Sound source positioning method and device, computer equipment and storage medium

Abstract

The invention provides a sound source positioning method and device, computer equipment and a storage medium. The method comprises the following steps: forming input data based on a space range corresponding to a target space and a prediction moment; inputting the input data into a pre-trained physical constraint neural network to obtain sound pressure information of the sound in the target space at the prediction moment; wherein the physical constraint neural network takes an acoustic wave equation as a physical equation, takes initial sound pressure distribution and change rate distribution of initial sound pressure as initial conditions, and takes an impedance boundary condition as a boundary condition; actual measurement data collected by a distributed acoustic sensing DAS system deployed in the target space is used as data constraint training; and based on the sound pressure information, determining position information of a target sound source emitting the sound in the target space.

Description

Technical Field

[0001] This disclosure relates to the field of acoustics, and more specifically, to a method, apparatus, computer device, and storage medium for locating a sound source. Background Technology

[0002] Sound source localization technology is a technique that uses sensors to receive sound wave signals and combines signal processing, array collaboration, or algorithm analysis to determine the three-dimensional position (or direction) of a sound source in space. The core of this technology is to use information such as the time difference, phase difference, and amplitude difference of sound wave propagation to infer the coordinates of the sound source.

[0003] Traditional sound source localization methods typically rely on time-of-arrival or beamforming techniques. These methods often require large-scale sensor networks, are sensitive to the number and location of sensors, have high computational complexity, and are difficult to meet usage requirements. Summary of the Invention

[0004] In view of this, the present disclosure provides a sound source localization method, apparatus, computer equipment, and storage medium.

[0005] Specifically, this disclosure is achieved through the following technical solution: In a first aspect, embodiments of this disclosure provide a sound source localization method, the method comprising: The input data is constructed based on the spatial range corresponding to the target space and the prediction time. The input data is fed into a pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the predicted time. The physical constraint neural network uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure rate of change distribution as the initial conditions, the impedance boundary conditions as the boundary conditions, and the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as data constraint training. Based on the sound pressure information, the location information of the target sound source emitting the sound is determined within the target space.

[0006] Optionally, the input data, based on the spatial range corresponding to the target space and the prediction time, includes: Based on the location density of the sound source localization, the target space is divided into multiple spatial grids, and corresponding location information is determined for each spatial grid. Based on the location information and prediction time corresponding to each spatial grid, input data corresponding to each spatial grid is constructed.

[0007] Optionally, the physical constraint neural network can be trained in the following manner: Based on the measured data collected and measured at the sampling points by the distributed acoustic sensing (DAS) system deployed in the target space; and by sampling the spatial domain and the target time domain of the target space to obtain the physical equation sampling points, boundary condition sampling points, and initial condition sampling points; wherein, the sampling points include: sampling location and sampling time; The sampling points are input into the physical constraint neural network to be trained to obtain the predicted sound pressure corresponding to the sampling points; The partial differential equation loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation; the boundary condition loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the boundary condition and the impedance boundary condition; the initial condition loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the initial condition and the initial condition; and the data loss, serving as a data constraint, is determined based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data. The physical constraint neural network to be trained is trained with the goal of minimizing the sum of the partial differential equation loss, the boundary condition loss, the initial condition loss, and the data loss, to obtain a trained physical constraint neural network.

[0008] Optionally, determining the partial differential equation loss as a physical constraint based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation includes: Substitute the predicted sound pressure corresponding to the sampling points of multiple physical equations into the acoustic wave equation to determine the wave equation residuals corresponding to the sampling points of multiple physical equations. The mean square loss of the wave equation residuals corresponding to the sampling points of multiple physical equations is determined, and the mean square loss of the wave equation residuals is used as the loss of the partial differential equation.

[0009] Optionally, determining the boundary condition loss as a physical constraint based on the predicted sound pressure corresponding to the boundary condition sampling point and the impedance boundary condition includes: Substitute the predicted sound pressure corresponding to multiple boundary condition sampling points into the impedance boundary equation used to describe the impedance boundary conditions, determine the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points, and determine the mean square loss of the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points. Use the mean square loss of the impedance boundary condition residuals as the boundary condition loss.

[0010] Optionally, the initial condition sampling points include: a first initial condition sampling point and a second initial condition sampling point; the initial condition loss includes: a first initial condition loss and a second initial condition loss; The step of determining the initial condition loss as a physical constraint based on the predicted sound pressure corresponding to the initial condition sampling point and the initial condition includes: The initial sound pressure information corresponding to the initial condition sampling point is determined based on the initial sound pressure distribution, and the rate of change information corresponding to the initial condition sampling point is determined based on the rate of change distribution of the initial sound pressure. The difference between the initial sound pressure information corresponding to the first initial condition sampling point and the predicted sound pressure corresponding to the first initial condition sampling point is determined as the first initial condition residual. The mean square loss of the first initial condition residuals corresponding to the multiple first initial condition sampling points is determined as the first initial condition loss. Furthermore, the difference between the derivative of the predicted sound pressure with respect to time corresponding to the second initial condition sampling point and the rate of change information corresponding to the second initial condition sampling point is determined as the second initial condition residual, and the mean square loss of the second initial condition residuals corresponding to the multiple second initial condition sampling points is determined as the second initial condition loss.

[0011] Optionally, determining the data loss as a data constraint based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data includes: Based on the predicted sound pressure corresponding to the measured sampling point, the Young's modulus of the optical fiber in the DAS system, and the Poisson's ratio, the predicted strain rate corresponding to the measured sampling point is determined under the predicted sound pressure corresponding to the measured sampling point. The difference between the predicted strain rate corresponding to the measured sampling point and the measured data is determined as the data residual; The mean square loss of the data residuals corresponding to the multiple measured sampling points is determined, and the mean square loss is determined as the data loss.

[0012] Optionally, the method further includes: In response to the achievement of model update conditions, the measured data collected by the DAS system for the target historical period at the current time is obtained; Based on the measured data, the physical constraint neural network at the current moment is updated.

[0013] Secondly, embodiments of this disclosure also provide a sound source localization device, the device comprising: The generation module is used to generate input data based on the spatial range corresponding to the target space and the prediction time. The processing module is used to input the input data into a pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the predicted time; wherein, the physical constraint neural network uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure rate of change distribution as the initial conditions, the impedance boundary conditions as the boundary conditions, and uses the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as data constraint training; The determination module is used to determine the location information of the target sound source emitting the sound within the target space based on the sound pressure information.

[0014] Thirdly, an optional implementation of this disclosure also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the first aspect above, or any possible implementation of the first aspect.

[0015] Fourthly, an optional implementation of this disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the first aspect or any possible implementation of the first aspect.

[0016] Fifthly, an optional implementation of this disclosure also provides a computer program product, the computer program product carrying program code, the program code including instructions that can be used to perform the steps of the sound source localization method as described in the first aspect or any one of the first aspects.

[0017] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure.

[0018] The sound source localization method provided in this disclosure uses the spatial range and predicted time corresponding to the target space as input data. This input data is then fed into a pre-trained physical constraint neural network to obtain sound pressure information of the sound within the target space at the target time. Based on this sound pressure information, the location information of the target sound source emitting the sound within the target space is determined. The physical constraint neural network used in the above sound source localization method uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the rate of change of the initial sound pressure as initial conditions, and impedance boundary conditions as boundary conditions. It is trained using measured data collected by a distributed acoustic sensing (DAS) system deployed in the target space as data constraints. Thus, during the training phase, the physical constraint neural network incorporates the sound wave equation and impedance boundary conditions into the learning process, achieving high-precision acoustic modeling that conforms to real-world physical laws. Simultaneously, by combining DAS measurement data with the PINN framework, synchronous inversion of the sound source and sound field is achieved. This allows acoustic physical laws to be embedded into the neural network training, and by utilizing DAS measured data, high-precision sound source localization and spatiotemporal sound field reconstruction are achieved, realizing the purpose of high-precision indoor activity monitoring and real-time sound source localization.

[0019] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a sound source localization method according to an exemplary embodiment of this disclosure; Figure 2 This is a flowchart illustrating a specific method for training a physically constrained neural network, as shown in an exemplary embodiment of this disclosure; Figure 3 This is a schematic diagram of a DAS data-driven physical information neural network, as illustrated in an exemplary embodiment of this disclosure. Figure 4 This is a schematic diagram illustrating continuous data recording of DAS data according to an exemplary embodiment of the present disclosure; Figure 5 This is a schematic diagram illustrating the different effects of the full-data simultaneous training strategy and the incremental time-step training strategy used in the embodiments of this disclosure, as shown in an exemplary embodiment of this disclosure. Figure 6 This is a schematic diagram illustrating the results of two different training strategies as shown in an exemplary embodiment of this disclosure; Figure 7 This is a specific example of a DAS system deployed in a library and receiving sound source signals, as illustrated in an exemplary embodiment of this disclosure. Figure 8This is a schematic diagram of a computer device illustrated in an exemplary embodiment of the present disclosure; Figure 9 This is a schematic diagram of the structure of a sound source localization device shown in an exemplary embodiment of the present disclosure. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0022] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0023] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0024] To facilitate understanding of this embodiment, the relevant technologies involved in this disclosure embodiment will first be described: Distributed Acoustic Sensing (DAS) is an emerging fiber optic sensing technology. Its core principle is to utilize a single optical fiber as a "distributed sensor." By detecting changes in Rayleigh scattering light, it achieves continuous, real-time localization and sensing of acoustic signals (including sound waves and vibrations) along the entire fiber optic route, without the need for additional sensor nodes. This technology can transform ordinary optical fibers into densely distributed acoustic sensor arrays, thus showing broad application prospects in environmental monitoring, urban infrastructure analysis, and indoor activity monitoring in smart cities. By capturing acoustic signals along the length of the optical fiber, DAS can detect and locate sound source events without deploying traditional sensor networks, offering advantages such as low cost and privacy protection. Beyond urban environments, DAS technology has also demonstrated its versatility in various fields. For example, significant progress has been made in ultra-long-distance cable monitoring without repeater amplification, volcanic event detection and shallow geological imaging, and seismic acoustic research using submarine communication cables.

[0025] Physics-Informed Neural Networks (PINNs), as a novel computational framework integrating physical laws and machine learning, have attracted widespread attention for solving both forward and inverse problems related to Partial Differential Equations (PDEs). PINNs introduce physical equation constraints during the training process of the neural network, enabling the network to utilize both experimental data and physical laws, thus simulating complex physical phenomena without requiring a large number of labeled samples. However, existing PINN frameworks still face several challenges in practical applications. For example, their soft constraint approach to boundary conditions (BCs) and initial conditions (ICs) requires extensive experimental adjustments to the constraint weights and sampling strategies, and even if the loss function converges, it is difficult to strictly satisfy the boundary or initial conditions. Furthermore, PINNs are sensitive to hyperparameters and suffer from scalability issues in large-scale or complex domains.

[0026] Traditional acoustic localization methods rely on a large number of sensor networks, which are sensitive to the number and location of sensors, have high computational complexity, and are difficult to meet the needs of real-time monitoring. On the other hand, existing DAS data-driven model training methods require large-dimensional data collection with complex noise, rely on data annotation, and have insufficient model generalization ability.

[0027] The existing PINN framework has the following problems: (1) The boundary conditions and initial conditions are not strict: When the boundary conditions and initial conditions are added as soft constraints to the loss function, PINN may have difficulty in accurately satisfying these conditions, which may lead to the generation of suboptimal solutions. (2) High hyperparameter sensitivity: Different weight combinations have a significant impact on training convergence and result accuracy; (3) Limited scalability: In large-scale complex domains, PINN training is time-consuming and optimization is difficult; The aforementioned issues result in insufficient accuracy, limited generalization ability, and low computational efficiency when combining the PINN model with DAS for sound source localization and sound field reconstruction in complex indoor environments. Therefore, there is an urgent need for a high-precision model that combines physical constraints with DAS data-driven approaches to improve indoor acoustic monitoring capabilities.

[0028] Based on this, the present disclosure provides a sound source localization method that deeply integrates a DAS fiber array with a PINN model, achieving the following key innovations: 1. Replacing traditional acoustic sensors with optical fibers: The DAS system senses external acoustic disturbances by detecting the phase change of the backscattered Rayleigh signal in the optical fiber, achieving high-density acoustic monitoring; deploying optical fiber networks in buildings or indoor spaces forms a multi-dimensional observation channel, which can continuously collect high sampling rate and high spatial resolution indoor acoustic data for subsequent sound source localization and sound field reconstruction.

[0029] 2. Simultaneously embed acoustic wave equations, impedance boundary conditions, and DAS observation data during neural network training: Under the constraints of DAS data, the PINN network is used to directly solve the field solution that satisfies the physical constraints of acoustic wave propagation.

[0030] 3. Adopt an incremental time-step training strategy: simulate the real-time data acquisition process, gradually optimize network parameters, and improve positioning stability and computational efficiency.

[0031] 4. Use the sound source location as a trainable parameter: Automatically invert the spatial coordinates of the sound source through gradient backpropagation, avoiding manual annotation.

[0032] To facilitate understanding of this embodiment, a detailed description of the sound source localization method disclosed in this disclosure is provided first. The execution entity of the sound source localization method provided in this disclosure is generally a computer device with certain computing capabilities. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a terminal, an in-vehicle device, a wearable device, etc. In some possible implementations, the sound source localization method can be implemented by a processor calling computer-readable instructions stored in memory.

[0033] The sound source localization method provided in the embodiments of this disclosure will be described below.

[0034] See Figure 1The diagram shows a flowchart of a sound source localization method provided in an embodiment of this disclosure. The method includes steps S101 to S103, wherein: S101: Input data is constructed based on the spatial range and prediction time corresponding to the target space; S102: Input the input data into a pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the predicted time; wherein, the physical constraint neural network uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure rate of change distribution as the initial conditions, the impedance boundary conditions as the boundary conditions, and uses the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as data constraint training; S103: Based on the sound pressure information, determine the location information of the target sound source emitting the sound within the target space.

[0035] In the sound source localization method provided in this embodiment, input data is constructed based on the spatial range corresponding to the target space and the predicted time. The input data is input into a pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the target time. Then, based on the sound pressure information, the location information of the target sound source emitting the sound in the target space is determined. The physical constraint neural network used in the above sound source localization method uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure change rate distribution as the initial conditions, and the impedance boundary conditions as the boundary conditions. It is trained using measured data collected by a distributed acoustic sensing (DAS) system deployed in the target space as data constraints. Thus, in the training phase, the physical constraint neural network achieves high-precision acoustic modeling that conforms to real physical laws by introducing the sound wave equation and impedance boundary conditions into the learning process. At the same time, by combining DAS measurement data and the PINN framework, synchronous inversion of the sound source and sound field is achieved. This allows acoustic physical laws to be embedded into the neural network training, and high-precision sound source localization and spatiotemporal sound field reconstruction can be achieved using DAS measured data, thus achieving the purpose of high-precision indoor activity monitoring and real-time sound source localization.

[0036] The following provides a detailed explanation of S101 to S103.

[0037] Regarding the above S101: In practical implementation, the target space is, for example, the space where sound source localization is to be performed. This target space is usually a relatively enclosed space, such as a room, factory, pipeline, tunnel, ship cabin, or storage tank. Specifically, depending on the actual needs, different spaces are designated as target spaces, and corresponding physical constraint neural networks are trained on the target spaces. This allows the physical constraint neural networks to learn the sound field propagation characteristics of the target spaces, thereby achieving high-precision sound source localization processing.

[0038] When constructing input data based on the spatial range and prediction time corresponding to the target space, for example, the target space can be divided into multiple spatial grids according to the positioning density of the sound source, and the corresponding position information can be determined for each spatial grid; then, input data corresponding to each spatial grid can be constructed based on the position information corresponding to each spatial grid and the prediction time.

[0039] Specifically, the location density of a sound source localization is typically used to describe the spatial granularity of the localization. A higher location density indicates a smaller spatial grid size, resulting in finer spatial granularity and higher localization accuracy. Conversely, a lower location density indicates a larger spatial grid size, resulting in larger spatial granularity and correspondingly lower accuracy in locating the sound source. For example, a higher location density can be set within a smaller space to achieve more accurate sound source localization. In practical applications, the specific location density can be set according to the requirements, and this disclosure does not limit this setting.

[0040] After the spatial grid is determined, the position information of each spatial grid in the target space is determined; wherein, the position information may be, for example, the position of the center point of each spatial grid in the target space, or the position of a vertex at a specific location in the target space, which can be set according to actual needs.

[0041] The prediction time is, for example, multiple times within a certain time slice. In practical applications, it is necessary to monitor the location of a sound source in a certain space in real time. The time slice for sound source prediction can be represented as [0,T]. Each time slice includes n prediction times, and the space is divided into m spatial grids.

[0042] For the k-th prediction time The input data corresponding to the k-th prediction time includes: , ... .

[0043] n prediction times can yield n sets of input data.

[0044] Regarding S102 and S103 above: In practice, after obtaining the input data for the prediction time, the corresponding input data is fed into the physical constraint neural network, which outputs the sound pressure information of each spatial grid at the prediction time. Based on the sound pressure information of each spatial grid at the prediction time, and according to the principle of sound propagation, the position of the target sound source at the prediction time can be calculated.

[0045] Through the above process, there are n prediction times within the time slice [0,T], and the target space is divided into m grids. Therefore, n sets of prediction results can be obtained for this time slice. By fitting the n sets of prediction results, the position information of the target sound source in the target space within the time slice [0,T] can be obtained.

[0046] In addition, the location information corresponding to multiple time slices can be combined to determine the movement trajectory of the target sound source in the target space.

[0047] The physical constraint neural network in this embodiment uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure rate of change distribution as the initial conditions, the impedance boundary conditions as the boundary conditions, and the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as the data constraint training.

[0048] Among them, the acoustic wave equation ensures that the sound pressure output by the model satisfies the physical laws of sound propagation; the initial sound pressure distribution and the initial sound pressure rate of change distribution are used as initial conditions to ensure that the sound pressure output by the model conforms to the real state at the initial moment; the impedance boundary condition constrains the sound pressure at the spatial boundary during sound propagation to satisfy the sound reflection condition, ensuring that the sound pressure output by the model conforms to the boundary constraints of the actual sound propagation at the boundary position in the actual scene.

[0049] Specifically, in the three-dimensional space region Internal sound pressure The propagation of satisfies the following wave equation (1): (1) in, Indicates the speed of sound in a medium; Represents the Laplace operator; Indicates time; To represent any point in three-dimensional space, it can be expressed as: This acoustic wave equation is used to describe how sound pressure varies with spatial location and time.

[0050] The initial conditions of the sound field satisfy the following formulas (2) and (3): (2) (3) in, This represents the initial sound pressure distribution; that is, at the initial moment. Spatial location The sound pressure distribution at a given location is determined by the function Sure.

[0051] This represents the first-order partial derivative of sound pressure with respect to time, which is also the distribution of the rate of change of the initial sound pressure.

[0052] In practical implementation, the initial sound pressure distribution of the sound source A Gaussian function approximation can be used, which is the initial sound pressure distribution. Satisfy the following formula (4): (4) in, Indicates the amplitude of the sound wave. Indicates the control pulse width. This indicates the location of an unknown sound source.

[0053] in, For example, it can be obtained using measured data acquired through a DAS system.

[0054] also, It can also be used as an optimizable parameter, with an initial value set for it, and continuously optimized during multiple iterations based on the initial value.

[0055] The impedance boundary condition satisfies the following formula (5): (5) in, Indicates the boundary acoustic impedance; Indicates air density; Representing the normal derivative, The formula represents the boundary region and constrains the normal phase change rate of sound pressure at the boundary, reflecting the reflection and transmission characteristics of sound waves at the boundary.

[0056] By using the aforementioned acoustic wave equation, initial conditions, and impedance boundary conditions as physical constraints to train a physical constraint neural network, the physical constraint neural network can learn the propagation characteristics of sound in the target space, making the sound pressure prediction results output by the physical constraint neural network conform to the actual physical relationship, thereby achieving high-precision acoustic modeling.

[0057] Furthermore, the physically constrained neural network in this embodiment also uses measured data collected by a distributed acoustic sensing (DAS) system deployed in the target space as data constraints. Thus, this method achieves high-precision sound source localization and spatiotemporal sound field reconstruction by embedding acoustic physical laws into the neural network training and utilizing DAS measured data, thereby achieving high-precision indoor activity monitoring and real-time sound source localization while protecting privacy.

[0058] In another embodiment, the physical constraint neural network described in this disclosure is trained using an incremental time-step training strategy. Under this strategy, according to the temporal order of acoustic observation data, at regular update cycles, the measured data collected by the DAS system can be used as incremental sample data and input into the physical constraint neural network trained in the previous update cycle. This updates the network parameters of the physical constraint neural network in real time and updates the sound source location in real time, thereby achieving online sound source localization and sound field updating. Compared to a one-time training strategy, this approach can quickly locate moving sound sources with a small amount of time-slice data, and can achieve real-time localization of sporadic sound sources in the target space.

[0059] Specifically, see Figure 2 As shown in the embodiments of this disclosure, a specific method for training the physical constraint neural network is also provided, including: S201: Based on the measured data corresponding to the sampling points collected and measured by the distributed acoustic sensing DAS system deployed in the target space; and to obtain the physical equation sampling points, boundary condition sampling points, and initial condition sampling points by sampling the spatial domain and target time domain of the target space; wherein, the sampling points include: sampling position and sampling time.

[0060] In specific implementations, the sampling points in this embodiment include: measured sampling points, physical equation sampling points, boundary condition sampling points, and initial condition sampling points. Each sampling point includes: a sampling location within the spatial domain corresponding to the target space, and a sampling time belonging to the target time domain.

[0061] Here, the spatial domain, also known as the target space, is represented, for example, as: .

[0062] The target time domain is, for example, a period of time with a preset duration, which is represented as [0, T]. Here, T represents the duration of the time domain, that is, the maximum time point.

[0063] a1: Measured sampling point, such as the sampling point determined when collecting measured data using a DAS system deployed within the target space. A DAS system typically includes multiple distributed optical fibers. These distributed optical fibers are deployed in different locations or areas of the target space according to a specific layout.

[0064] Specifically, the fiber optic cable layout in the target space can adopt a planar layout, an edge layout, a random Halton sequence distribution, or other distribution methods. The specific method can be selected according to actual needs, and this disclosure does not limit the specific arrangements. Multiple locations can be pre-defined on the distributed optical fiber as channels for sampling and measuring data. Each channel corresponds to a location coordinate in the target space, which can be represented as follows: .

[0065] A narrow-pulse laser (pulse width determines spatial resolution) is emitted into an optical fiber. As the laser propagates through the fiber, each channel generates Rayleigh scattered light, which is reflected back to the receiver. When a sound source is present near the fiber, the sound wave vibration causes minute deformation (nanometer-scale) in the fiber, resulting in a shift in the phase / intensity of the scattered light in that region of the channel. This shift causes a change in the Rayleigh scattered light. The DAS system receives the scattered light signals from each channel in real time and extracts the phase / intensity change of the signal through demodulation technology. This is the rate of change of the mechanical strain along the fiber axis over time (strain rate, i.e., the sound-sensing signal). Based on this strain rate, the range of channels affected by the sound source is identified. The arrival time of sound waves from the same source varies between different channels: the closer the channel is to the sound source, the earlier the signal change is detected. The DAS system compares the signal trigger time difference of the affected channels and combines this with the propagation speed of sound waves in the medium (e.g., 343 m / s in air, faster in solids) to use physical relationships to deduce the vertical distance from the sound source to the fiber and its projected position along the fiber. If the sound source moves along the fiber, the system can track the sound source trajectory in real time by continuously sampling the time difference changes.

[0066] In the process of acquiring measured sampling points using a DAS system, the target time domain can be divided into multiple time slices, and measured data can be acquired at multiple sampling time points t within each time slice. For each channel, for example, n measured data points can be acquired within each time slice. These n measured data points, arranged in the order of acquisition, constitute a sequence of measured data, i.e., the measured data for each channel within its corresponding time slice. For each channel, the sampling time point corresponding to that channel constitutes a measured sampling point. This measured sampling point corresponds to measured data obtained by the DAS system, such as the strain rate of the optical fiber.

[0067] For the k-th channel out of m channels, using a DAS system deployed in the target space, it is possible to acquire the measured data corresponding to n sampling time points of the k-th channel within the current time slice; the measured data corresponding to a certain channel at a certain sampling time point can be represented as: ,in, This indicates the location information of the channel within the target space; Indicates the sampling time point; This represents the strain rate at the fiber position corresponding to that channel at that sampling time point. This constitutes a measured sampling point, which, for simplicity, is represented as: , .

[0068] Using this measured data as a data constraint during the training of the physical constraint neural network As input to the model, This represents the label corresponding to the sampled data. After being input into the physical constraint neural network to be trained, the predicted sound pressure corresponding to that channel can be obtained. Based on this predicted sound pressure and relevant parameters of the optical fiber, such as Poisson's ratio and Young's modulus of the optical fiber, the predicted strain rate of the optical fiber in that channel can be calculated.

[0069] a2: Sampling points for the physical equations, typically obtained by sampling across the entire spatial and temporal domains. Generally, it's necessary to perform sampling covering the entire defined spatial and temporal domains. Sampling methods can include, for example, uniform random sampling or LHS sampling. It is usually represented as: Simplified to: , .

[0070] a3: Initial condition sampling points, typically obtained by full-domain sampling of the spatial domain at the initial moment. Usually, only full-domain sampling of the spatial domain at the initial moment is required. The sampling method is similar to that of the sampling points for the physical equations. It is usually represented as: Simplified to: , .

[0071] a4: Boundary condition sampling points, typically used for temporal domain sampling covering spatial boundaries. During sampling, the spatial boundary coordinates are usually fixed, such as by fixing... These are the boundary coordinates of the spatial domain, and sampling is performed in the time domain to cover the entire domain. It is typically represented as: Simplified to: , .

[0072] After obtaining the above four types of sampling points, other methods for training physically constrained neural networks include: S202: Input the sampling points into the physical constraint neural network to be trained to obtain the predicted sound pressure corresponding to the sampling points.

[0073] S203: Determine the partial differential equation loss as a physical constraint based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation; determine the boundary condition loss as a physical constraint based on the predicted sound pressure corresponding to the sampling point of the boundary condition and the impedance boundary condition; determine the initial condition loss as a physical constraint based on the predicted sound pressure corresponding to the sampling point of the initial condition and the initial condition; and determine the data loss as a data constraint based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data.

[0074] In specific implementation, such as Figure 3 The diagram shows a schematic of a DAS data-driven physical information neural network.

[0075] Among them, the network structure of the physically constrained neural network is as follows: Figure 3 The upper half of the diagram shows the following components: input layer, hidden layer, output layer, and physical constraint module.

[0076] The input layer includes time ( ), spatial coordinates , representing the spatiotemporal location of the sound field to be predicted.

[0077] The hidden layers are multi-layer fully connected networks used to learn the mapping relationship between spatiotemporal coordinates and sound pressure.

[0078] The output layer outputs sound pressure (p), which is the predicted sound field distribution.

[0079] Physical constraint module: Calculates sound pressure versus time using automatic differentiation. ),space( , The partial derivatives of (etc.) are substituted into the acoustic wave equation, initial conditions, and boundary conditions to calculate the residuals, which are finally expressed through the loss function. Implement physical constraints to ensure that the predicted sound field conforms to the physical laws of acoustics.

[0080] Should Figure 3 The lower half represents DAS data-driven PINN, where Figure 3 In the diagram 'a', the horizontal and vertical axes represent the x and y coordinates, respectively. Figure 3 In the middle b, the horizontal and vertical axes represent time and channel number, respectively.

[0081] When the sampling points are input into the physical constraint neural network to be trained, for example, the above four types of sampling points are input into the physical constraint neural network to be trained, and the predicted sound pressure corresponding to each sampling point is obtained.

[0082] For example, regarding the measured sampling points The corresponding predicted sound pressure level is expressed as follows: .

[0083] For the sampling points of the physical equation The corresponding predicted sound pressure level is expressed as follows: .

[0084] For initial condition sampling points The corresponding predicted sound pressure level is expressed as follows: .

[0085] Sampling points for boundary conditions The corresponding predicted sound pressure level is expressed as follows: .

[0086] After obtaining the predicted sound pressure levels corresponding to various sampling points, the model's various losses can be determined using the following method: (1): The partial differential equation loss, which serves as a physical constraint, is determined using the following method based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation: Substitute the predicted sound pressure corresponding to the sampling points of multiple physical equations into the acoustic wave equation to determine the wave equation residuals corresponding to the sampling points of multiple physical equations, and determine the mean square loss of the wave equation residuals corresponding to the sampling points of multiple physical equations. Use the mean square loss of the wave equation residuals as the partial differential equation loss.

[0087] Among them, the partial differential equation loss Satisfy the following formula (6): (6) in, This indicates the predicted sound pressure level corresponding to the sampling point of the physical equation at the sampling point. The spatial second-order partial derivative at a given point is used to describe the distribution variation of sound pressure.

[0088] This indicates the predicted sound pressure level corresponding to the sampling point of the physical equation at the sampling point. The second-order partial derivative at time is used in conjunction with the speed of sound to describe the evolution of sound pressure over time.

[0089] This is the residual of the wave equation, which measures the degree of deviation between the model's predicted sound pressure and the wave equation; if the difference between the above two items is 0, it means that the predicted sound pressure does not deviate from the wave equation.

[0090] Indicates to The L2 norm of the wave equation residuals corresponding to the sampling points of each physical equation is calculated and averaged to obtain the mean square loss of the wave equation residuals, which is also the partial differential equation loss.

[0091] In PINN training, by minimizing the partial differential equation loss, the sound pressure field learned by the model is forced to strictly satisfy the wave equation, ensuring that the predicted sound field not only meets the data-driven fitting accuracy but also follows the physical laws of sound wave propagation.

[0092] (2): The boundary condition loss as a physical constraint is determined based on the predicted sound pressure corresponding to the boundary condition sampling point and the impedance boundary condition in the following manner: Substitute the predicted sound pressure corresponding to multiple boundary condition sampling points into the impedance boundary equation used to describe the impedance boundary conditions, determine the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points, and determine the mean square loss of the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points. Use the mean square loss of the impedance boundary condition residuals as the boundary condition loss.

[0093] The boundary condition loss Satisfy the following formula (7): (7) in, Indicates the boundary acoustic impedance; Indicates air density; Representing the normal derivative, Indicates the boundary region.

[0094] This represents the normal impedance term of the predicted sound pressure at the boundary condition sampling point, describing the barrier characteristics of the boundary to the normal change of sound pressure.

[0095] This represents the time variation term of the sound pressure boundary condition sampling point, which, combined with the medium density, reflects the time evolution of sound pressure and the coupling characteristics of the medium.

[0096] This represents the impedance boundary condition residual, measuring the degree of deviation between the model's predicted value and the boundary conditions. If the sum of the above two items is 0, it indicates that the predicted sound pressure level does not deviate from the boundary condition equation.

[0097] This means that the L2 norm is calculated and averaged for the residuals of all boundary condition sampling points to obtain the mean square loss of the impedance boundary condition residuals.

[0098] In PINN training, by minimizing this boundary condition loss, the sound pressure field learned by the model is forced to strictly satisfy the impedance constraint at the boundary, ensuring that the behavior of the predicted sound field in the boundary region (such as sound-absorbing walls, acoustic barriers, etc.) conforms to the actual physical scene.

[0099] (3): In this embodiment of the disclosure, the initial conditions include the initial sound pressure distribution and the initial sound pressure rate of change distribution. The initial condition loss may include, for example, a first initial condition loss corresponding to the initial sound pressure distribution and a second initial condition loss corresponding to the initial sound pressure rate of change distribution. The first initial condition loss is determined based on a first initial condition sampling point; the second initial condition loss is determined based on a second initial condition sampling point.

[0100] Specifically, the initial condition loss, which serves as a physical constraint, can be determined using the following method based on the predicted sound pressure corresponding to the initial condition sampling point and the initial conditions: The initial sound pressure information corresponding to the initial condition sampling point is determined based on the initial sound pressure distribution, and the rate of change information corresponding to the initial condition sampling point is determined based on the rate of change distribution of the initial sound pressure. The difference between the initial sound pressure information corresponding to the first initial condition sampling point and the predicted sound pressure corresponding to the first initial condition sampling point is determined as the first initial condition residual. The mean square loss of the first initial condition residuals corresponding to the multiple first initial condition sampling points is determined as the first initial condition loss. Furthermore, the difference between the derivative of the predicted sound pressure with respect to time corresponding to the second initial condition sampling point and the rate of change information corresponding to the second initial condition sampling point is determined as the second initial condition residual, and the mean square loss of the second initial condition residuals corresponding to the multiple second initial condition sampling points is determined as the second initial condition loss.

[0101] Specifically, the first initial condition loss For example, it satisfies the following formula (8): (8) Second initial condition loss For example, it satisfies the following formula (9): (9) in, The distribution function represents the initial sound pressure.

[0102] This represents the predicted sound pressure level and the preset initial sound pressure level distribution at the first initial condition sampling point at the initial moment of the prediction. The residual, also known as the first initial condition residual, is used to ensure that the initial sound pressure field conforms to the physical starting point, such as the sound pressure distribution of the initial excitation of the sound source.

[0103] This represents the rate of change at the initial time of prediction, based on the first initial condition sampling point and the preset initial rate of change. The residual, also known as the second initial condition residual, is used to ensure that the change trend of sound pressure at the initial moment conforms to physical logic, such as the sound pressure growth rate of the initial excitation of the sound source.

[0104] and The L2 norm is calculated and averaged for the first initial condition residuals corresponding to multiple first initial condition sampling points and the second initial condition residuals corresponding to multiple second initial condition sampling points, respectively, to obtain the mean square loss of the initial condition residuals, which is also the first initial condition loss and the second initial condition loss.

[0105] In PINN training, by minimizing this loss, the model is forced to strictly match the sound pressure and its rate of change at the initial moment with the preset initial state, ensuring that the "starting point" of the sound field evolution conforms to the physical reality (such as the initial excitation state of a sudden sound source), providing the correct initial constraints for subsequent spatiotemporal evolution based on the wave equation, which is suitable for scenarios such as sporadic sound source localization and transient sound field simulation.

[0106] In the above embodiments, the first initial condition sampling point and the second initial condition sampling point may be the same sampling point, different sampling points, or partially the same sampling point. The specific embodiments disclosed herein are not limited.

[0107] (4): The data loss, which serves as a data constraint, is determined using the following method, based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data: Based on the predicted sound pressure corresponding to the measured sampling point, the Young's modulus of the optical fiber in the DAS system, and the Poisson's ratio, the predicted strain rate corresponding to the measured sampling point is determined under the predicted sound pressure corresponding to the measured sampling point. The difference between the predicted strain rate corresponding to the measured sampling point and the measured data is determined as the data residual; The mean square loss of the data residuals corresponding to the multiple measured sampling points is determined, and the mean square loss is determined as the data loss.

[0108] In practice, data loss For example, it satisfies the following formula (10): (10) in, Indicates the number of actual sampling points; Represents the i-th measured sampling point The corresponding measured data.

[0109] This represents the Young's modulus of the optical fiber material; Represents Poisson's ratio. This represents the predicted sound pressure corresponding to the measured sampling point; This represents the predicted strain rate corresponding to the measured sampling point.

[0110] , that is This represents the residual data corresponding to the measured sampling points.

[0111] This means that the mean square loss of the data residuals is obtained by calculating the L2 norm of the data residuals corresponding to multiple measured sampling points and averaging them.

[0112] Following S203 above, the method for training a physically constrained neural network provided in this embodiment of the disclosure further includes: S204: Train the physical constraint neural network to be trained with the goal of minimizing the sum of the partial differential equation loss, the boundary condition loss, the initial condition loss and the data loss, to obtain a trained physical constraint neural network.

[0113] In practical implementation, a physically constrained neural network can be represented as follows: in, This represents the trainable parameters.

[0114] The acoustic wave equation and impedance boundary conditions are embedded as physical constraints into the loss function to ensure that the training results of the neural network satisfy the basic laws of acoustics.

[0115] Total loss function Includes the following parts: Partial Differential Equation Loss Terms : Ensure that the acoustic wave equation is satisfied at the selected space-time point; Initial condition loss term : Ensure that the initial sound pressure distribution and time derivative meet the set conditions; Boundary condition loss term : Ensure that the impedance boundary conditions are effectively satisfied; Data loss items The output of the constraint network is consistent with the strain signal measured by DAS. Each item is weighted by a coefficient , , , Adjusting the contribution ratio to achieve a balance between physical constraints and data-driven approaches will reduce the total loss. Satisfy the following formula: With the goal of minimizing the total loss mentioned above, the physical constraint neural network to be trained is iteratively trained to obtain a physical constraint neural network for predicting sound pressure in the target space.

[0116] In the iterative training process, there may be multiple iteration cycles; the sampling points corresponding to different iteration cycles may be exactly the same, completely different, or partially the same.

[0117] Furthermore, in another embodiment of this disclosure, an incremental training strategy can be used to train the physical constraint neural network. The core of the incremental training strategy is based on a pre-trained model, which is fine-tuned to adapt to new PDE problems or data. This avoids the high cost of training from scratch and can solve the problem of quickly solving the same family of equations.

[0118] When training a physically constrained neural network (PCN) using an incremental training strategy, the initial parameters of the PCN to be trained can be, for example, those of a PCN trained for spaces other than the target space, used to predict sound pressure levels in those other spaces. Alternatively, a reference space can be used, and a PCN for that reference space can be trained first. When training the PCN for the target space, the network parameters of the PCN corresponding to the reference space are used to initialize the parameters of the PCN for the target space. Based on this, the relevant sampling points collected from the target space are then used to train the PCN for the target space.

[0119] Furthermore, for spaces of different scales or complexities, different physical constraint neural network structures can be set, such as different numbers of network layers, neurons, activation functions, derivatives, etc., to determine the corresponding physical constraint neural network structure for different spaces.

[0120] In another embodiment of this disclosure, it further includes: in response to reaching the model update condition, acquiring measured data collected by the DAS system for the target historical time period at the current moment; Based on the measured data, the physical constraint neural network at the current moment is updated.

[0121] In practice, physically constrained neural networks can learn the relevant characteristics of sound propagation within a target space, thus enabling them to predict sound sources within that space. However, over time, sporadic sound sources may appear within the target space; for example, in an indoor setting, an unknown sound source might accidentally enter. This can cause the accuracy of the physically constrained neural network's prediction of the sound field within the target space to gradually decrease over time.

[0122] Therefore, this embodiment sets model update conditions. When the model update conditions are met, the physical constraint neural network at the current moment needs to be retrained based on the measured data collected by the DAS system during the historical period corresponding to the current moment, so as to update the physical constraint neural network at the current moment. This allows the updated physical constraint neural network to learn the propagation characteristics of the sound from the incidental sound source in the target space, thereby improving the prediction accuracy of the model.

[0123] Here, the historical period corresponding to the current moment is, for example, a historical period with the current moment as the latest moment. Figure 4 The diagram shows a continuous data record of DAS data. The vertical axis represents the DAS channel number, and the horizontal axis represents time (month-day hour:minute). For example, the current time is "10-08 16:00" (October 8th, 16:00), and the corresponding target historical time period is, for example, the period from "10-08 08:00" (October 8th, 8:00) to "10-08 16:00" (October 8th, 16:00).

[0124] Here, the model update conditions may include, for example, reaching a preset update cycle, or detecting a new sound source entering the target scene based on DAS measured data. Alternatively, other update conditions may be set, which are not limited in this embodiment.

[0125] Thus, by employing an incremental time-stepping training strategy, acoustic observation data is gradually input into PINN in chronological order, updating network parameters and sound source locations in real time, thereby achieving online sound source localization and sound field updates. Compared to a one-time training strategy, this approach can quickly locate moving sound sources with a limited amount of time-slice data, while simultaneously improving training stability and real-time performance.

[0126] like Figure 5 The diagram illustrates the different effects of the simultaneous training strategy with full data and the incremental time-step training strategy used in the embodiments of this disclosure.

[0127] This figure illustrates a comparison between the sound field reconstruction results and the true values ​​after 1000-5000 training rounds under two different training strategies. The horizontal and vertical axes of each grid represent the x and y coordinates of the result, with the output results corresponding to different timestamps from left to right. It is evident that the incremental time-step training strategy achieves better training results for physically constrained neural networks.

[0128] like Figure 6The diagram illustrates the results of two different training strategies. It compares the sound field reconstruction results with the true values ​​after 1000-5000 training rounds under each strategy. The horizontal and vertical axes of each grid represent the x and y coordinates of the result, with the output results corresponding to different timestamps from left to right.

[0129] The image in the top left corner shows a 15m x 15m room with 28 DAS channels. The red line indicates the fiber optic deployment. Footsteps are used as the acoustic signal source, and the channels are labeled from 0 to 27. The remaining images show the convergence of the normalized x and y coordinates of the target location over 10,000 training epochs using an incremental time-step training strategy. The horizontal axis represents the training epoch, and the vertical axis represents the x and y values. Each graph corresponds to a specific segment recorded by the DAS; the blue line represents x, and the orange line represents y.

[0130] This disclosure also provides specific examples of verifying the functionality of a physical constraint neural network trained based on the above method.

[0131] Example 1: Two-dimensional verification: 320 DAS sampling points were deployed in a square room with dimensions of 1m × 1m. Velocity of sound. ,density The initial position of the sound source is set to .

[0132] A PINN network with two layers and 64 neurons per layer was used, with tanh as the activation function.

[0133] The results show that the incremental time step training strategy can effectively suppress the convergence error of the symmetric solution, and the positioning error is less than 0.02m.

[0134] Example 2: 3D verification: 512 DAS channels were deployed in a 1m×1m×1m cubic room, using three fiber array layouts: edge, ground, and volume distribution.

[0135] After 2000 rounds of training, the volume distribution array combined with the incremental time step strategy can achieve sound source localization with an error of less than 1% in three-dimensional space, verifying the scalability of the scheme.

[0136] Example 3: A 2 km fiber optic network was deployed in a real building, and sampling was performed with a spatial resolution of 1.6 m. Using this method, the footsteps of pedestrians walking in a 15 m × 15 m room can be accurately reconstructed within 10,000 training cycles, achieving privacy-friendly indoor activity monitoring.

[0137] like Figure 7 As shown, where Figure 7 The left side of the diagram shows a schematic of how fiber optic cables are deployed in a real building. The total length of the fiber optic cable is 2000 meters, with a gauge length of 3.2 meters, a spatial resolution of 1.6 meters, 1250 channels, and a sampling rate of 1000Hz. In this diagram, the red lines represent the distributed fiber optic cables in the DAS system; the blue dots on the distributed fiber optic cables represent different channels. Figure 7 The right side of the image records the sound source signal received by the optical fiber. The horizontal axis represents the time corresponding to the data, and the vertical axis represents the optical fiber position corresponding to the data.

[0138] This invention proposes a sound source localization and sound field reconstruction method based on distributed optical fiber acoustic sensing (DAS) and physically constrained neural networks (PINN), which has the following significant advantages compared to existing technologies: (a) Achieving high-precision sound source localization: By embedding the acoustic wave equation and impedance boundary condition into the loss function of a neural network, sound source localization and spatiotemporal sound field reconstruction are achieved. Unlike existing purely data-driven or traditional algorithmic methods, the network simultaneously satisfies physical constraints, improving prediction accuracy and physical consistency.

[0139] By embedding acoustic wave equations and impedance boundary conditions into the PINN network, the predicted sound field strictly follows physical laws, thereby significantly improving the accuracy of sound source localization.

[0140] Compared with existing methods that rely solely on dense sensors, this invention can still accurately locate single or multiple sound sources in two-dimensional or three-dimensional space under limited DAS channels, reducing positioning errors and improving reliability.

[0141] (II) High-fidelity spatiotemporal sound field reconstruction: By using PINN to train DAS observation data with physical constraints, the spatial distribution and temporal evolution of the sound field can be recovered simultaneously, enabling accurate reconstruction of sound wave propagation, reflection, and multi-source interference.

[0142] By employing an incremental time-stepping training strategy, the continuous sound field reconstruction performance in complex indoor environments outperforms traditional data-driven networks, avoiding reconstruction errors caused by boundary reflections.

[0143] (iii) Reduce dependence on labeled data and improve generalization ability: Within a training framework that integrates distributed fiber acoustic sensing (DAS) data, network training directly utilizes fiber strain rate measurements acquired by the DAS system, reducing reliance on manually labeled data. Compared to existing methods, this significantly improves the model's generalization ability and adaptability in complex indoor environments. Furthermore, embedding physical constraints into the loss function allows the network to converge during training without requiring extensive manually labeled sound source data.

[0144] Experiments show that even with limited observation channels and a small amount of time-slice data, PINN can still accurately predict sound pressure distribution and locate sound sources, significantly improving the model's generalization ability and robustness compared to traditional methods.

[0145] (iv) Improve computational efficiency and real-time performance: Acoustic observation data is sequentially input into PINN over time, updating network parameters and sound source locations in real time to achieve online sound source localization and sound field updates. Compared to a one-time training strategy, this approach can quickly locate moving sound sources with limited time-slice data, while improving training stability and real-time performance.

[0146] By using an incremental time-step training strategy, new data can be gradually input and the network parameters θ and sound source location X can be optimized in real time, enabling online sound source localization and sound field updates.

[0147] For moving sound sources (such as human walking), tracking can be completed in fewer training rounds, significantly reducing computation time compared to traditional one-time training methods, thus meeting the needs of real-time monitoring.

[0148] (v) Adapting to diverse fiber optic deployment schemes: This method is adaptable to fiber optic deployment along walls, floors, or in three-dimensional space (volume array), achieving effective sound source localization and sound field reconstruction. Optimizing fiber optic layout can further improve localization accuracy and reconstruction quality, enhancing the flexibility and feasibility of system deployment.

[0149] (vi) Technological innovation and application prospects: By combining DAS data with physically constrained neural networks, a new indoor acoustic monitoring technology has been developed, which breaks through the high dependence of traditional methods on sensor density and deployment location, and solves the problem of multi-source localization and sound field reconstruction in complex environments.

[0150] The technology can be widely applied to scenarios such as intelligent building monitoring, structural health monitoring, indoor positioning, security and crowd activity sensing, and has significant technical, economic and social benefits.

[0151] In summary, by combining physically constrained PINN and DAS data, this invention not only solves the technical problems of multi-source localization and sound field reconstruction in complex indoor environments, but also significantly improves accuracy, efficiency, data dependence, and deployment flexibility.

[0152] Corresponding to the embodiments of the aforementioned sound source localization method, this disclosure also provides embodiments of a sound source localization device.

[0153] Embodiments of the sound source localization device disclosed herein can be applied to computer devices. The device embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the computer device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 8 The diagram shown is a hardware structure diagram of a computer device containing the sound source localization device of this disclosure, except... Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, the computer device in which the device is located in the embodiment may also include other hardware depending on the actual function of the sound source localization device, which will not be described in detail here.

[0154] Please refer to Figure 9 The sound source localization device provided in this embodiment includes: The generation module 91 is used to generate input data based on the spatial range corresponding to the target space and the prediction time. Processing module 92 is used to input the input data into a pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the predicted time; wherein, the physical constraint neural network uses the acoustic wave equation as the physical equation, the initial sound pressure distribution and the initial sound pressure rate of change distribution as the initial conditions, the impedance boundary conditions as the boundary conditions, and uses the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as data constraint training; The determination module 93 is used to determine the location information of the target sound source emitting the sound within the target space based on the sound pressure information.

[0155] Optionally, the generation module 91, when constructing input data based on the spatial range corresponding to the target space and the prediction time, is used to: Based on the location density of the sound source localization, the target space is divided into multiple spatial grids, and corresponding location information is determined for each spatial grid. Based on the location information and prediction time corresponding to each spatial grid, input data corresponding to each spatial grid is constructed.

[0156] Optionally, it also includes: a training module 94 for training the physical constraint neural network in the following manner: Based on the measured data collected and measured at the sampling points by the distributed acoustic sensing (DAS) system deployed in the target space; and by sampling the spatial domain and the target time domain of the target space to obtain the physical equation sampling points, boundary condition sampling points, and initial condition sampling points; wherein, the sampling points include: sampling location and sampling time; The sampling points are input into the physical constraint neural network to be trained to obtain the predicted sound pressure corresponding to the sampling points; The partial differential equation loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation; the boundary condition loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the boundary condition and the impedance boundary condition; the initial condition loss, serving as a physical constraint, is determined based on the predicted sound pressure corresponding to the sampling point of the initial condition and the initial condition; and the data loss, serving as a data constraint, is determined based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data. The physical constraint neural network to be trained is trained with the goal of minimizing the sum of the partial differential equation loss, the boundary condition loss, the initial condition loss, and the data loss, to obtain a trained physical constraint neural network.

[0157] Optionally, the training module 94, when determining the partial differential equation loss as a physical constraint based on the predicted sound pressure corresponding to the sampling point of the physical equation and the acoustic wave equation, is used to: Substitute the predicted sound pressure corresponding to the sampling points of multiple physical equations into the acoustic wave equation to determine the wave equation residuals corresponding to the sampling points of multiple physical equations. The mean square loss of the wave equation residuals corresponding to the sampling points of multiple physical equations is determined, and the mean square loss of the wave equation residuals is used as the loss of the partial differential equation.

[0158] Optionally, the training module 94, when determining the boundary condition loss as a physical constraint based on the predicted sound pressure corresponding to the boundary condition sampling point and the impedance boundary condition, is used to include: Substitute the predicted sound pressure corresponding to multiple boundary condition sampling points into the impedance boundary equation used to describe the impedance boundary conditions, determine the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points, and determine the mean square loss of the impedance boundary condition residuals corresponding to the multiple boundary condition sampling points. Use the mean square loss of the impedance boundary condition residuals as the boundary condition loss.

[0159] Optionally, the initial condition sampling points include: a first initial condition sampling point and a second initial condition sampling point; the initial condition loss includes: a first initial condition loss and a second initial condition loss; The training module 94, when determining the initial condition loss as a physical constraint based on the predicted sound pressure corresponding to the initial condition sampling point and the initial conditions, is used for: The initial sound pressure information corresponding to the initial condition sampling point is determined based on the initial sound pressure distribution, and the rate of change information corresponding to the initial condition sampling point is determined based on the rate of change distribution of the initial sound pressure. The difference between the initial sound pressure information corresponding to the first initial condition sampling point and the predicted sound pressure corresponding to the first initial condition sampling point is determined as the first initial condition residual. The mean square loss of the first initial condition residuals corresponding to the multiple first initial condition sampling points is determined as the first initial condition loss. Furthermore, the difference between the derivative of the predicted sound pressure with respect to time corresponding to the second initial condition sampling point and the rate of change information corresponding to the second initial condition sampling point is determined as the second initial condition residual, and the mean square loss of the second initial condition residuals corresponding to the multiple second initial condition sampling points is determined as the second initial condition loss.

[0160] Optionally, the training module 94, when determining the data loss as a data constraint based on the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data, is used to: Based on the predicted sound pressure corresponding to the measured sampling point, the Young's modulus of the optical fiber in the DAS system, and the Poisson's ratio, the predicted strain rate corresponding to the measured sampling point is determined under the predicted sound pressure corresponding to the measured sampling point. The difference between the predicted strain rate corresponding to the measured sampling point and the measured data is determined as the data residual; The mean square loss of the data residuals corresponding to the multiple measured sampling points is determined, and the mean square loss is determined as the data loss.

[0161] Optionally, the training module 94 is further configured to: In response to the achievement of model update conditions, the measured data collected by the DAS system for the target historical period at the current time is obtained; Based on the measured data, the physical constraint neural network at the current moment is updated.

[0162] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0163] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0164] This disclosure also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the sound source localization method described in the above method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0165] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the sound source localization method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0166] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0167] The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, the computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; or an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium may be a volatile or non-volatile storage medium, or may include both volatile and non-volatile types of storage media.

[0168] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method of acoustic source localization, the method comprising: The method comprises: based on the corresponding space range of the target space and the prediction time, the input data is constituted; input the input data into the pre-trained physical constraint neural network to obtain the sound pressure information of the sound in the target space at the prediction time; wherein, the physical constraint neural network takes the acoustic wave equation as the physical equation, takes the initial sound pressure distribution and the change rate distribution of the initial sound pressure as the initial condition, takes the impedance boundary condition as the boundary condition, and utilizes the measured data collected by the distributed acoustic sensing DAS system deployed in the target space as the data constraint to obtain the trained result; based on the sound pressure information, the position information of the target sound source in the target space is determined.

2. The method of claim 1, wherein, The input data is constituted based on the corresponding space range of the target space and the prediction time, which comprises: according to the positioning density of the sound source positioning, the target space is divided into a plurality of space grids, and the corresponding position information of each space grid is determined; according to the position information of each space grid and the prediction time, the input data corresponding to each space grid is constituted.

3. The method according to claim 1 or 2, characterized in that, The physical constraint neural network is trained in the following manner: based on the measured data corresponding to the measured sampling points collected by the distributed acoustic sensing DAS system deployed in the target space; and sampling the spatial domain and the target time domain of the target space to obtain physical equation sampling points, boundary condition sampling points, and initial condition sampling points; wherein, the sampling points include sampling positions and sampling times; input the sampling points into the physical constraint neural network to be trained to obtain the predicted sound pressure corresponding to the sampling points; determine the partial differential equation loss as the physical constraint according to the predicted sound pressure corresponding to the physical equation sampling points and the acoustic wave equation, determine the boundary condition loss as the physical constraint according to the predicted sound pressure corresponding to the boundary condition sampling points and the impedance boundary condition, determine the initial condition loss as the physical constraint according to the predicted sound pressure corresponding to the initial condition sampling points and the initial condition, and determine the data loss as the data constraint according to the predicted sound pressure corresponding to the measured sampling points and the corresponding measured data; train the physical constraint neural network to be trained to obtain the trained physical constraint neural network, aiming at minimizing the sum of the partial differential equation loss, the boundary condition loss, the initial condition loss, and the data loss.

4. The method of claim 3, wherein, The partial differential equation loss as the physical constraint is determined according to the predicted sound pressure corresponding to the physical equation sampling points and the acoustic wave equation, which comprises: substitute the predicted sound pressure corresponding to a plurality of physical equation sampling points into the acoustic wave equation respectively to determine the wave equation residual error corresponding to each of the plurality of physical equation sampling points; determine the mean square loss of the wave equation residual error corresponding to each of the plurality of physical equation sampling points, and take the mean square loss of the wave equation residual error as the partial differential equation loss.

5. The method of claim 3, wherein, The boundary condition loss as the physical constraint is determined according to the predicted sound pressure corresponding to the boundary condition sampling points and the impedance boundary condition, which comprises: The predicted sound pressures corresponding to the plurality of boundary condition sampling points are substituted into an impedance boundary equation for describing an impedance boundary condition, respectively, to determine impedance boundary condition residuals corresponding to the plurality of boundary condition sampling points, respectively, and to determine a mean square loss of the impedance boundary condition residuals corresponding to the plurality of boundary condition sampling points, respectively, and the mean square loss of the impedance boundary condition residuals is taken as the boundary condition loss.

6. The method of claim 3, wherein, The initial condition sampling points include first initial condition sampling points and second initial condition sampling points, and the initial condition loss includes first initial condition loss and second initial condition loss. The initial condition loss as a physical constraint is determined according to the predicted sound pressure corresponding to the initial condition sampling point and the initial condition, including: Initial sound pressure information corresponding to the initial condition sampling point is determined according to the initial sound pressure distribution, and rate information corresponding to the initial condition sampling point is determined according to the rate of change of the initial sound pressure distribution. A difference between the initial sound pressure information corresponding to the first initial condition sampling point and the predicted sound pressure corresponding to the first initial condition sampling point is determined as a first initial condition residual, and a mean square loss of the first initial condition residuals corresponding to the plurality of first initial condition sampling points is determined, and the mean square loss of the first initial condition residuals corresponding to the plurality of first initial condition sampling points is determined as the first initial condition loss. A difference between a derivative of the predicted sound pressure with respect to time corresponding to the second initial condition sampling point and the rate information corresponding to the second initial condition sampling point is determined as a second initial condition residual, and a mean square loss of the second initial condition residuals corresponding to the plurality of second initial condition sampling points is determined, and the mean square loss of the second initial condition residuals corresponding to the plurality of second initial condition sampling points is determined as the second initial condition loss.

7. The method of claim 3, wherein, The data loss as a data constraint is determined according to the predicted sound pressure corresponding to the measured sampling point and the corresponding measured data, including: A predicted strain rate corresponding to the measured sampling point under the predicted sound pressure corresponding to the measured sampling point is determined according to the predicted sound pressure corresponding to the measured sampling point, a Young's modulus of the optical fiber in the DAS system, and a Poisson's ratio; A difference between the predicted strain rate corresponding to the measured sampling point and the measured data is determined as a data residual; A mean square loss of the data residuals corresponding to the plurality of measured sampling points is determined, and the mean square loss is determined as the data loss.

8. The method of claim 1, wherein, The method further includes: In response to reaching a model update condition, obtaining measured data collected by the DAS system corresponding to a target historical period at a current time; Based on the measured data, updating the physical constraint neural network at the current time.

9. A sound source positioning apparatus characterized by comprising: The device includes: The generation module is configured to generate input data based on a spatial range corresponding to a target space and a predicted time. The processing module is configured to input the input data into a pre-trained physically-constrained neural network to obtain sound pressure information of the sound in the target space at the predicted time; wherein the physically-constrained neural network takes an acoustic wave equation as a physical equation, takes an initial sound pressure distribution and a change rate distribution of the initial sound pressure as initial conditions, takes an impedance boundary condition as a boundary condition, and is trained by using measured data collected by a distributed acoustic sensing (DAS) system deployed in the target space as data constraints; The determining module is configured to determine, based on the sound pressure information, position information of a target sound source emitting the sound in the target space.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps of the method of any one of claims 1-8.

11. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor, when executing the program, implements the following steps: constructing input data based on a spatial range corresponding to a target space and a predicted time; The processing module is configured to input the input data into a pre-trained physically-constrained neural network to obtain sound pressure information of the sound in the target space at the predicted time; wherein the physically-constrained neural network takes an acoustic wave equation as a physical equation, takes an initial sound pressure distribution and a change rate distribution of the initial sound pressure as initial conditions, takes an impedance boundary condition as a boundary condition, and is trained by using measured data collected by a distributed acoustic sensing (DAS) system deployed in the target space as data constraints; The determining module is configured to determine, based on the sound pressure information, position information of a target sound source emitting the sound in the target space.

Images