Audio indoor positioning base station system based on multi-parameter fusion

Abstract

The application relates to the technical field of electronic communication, and particularly provides an audio indoor positioning base station system based on multi-parameter fusion. The system comprises an audio signal collection, preprocessing and feature extraction module, a local lightweight fusion module, a cloud collaborative computing engine and a positioning result output module. Through dynamic unloading of a computing load to the cloud and adoption of a lightweight neural network model for local preliminary positioning, intelligent switching of a cloud collaborative mode and a local autonomous mode is realized in combination with a dynamic load decision unit, so that the positioning precision and system robustness are guaranteed while the base station computing complexity and power consumption are reduced.

Description

Technical Field

[0001] This invention belongs to the field of electronic communication technology, and specifically relates to an audio indoor positioning base station system based on multi-parameter fusion. Background Technology

[0002] Indoor positioning technology, as one of the core technologies in the fields of IoT and smart services, plays a crucial role in scenarios such as smart warehousing, personnel guidance, and equipment tracking. Audio signal-based positioning methods have become a research hotspot due to their low deployment cost, immunity to light conditions, and high accuracy. Audio-based indoor positioning systems typically deploy multiple positioning base stations in an indoor environment, utilizing parameters such as the time difference of arrival (TDO) or signal strength of audio signals to calculate the target location.

[0003] The multi-parameter fusion-based audio localization method aims to improve localization accuracy and robustness in complex indoor environments by comprehensively processing various features such as sound wave arrival time, phase difference, and signal strength. The core of this method lies in designing an efficient data fusion algorithm to jointly optimize and estimate the state of multi-source heterogeneous audio observation information.

[0004] In existing technologies, multi-parameter fusion algorithms typically involve high-dimensional state space modeling and iterative optimization processes, such as using extended Kalman filtering or particle filtering for position tracking, which leads to increased computational complexity.

[0005] However, existing positioning base stations generally use resource-constrained embedded processors, whose limited computing power and memory bandwidth make it difficult to support the real-time operation of complex fusion algorithms. Especially in scenarios that require high-frequency location updates or simultaneous processing of multiple target signals, system latency is aggravated, making it impossible to meet the response requirements of real-time interactive applications.

[0006] Furthermore, high computational loads lead to increased base station power consumption and heat dissipation issues, further limiting the feasibility of system deployment and long-term stability in real-world environments. Therefore, achieving efficient, low-latency multi-parameter audio fusion positioning under embedded resource constraints has become a pressing technical challenge in this field. Summary of the Invention

[0007] The technical problem this invention aims to solve is to overcome the shortcomings of existing audio indoor positioning base stations, which suffer from high system latency, high power consumption, and limited deployment feasibility due to the difficulty in running complex multi-parameter fusion algorithms in real time under embedded resource constraints. To achieve the above objective, this invention provides an audio indoor positioning base station system based on multi-parameter fusion. This system reduces the real-time computational burden on the core processor while ensuring positioning accuracy by constructing a dynamic offloading and hierarchical fusion architecture.

[0008] This invention provides an indoor audio positioning base station system based on multi-parameter fusion. The system includes an audio signal acquisition module, a signal preprocessing and feature extraction module, a local lightweight fusion module, a cloud collaborative computing engine, and a positioning result output module.

[0009] The audio signal acquisition module is responsible for receiving audio signals emitted from the positioning tag through a microphone array.

[0010] The signal preprocessing and feature extraction module is connected to the audio signal acquisition module. It is used to filter and reduce noise in the received raw audio signal and extract three core observation features in parallel: sound wave arrival time, carrier phase difference, and signal strength indication value.

[0011] The local lightweight fusion module is connected to the signal preprocessing and feature extraction module. It integrates a pre-trained lightweight neural network model, which takes the three types of observation features extracted above as input and directly outputs the preliminary position estimate coordinates of the target and its corresponding confidence level.

[0012] The cloud-based collaborative computing engine establishes a communication connection with the local lightweight fusion module via wired or wireless networks. The engine is configured to receive preliminary location estimates and confidence data from multiple base stations and run a high-precision fusion optimization algorithm. This algorithm performs global optimization and correction of the preliminary location estimates based on the observation data of all base stations.

[0013] The positioning result output module connects to the cloud-based collaborative computing engine to receive the final optimized positioning coordinates and encapsulate them into a data packet of a specified format for external transmission.

[0014] Furthermore, the specific workflow of the signal preprocessing and feature extraction module includes: First, a high-pass filter with a cutoff frequency of 8 kHz is used to preprocess the original audio signal to remove low-frequency environmental noise; Next, the time difference of arrival of the sound wave is calculated using the generalized cross-correlation phase transform algorithm, with the calculation window length of the algorithm set to 512 sampling points; Simultaneously, a digital phase-locked loop structure is used to track and extract the phase difference information of the audio carrier; The signal strength indication value is obtained by calculating the average power of the signal over the entire cycle.

[0015] Furthermore, the lightweight neural network model integrated in the local lightweight fusion module is a 4-layer fully connected network optimized through pruning and quantization operations.

[0016] The network has 9 input layer nodes, corresponding to the observation values ​​of 3 observation features on 3 different base stations; 32 and 16 nodes in the 2 hidden layers, respectively; and 3 output layer nodes, corresponding to the target's two-dimensional plane coordinates x and y and the confidence score, respectively.

[0017] All weight parameters of the network model are stored in 8-bit integer format, and the inference process is completed entirely within the microcontroller unit of the base station without calling external floating-point arithmetic units.

[0018] Furthermore, the high-precision fusion optimization algorithm run by the cloud-based collaborative computing engine is specifically an iterative nearest neighbor matching algorithm based on a sliding window.

[0019] The algorithm maintains a sliding window of length 10 to store the preliminary location estimation sequence from all base stations for the most recent 10 time steps; For each new data moment, the algorithm first associates and matches it with the historical data within the sliding window, then constructs a 6-dimensional state vector about the target's position and velocity, and updates the optimal estimate of the state vector by minimizing the weighted sum of squared residuals of all associated observations. The weighting coefficients are dynamically determined by the confidence scores reported by each base station.

[0020] Furthermore, the system also includes a dynamic load decision unit.

[0021] This unit monitors the processor load and network communication latency of the local lightweight fusion module in real time.

[0022] When the processor load rate is less than 70% and the network latency is less than 50 milliseconds, the system defaults to running the above-mentioned cloud collaboration fusion mode. When the processor load rate is greater than 70% or the network latency is greater than 50 milliseconds, the dynamic load decision unit will trigger the local autonomous mode. In this mode, the positioning result output module will directly use the preliminary position estimate output by the local lightweight fusion module as the final result for output, and at the same time, it will suspend sending data to the cloud collaborative computing engine.

[0023] Furthermore, in the local autonomous mode, the local lightweight fusion module also performs a Kalman filter smoothing operation based on historical trajectories.

[0024] This operation utilizes the target's position history over the last five moments to construct a uniform motion model, and performs a one-step prediction and correction on the preliminary position estimate at the current moment, in order to reduce positioning jumps caused by observation noise at a single moment.

[0025] Furthermore, the microphone array used in the audio signal acquisition module consists of four digital microphones arranged in a 2x2 square array, with a physical aperture of 8 centimeters.

[0026] The array configuration supports preliminary sound source direction estimation based on beamforming technology, and the direction estimation result is sent as an auxiliary feature to the signal preprocessing and feature extraction module.

[0027] Furthermore, the positioning result output module supports two data output protocols: User Datagram Protocol (UDP) and Transmission Control Protocol (TCP).

[0028] The data packet is encapsulated using a lightweight data exchange format based on JavaScript object representation. The packet contains a target identifier, timestamp, two-dimensional coordinate data, and a data source identifier. The data source identifier is used to distinguish whether the result is a cloud-optimized result or a locally autonomous result.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention deploys a computationally intensive, high-precision fusion optimization algorithm on a cloud-based collaborative computing engine, while retaining only a deeply optimized, lightweight neural network model on the local base station for initial positioning calculations. This achieves an effective distribution of computational load between the edge and the cloud. This layered processing architecture eliminates the need for resource-constrained embedded base stations to undertake complex global optimization calculations, thereby reducing the real-time computational complexity and power consumption of individual base stations and ensuring low-latency response capabilities in high-frequency location update scenarios.

[0030] 2. The dynamic load decision unit introduced in this invention can intelligently switch between cloud-based collaborative mode and local autonomous mode based on the real-time system status. This design endows the system with the ability to adapt to network fluctuations and local computing resource constraints. Even when the network is interrupted or cloud services are unavailable, the system can still provide available location services through local autonomous mode, greatly enhancing the robustness and reliability of the system in complex real-world deployment environments.

[0031] 3. The local lightweight fusion module employs an 8-bit integer quantization neural network model, which significantly reduces the model's memory footprint and computational resource requirements, enabling high-performance data fusion algorithms to run efficiently on low-cost microcontrollers. Meanwhile, the cloud-based collaborative computing engine utilizes a sliding window iterative optimization algorithm that comprehensively leverages historical and current observation data from multiple base stations to effectively smooth observation noise and correct local errors. This achieves positioning accuracy at the system level comparable to complex filtering algorithms, striking a good balance between accuracy and efficiency. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall technical architecture of the audio indoor positioning base station system based on multi-parameter fusion proposed in this invention; Figure 2This is a schematic diagram of the core principle framework of the dynamic offloading and layered fusion architecture of computing load in this invention; Figure 3 This is a logical flowchart of the signal preprocessing and feature extraction module in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the local lightweight fusion module and the cloud collaborative computing engine in this invention; Figure 5 This is a flowchart illustrating the logical flow of the dynamic load decision-making unit in this invention, which switches between cloud-based collaborative mode and local autonomous mode. Detailed Implementation

[0033] Please refer to the attached document. Figures 1 to 5 This embodiment details a specific implementation of an indoor audio positioning base station system based on multi-parameter fusion. The system comprises an audio signal acquisition module, a signal preprocessing and feature extraction module, a local lightweight fusion module, a cloud-based collaborative computing engine, a positioning result output module, and a dynamic load decision unit. These modules achieve functional collaboration through electrical connections or data communication links.

[0034] The audio signal acquisition module uses a microphone array consisting of four digital microphones arranged in a 2x2 square array, with a physical aperture of 8 centimeters.

[0035] The array is connected to the base station's microcontroller unit via an integrated circuit bus interface. Each microphone in the array has a sampling rate of 48 kHz and a quantization bit depth of 16 bits.

[0036] The array receives linear frequency modulated audio signals in the 19 kHz to 21 kHz frequency band transmitted from the positioning tag and generates preliminary sound source direction estimates through a beamforming algorithm.

[0037] The direction estimate is transmitted as an auxiliary feature to the signal preprocessing and feature extraction module.

[0038] The array continuously monitors the ambient background noise intensity during operation, and automatically triggers the gain adjustment circuit when the noise power exceeds -65 dB.

[0039] Please refer to the appendix. Figure 3 The signal preprocessing and feature extraction module contains a three-stage parallel processing pipeline.

[0040] The first stage is the signal conditioning unit, which uses a Butterworth high-pass filter with a cutoff frequency of 8 kHz to preprocess the original audio signal. The filter order is 4th and the stopband attenuation rate is 24 dB per octave.

[0041] The second level is the feature extraction unit, which contains three independent computational cores: The core of the sound wave arrival time calculation adopts the generalized cross-correlation phase transformation algorithm, with a calculation window length of 512 sampling points and a time delay search range of -256 to +256 sampling points, corresponding to a time resolution of 10.67 microseconds; The carrier phase difference extraction core adopts a digital phase-locked loop structure with a phase-locked loop bandwidth of 200 Hz, a damping coefficient of 0.707, and a maximum trackable frequency deviation of ±500 Hz. The core of the signal strength indicator calculation is to measure the average power of the signal over 128 consecutive sampling points. The power calculation uses a true RMS detection algorithm.

[0042] The third level is the feature calibration unit, which performs temperature compensation and nonlinear correction on the three types of features. The ambient temperature data collected by the temperature sensor is used to correct the sound wave propagation speed, and the correction coefficients are stored in a lookup table in the read-only memory.

[0043] The local lightweight fusion module integrates a 4-layer fully connected neural network model that has been pruned and quantized for optimization.

[0044] The model input layer has 9 nodes, corresponding to the observation values ​​of 3 observation features on 3 different base stations; The first hidden layer has 32 nodes and the activation function is the rectified linear unit. The second hidden layer has 16 nodes and uses the hyperbolic tangent function as the activation function; the output layer has 3 nodes, which correspond to the target's two-dimensional plane coordinates x-axis, y-axis, and confidence score, respectively.

[0045] All weight parameters of the model are stored in flash memory in 8-bit integer format, and the inference process is completed by the fixed-point digital signal processor core of the microcontroller unit.

[0046] When the model is executed, the input features are first normalized. The normalization parameters include the mean and standard deviation of each feature. These parameters are pre-stored in a specific area of ​​random access memory.

[0047] The model inference cycle is strictly controlled within 5 milliseconds, and the coordinate data and confidence score in the output buffer are updated after each inference is completed.

[0048] Please refer to the attached document. Figure 4 The cloud-based collaborative computing engine is deployed on a remote server cluster and establishes network connections with multiple base stations through Transmission Control Protocol or User Datagram Protocol.

[0049] The engine runs an iterative nearest neighbor matching algorithm based on a sliding window, which maintains a sliding window of length 10 to store the historical position sequence.

[0050] For each newly arrived preliminary location estimation packet, the algorithm performs the following operations: First, perform data validity verification, checking the continuity of timestamps and the reasonableness of coordinates; Next, data association and matching are performed, using Mahalanobis distance to measure the similarity between the current observation and the historical trajectory, with a distance threshold set to 2.5; then, a 6-dimensional state vector is constructed, including the x-axis component of the position coordinate, the y-axis component of the position coordinate, the x-axis component of the velocity vector, the y-axis component of the velocity vector, the x-axis component of the acceleration vector, and the y-axis component of the acceleration vector. Finally, the state estimate is updated by minimizing the weighted sum of squared residuals, and the objective function is: ; It is the first The weighting coefficients corresponding to each observation data point. It is the first One set of raw observation data. It is a state vector The observation function.

[0051] The weighting coefficients are obtained by linearly mapping the confidence scores reported by each base station, with the mapping relationship being the confidence score divided by 100. The optimization process uses the Levenburg-Marquardt algorithm, and the iteration terminates when the residual change is less than 1e-6 or the number of iterations reaches 50. The optimized coordinate data is returned to the positioning result output module through a secure socket layer encrypted channel.

[0052] The positioning result output module includes a data encapsulation unit and a protocol adaptation unit.

[0053] The data encapsulation unit organizes the received coordinate data according to the JavaScript object representation lightweight data exchange format. The data packet includes a target identifier field, a timestamp field, a coordinate x-axis value field, a coordinate y-axis value field, and a data source identifier field.

[0054] The target identifier is 8 bytes long, and the timestamp uses Coordinated Universal Time (UTC) format with millisecond precision.

[0055] The protocol adaptation unit supports two transmission modes: User Datagram Protocol (UDP) and Transmission Control Protocol (TCP). In UDP mode, the data packet transmission interval is 20 milliseconds, while in TCP mode, a persistent connection is established and a data acknowledgment mechanism is enabled.

[0056] The module has an internal sending buffer with a capacity of 100 data packets. When the buffer usage rate is greater than 80%, the data compression algorithm is automatically triggered.

[0057] The dynamic load decision unit monitors the system's operating status in real time, collecting metrics including processor load rate, network round-trip latency, and memory usage of the local lightweight fusion module.

[0058] The processor load rate is obtained through the performance counter of the microcontroller unit, with a sampling period of 100 milliseconds; Network latency is measured by sending heartbeat packets at a frequency of 10 times per second; memory usage is calculated by monitoring the stack pointer position.

[0059] When the three conditions of processor load rate being less than 70%, network latency being less than 50 milliseconds, and memory utilization being less than 85% are met simultaneously, the system runs in cloud collaboration mode. When any condition is not met, the system switches to local autonomous mode.

[0060] The mode switching instruction is implemented through the system interrupt mechanism, and the switching process can be completed within 2 milliseconds.

[0061] In local autonomous mode, the local lightweight fusion module enables Kalman filtering smoothing based on historical trajectories.

[0062] This operation maintains a historical position queue of 5 frames and uses a uniform motion model for state prediction.

[0063] The state vector contains four components: position and velocity. The observation vector is the 2D coordinates output by the local neural network.

[0064] The diagonal elements of the process noise covariance matrix are set to 0.1, while the diagonal elements of the observation noise covariance matrix are dynamically adjusted based on the confidence score, with the adjustment coefficient being 1 minus the confidence score divided by 100.

[0065] The filtered output is updated every 10 milliseconds. When the confidence level of three consecutive observations is less than 0.6, the process noise covariance is automatically increased to enhance the system robustness.

[0066] The system power supply management adopts a hierarchical power consumption control strategy.

[0067] The audio signal acquisition module automatically enters a low-power mode when there is no signal, reducing power consumption to 20% of normal operation. The signal preprocessing and feature extraction module dynamically adjusts the clock frequency according to the feature update frequency, with a frequency adjustment range of 48 MHz to 12 MHz. The local lightweight fusion module enters a dormant state immediately after completing the inference, until the next processing cycle begins.

[0068] The average power consumption of the entire base station under typical operating conditions is 350 milliwatts, and the peak power consumption is less than 850 milliwatts.

[0069] Site-specific calibration is required during system deployment.

[0070] The calibration process includes collecting environmental background noise samples, measuring sound wave propagation characteristics, and establishing temperature and humidity compensation models.

[0071] Calibration data is stored in encrypted non-volatile memory and the calibration parameters are automatically loaded each time the system starts.

[0072] The base stations achieve time synchronization through a precision clock protocol, with clock deviation controlled within ±1 microsecond, ensuring the time consistency of observation data from multiple base stations.

[0073] This embodiment provides another implementation of an indoor audio positioning base station system based on multi-parameter fusion, focusing on the system optimization scheme under extreme network conditions. When the system detects that the network latency is consistently greater than 100 milliseconds or the network packet loss rate is greater than 10%, it automatically activates the enhanced local autonomous mode.

[0074] In the enhanced local autonomy mode, the local lightweight fusion module is expanded into a dual-model parallel architecture. The main model is the aforementioned 4-layer fully connected network, and the auxiliary model is a 2-layer radial basis function network.

[0075] Both models receive the same input features but employ different fusion strategies. The main model's output serves as the base localization result, while the auxiliary model's output is used for reliability verification.

[0076] When the Euclidean distance between the output coordinates of the two models is less than 0.5 meters, the weighted average of the two results is taken as the final output, with the weight ratio being the ratio of the confidence of the main model to the confidence of the auxiliary model; when the distance is greater than or equal to 0.5 meters, only the output result of the main model is used, and the data is marked as a low reliability observation.

[0077] The signal preprocessing and feature extraction module adds a feature redundancy verification mechanism in enhanced mode.

[0078] Calculate statistical consistency indices for each type of feature, including mean, variance, and autocorrelation coefficient.

[0079] When the statistical indicators of a certain type of feature exceed the preset range, the backup feature extraction algorithm is automatically activated.

[0080] The backup algorithms employ a minimum mean square error adaptive filter for calculating the sound wave arrival time, a Costas ring structure for carrier phase difference extraction, and peak hold detection for signal strength indication calculation. While these backup algorithms are computationally complex, they ensure basic system functionality should the main algorithm fail.

[0081] The dynamic load decision unit introduces a predictive switching mechanism in enhanced mode. An integrated time-series forecasting model predicts processor load and network latency trends for the next 3 seconds based on 120 seconds of historical system state data. The forecasting model employs an autoregressive integral moving average model, with the model order determined using the Akaike information content criterion. When the forecast indicates an impending deterioration in the system state, a mode switching preparation process is initiated 50 milliseconds in advance, including pre-allocating memory resources, establishing local caches, and pausing non-critical tasks.

[0082] The location results output module supports local data caching and delayed upload in enhanced mode. An internal circular data buffer with a capacity of 500 data packets is added. When the network connection is restored, the module retransmits the cached data in chronological order, while also attaching a network interruption time stamp. The data packet encapsulation format adds a sequence number field and a retransmission flag to ensure that the cloud-based collaborative computing engine can correctly reassemble the data sequence.

[0083] In enhanced mode, the system employs an adaptive sampling rate adjustment strategy. When the average confidence level of 10 consecutive observations is less than 0.7, the sampling rate of the audio signal acquisition module is automatically reduced from 48 kHz to 24 kHz, and the computation window length of the signal preprocessing and feature extraction modules is adjusted accordingly to 256 sampling points. This sampling rate adjustment reduces the computational load by approximately 40% while maintaining basic positioning functionality. Once the system returns to normal, the sampling rate gradually recovers to the standard value within 5 seconds.

[0084] Please refer to the appendix. Figure 5 The enhanced local autonomy mode also includes a cross-base station cooperation mechanism. Neighboring base stations establish direct communication links via wireless personal area networks (PANs) to exchange local positioning results and confidence information. Each base station maintains a neighbor base station status table, recording the quality indicators of the last 10 communications. When the confidence of a base station remains consistently low, it can request neighbor base stations to share their intermediate feature extraction results, improving positioning reliability through data fusion. Cooperative communication uses a time-division multiple access (TDMA) approach, with time slot allocation dynamically scheduled by the master base station.

[0085] The system employs a tiered alarm mechanism in enhanced mode. Different levels of alarm information are issued based on the severity of system degradation. Level 1 alarms indicate a slight performance degradation and are only logged in the system log; Level 2 alarms indicate limited functionality and notify maintenance personnel for inspection; Level 3 alarms indicate the system is on the verge of failure and automatically trigger a hardware reset procedure. Alarm thresholds are dynamically adjusted using a machine learning algorithm with a learning cycle of 24 hours.

[0086] The power management system implements emergency power consumption control in enhanced mode. When the battery level is detected to be less than 20%, the cloud-based collaborative communication function is automatically shut down, the monitoring frequency of the dynamic load decision unit is reduced from 10 times per second to 2 times per second, and the historical trajectory smoothing operation is paused. These measures can further reduce system power consumption by 60% and extend the battery life by 2 hours in emergency working conditions.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An indoor audio positioning base station system based on multi-parameter fusion, characterized in that, include: An audio signal acquisition module is used to receive audio signals emitted from positioning tags via a microphone array; The signal preprocessing and feature extraction module is connected to the audio signal acquisition module. It is used to filter and reduce noise in the received raw audio signal and extract three core observation features in parallel: sound wave arrival time, carrier phase difference, and signal strength indication value. The local lightweight fusion module is connected to the signal preprocessing and feature extraction module. It integrates a pre-trained lightweight neural network model. This model takes the extracted three types of observation features as input and directly outputs the preliminary estimated coordinates of the target's position and its corresponding confidence level. The cloud-based collaborative computing engine establishes a communication connection with the local lightweight fusion module via wired or wireless networks. It is used to receive preliminary location estimates and confidence data from multiple base stations and run a high-precision fusion optimization algorithm to globally optimize and correct the preliminary location estimates based on the observation data of all base stations. The positioning result output module connects to the cloud-based collaborative computing engine to receive the final optimized positioning coordinates and encapsulate them into a data packet of a specified format for external transmission. The dynamic load decision unit monitors the processor load rate and network communication latency of the local lightweight fusion module in real time, and switches between cloud collaborative mode and local autonomous mode based on the monitoring results.

2. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The workflow of the signal preprocessing and feature extraction module includes: A high-pass filter is used to preprocess the original audio signal to remove low-frequency ambient noise; The time difference of sound wave arrival is calculated using a generalized cross-correlation phase transformation algorithm. The phase difference information of the audio carrier is tracked and extracted using a digital phase-locked loop structure; The signal strength indication value is obtained by calculating the average power of the signal over a complete cycle.

3. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The lightweight neural network model integrated in the local lightweight fusion module is a 4-layer fully connected network optimized by pruning and quantization operations. The network has 9 input layer nodes, corresponding to the observation values ​​of 3 observation features on 3 different base stations; The two hidden layers have 32 and 16 nodes respectively; The output layer has 3 nodes, which correspond to the x-axis component, y-axis component, and confidence score of the target's two-dimensional planar coordinates, respectively. All weight parameters of the network model are stored in 8-bit integer format, and the inference process is completed entirely within the microcontroller unit of the base station.

4. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The high-precision fusion optimization algorithm run by the cloud-based collaborative computing engine is an iterative nearest neighbor matching algorithm based on a sliding window. The algorithm maintains a sliding window of length 10 to store the preliminary location estimation sequence from all base stations for the most recent 10 time steps; For each new data moment, the algorithm first associates and matches it with the historical data within the sliding window, then constructs a 6-dimensional state vector about the target's position and velocity, and updates the optimal estimate of the state vector by minimizing the weighted sum of squared residuals of all associated observations. The weighting coefficients are dynamically determined by the confidence scores reported by each base station.

5. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The control logic of the dynamic load decision unit includes: When the processor load rate is less than 70% and the network latency is less than 50 milliseconds, the system runs in cloud-based collaborative fusion mode. When the processor load rate is greater than 70% or the network latency is greater than 50 milliseconds, the dynamic load decision unit triggers the local autonomous mode. In local autonomous mode, the positioning result output module directly uses the preliminary position estimate output by the local lightweight fusion module as the final result, while suspending the sending of data to the cloud collaborative computing engine.

6. The audio indoor positioning base station system based on multi-parameter fusion according to claim 5, characterized in that, In local autonomous mode, the local lightweight fusion module also performs Kalman filtering smoothing operation based on historical trajectories; This operation uses the target's position history over the last 5 moments to construct a uniform motion model, and performs a one-step prediction and correction on the preliminary position estimate at the current moment.

7. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The audio signal acquisition module uses a microphone array consisting of four digital microphones arranged in a 2x2 square array, with a physical aperture of 8 centimeters. The array configuration supports preliminary sound source direction estimation based on beamforming technology, and the direction estimation results are sent as auxiliary features to the signal preprocessing and feature extraction module.

8. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The positioning result output module supports two data output protocols: User Datagram Protocol (UDP) and Transmission Control Protocol (TCP). The data packet encapsulation format adopts a lightweight data exchange format based on JavaScript object representation; The package contains a target identifier, timestamp, two-dimensional coordinate data, and a data source identifier. The data source identifier is used to distinguish whether the result is a cloud-optimized result or a local autonomous result.

9. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The dynamic load decision unit also includes a predictive switching mechanism; This mechanism predicts future processor load and network latency trends based on historical system state data. When the forecast indicates that the system state is about to deteriorate, the mode switching preparation process is initiated in advance.

10. The audio indoor positioning base station system based on multi-parameter fusion according to claim 1, characterized in that, The system employs an adaptive sampling rate adjustment strategy in the enhanced local autonomous mode; When the average confidence level of multiple consecutive observations is less than a preset threshold, the sampling rate of the audio signal acquisition module is automatically reduced. Once the system returns to normal, the sampling rate gradually recovers to the standard value.

Images