

[0179] In some other embodiments, obtain the phase compensation coefficient, and convert it into FIR filter parameters based on the phase compensation coefficient to correct the phase error of the parallel beam, and further include the following steps:

[0180] S610, obtain a phase compensation table containing the results of several tested array elements, where the phase compensation table contains the phase deviation amounts of different array elements relative to the reference array element.

[0181] What is estimated through calibration or off-line measurement in the laboratory are the phase compensation parameters of each channel (such as the target phase difference), which need to be converted into time delay parameters that can be implemented in digital signal processing hardware (such as FPGA), and then design an FIR filter to complete the delay compensation.

[0182] Integrate a number of phase compensation parameters obtained in the laboratory into a phase compensation table, which contains the phase deviation amounts of each array element relative to the reference array element.

[0183] Ideally, the phase deviation should satisfy:

[0184] 。

[0185] Represents the time delay coefficient of the i-th array element relative to the reference array element. If there is an error between the actually measured phase deviation and the theoretical value ,then phase compensation needs to be performed through a FIR filter to restore the total phase deviation to 。

[0186] S620. Based on the target frequency point k corresponding to each array element and the number of frequency point requirements x1, determine the phase compensation coefficients corresponding to each array element in the phase compensation table and integrate them into the target frequency domain response.

[0187] Based on the frequency range where each target frequency point is located and the selected number of frequency points, determine the phase compensation coefficients in the phase compensation table. Specifically, input the frequency f into the above formula. After determining the target ,the time delay coefficient can be calculated ,and based on the linear phase characteristic of the constant group delay of the FIR filter, realize the conversion between the time delay and the phase characteristic to determine the phase compensation coefficients.

[0188] After obtaining the phase compensation coefficients, further calculate the target frequency domain response based on the impulse response of the FIR filter characteristic combined with the phase compensation coefficients.

[0189] S630. Perform an inverse Fourier transform on the target frequency domain response and calculate the FIR filter parameters in combination with the filter order.

[0190] Furthermore, in order to further perform the conversion of the time domain coefficients of the target frequency domain response and realize the adjustment of the phase compensation delay through the pulse time coding of the neuron cluster to achieve phase correction, it is also necessary to perform an inverse Fourier transform on the target frequency domain response and combine it with the corresponding order of the FIR filter to obtain the time domain impulse response. Specifically:

[0191] df.

[0192] Represents the filter parameters of the FIR filter, Represents the target frequency domain response, Represents the frequency, Represents the phase compensation factor, The kernel function characterized as the inverse Fourier transform (establishing the mapping relationship between the frequency domain and the time domain), and df is characterized as the integration variable.

[0193] S640, perform a convolution operation on the steering vectors corresponding to each array element based on the fir filter parameters and output to complete the phase error correction.

[0194] Configure the filter with the fir filter parameters. When there are phase differences among multiple actually measured signals, adjust the corresponding phase delay time through the filter, so that the leading channel signal enters a positive delay to wait for a certain time. At the same time, reduce the delay of the lagging channel signal. Precisely adjust the time difference between the channel signals.

[0195] In some other embodiments, after performing phase error correction on the parallel beams, perform a parallel-to-serial conversion process to generate a serial output, and further include the following steps:

[0196] S650, set several ping-pong buffers and configure the read and write order for each path based on the number of parallel data streams.

[0197] In order to finally output the data processed in parallel in a serial manner, the present application is implemented through the dynamic correction method of the ping-pong buffer, using the ping-pong buffer to achieve pipelined data buffering and calibration to ensure real-time performance and accuracy.

[0198] First, set several ping-pong buffers. Each ping-pong buffer can write data and read out the corresponding data after the space in the buffer is full.

[0199] At the same time, determine the read and write order of each path based on the number of parallel data streams. The more parallel data streams there are, the earlier the path with a smaller data processing volume performs read and write operations.

[0200] S660, based on the read and write order, write the array elements corresponding to each parallel data stream into the ping-pong buffer in each frame to perform phase error correction of the steering vector.

[0201] S670, when the depth corresponding to a ping-pong buffer is fully written, read and output the corresponding parallel data stream, and replace the ping-pong buffer to continue writing the array elements corresponding to each parallel data stream to perform phase error correction of the steering vector.

[0202] S680, splice the continuously read parallel data streams into a serial data stream to generate a serial output.

[0203] The depths of multiple ping-pong buffers are the same, which are used to alternately store parallel data. At the same time, through the state machine, realize the ping-pong alternating operation of reading data from area b when writing to area a and reading data from area a when writing to area b, to avoid data waiting.

[0204] First, based on the read-write order, the array elements corresponding to the parallel data stream are written into one of the ping-pong buffers for phase error correction. When the buffer is full, the writing switches to the next buffer, and at the same time, the data after phase error correction in the previous buffer is read. This ensures that the data in each buffer is calibrated before the other buffers are written, avoiding timing misalignment caused by lagging correction.

[0205] One frame of data is read from the buffer in each clock cycle and split into serial bitstreams according to the channel order. The dynamic calibration scheme based on ping-pong buffers realizes real-time calibration of multi-channel parallel beams and efficient data format conversion through the "calibration-buffering-serialization" architecture. Its core lies in using the ping-pong mechanism to decouple data writing and reading, while embedding dynamic calibration logic to correct channel errors, and finally generating high-precision serial output, which is suitable for high-speed signal processing scenarios with strict requirements for real-time performance and reliability.

[0206] This application also discloses a low-power composite multi-channel acoustic imaging system for implementing the above method.

[0207] The implementation principle is as follows:

[0208] The acoustic signal processing flow is reconstructed through a three-stage pipeline of the serial-parallel conversion architecture, and the computing bottleneck is broken through by optimizing matrix features. Different numbers of parallel channels and processing precisions are selected according to the signal features of different sampled acoustic signals and different acoustic environments. The high computational load in the full frequency band brought by the traditional FFT operation is optimized through DFT operations, and only the target frequency points are selected for individual operations according to the corresponding environment and features, further reducing the load power during beamforming and achieving real-time high-precision acoustic positioning with micro-watt level power consumption.

[0209] It should be understood that although the steps in the flowchart of the accompanying drawings are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit and can be executed in other orders.

[0210] The above are all the preferred embodiments of this application, and the protection scope of this application is not limited accordingly. Therefore, all equivalent changes made according to the structure, shape, and principle of this application should be covered within the protection scope of this application.

Claims

1. A low-power composite multi-channel acoustic imaging method, characterized in that: The following steps are involved: Acquire the sampled acoustic signal and send it to the configured three-stage pipeline structure including the front end, middle end and back end; At the front end, Acquiring signal characteristics of the sampled acoustic signal to select a corresponding deserialization path, and performing serial conversion and processing on the sampled acoustic signal based on the deserialization path to obtain a plurality of parallel data streams; In the middle, Retrieve a parallel beamforming array to configure a plurality of processing units and distribute a plurality of parallel data streams to each of the processing units, obtain acoustic environment information and distribute a corresponding target frequency point k and a required frequency point number x1 to each of the processing units; Based on the target frequency k and the required frequency number x1, a DFT operation is performed on each of the parallel data streams to obtain a corresponding covariance matrix; Obtaining a steering vector matrix corresponding to the parallel beamforming array and combining it with a plurality of the covariance matrices to generate parallel beams; On the backend, A phase compensation coefficient is obtained, and based on the phase compensation coefficient, it is converted into a FIR filter parameter to correct the phase error of the parallel beam, and then a parallel-to-serial process is performed to generate a serial output.

2. The low-power composite multi-channel acoustic imaging method according to claim 1, characterized in that: Acquiring the signal characteristics of the sampled acoustic signal to select a corresponding deserialization path, and performing serial conversion and processing on the sampled acoustic signal based on the deserialization path to obtain a plurality of parallel data streams, including the following steps: Acquiring a high-speed serial signal corresponding to the sampled acoustic signal; Collecting the corresponding data rate and processing requirements in the signal characteristics, and matching the corresponding deserialization ratio n based on the data rate and the processing requirements to select the deserialization path, wherein the deserialization ratio includes 1:4, 1:8, and 1:16; The high-speed serial signal is deserialized into a plurality of parallel data streams based on the deserialization path.

3. The low-power composite multi-channel acoustic imaging method according to claim 2, characterized in that: After deserializing the high-speed serial signal into a plurality of parallel data streams based on the deserialization path, the method further includes the following steps: Obtain the number of microphone channels x2, and calculate the number of multiplexed signals m required to be configured for each parallel data stream in combination with the deserialization ratio n, wherein the number of multiplexed signals is inversely proportional to the inverse of the deserialization ratio; Obtain x2 1-bit PDM signals, and double the number of the 1-bit PDM signals according to the left and right channels to obtain a sampling processing amount of x3; The x3 1-bit PDM signals are converted into multi-bit PCM signals through a CIC filter.

4. The low-power composite multi-channel acoustic imaging method according to claim 3, characterized in that: Retrieving a parallel beamforming array to configure a plurality of processing units and allocating a plurality of parallel data streams to each of the processing units, acquiring acoustic environment information and allocating a corresponding target frequency k and a required frequency number x1 to each of the processing units, includes the following steps: Obtaining the number x4 of the processing units included in the parallel beamforming array, and calculating the number x5 of paths that each processing unit needs to process based on x3 / x4 to distribute a plurality of multi-bit PCM signals to each processing unit; Determining an environment type based on the acoustic environment information, and determining a target frequency range according to the environment type; Determine the environmental complexity based on the acoustic environment information, and determine the required number of frequency points x1 according to the environmental complexity; In each of the processing units, a corresponding number of frequency feature points are selected in the target frequency range corresponding to each of the multi-bit PCM signals based on the required number of frequency points x1 as the target frequency point k.

5. The low-power composite multi-channel acoustic imaging method according to claim 3, characterized in that: Based on the target frequency k and the required frequency number x1, a DFT operation is performed on each of the parallel data streams to obtain a corresponding covariance matrix, including the following steps: Analyze the frequency domain distribution of the multi-bit PCM signal to generate a first coefficient, obtain imaging accuracy requirements to generate a second coefficient, and obtain signal quality to generate a third coefficient; Dynamically configure a corresponding initial bit width based on the first coefficient, the second coefficient, and the third coefficient; Selecting the corresponding target frequency point k based on the initial bit width to perform DFT operation to obtain frequency domain data of each multi-bit PCM signal; A frequency domain vector is generated based on the frequency domain data of the plurality of multi-bit PCM signals, and an upper triangular covariance of each of the multi-bit PCM signals is calculated through conjugate characteristics according to the frequency domain vector to generate a covariance matrix.

6. The low-power composite multi-channel acoustic imaging method according to claim 5, characterized in that: Obtaining a steering vector matrix corresponding to the parallel beamforming array and combining a plurality of the covariance matrices to generate parallel beams comprises the following steps: Determining a steering vector corresponding to each array element in the parallel beamforming array to form the steering vector matrix; The covariance matrix is ​​reduced in dimension to obtain a plurality of sub-matrices, and each of the array elements is associated with each of the multi-bit PCM signals to respectively multiply each of the sub-matrices of the steering vector matrix to obtain a plurality of sub-results, wherein each of the sub-matrices reuses the same IP core when multiplied with the steering vector matrix; The plurality of sub-results are combined into the complete parallel beam in the form of a block matrix.

7. The low-power composite multi-channel acoustic imaging method according to claim 4, characterized in that: Obtaining a steering vector matrix corresponding to the parallel beamforming array and combining a plurality of the covariance matrices to generate parallel beams also includes the following steps: generating a plurality of neuron clusters of the same size based on the target frequency range, each of the neuron clusters being configured with a processing strategy including synaptic weights and pulse time coding; Based on the target frequency point k, a corresponding frequency is obtained, and the frequency is mapped to the corresponding neuron cluster to obtain the corresponding processing strategy to adjust the output of the parallel beam, specifically; Generate a probability of firing a pulse signal based on the synaptic weight in the processing strategy for random weighting, wherein the pulse signal includes 0 or 1, wherein when the pulse signal is 0, the covariance corresponding to the target frequency point k does not participate in the calculation of the covariance matrix; The phase complement delay time is adjusted based on the pulse time encoding in the processing strategy.

8. The low-power composite multi-channel acoustic imaging method according to claim 3, characterized in that: Obtaining a phase compensation coefficient, and converting the phase compensation coefficient into a FIR filter parameter to correct the phase error of the parallel beam, further comprising the following steps: Acquire a phase compensation table including results of a number of tested array elements, wherein the phase compensation table includes phase deviations of different array elements relative to a reference array element; Based on the target frequency point k and the required frequency point number x1 corresponding to each array element, the phase compensation coefficient corresponding to each array element is determined in the phase compensation table and integrated into a target frequency domain response; Performing an inverse Fourier transform on the target frequency domain response and calculating the FIR filter parameters in combination with the filter order; Based on the FIR filter parameters, a convolution operation is performed on the steering vector corresponding to each array element and the convolution operation is output to complete phase error correction.

9. The low-power composite multi-channel acoustic imaging method according to claim 1, characterized in that: After performing phase error correction on the parallel beams, performing parallel-to-serial processing to generate a serial output, the method further includes the following steps: Setting a number of ping-pong buffers and configuring a read and write order for each path based on the number of the parallel data streams; Based on the read-write sequence, the array elements corresponding to each of the parallel data streams are written into the ping-pong buffer area in each frame to perform phase error correction of the steering vector; When the depth corresponding to one of the ping-pong buffers is fully written, the corresponding parallel data stream is read and output, and the ping-pong buffer is replaced to continue writing the array elements corresponding to each of the parallel data streams to correct the phase error of the steering vector; The parallel data streams read out continuously are spliced ​​into a serial data stream to generate a serial output.

10. A low-power composite multi-channel acoustic imaging system, characterized in that: Used to implement the method according to any one of claims 1 to 9.

Images