Sensing data chip method based on multi-source information fusion and dynamic parameter adjustment

Abstract

This invention relates to the field of sensor data processing technology and discloses a sensor data chip method based on multi-source information fusion and dynamic parameter adjustment. The method includes acquiring raw data streams from multiple heterogeneous sensors, performing timestamp alignment and format standardization to generate synchronized sensor data; extracting environmental noise components and dynamically suppressing noise using an adaptive filtering algorithm to obtain denoised sensor data; performing multi-dimensional feature decomposition on the denoised data, calculating time-domain, frequency-domain, and spatial-domain feature vectors, and generating a multi-dimensional feature set; dynamically adjusting the feature weight allocation strategy based on the current system load state and the historical distribution pattern of the multi-dimensional feature set to generate an optimized fusion weight matrix; and weighted aggregation of the multi-dimensional feature set and the optimized fusion weight matrix to output a multi-source fused feature vector. This method can solve the problems of data synchronization, noise suppression, and feature fusion adaptability in multi-source heterogeneous sensor data processing, meeting the processing needs of different scenarios.

Description

Technical Field

[0001] This invention relates to the field of sensor data processing technology, specifically to a sensor data chip method based on multi-source information fusion and dynamic parameter adjustment. Background Technology

[0002] In current sensor technology applications, whether in industrial monitoring, environmental sensing, or smart device interaction, the reliance on the accuracy and effectiveness of sensor data is constantly increasing. However, in real-world applications, sensor systems often need to integrate data from multiple heterogeneous sensors. These sensors differ significantly in their working principles, data formats, and sampling frequencies, making it difficult to directly use the raw data streams collaboratively. For example, temperature sensors and vibration sensors have different output formats, and accelerometers and optical sensors have different sampling periods. If these data are not processed uniformly, subsequent data applications will face serious accuracy issues. Sensor data is highly susceptible to environmental noise interference during acquisition. The types and intensities of noise vary significantly across different application scenarios. Traditional fixed filtering algorithms struggle to dynamically adjust filtering parameters based on noise changes, resulting in unstable denoising performance. Some useful sensor information may be over-filtered, or noise may not be effectively suppressed, thereby affecting the reliability of subsequent data feature extraction.

[0003] In the feature processing stage, existing technologies typically employ fixed feature weight allocation methods, failing to consider real-time changes in system load and historical distribution patterns of feature data. When the system load is high, fixed weight allocation may lead to decreased data processing efficiency and increased system latency. Furthermore, when the distribution patterns of feature data change, fixed weights cannot adapt to these changes in a timely manner, making it difficult for the fused feature vector to accurately reflect the actual sensing situation, thus affecting the final decision-making and application effectiveness. In addition, single-dimensional feature extraction cannot comprehensively cover the effective information in sensor data; features from different dimensions such as the time domain, frequency domain, and spatial domain are often processed in isolation, failing to form a synergistic effect, resulting in low utilization of feature information and further limiting the performance improvement of the sensing system. Summary of the Invention

[0004] The purpose of this invention is to provide a sensor data chip method based on multi-source information fusion and dynamic parameter adjustment to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a sensor data chip method based on multi-source information fusion and dynamic parameter adjustment, the method comprising: Raw data streams from multiple heterogeneous sensors are collected, and the raw data streams are timestamped and standardized to generate synchronized sensor data. The environmental noise component is extracted from the synchronized sensing data, and an adaptive filtering algorithm is used to dynamically suppress the environmental noise component to obtain the denoised sensing data. The denoised sensor data is subjected to multi-dimensional feature decomposition, and time-domain feature vector, frequency-domain feature vector and spatial-domain feature vector are calculated respectively to generate a multi-dimensional feature set; Based on the current system load status and the historical distribution pattern of the multidimensional feature set, the feature weight allocation strategy is dynamically adjusted to generate an optimized fusion weight matrix. The multi-dimensional feature set is weighted and aggregated with the optimized fusion weight matrix to output a multi-source fusion feature vector.

[0006] Preferably, the timestamp alignment and format normalization processing includes: The sampling frequency difference of each sensor data stream is detected, and the low-frequency data stream is interpolated and compensated based on the global clock signal, while the high-frequency data stream is downsampled and truncated. The data encoding protocols of different sensors are analyzed, the dimensions of analog signals are uniformly converted into standard physical units, and the digital signals are reassembled into structured data packets after being aligned according to their bit width.

[0007] Preferably, the adaptive filtering algorithm's dynamic suppression process for the environmental noise components includes: A time-varying power spectral density model of the environmental noise component is established, and the noise floor threshold is updated in real time through a sliding window mechanism; When the instantaneous signal-to-noise ratio of the denoised sensor data is detected to be lower than the noise floor threshold, the wavelet packet decomposition and reconstruction algorithm is triggered to enhance the target frequency band.

[0008] Preferably, the multi-dimensional feature decomposition is performed in the following ways: The zero-crossing rate and peak-valley interval joint analysis method is used for the time domain eigenvectors, the Mel-Cepstral coefficient transform is performed on the frequency domain eigenvectors, and the principal component projection of the covariance matrix is ​​calculated for the spatial domain eigenvectors.

[0009] Preferably, the dynamic adjustment feature weight allocation strategy includes: Monitor the temperature change curve of the processor core and memory usage, and reduce the calculation accuracy of frequency domain feature vectors when system resources are strained; Based on the stability index of the multidimensional feature set over a set period in the past, an exponential decay constraint is applied to feature dimensions whose fluctuations exceed a threshold.

[0010] Preferably, the implementation process of the weighted aggregation operation includes: The optimized fusion weight matrix is ​​decomposed into a diagonal matrix and a rotation matrix. First, a rotation transformation is performed on the multidimensional feature set, and then a linear weighted sum is performed according to the feature dimension.

[0011] Preferably, the method further includes: After generating the multi-source fusion feature vector, a closed-loop feedback mechanism is activated to collect execution error data in actual application scenarios; The values ​​of the diagonal elements in the optimized fusion weight matrix are corrected in reverse based on the execution error data.

[0012] Preferably, the triggering conditions for the closed-loop feedback mechanism include: When the fusion result exceeds the preset confidence interval after a set number of consecutive iterations, it automatically switches to incremental learning mode. In the incremental learning mode, the parameters of the rotation matrix are iteratively updated using stochastic gradient descent.

[0013] Preferably, the method further includes: Establish a sensor health status assessment model and regularly analyze the data offset and response delay time of each sensor; When the evaluation score of a specific sensor is lower than the reliability threshold, the feature dimension corresponding to that sensor is masked in the multidimensional feature set.

[0014] Preferably, the method further includes: During the system initialization phase, a preset benchmark dataset is loaded, and the initial parameters of the adaptive filtering algorithm are calibrated by comparing the matching degree between the multi-source fusion feature vector and the standard label. After calibration, the clock frequency of the hardware accelerator is locked until the ambient temperature change exceeds the tolerance range, at which point the calibration process is retried.

[0015] Compared with the prior art, the beneficial effects of the present invention are: To address the discrepancies in raw data streams from multiple heterogeneous sensors, synchronized sensor data is generated through timestamp alignment and format standardization. This effectively resolves the inconsistencies in time and format between different sensor data, enabling efficient collaboration among various sensor data. This provides a unified and standardized data foundation for subsequent data processing and applications, avoiding processing errors caused by data asynchrony or inconsistent formats. It also allows subsequent noise suppression, feature extraction, and other processes to be conducted based on a more reliable data source. In terms of environmental noise processing, this method uses an adaptive filtering algorithm to dynamically suppress environmental noise components. Compared with the traditional fixed filtering algorithm, it can adjust the filtering parameters in real time according to the changes in noise in the actual environment. This not only allows for more accurate identification and suppression of noise, but also preserves the useful information in the sensor data to the greatest extent possible. It avoids the over-filtering of useful information or the impact of noise residue on data quality, thus obtaining purer denoised sensor data. This provides high-quality data support for subsequent multi-dimensional feature decomposition, enabling the extracted features to more realistically and accurately reflect the actual state of the sensing object.

[0016] For the multi-dimensional feature decomposition stage, this method calculates time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors separately to generate a multi-dimensional feature set. This breaks through the limitations of traditional single-dimensional feature extraction and comprehensively covers the effective information of sensor data in different dimensions. Time-domain features can reflect the pattern of data changes over time, frequency-domain features can reveal the frequency distribution characteristics of data, and spatial-domain features can reflect the spatial correlation of data. The combination of the three forms a more comprehensive and richer feature system, allowing the effective information in sensor data to be fully mined and utilized, laying a comprehensive feature foundation for subsequent feature fusion. In terms of feature weight allocation and fusion, this method dynamically adjusts the feature weight allocation strategy based on the current system load status and the historical distribution patterns of the multidimensional feature set, generating an optimized fusion weight matrix. This dynamic adjustment method can adapt to changes in system load in real time. When the system load is high, the data processing flow can be optimized by adjusting the weights, reducing unnecessary computational overhead, ensuring system operating efficiency, and avoiding processing delays. At the same time, adjusting the weights in conjunction with the historical distribution patterns of feature data allows the weight allocation to better match the actual changing trends of the feature data, making the fusion process more targeted. This enables the multi-source fusion feature vector output by the weighted aggregation operation to more accurately integrate various feature information, improving the reliability and effectiveness of the feature vector. Consequently, subsequent applications based on this feature vector, such as status monitoring and data-driven decision-making, will have better performance and be suitable for more complex and ever-changing sensing application scenarios. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the sensor data chip method based on multi-source information fusion and dynamic parameter adjustment described in this invention. Figure 2 A flowchart for timestamp alignment and format normalization. Figure 3 A flowchart for dynamically adjusting feature weight allocation strategies. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a sensor data chip method based on multi-source information fusion and dynamic parameter adjustment. The method includes: acquiring raw data streams through multiple heterogeneous sensors, which may include various types such as temperature, pressure, and motion sensors, generating continuous data sequences. The raw data streams are then timestamped and aligned using a high-precision clock signal to coordinate the sampling times of each sensor, eliminating time deviations caused by inconsistent sampling frequencies; simultaneously, format standardization is performed, converting the outputs of different sensors into a unified data structure and physical units. After the synchronized sensor data is generated, a noise processing stage is initiated, extracting environmental noise components from the data. These components may originate from external factors such as electromagnetic interference or mechanical vibration; an adaptive filtering algorithm is used to dynamically suppress noise. This algorithm adjusts the filtering parameters in real time according to changes in noise characteristics, ensuring that the denoised sensor data maintains a high signal-to-noise ratio. Next, multi-dimensional feature decomposition is performed, calculating time-domain feature vectors, frequency-domain feature vectors, and spatial-domain feature vectors for the denoised data; time-domain analysis focuses on the statistical characteristics of the data waveform, frequency-domain analysis reveals the periodic components of the signal, and spatial-domain analysis captures the correlation between sensor arrays, thereby generating a feature set containing multiple information dimensions. Based on the current system load status, such as processor computing resource utilization, and the historical distribution patterns of the multidimensional feature set, the feature weight allocation strategy is dynamically adjusted. For example, when system resources are strained, the weights of computationally complex features are reduced, generating an optimized fusion weight matrix. In the final stage, a weighted aggregation operation is performed, combining the multidimensional feature set with the optimized fusion weight matrix. Through linear combination or nonlinear transformation, a multi-source fusion feature vector is output, which serves as the basis for subsequent decision-making or control.

[0020] Example 1: See Figure 2In the timestamp alignment and format standardization process, detecting the sampling frequency differences of each sensor data stream is the primary step. This process is handled by the sampling rate detection module integrated within the chip, which continuously monitors the clock frequency and data arrival interval of each sensor data channel connected to the chip. For low-frequency data streams with sampling frequencies significantly lower than the system reference clock, the system initiates an interpolation compensation algorithm. This algorithm, based on a high-precision time reference provided by the global clock signal, calculates approximate data values ​​for the intermediate moments between two adjacent valid low-frequency data points using a preset interpolation algorithm. This compensation effectively fills in the time point data missing due to the low sampling rate, enabling the low-frequency data stream to synchronize with the high-frequency data stream in the time dimension. For high-frequency data streams with sampling frequencies higher than the system processing capacity or the reference clock, a downsampling truncation operation is performed. The downsampling process does not simply discard data points; instead, the data first passes through a digital anti-aliasing filter to remove high-frequency components that may introduce spectral aliasing. Subsequently, it is sampled regularly according to the system's processing rate, and the length of the truncated data sequence matches the system's processing frame length.

[0021] Deciphering the data encoding protocols of different sensors is a complex task, requiring the chip to have multi-protocol parsing capabilities. Analog signal sensors typically output continuously varying voltage or current signals. These signals are quantized into discrete digital values ​​by an analog-to-digital converter, but their dimensions represent different physical meanings. For example, millivolts may correspond to temperature, while volts may correspond to pressure. The dimension unification and conversion module maps the raw digital readings to standard physical units based on a sensor parameter lookup table pre-programmed into the chip's read-only memory. For example, it unifies all readings into Pascals, Kelvins, or meters per second squared in the International System of Units (SI). For sensors that directly output digital signals, their data formats may vary from manufacturer to manufacturer, with different bit widths, byte orders, and encoding protocols. The digital signal bit width alignment module identifies the effective bit width of the input data. For data with insufficient bit width, high-bit zeros are padded; for data with excessive bit width, the effective bit fields are truncated. The reassembly process packages the aligned bit data, sensor identifier, status code, and synchronized timestamp into a structured data packet with a fixed format.

[0022] The global clock signal is generated by a high-stability crystal oscillator inside the chip or provided by an external high-precision clock reference source. This signal serves as the absolute time base for the entire data processing flow. Based on this base, the timestamp alignment module assigns a uniform timestamp to each arriving data sample, regardless of its sensor origin. When processing low-frequency data, the interpolation compensation algorithm selects an appropriate interpolation method based on the trend of data changes. For signals with gradual changes, linear interpolation may be used to prioritize computational efficiency, while for signals with complex changes, cubic spline interpolation may be used to better approximate the true curve. The anti-aliasing filter in the downsampling truncation operation is typically designed as a linear-phase finite-length unit impulse response filter or an infinite-length unit impulse response filter. Its cutoff frequency is carefully set according to the downsampling ratio to prevent aliasing distortion in the useful signal band.

[0023] The format standardization process also includes a data integrity verification mechanism. For example, after generating a structured data packet, a cyclic redundancy check (CRC) code is calculated and appended to the packet's end. When subsequent modules receive data, they recalculate the CRC code and compare it with the appended CRC code. If they do not match, a retransmission is requested or the data packet is discarded, thus maintaining the reliability of data transmission. The final synchronized sensor data is temporarily stored in the chip's cache in the form of a two-dimensional matrix or a higher-dimensional tensor. The rows of the matrix typically represent time series, and the columns represent different sensor channels or feature dimensions. This well-organized data structure greatly facilitates vectorized operations and parallel processing in subsequent modules.

[0024] The sampling rate detection module relies on the collaborative work of a hardware counter and a software monitoring program. The hardware counter precisely measures the pulse interval of each sensor data channel, while the software program analyzes the statistical distribution of these intervals to determine a stable sampling frequency. For sensors with jitter or non-uniform sampling, the system employs more advanced prediction algorithms, such as Kalman filters, to estimate the optimal interpolation point. Interpolation compensation not only fills in the data but also assesses the uncertainty introduced by interpolation and adds a confidence field to the data packet for reference by subsequent processing modules. Digital signal protocol parsing involves decoding common hardware communication protocols such as serial peripheral interfaces and internal integrated circuits. The protocol processing unit inside the chip samples the bit stream from the data lines according to preset timing rules and combines it into byte data according to a specified byte order. Bit width alignment ensures that data from different sources occupy the same storage space in memory, avoiding calculation errors caused by inconsistent data types in the subsequent feature extraction stage. The structured data packet format is carefully designed, including a header, payload, and trailer. The header contains a synchronization word and packet length information, while the trailer contains a checksum. This encapsulation format enhances data manageability.

[0025] The accuracy of the timestamp directly affects the fusion effect, therefore the stability of the global clock is crucial. The chip may use phase-locked loop (PLL) technology to synchronize the internal clock with a more stable external clock source, thereby controlling clock drift at the nanosecond level. The choice of interpolation algorithm is not static; the system can dynamically switch interpolation strategies based on the short-term statistical characteristics of the data stream. For example, linear interpolation is used when data stability is detected, and higher-order interpolation is switched to preserve detail when data abrupt changes are detected. The design of the downsampling filter needs to strike a balance between stopband attenuation and transition band width, typically optimized using window function methods or equiripple methods. Data integrity verification is a critical step in ensuring information reliability. In addition to cyclic redundancy check (CRC), for highly demanding applications, forward error correction codes such as Hamming codes may be used to automatically correct for a small number of errors. Before being sent to the next stage, the synchronized data matrix typically undergoes preliminary outlier detection, such as using interquartile range (INR) or Z-score methods. Data points that significantly deviate from the normal range are marked as invalid but not deleted immediately, leaving them for further judgment by subsequent modules based on context information. The entire timestamp alignment and format standardization process is implemented on the chip in a pipeline manner, with multiple processing stages operating in parallel, thereby meeting the low latency requirements for real-time synchronous processing of multi-channel sensor data.

[0026] Example 2: The adaptive filtering algorithm for dynamically suppressing environmental noise components begins with establishing a time-varying power spectral density model of the environmental noise components. This model relies on performing frequency domain analysis on a continuously acquired segment of raw sensor data. A Fast Fourier Transform (FFT) is used to convert the time-domain signal into frequency components, thereby estimating the energy distribution of background noise in different frequency bands. This distribution is not fixed but fluctuates in real time with environmental interference factors. A sliding window mechanism is used to update the noise floor threshold in real time. This window selects the noise spectrum estimation results over a recent period, calculates its statistical average, and adds several times the standard deviation as the dynamic threshold. The window size can be adaptively adjusted according to the severity of environmental changes. A longer window is used during stable periods to obtain a stable estimate, while a shorter window is used during abrupt changes for rapid response. When the detection module detects that the instantaneous signal-to-noise ratio of the denoised sensor data is lower than the currently updated noise floor threshold, it triggers a wavelet packet decomposition and reconstruction algorithm to enhance the target frequency band. Wavelet packet decomposition provides a finer frequency band division than traditional wavelet decomposition, decomposing the signal into a series of overlapping sub-bands with uniform frequency resolution.

[0027] The multi-dimensional feature decomposition method involves a joint analysis of the time-domain feature vector using zero-crossing rate and peak-valley interval. Zero-crossing rate calculates the number of times a signal value crosses the zero or average level per unit time, reflecting the signal's oscillation frequency. Peak-valley interval, on the other hand, identifies local maxima and minima and measures the time or amplitude difference between consecutive peaks and troughs to characterize the signal's envelope shape and modulation characteristics. Joint analysis of both methods allows for a more comprehensive capture of the signal's dynamic behavior in the time domain. A Mel-Cepstral Transform is performed on the frequency-domain feature vector. This transform first performs a nonlinear mapping on the signal's short-time Fourier transform spectrum, converting it from a linear frequency scale to a Mel scale more consistent with human auditory perception. Then, the logarithm of the Mel spectrum is taken to compress the dynamic range. Finally, a discrete cosine transform is performed to remove correlations between components, resulting in cepstral coefficients representing the shape of the spectral envelope. Principal component projection of the covariance matrix is ​​calculated for the spatial domain eigenvectors. When processing data from a sensor array, a covariance matrix with different sensors as variables is first constructed. This matrix reflects the degree of linear correlation between the signals of each sensor. Then, eigenvalue decomposition is performed on the matrix to extract the principal component directions with the largest eigenvalues. The original spatial domain vectors are then projected onto these principal component directions to achieve data dimensionality reduction and highlight the main patterns of spatial distribution.

[0028] The update process of the time-varying power spectral density model is strictly synchronized with data acquisition. Each new data frame triggers a fine-tuning of the model. Recursive estimation methods, such as exponentially weighted moving averages, are used to balance the influence of historical information and current observations, ensuring that noise estimation has both a certain degree of memory and can track non-stationary changes. Wavelet packet decomposition and reconstruction algorithms need to consider signal characteristics when selecting wavelet basis functions. For example, Daubechies wavelets have tight support and smoothness, making them suitable for processing smooth signals, while Symlets wavelets offer better symmetry and are more advantageous in signal boundary processing. The choice of the decomposition level depends on the required fineness of the frequency band to be analyzed and the constraints of computational resources. Joint analysis methods do not calculate the zero-crossing rate and peak-to-valley interval in isolation in the time domain. Instead, they combine their statistical characteristics, such as mean, variance, and higher-order moments, and may introduce the difference features of the time series, i.e., calculating the rate of change of these features between consecutive signal frames, thus forming a set of high-dimensional time-domain feature vectors that more meticulously describe the time-varying characteristics of the signal.

[0029] The implementation of Mel-Cepstral Coefficient Transform (MCT) involves a series of standardized steps. A pre-emphasis filter is used to boost the high-frequency components of the signal to compensate for the attenuation of high-frequency energy during sound production. Frame segmentation divides the long signal into short segments, assuming quasi-stationarity within each segment. Windowing operations typically use Hamming or Hanning windows to reduce spectral leakage. The Mel filter design simulates the critical bandwidth effect of the human cochlea. The lower-order coefficients in the final cepstral coefficients represent the spectral envelope, i.e., the formant structure, while the higher-order coefficients represent the excitation source information. Principal component projection calculation involves eigenvalue decomposition of the covariance matrix, a computationally intensive operation. For applications with high real-time requirements, incremental principal component analysis can be used, or hardware accelerators such as tensor processing units can be used to compute eigenvalues ​​and eigenvectors in parallel. The new features after projection, i.e., the principal component scores, have a dimension much lower than the original sensor channel count but retain most of the variance information. The sliding window mechanism is managed by a dedicated memory controller, which maintains a first-in, first-out data buffer. New noise spectrum estimation results are written to the end of the buffer while the oldest results are removed. Statistics of the buffer contents, such as the mean and standard deviation, are calculated in real time and used for threshold updates.

[0030] The wavelet packet decomposition and reconstruction algorithm enhances the target frequency band by scaling or thresholding the coefficients of specific sub-bands. For sub-bands heavily contaminated with noise, the coefficient amplitude may be attenuated or even zeroed out, while for sub-bands containing useful signals, the coefficients may be appropriately amplified to improve significance. The reconstruction process uses the processed sub-band coefficients for inverse transformation to synthesize the enhanced time-domain signal. Joint analysis of zero-crossing rate and peak-valley interval requires a robust extremum detection algorithm to avoid misjudging extrema due to minor fluctuations or noise interference. This might involve smoothing filtering before detecting extrema, or setting amplitude and interval thresholds to filter effective peak-valley points. The dimension of the Mel-frequency cepstral coefficient transform needs to be determined based on the specific application. In speech recognition, 12-13 coefficients are typically used, while in mechanical vibration analysis, more coefficients may be needed to capture rich spectral details. The calculated coefficients can also be further subjected to first-order and second-order differences to obtain dynamic features. The calculation of the spatial domain covariance matrix is ​​typically based on observation data from multiple moments in a sensor array. The elements of the matrix represent the temporal covariance of different sensor signals. Principal component projection aims to find the projection direction with the largest data variance; these directions often correspond to the directions of the main sound or vibration sources, thus achieving spatial feature extraction and dimensionality reduction. The entire noise suppression and feature decomposition process is designed as a highly pipelined operation on the chip. Modules such as time-domain filtering, frequency-domain transformation, and spatial computation can work in parallel. Intermediate results are exchanged via an on-chip high-speed bus or shared memory to meet the stringent timing requirements of real-time processing of multi-sensor data streams. The generation of eigenvectors is the foundation for subsequent fusion decisions, and their quality directly affects the performance of the entire system. Therefore, the feature decomposition module often includes a self-checking function to monitor the rationality of eigenvalues ​​and send an interrupt request to the control core when an anomaly is detected.

[0031] Example 3: See Figure 3The implementation of the dynamic feature weight allocation strategy relies on continuous monitoring of the processor core temperature change curve and system memory usage. The temperature change curve is collected by digital temperature sensors integrated near each computing core of the chip at a fixed sampling period. These sensors convert the resistance changes of the thermistors into digital readings, forming a temperature sequence that changes over time. The memory usage is provided by the performance monitoring counter of the operating system kernel or a dedicated hardware register, which provides real-time feedback on the ratio of currently allocated memory blocks to total available memory. When system resources are strained, such as when the temperature exceeds a preset threshold or the memory usage approaches the upper limit, the weight adjustment logic will trigger a computational precision degradation mechanism. Precision control is implemented for frequency domain feature vectors with high computational complexity. For example, the number of Mel-frequency cepstral coefficients extracted is reduced from the standard thirteen dimensions to eight dimensions, or fixed-point arithmetic is used instead of floating-point arithmetic to approximate the calculation of the fast Fourier transform. The weights are dynamically adjusted based on the stability index of the multidimensional feature set over a set period. The stability index is obtained by calculating the variance or information entropy of each feature dimension through a sliding window. The variance reflects the fluctuation range of the feature value, while the entropy value characterizes the uncertainty of the feature distribution. For feature dimensions whose fluctuation exceeds a preset threshold, the system applies an exponential decay constraint, and its weight decreases over time according to an exponential function. The decay factor is adaptively set according to the degree of fluctuation; the greater the fluctuation, the faster the decay.

[0032] The implementation of weighted aggregation begins by decomposing the optimized fusion weight matrix into a diagonal matrix and a rotation matrix. The diagonal matrix is ​​a square matrix where the elements on the main diagonal represent the independent weight coefficients of each feature dimension, and all off-diagonal elements are zero. The rotation matrix is ​​an orthogonal matrix used to perform a linear transformation on the feature space, changing the coordinate basis of the feature vectors to eliminate the correlation between different features. First, a rotation transformation is performed on the multidimensional feature set, that is, the original feature vectors are multiplied by the rotation matrix, projecting them into a new orthogonal coordinate system. In this new coordinate system, the covariance between features is minimized. Then, linear weighted accumulation is performed according to the feature dimension. Each transformed feature component is multiplied by the corresponding weight coefficient in the diagonal matrix. Finally, all weighted components are summed and output as the multi-source fusion feature vector. This linear weighted accumulation operation can be expressed mathematically as follows: Where: symbol The multi-source fusion feature vector represents the final output. It is an aggregation result in scalar or vector form, and its symbol is... This represents a summation operation, iterating through all feature dimensions, with the following symbol... It is the summation index, representing the j-th feature dimension, symbol... The symbol represents the total dimension of the feature space, i.e., the length of the feature vector. It is the value of the j-th diagonal element in the diagonal matrix, representing the weight coefficient of the j-th feature dimension in the weighted accumulation, with the sign... It is a rotation matrix used to perform linear transformations on eigenvectors, with the symbol... It represents the eigenvalue of the j-th feature dimension in the original multidimensional feature set. The parentheses indicate the order of matrix multiplication operations, i.e., rotation transformation is performed first, followed by weighting.

[0033] Processor core temperature monitoring focuses not only on instantaneous values ​​but also on trends. For example, it calculates the first-order difference of temperature to determine the rate of temperature rise or fall. When a sustained temperature increase with a steep slope is detected, the system may trigger precision degradation in advance to avoid overheating. Memory usage monitoring distinguishes between different types of memory, such as cache and main memory, and considers memory access frequency. Frequent cache misses or page swapping, even with low usage, may be considered as resource strain. The precision adjustment strategy for frequency domain feature vector calculation is tiered. For example, under normal conditions, a high-precision mode is used to calculate all Mel-frequency cepstral coefficients. Under light load, the number of coefficients is reduced while floating-point operations are maintained. Under heavy load, it further switches to fixed-point operations and only calculates core frequency band coefficients. Precision degradation decisions are made in real time by the hardware state machine or microcontroller firmware.

[0034] The calculation cycle of the feature stability index can be dynamically set according to application requirements. For example, a longer window, such as several seconds, can be used to calculate the variance under stable environmental conditions, while the window can be shortened to milliseconds when the environment changes drastically. The exponential decay constraint is implemented by multiplying the weights by a decay factor less than one after each processing cycle. The magnitude of the decay factor is proportional to the normalized value of the feature variance, ensuring that the weights of features with large fluctuations decrease rapidly. The decomposition of the fusion weight matrix adopts numerically stable methods such as singular value decomposition, ensuring the orthogonality of the rotation matrix and the non-negativity of the diagonal matrix. The decomposition process may be iterative until the matrix condition number meets the requirements. The purpose of the rotation transformation is to whiten the feature space, so that the transformed feature components have zero mean and unit variance and are uncorrelated. This helps to improve the robustness of the fusion result. The calculation of the rotation transformation may utilize hardware accelerators, such as matrix multiplication units, to improve efficiency. The linear weighted accumulation operation can be implemented at the hardware level through a multiply-accumulate instruction pipeline, completing one multiply-accumulate operation per clock cycle, thereby efficiently processing high-dimensional feature vectors. The entire dynamic adjustment and weighted aggregation process is designed to be configurable, with key parameters such as temperature threshold, memory utilization threshold, stability calculation window size, and decay factor stored in the chip's programmable registers, allowing for calibration according to specific application scenarios.

[0035] The system resource monitoring module and the feature extraction module exchange status information via a high-speed bus. When the weight adjustment strategy changes, a control signal is sent to the feature extraction module to adjust its calculation parameters. Conversely, changes in feature stability are also fed back to the resource monitoring module for optimizing threshold settings. The decomposition and updating of the weight matrix are not necessarily performed in every processing cycle, but are triggered only when a significant change in system state or feature distribution is detected, thus reducing computational overhead. The rotation matrix required for rotation transformation may be pre-calculated and cached in on-chip memory, recalculated only when the feature correlation structure drifts. The result of linear weighted accumulation undergoes range checking before output to prevent overflow or underflow, and may be normalized to ensure numerical stability. During implementation, the chip's power management unit coordinates with dynamic adjustment strategies, such as synchronously adjusting processor voltage and frequency to further save energy when reducing computational precision. The adjustment of the feature weight allocation strategy follows a predefined strategy table, which defines the precision level and attenuation parameters under different resource states. The hardware path for weighted aggregation is optimized to reduce latency, for example, by using parallel multipliers and tree-structured adders to accelerate dot product operations. The final multi-source fused feature vector is then fed into subsequent decision or classification modules for further analysis or control. The implementation of this embodiment emphasizes adaptability and efficiency, enabling stable operation in resource-constrained embedded environments.

[0036] Example 4: After generating the multi-source fusion feature vector, the system initiates a closed-loop feedback mechanism. This mechanism is triggered by continuous monitoring of the execution results in actual application scenarios. For example, in a robotic arm trajectory control scenario, the multi-source fusion feature vector is used to estimate the spatial position of the robotic arm's end effector, while the execution error data is obtained by measuring the difference between the actual position measured by a high-precision laser tracker and the target position predicted by the fusion vector. The collected execution error data is temporarily stored in a buffer in the form of a time series. Each data point includes a timestamp, error magnitude, and the corresponding fusion feature vector identifier. The system periodically analyzes these error data, calculating their statistical characteristics such as mean, standard deviation, and autocorrelation function to determine whether there is a systematic bias or fluctuation in a specific pattern. Based on the analysis results of the execution error data, the system reverse-corrects the diagonal element values ​​in the optimized fusion weight matrix. The correction principle is to reduce the weights of feature dimensions associated with larger errors. The correction amount is calculated using the gradient information of the error, and a small-step iterative update method is used to avoid drastic weight oscillations. The closed-loop feedback mechanism's trigger condition includes a crucial judgment logic: when the fusion results exceed a preset confidence interval for a set number of consecutive times, it automatically switches to incremental learning mode. The confidence interval is typically determined based on the distribution range of historical fusion results under normal operating conditions; for example, the mean plus or minus three standard deviations can be used as the boundary. The threshold for the number of consecutive times exceeding this boundary can be set according to the system's tolerance for false alarms. In incremental learning mode, the system no longer retrains the model using batch data. Instead, it iteratively updates the rotation matrix parameters using stochastic gradient descent. Each time a new data sample is processed, the gradient of the loss function with respect to the rotation matrix parameters is calculated, and then the parameters are updated in small steps in the opposite direction of the gradient. This mode can continuously adapt to slow changes in system characteristics.

[0037] The frequency of error data acquisition is consistent with the frequency of the main control loop to ensure that the latest feedback information is obtained in each control cycle. Error data is filtered before being used to correct weights to suppress measurement noise interference; for example, a first-order low-pass filter is used to smooth random fluctuations. The correction of the diagonal elements of the weight matrix is ​​selective. The system analyzes the correlation between the error and each feature dimension, adjusting only the weights of dimensions significantly correlated with the current error pattern to avoid overfitting to accidental error fluctuations. Switching to incremental learning mode is accompanied by adjustments to the learning rate. Initially, the learning rate is set high for rapid convergence, and then gradually decreases over time to improve parameter stability.

[0038] A specific implementation process occurred in an indoor UAV hovering control system based on multi-sensor fusion. The UAV estimates its attitude and position by fusing data from visual sensors, an inertial measurement unit, and an ultrasonic ranging module. The multi-source fused feature vector is used to generate control commands to maintain hovering. The system records the error between the fused estimated position and the actual position (provided by an indoor ultra-wideband positioning system) over a period of time. When the vertical altitude estimation error exceeds the normal range for several consecutive control cycles, closed-loop feedback and incremental learning are triggered. The table below shows the adjustment of some diagonal elements (corresponding to different sensor feature dimensions) in the weight matrix before and after triggering incremental learning.

[0039] Table 1: Feature weight adjustment before and after incremental learning

[0040] Error data analysis revealed a cumulative bias in the vertical acceleration estimation, while the absolute height information provided by vision and ultrasound was more reliable. Therefore, the system reduced its reliance on the Z-axis acceleration features of the inertial measurement unit (IMU) while increasing the weights of visual and ultrasound-related features. Updating the rotation matrix parameters focused on adjusting elements that affect the spatial combination of information from different sensors, such as changing the coupling coefficient between visual and inertial features in attitude estimation. Each iteration of the stochastic gradient descent method used only the latest data sample, calculated the estimation error caused by that sample, and then updated the value of each element based on the partial derivative of the error with respect to that element in the rotation matrix. The update step size was controlled by an adaptive learning rate scheduler. In incremental learning mode, the system continuously monitored the error trend. If the error gradually decreased and returned to the confidence interval, this mode was maintained for a period to consolidate the learning effect. If the error failed to improve effectively or diverged, a higher-level fault handling procedure was triggered, such as attempting to switch to a backup fusion strategy or requesting external intervention. The entire closed-loop feedback process was designed with robustness in mind, limiting the update magnitude of weights and parameters to prevent a sharp deterioration in system performance due to a single anomalous data point. Historical error data is stored cyclically, with the latest data overwriting the oldest data, thus ensuring that the learning process is always based on recent system performance.

[0041] In the drone hovering example, after triggering incremental learning, the system gradually adjusts the weights and rotation matrix over the next few dozen control cycles, gradually improving the drone's hovering accuracy and narrowing the altitude fluctuation range. After the learning process ends, the system exits the incremental learning mode and returns to normal monitoring, but the newly learned parameters are retained and used for subsequent fusion calculations until the next trigger condition is met. This mechanism enables the sensor data chip to adapt to the slow degradation of sensor performance, changes in environmental characteristics, and alterations in external interference patterns. Implementing a closed-loop feedback mechanism requires additional computing resources for error analysis and parameter updates. Therefore, in practical designs, this function may be handled by a dedicated low-power coprocessor, working in parallel with the main fusion computing core to minimize the impact on real-time processing performance. All parameters related to feedback learning, such as confidence interval boundaries, number of consecutive triggers, initial learning rate, and its decay strategy, are stored in non-volatile memory, allowing for configuration and optimization according to different application scenarios. Through this closed-loop feedback and incremental learning, the sensor data chip based on multi-source information fusion achieves a certain degree of adaptability, maintaining good sensing performance over long periods of operation.

[0042] Example 5: The process of establishing a sensor health status assessment model begins with periodically analyzing the data offset and response delay time of each sensor. The calculation of data offset typically requires a known and stable reference benchmark. In industrial robot applications, multiple torque and position sensors at the robot joints are monitored simultaneously. When the robot arm is stationary at a preset zero point, theoretically, the sensor readings should stabilize near their calibrated values. The assessment model continuously records the long-term average value of each sensor's output and compares it with the factory calibration value or a theoretically calculated value based on a physical model. The systematic differences between them are quantified as data offset. The measurement of response delay time is accomplished by injecting test signals or utilizing the inherent periodic movements of the system. For example, when the robot performs uniform rotation, different sensors at the same joint should detect synchronous periodic changes. By cross-correlating the rising edge or feature point occurrence time of each sensor signal using a high-precision timer, the delay time of each sensor relative to the trigger event or relative to other sensors can be calculated. The frequency of periodic analysis is set according to the criticality of the application. For high-precision machining robots, it may be set to once every few minutes, while for ordinary monitoring systems, it may be once per hour. The analysis results are quantified into a comprehensive evaluation score, which is a weighted function of data offset and response latency, and may also include other factors such as signal noise level or packet loss rate. When the evaluation score of a specific sensor is lower than a preset reliability threshold, the system does not immediately disable it completely. Instead, it masks the corresponding feature dimension of the sensor in the multi-dimensional feature set. The masking operation means that in subsequent feature fusion calculations, this dimension is assigned zero or very low weight, and its data no longer participates in the weighted aggregation operation. However, the sensor's raw data acquisition and health monitoring continue, in preparation for re-enabling it once its performance recovers.

[0043] During system initialization, the chip loads a pre-defined benchmark dataset from read-only memory or external storage. This dataset contains rigorously calibrated sensor data collected in a standard laboratory environment, along with their corresponding standard labels. For example, for an inertial measurement unit (IMU), the benchmark data might include triaxial acceleration and triaxial angular velocity data under various standard motion modes such as platform stationary, uniform rotation, and sinusoidal vibration, as well as real trajectories acquired through a high-precision motion capture system as standard labels. The initial calibration process adjusts the initial parameters of the adaptive filtering algorithm by comparing the matching degree between the multi-source fused feature vector generated by the current chip processing and the standard labels in the benchmark dataset. The matching degree is typically measured using metrics such as mean square error or correlation coefficient. The calibration algorithm iteratively optimizes the internal parameters of the adaptive filtering algorithm, such as filter order, cutoff frequency, or convergence factor, to achieve the optimal overall matching degree of the fused feature vector on the benchmark set. After calibration, the system locks the clock frequency of the hardware accelerator to a fixed value that meets real-time processing requirements while balancing power consumption and heat dissipation. This locking avoids computational latency jitter that might be introduced by dynamic clock frequency adjustments, ensuring the predictability of data processing timing. The frequency lock state will continue until the system detects that the ambient temperature change exceeds the set tolerance range, such as the chip junction temperature change exceeding ±10 degrees Celsius. The temperature change is monitored by the built-in temperature sensor. Once the limit is exceeded, the system determines that the current working environment is significantly different from that at the time of initial calibration, and the original calibration parameters may no longer be optimal, thus re-triggering the entire calibration process.

[0044] Taking a sensor chip used in automotive chassis centralized control as an example, this chip integrates data from multiple wheel speed sensors, lateral accelerometers, yaw rate sensors, and steering angle sensors to evaluate the vehicle's driving status. Its health assessment model continuously monitors the signals from each sensor. For example, under specific conditions of constant speed driving in a straight line, it compares the consistency of readings from different wheel speed sensors and whether the lateral accelerometer reading is close to zero. If a wheel speed sensor consistently outputs readings significantly different from other sensors, or if its signal response to vehicle acceleration commands is noticeably slow, its evaluation score will decrease. If the score is below a threshold, the data contribution from that abnormal sensor will be masked when calculating the overall vehicle speed—a fusion characteristic—and the system will primarily rely on other normal wheel speed sensors and signals from the transmission system. During system power-on initialization, the chip loads a set of benchmark data. This data may originate from standardized operation records of the vehicle model performed on a standard test track, such as slalom and straight-line braking. The data includes raw signals from each sensor and real vehicle states (such as center of gravity speed and yaw angle) measured by specialized equipment. The calibration process adjusts the adaptive filtering parameters used to process these standard data streams, minimizing the error between the fused features (such as vehicle yaw rate) calculated by the chip and the standard labels. After successful calibration, the clock frequency of the processing core is locked. Subsequently, when a vehicle moves from a cold outdoor environment into a warmer underground parking lot, the chip temperature rises significantly, exceeding the tolerance range. The system automatically initiates recalibration to adapt to potential sensor characteristic drift and circuit performance changes under the new temperature conditions.

[0045] The health status assessment model is designed to account for the diversity of failure modes. Data offsets may manifest as a constant deviation, a gain error proportional to the measured value, or a drift that gradually increases over time. The model attempts to identify the type of offset and incorporate it into the assessment. Measuring response latency requires a sophisticated time synchronization mechanism, typically by embedding high-resolution timestamps in the sensor data packet to achieve accurate latency calculation. The model for calculating evaluation scores may also require calibration to ensure fairness and validity in different application scenarios. The construction of benchmark datasets is crucial, needing to cover the entire expected operating range of the sensor and various typical operating conditions. The quality of the dataset directly determines the accuracy of the initial calibration. The calibration process is usually a constrained optimization problem, requiring consideration of the realistic physical meaning of parameters and the stability requirements of the algorithm while pursuing a high degree of fit. Locking the hardware clock frequency involves interaction with the chip's power management unit to ensure voltage and frequency stability during the lockout period. Monitoring ambient temperature changes is dynamic; the system calculates a moving average or trend of temperature change rather than triggering recalibration based on a single instantaneous reading, avoiding unnecessary calibration overhead due to brief temperature fluctuations. The implementation of this embodiment enhances the reliability of the sensor data chip during long-term operation and its adaptability to environmental changes.

[0046] 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.

[0047] 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. A sensor data chip method based on multi-source information fusion and dynamic parameter adjustment, characterized in that, Includes the following steps: Raw data streams from multiple heterogeneous sensors are collected, and the raw data streams are timestamped and standardized to generate synchronized sensor data. The environmental noise component is extracted from the synchronized sensing data, and an adaptive filtering algorithm is used to dynamically suppress the environmental noise component to obtain the denoised sensing data. The denoised sensor data is subjected to multi-dimensional feature decomposition, and time-domain feature vector, frequency-domain feature vector and spatial-domain feature vector are calculated respectively to generate a multi-dimensional feature set; Based on the current system load status and the historical distribution pattern of the multidimensional feature set, the feature weight allocation strategy is dynamically adjusted to generate an optimized fusion weight matrix. The multi-dimensional feature set is weighted and aggregated with the optimized fusion weight matrix to output a multi-source fusion feature vector. The dynamic adjustment strategy for feature weight allocation includes: Monitor the temperature change curve of the processor core and memory usage, and reduce the calculation accuracy of frequency domain feature vectors when system resources are strained; Based on the stability index of the multidimensional feature set over a set period in the past, an exponential decay constraint is applied to feature dimensions whose fluctuations exceed a threshold.

2. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 1, characterized in that, The timestamp alignment and format normalization process includes: The sampling frequency difference of the data streams from each sensor is detected. The low-frequency data stream is interpolated and compensated based on the global clock signal, and the high-frequency data stream is downsampled and truncated. The data encoding protocols of different sensors are analyzed, the dimensions of analog signals are uniformly converted into standard physical units, and the digital signals are reassembled into structured data packets after being aligned according to their bit width.

3. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 2, characterized in that, The adaptive filtering algorithm's dynamic suppression process for the environmental noise components includes: A time-varying power spectral density model of the environmental noise component is established, and the noise floor threshold is updated in real time through a sliding window mechanism; When the instantaneous signal-to-noise ratio of the denoised sensor data is detected to be lower than the noise floor threshold, the wavelet packet decomposition and reconstruction algorithm is triggered to enhance the target frequency band.

4. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 3, characterized in that, The execution methods of the multi-dimensional feature decomposition include: The zero-crossing rate and peak-valley interval joint analysis method is used for the time domain eigenvectors, the Mel-Cepstral coefficient transform is performed on the frequency domain eigenvectors, and the principal component projection of the covariance matrix is ​​calculated for the spatial domain eigenvectors.

5. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 4, characterized in that, The implementation process of the weighted aggregation operation includes: The optimized fusion weight matrix is ​​decomposed into a diagonal matrix and a rotation matrix. First, a rotation transformation is performed on the multidimensional feature set, and then a linear weighted sum is performed according to the feature dimension.

6. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 5, characterized in that, Also includes: After generating the multi-source fusion feature vector, a closed-loop feedback mechanism is activated to collect execution error data in actual application scenarios; The values ​​of the diagonal elements in the optimized fusion weight matrix are corrected in reverse based on the execution error data.

7. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 6, characterized in that, The triggering conditions for the closed-loop feedback mechanism include: When the fusion result exceeds the preset confidence interval after a set number of consecutive iterations, it automatically switches to incremental learning mode. In the incremental learning mode, the parameters of the rotation matrix are iteratively updated using stochastic gradient descent.

8. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 7, characterized in that, Also includes: Establish a sensor health status assessment model and regularly analyze the data offset and response delay time of each sensor; When the evaluation score of any sensor is lower than the reliability threshold, the feature dimension corresponding to that sensor is masked in the multidimensional feature set.

9. The sensor data chip method based on multi-source information fusion and dynamic parameter adjustment according to claim 8, characterized in that, Also includes: During the system initialization phase, a preset benchmark dataset is loaded, and the initial parameters of the adaptive filtering algorithm are calibrated by comparing the matching degree between the multi-source fusion feature vector and the standard label. After calibration, the clock frequency of the hardware accelerator is locked until the ambient temperature change exceeds the tolerance range, at which point the calibration process is retried.

Images