An edge-computing-based sensor data fusion anomaly detection system and method

Abstract

The application relates to the field of electric digital data processing and discloses a sensing data fusion abnormality detection system and method based on edge computing, which comprises a data acquisition module, a feature storage module and a data analysis module. The data analysis module acquires N-path heterogeneous digital signals of the real-time state of a controlled object, constructs an observation vector mapped to an N-dimensional feature space through normalization processing, calls a pre-stored coupling feature matrix, projects the observation vector to a stable manifold space, extracts an orthogonal residual vector of the observation vector and calculates the module length, and when the module length continuously exceeds a judgment threshold for a period reaching a time threshold, it is judged that the controlled object has nonlinear structural decoupling. The application identifies abnormalities by monitoring the topological offset of the observation vector relative to the stable manifold, realizes deep mining of the physical coupling logic among multi-source signals, and effectively resists signal slow drift caused by environmental fluctuations.

Description

Technical Field

[0001] This invention belongs to the field of electronic digital data processing technology, and in particular relates to a sensor data fusion anomaly detection system and method based on edge computing. Background Technology

[0002] In current distributed industrial computing architectures, edge nodes are used to process multiple digital signal streams generated by heterogeneous sensors to achieve real-time monitoring of the status of production units. In typical operating conditions such as offshore wind power pitch control systems or collaborative robot clusters, the processing unit collects sensor data in dimensions such as vibration, current, and temperature, and uses methods such as weighted fusion or linear feature superposition to convert the original signals into feature vectors. By comparing them with preset thresholds, abnormalities in the system status can be identified.

[0003] The aforementioned conventional processing methods rely on the absolute offset of signal amplitude as the criterion, implicitly assuming that anomalies must manifest as numerical range overflows. However, in the early stages of industrial equipment failure, anomalies often manifest as superficial decoupling of the physical correlation logic between multiple source signals, while the value of a single channel remains within the nominal range. If the dimension of feature extraction is increased to capture such nonlinear deviations, the computational load on edge nodes is significantly increased, causing processing delays to exceed the 10ms threshold required for industrial real-time control. If linear dimensionality reduction is used to meet real-time requirements, the system will suffer from low detection sensitivity when facing hidden structural faults. In addition to hardware sensing limitations, multi-sensor fusion control methods also have limitations. Logical flaws exist, for example, Chinese invention patent application CN115577320A discloses a multi-sensor asynchronous data fusion processing method based on data interpolation. It achieves heterogeneous data alignment on the time axis through quaternion interpolation algorithm and smooths the core anchored observation values. Based on the mathematical fitting preset premise, the system easily ignores the phase misalignment or nonlinear distortion caused by the evolution of the physical topology between sensors. When processing large-scale high-dimensional signal streams, the interpolation operation produces data artifacts, annihilating the weak correlation characteristics that reflect the essential degradation of the equipment. When facing common-mode interference in non-steady-state operating conditions, it lacks an endogenous noise shielding mechanism and it is difficult to accurately measure the degree of physical constraint deviance under the limited computing power on the edge side.

[0004] Therefore, the technical problem to be solved by this invention is how to provide an edge computing-based sensor data fusion anomaly detection system that can realize the real-time characterization of nonlinear collaborative logic between heterogeneous signals by constructing a dynamic manifold projection processing architecture, and simultaneously solve the problem of imbalance between processing latency and detection accuracy under computing power-limited environment. Summary of the Invention

[0005] This invention provides an anomaly detection system based on edge computing-driven sensor data fusion, comprising:

[0006] Data acquisition module, feature storage module, and data analysis module;

[0007] The data analysis module is connected to both the data acquisition module and the feature storage module; the data analysis module detects the controlled object through the following processing steps:

[0008] Step S1: Obtain N multi-source heterogeneous digital signals representing the real-time operating state of the controlled object, and perform dimensional normalization processing on the N multi-source heterogeneous digital signals to construct a real-time observation vector mapped to the N-dimensional feature space.

[0009] Step S2: Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors; the coupled feature matrix is ​​generated by singular value decomposition of historical sample data of the controlled object under stable operating conditions.

[0010] Step S3: Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition.

[0011] Step S4: Calculate the magnitude of the orthogonal residual vector and compare the magnitude with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold.

[0012] Step S5: When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module.

[0013] Preferably, it also includes a logic compensation module for maintaining the continuity of the evolution of the N-dimensional feature space on the time axis; the logic compensation module completes the real-time observation vector by refining the following sub-steps of step S1: step S11, monitor the real-time data integrity of the data acquisition module, and when a single data packet loss is detected in the N-channel multi-source heterogeneous digital signals, extract the associated prediction value of the previous sampling period; step S12, use a linear interpolation algorithm to fill in the missing signal bits to correct the topological offset of the real-time observation vector.

[0014] Preferably, the data analysis module is also used to run the residual discrimination mechanism; the residual discrimination mechanism uses orthogonal residual vectors to shield the linear common-mode component interference in the N-channel multi-source heterogeneous digital signals, so as to resist the level drift caused by ambient temperature fluctuations in the N-channel multi-source heterogeneous digital signals.

[0015] Preferably, the data analysis module is also used to adjust the parallel granularity of the processing tasks according to the available computing power resources on the edge side; the data analysis module adjusts the iteration step size of the orthogonal projection transformation according to the value of the available computing power resources, and maintains a dynamic balance between computing power allocation and anomaly detection accuracy while maintaining a processing latency of less than 10ms.

[0016] Preferably, the data acquisition module includes an asynchronous sampling calibration unit for timestamp alignment of the sampling frequency differences of N multi-source heterogeneous digital signals; the asynchronous sampling calibration unit uses the high-frequency signal as the sampling reference to fit the sampling value of the low-frequency signal to ensure the consistency of each component in the real-time observation vector in the time domain.

[0017] Preferably, the feature storage module is also used to receive a global manifold update strategy from the cloud platform; the feature storage module performs online incremental correction on the pre-stored coupling feature matrix according to the aging model of the controlled object, so as to compensate for the drift of the reference manifold caused by the evolution of the physical properties of the controlled object.

[0018] Preferably, the data analysis module is also used to perform multi-dimensional information entropy measurement; the data analysis module calculates the information distribution entropy value of the orthogonal residual vector in different component dimensions, and locates the fault signal path based on the information distribution entropy value.

[0019] Preferably, the preset judgment threshold is set as an adaptive threshold that dynamically shifts with the complexity of the environment; the data analysis module dynamically sets the judgment boundary of the preset judgment threshold based on the variance envelope of the orthogonal residual vector within the historical time window.

[0020] Preferably, it also includes a feedback verification module for verifying the abnormal warning command; the feedback verification module extracts the sampling slices before and after the output time of the abnormal warning command, and uploads the sampling slices to the server for failure mode analysis, while the command data analysis module adjusts the sampling frequency parameters.

[0021] An anomaly detection method implemented by an edge computing-based sensor data fusion anomaly detection system includes the following steps:

[0022] N multi-source heterogeneous digital signals representing the real-time operating status of the controlled object are acquired, and the N multi-source heterogeneous digital signals are normalized to construct a real-time observation vector mapped to an N-dimensional feature space.

[0023] Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform an orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors;

[0024] Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition;

[0025] Calculate the magnitude of the orthogonal residual vector and compare it with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold;

[0026] When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module.

[0027] Compared with existing technologies, the edge computing-based sensor data fusion anomaly detection system of the present invention has the following advantages:

[0028] 1. In anomaly detection of sensor data fusion, a dynamic manifold projection processing architecture is constructed. This architecture maps the raw electrical digital signals output by multi-source heterogeneous sensors into real-time state vectors in a high-dimensional feature space, and performs orthogonal projection mapping using a preset stable subspace basis matrix. This transforms the system's anomaly judgment logic from monitoring the numerical range of a single physical quantity to monitoring the topological stability of the logical coupling relationship between multiple signals. This mechanism utilizes the geometric structural characteristics of signals in digital space. Even if the output values ​​of each sensor channel are within the nominal range, as long as the physical constraint relationship between signals undergoes surface decoupling, the observation vector deviates from the stable subspace in digital space and generates orthogonal residual transitions. This enables the capture of hidden correlated faults and solves the problem of early weak fault features being annihilated due to the smoothing effect in traditional linear fusion methods.

[0029] 2. By integrating lightweight orthogonal projection logic and recursive subspace tracking algorithm into the processing unit, high-dimensional sensing data streams are projected onto low-dimensional feature manifolds in real time, significantly reducing the dimensionality and computational complexity of the data to be processed while ensuring feature fidelity. This processing path uses matrix transformations in linear algebra to replace highly complex nonlinear modeling and reasoning, enabling edge computing nodes to complete millisecond-level real-time responses to large-scale heterogeneous signal streams within limited instruction cycles and memory space. This effectively resolves the contradiction between scarce edge computing resources and massive data throughput, improving the processing performance of industrial monitoring systems under extreme resource boundaries.

[0030] 3. The dynamic residual manifold arbitration mechanism adopted monitors the geodesic offset of the observation vector relative to the stable manifold surface, rather than the absolute coordinates of the signal, thus enabling the system to have intrinsic noise immunity. Since external interference such as ambient temperature fluctuations and level drift usually manifest as linear common-mode components in each signal channel, such interference is naturally shielded by orthogonal space during the projection onto the low-dimensional manifold, thereby significantly reducing the false alarm frequency of the system in complex electromagnetic environments or non-steady-state conditions, and improving the logical reliability of anomaly detection results without relying on complex filtering algorithms. Attached Figure Description

[0031] Figure 1 This is a flowchart of the sensor data fusion anomaly detection process for dynamic manifold projection of the present invention;

[0032] Figure 2 This is a system logic architecture diagram of the present invention, which features cloud-edge collaboration and adaptive computing power. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0034] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0035] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0036] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0037] An edge computing-based sensor data fusion anomaly detection system includes:

[0038] The system includes a data acquisition module, a feature storage module, and a data analysis module; the data analysis module is connected to both the data acquisition module and the feature storage module.

[0039] The data analysis module detects the controlled object through the following processing steps:

[0040] Step S1: Obtain N multi-source heterogeneous digital signals representing the real-time operating state of the controlled object, and perform dimensional normalization processing on the N multi-source heterogeneous digital signals to construct a real-time observation vector mapped to the N-dimensional feature space.

[0041] Step S2: Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors; the coupled feature matrix is ​​generated by singular value decomposition of historical sample data of the controlled object under stable operating conditions.

[0042] Step S3: Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition.

[0043] Step S4: Calculate the magnitude of the orthogonal residual vector and compare the magnitude with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold.

[0044] Step S5: When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module.

[0045] Preferably, it also includes a logic compensation module for maintaining the continuity of the evolution of the N-dimensional feature space on the time axis; the logic compensation module completes the real-time observation vector by refining the following sub-steps of step S1: step S11, monitor the real-time data integrity of the data acquisition module, and when a single data packet loss is detected in the N-channel multi-source heterogeneous digital signals, extract the associated prediction value of the previous sampling period; step S12, use a linear interpolation algorithm to fill in the missing signal bits to correct the topological offset of the real-time observation vector.

[0046] Preferably, the data analysis module is also used to run the residual discrimination mechanism; the residual discrimination mechanism uses orthogonal residual vectors to shield the linear common-mode component interference in the N-channel multi-source heterogeneous digital signals, so as to resist the level drift caused by ambient temperature fluctuations in the N-channel multi-source heterogeneous digital signals.

[0047] Preferably, the data analysis module is also used to adjust the parallel granularity of the processing tasks according to the available computing power resources on the edge side; the data analysis module adjusts the iteration step size of the orthogonal projection transformation according to the value of the available computing power resources, and maintains a dynamic balance between computing power allocation and anomaly detection accuracy while maintaining a processing latency of less than 10ms.

[0048] Preferably, the data acquisition module includes an asynchronous sampling calibration unit for timestamp alignment of the sampling frequency differences of N multi-source heterogeneous digital signals; the asynchronous sampling calibration unit uses the high-frequency signal as the sampling reference to fit the sampling value of the low-frequency signal to ensure the consistency of each component in the real-time observation vector in the time domain.

[0049] Preferably, the feature storage module is also used to receive a global manifold update strategy from the cloud platform; the feature storage module performs online incremental correction on the pre-stored coupling feature matrix according to the aging model of the controlled object, so as to compensate for the drift of the reference manifold caused by the evolution of the physical properties of the controlled object.

[0050] Preferably, the data analysis module is also used to perform multi-dimensional information entropy measurement; the data analysis module calculates the information distribution entropy value of the orthogonal residual vector in different component dimensions, and locates the fault signal path based on the information distribution entropy value.

[0051] Preferably, the preset judgment threshold is set as an adaptive threshold that dynamically shifts with the complexity of the environment; the data analysis module dynamically sets the judgment boundary of the preset judgment threshold based on the variance envelope of the orthogonal residual vector within the historical time window.

[0052] Preferably, it also includes a feedback verification module for verifying the abnormal warning command; the feedback verification module extracts the sampling slices before and after the output time of the abnormal warning command, and uploads the sampling slices to the server for failure mode analysis, while the command data analysis module adjusts the sampling frequency parameters.

[0053] An anomaly detection method implemented by an edge computing-based sensor data fusion anomaly detection system includes the following steps:

[0054] N multi-source heterogeneous digital signals representing the real-time operating status of the controlled object are acquired, and the N multi-source heterogeneous digital signals are normalized to construct a real-time observation vector mapped to an N-dimensional feature space.

[0055] Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform an orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors;

[0056] Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition;

[0057] Calculate the magnitude of the orthogonal residual vector and compare it with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold;

[0058] When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module.

[0059] Example 1: In the monitoring of a high-speed CNC machine tool spindle equipped with an embedded edge computing node with a main frequency of 480MHz, the data acquisition module inside the edge computing node acquires 12 multi-source heterogeneous digital signals in parallel, representing the real-time operating status of the spindle, including vibration, current, and temperature. When the physical correlation between the multi-source signals is decoupled due to early contact fatigue pitting, the absolute value of the digital level output by a single sensor channel is within the nominal safe range. Based on this, the data analysis module normalizes the 12 multi-source heterogeneous digital signals, transforming the original amplitudes of different dimensions to a unified processing space to construct a real-time observation vector mapped to a 12-dimensional feature space. On this basis, the data analysis module retrieves the pre-stored coupling feature matrix in the feature storage module. The coupling feature matrix is ​​composed of manifold basis vectors and is generated by singular value decomposition of 512 historical sliding window sample data collected by the spindle under stable operating conditions. This addresses the impact of mechanical aging of the controlled object during long-term service. The quasi-manifold drift feature storage module embeds an aging evolution prediction model based on the statistical distribution mechanism of Weibull fatigue life and dynamically receives global manifold update strategies from the cloud. The data analysis module establishes an adaptive feature tracking time window based on the degradation rate of the current physical structure assessed by the aging model, continuously extracts the topological normal projection variance between the latest real-time observation data points and the original stable manifold space, and then uses a first-order perturbation approximation matrix decomposition algorithm based on subspace tracking to extract incremental singular values. It performs periodic online orthogonal rotation displacement compensation on the manifold basis vectors that have undergone energy weight decay within the pre-stored coupled feature matrix, thereby completing the incremental update of the feature matrix without performing full high-order recalculation. The data analysis module orthogonally projects and transforms the real-time observation vectors into the stable manifold space spanned by the manifold basis vectors, thus transforming the absolute range check operation for the numerical range of a single physical quantity into a continuous calculation process for measuring the rank deficiency state of a multidimensional digital matrix within the digital manifold space.

[0060] During the orthogonal projection transformation, the logic compensation module continuously monitors the integrity of the digital bitstream received by the data acquisition module. When data packet loss is detected in a digital signal due to non-stationary electromagnetic interference, the logic compensation module extracts the associated predicted value from the previous sampling period and fills in the missing signal bits using a linear interpolation algorithm, thereby maintaining the topological continuity of the 12-dimensional feature space evolving along the time axis. Under these parameters, the data analysis module dynamically adjusts the iteration step size of the orthogonal projection transformation based on the available computing resources at the edge, extracting the normal projection component of the real-time observation vector relative to the stable manifold space while maintaining a computational delay of 8.5ms, and generating an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition. The normal extraction process of the orthogonal residual vector uses the orthogonal complement mapping mechanism of the digital space to shield the linear common-mode component interference caused by the overall fluctuation of the ambient temperature in each channel's digital signal. The data analysis module calculates the orthogonal residual vector. The Euclidean modulus of the quantity is measured and numerically compared with a preset judgment threshold. The preset judgment threshold is calculated and generated according to the system's pre-set engineering calibration procedure. The specific calculation procedure is to extract the boundary value on the variance envelope of the orthogonal residual vector modulus sequence within 30 historical communication cycles under fault-free baseline operating conditions of the spindle, and multiply it by a margin coefficient of 1.25 to set it as the current dynamic judgment boundary. When the data analysis module continuously compares and finds that the modulus of the orthogonal residual vector exceeds the preset judgment threshold for a period of 3 sampling cycles, the data analysis module determines that the spindle has undergone nonlinear structural decoupling and outputs an abnormal warning command to the external execution terminal module. Accompanying this abnormal warning action, the feedback verification module synchronously triggers the failure feature tracing mechanism of the lower-level physical control unit, and immediately instructs the data analysis module to temporarily increase the sampling frequency parameter of a specific high-frequency vibration channel to capture the nanosecond-level transient fracture acoustic emission high-frequency scattered waves excited by the very early surface crack.

[0061] To prevent abrupt changes in the local physical sampling clock band, which could lead to a destructive dimensionality mismatch between the newly generated real-time observation vector's span on the time axis and the preset storage feature matrix, the underlying bus routing of the data acquisition system synchronously activates a multi-rate interpolation decimation cascaded dimensionality reduction digital filter module. This module utilizes band-limited multiphase decimation core logic to dynamically and dynamically smooth and shrink the dense data stream of high-frequency burst resampling back to the system's original reference frequency grid scale in real time, thus maintaining the absolute physical constancy of the spatial dimension of the observation vector fed into the orthogonal reconstruction stage. In multi-source signal systems, the decoupling of nonlinear physical constraints on a single physical path causes abrupt changes in the probability distribution characteristics of the channel residual signal in the time domain. The data analysis module outputs an anomaly warning command to activate the multi-dimensional information entropy positioning logic. The data analysis module then establishes durations for each independent dimension of the acquired orthogonal residual vector. A millisecond sliding data window is used to extract the amplitude range of each dimension component within the window. The data analysis module calculates the discrete probability density by counting the frequency of each dimension component falling within each sub-interval and analyzing the sampling points within each sub-interval. The data analysis module is based on the formula Calculate the orthogonal residual vector in the th case. Information distribution entropy value in dimensions , To characterize the topological divergence of a single signal outside the manifold space, the data analysis module retrieves the initial baseline entropy value of the equipment under baseline operating conditions and calculates the distribution entropy value of the currently acquired information. If the absolute difference between the initial baseline entropy value and the absolute difference of a certain dimension exceeds the baseline difference tolerance of 0.15 times the initial baseline entropy value for five consecutive sampling periods, the data analysis module locks the feature dimension and outputs the specific fault signal path physical port number based on the feature space to sensor physical port address mapping table pre-written by the system. The 0.15 times constant entropy increase jump threshold used as the final arbitration benchmark is a conclusion drawn from the first-level empirical fitting curve point feature boundary obtained by long-term cumulative mapping of a full-scale spindle fatigue life cyclic accelerated tearing failure test dynamometer cluster deployed in the laboratory. The irrefutable evidence data from the accelerated failure peeling test shows that even under extremely harsh stable electrical conditions, various electromechanical equipment can still withstand normal healthy service conditions. The combined effects of magnetic radiation field distortion and the severe parasitic cross-interference of common modes such as thermal drift of conventional environmental sensors mean that the peak height of the total entropy of the system information distribution, which is forcibly pushed up by unforced external random interference, will be suppressed and imprisoned below the margin limit of the low-dimensional safe background noise quagmire, which is no more than 0.08 times the increase rate. It is absolutely impossible to exceed this limit. Only when the multi-channel sensor forced rigid covariance correlation constraint network of the structural body mechanics and physics level, which is irreversibly healable, is instantly and violently torn and brittlely broken, can the massive random eruption effect of the asymmetric high-order discrete signal released by the internal mechanical topology disintegration truly generate and easily cross the 0.15 times normal defense statistical jump barrier on this overall information measurement scale with overwhelming geometric potential energy.

[0062] Example 2: In the hardware-in-the-loop test environment for verifying multi-source heterogeneous digital signal processing logic, a rotor dynamics test bench equipped with a multi-channel synchronous data acquisition card and providing physical experimental data characterizing the rotor's operating state is used to construct the test environment. The length of the sliding window sample data is determined to balance the frequency domain resolution of signal feature extraction and the real-time memory consumption of the edge computing node. The parameter calibration procedure includes extracting the lowest frequency fundamental component from 12 multi-source heterogeneous digital signals, setting the time span of the sliding window to cover at least 3 complete cycles of the fundamental component, and simultaneously constraining the volume of the data matrix generated within this time span to be less than the first-level buffer of the edge computing node. Based on the above procedures and under the condition that the rotor fundamental frequency is 120Hz, 512 sampling points were selected as the length of the sliding window sample data. The length of 512 sampling points satisfies the convergence condition of the singular value decomposition algorithm for extracting the manifold basis vector. The process of verifying the system's ability to cope with disturbances includes injecting Gaussian white noise with a signal-to-noise ratio of 20dB into the original vibration digital signal collected by the test bench, and injecting power frequency harmonics with a frequency of 50Hz into the current digital signal; the external heating device is started simultaneously to make the temperature of the sensor base rise continuously at a rate of 0.5℃ per minute, generating linear common-mode component interference that causes baseline drift of multi-source signals.

[0063] The experiment selected two sets of identical input signal sequences. The control group used static threshold comparison logic based on historical averages, while the experimental group used a data analysis module that included orthogonal complement mapping and dynamic residual discrimination mechanism. In the initial operation phase when the rotor had no structural damage and a continuous temperature rise disturbance was applied, the monitoring parameters of the control group continuously and linearly accumulated due to temperature drift, and the comprehensive characteristic modulus shifted from the initial 1.02 to 2.45, outputting an error warning command. The data analysis module of the experimental group orthogonally projected the real-time observation vector containing temperature drift characteristics into the stable manifold space, stripping away the data residing in the stable manifold space. The environmental heating common mode characteristics within the fixed manifold space were analyzed. The calculated orthogonal residual vector magnitude remained within the range of 0.04 to 0.07. No abnormal warning commands were output. A communication interruption fault with a packet loss rate of 5% was randomly injected into the test bench data stream. The monitoring parameters of the control group showed a transient jump with an amplitude of 4.8. The logic compensation module of the test group extracted historical correlation prediction values ​​and performed linear interpolation algorithm to fill missing signal bits, maintaining the temporal continuity of the rank-deficient state operation of the multidimensional digital matrix. The highest peak value of the orthogonal residual vector magnitude was 0.12, filtering out data artifacts caused by communication packet loss.

[0064] The optimal working range of the margin coefficient within the preset judgment threshold was determined by simulating spalling faults at different depths on the outer ring of a rotor bearing. The numerical measurement system response characteristics of the margin coefficient were adjusted. Measurement data showed that when the margin coefficient was below the lower limit of 1.10, transient high-frequency components caused by speed fine-tuning under normal operating conditions led to the magnitude of the orthogonal residual vector crossing the dynamic judgment boundary, resulting in an 18% false alarm rate. When the margin coefficient exceeded the upper limit of 1.40, the data analysis module exhibited nonlinear hysteresis in its response to early spalling at a depth of 0.1 mm, and the time window required for the magnitude of the orthogonal residual vector to cross the judgment threshold was extended to over 45 ms. With a margin coefficient set at 1.25, as the bearing spalling depth increased from 0.1 mm to 0.3 mm and 0.5 mm, the experimental... The orthogonal residual vector magnitudes of the group calculation outputs show monotonically increasing values ​​of 1.45, 2.86, and 5.32, respectively. The gradient evolution of the magnitude data directly reflects the physical degradation degree of the nonlinear structural decoupling of the controlled object. After the continuous operation cycle of the above-mentioned composite noise interference and nonlinear fault injection, the experimental results confirm that the data analysis process based on the digital space orthogonal complement mapping mechanism removes the linear common mode component and accurately quantifies the nonlinear structural offset of the device. The data processing mechanism transforms the abnormal characterization of multi-source sensor signals from absolute amplitude exceeding the limit to a dynamic measurement of the topological distance of the manifold space. Under the physical condition of communication packet loss, it maintains the topological continuity of the feature space and drives the edge computing nodes to output quantitative early warning commands that match the physical damage depth based on the dynamic residual discrimination mechanism.

[0065] Example 3: This example uses a multi-core processor-based offshore wind turbine pitch control monitoring node to form the operating condition. The data acquisition module receives high-frequency vibration digital signals with a sampling frequency of 10kHz and low-frequency temperature digital signals with a sampling frequency of 1Hz in parallel. The asynchronous deviation in the physical sampling clocks of the two heterogeneous multi-source digital signals causes a topological misalignment in the time dimension of the real-time observation vector constructed from the image. The asynchronous sampling calibration unit inside the data acquisition module generates a globally increasing timestamp sequence using the sampling period of the 10kHz high-frequency vibration signal as the reference clock period. The asynchronous sampling calibration unit obtains the measured temperature value of the 1Hz low-frequency temperature digital signal in the current sampling period and the measured temperature value in the previous sampling period, calculates the linear change slope between the two points, and fills the gap between the two low-frequency sampling points with the fitted measured value based on the globally increasing timestamp sequence. To avoid the aforementioned low-dimensional linear interpolation... The value operation generates a large amount of rigid, absolutely collinear data, causing the covariance matrix of the subsequent multidimensional digital matrix to directly fall into mathematical singularity. When the asynchronous sampling calibration unit fills in the fitted measured values, it synchronously superimposes surface thermodynamic random shot noise that follows a zero-mean Gaussian distribution onto the interpolation sequence points. The physical variance of this injected shot noise is dynamically clamped to be equivalent to the variance of the background intrinsic white noise of the corresponding temperature sensor within the 1Hz effective acquisition bandwidth. Thus, a true physical dispersion that conforms to the objective thermodynamic small fluctuation characteristics is reconstructed on the 10kHz high-frequency time scale. This mechanism ensures the full-rank absolute validity of the multi-source digital signal matrix when participating in the subsequent dimension reduction orthogonal decomposition mapping. After the timestamp sequence is aligned, the data analysis module applies the range normalization algorithm to constrain the values ​​of the multi-source heterogeneous digital signals to the dimensionless interval of 0 to 1, constructing a real-time observation vector in the N-dimensional feature space matched at a single moment.

[0066] The data analysis module retrieves the pre-stored coupled feature matrix from the feature storage module to perform orthogonal projection transformation; the data analysis module reads the system status register of the embedded operating system to obtain the number of active cores and the remaining L1 cache space of the current available computing power resources on the edge side; the data analysis module dynamically adjusts the iteration step size of the orthogonal projection transformation based on the extracted physical hardware resource indicators; when the remaining L1 cache space drops below 20% of the total capacity threshold, the data analysis module truncates the tail high-order components of the singular value decomposition matrix and expands the iteration step size to twice that under standard operating conditions, performs low-rank approximation on the high-order feature matrix to maintain the main features of the orthogonal projection topology of the space, proportionally reduces the number of multiply-accumulate instructions of the underlying controller, and the data analysis module monitors the proportion of the remaining L1 cache space. The warning threshold is reached based on the calculation formula. Calculate the upper limit of the allowed manifold projection dimension. ,in, The data analysis module is based on the number of columns of the full-rank matrix under the baseline operating conditions. Numerical truncation of the coupling feature matrix and low-order manifold basis vectors is performed. The underlying embedded hardware executes real-time observation of the vectors' mapping matrix multiplication operations in the stable manifold space after truncation. The data analysis module adjusts the microprocessor's underlying loop accumulation control logic from a row-by-row traversal single-step iteration mode to a row-by-row read double-step skip iteration mode. This allows the hardware instruction pipeline to skip the inner product operation steps of the truncated low-weight components. This underlying double-step skip iteration mode does not forcibly omit the absolutely orthogonal dimensions in the original topological space, thus destroying basic geometric measures. Instead, it is based on the inherent physical characteristics of a large number of approximately minimal zero elements sparsely distributed within the basis vectors of the high-order manifold after singular value decomposition. The data analysis module guides the microprocessor multiplier instruction set to extract and perform accumulation operations only on the non-sparse coordinate domains of the matrix that emphasize topological energy weights, thereby upholding the orthogonal projection property of multidimensional vector inner products and the rigorous mathematical and physical meaning of Euclidean distance quantification. Under the premise of ensuring basic requirements, a non-destructive physical reduction of the limited computing power resources of the underlying microcontroller is achieved; the number of floating-point operations in matrix multiplication is reduced by reducing the projection accuracy within a preset tolerance range; the hardware computing power adaptive allocation logic limits the single operation latency of orthogonal complement mapping to within 10ms, maintaining the real-time physical response capability of the residual discrimination mechanism; the data analysis module calculates the Euclidean magnitude of the orthogonal residual vector and compares the magnitude with a preset judgment threshold; when the continuous comparison confirms that the magnitude exceeds the preset judgment threshold for a period of 3 sampling periods, the data analysis module cuts off the local control loop of the abnormal physical node and outputs an abnormal warning command to the external execution terminal module; the asynchronous data alignment and computing power adaptive collaborative processing logic solves the technical problems of heterogeneous data fusion timing reconstruction error and limited computing power resource allocation in the edge computing environment, ensuring the continuous output of the abnormal detection sequence within the physical limits of environmental fluctuations and sudden changes in device communication load.

[0067] Example 4: When the system faces the operating conditions of deploying a brand-new offshore wind turbine without a prior physical model, the data analysis module performs a pre-baseline calibration process before activating the residual discrimination mechanism. The data acquisition module monitors the main shaft speed and output power of the wind turbine. Under the condition that the fluctuation rate of the main shaft speed and output power is less than 5% for 24 consecutive hours, the data acquisition module continuously acquires the fault-free prior digital signal sequence. The data analysis module applies the covariance matrix decomposition algorithm to the fault-free prior digital signal sequence to extract the principal component eigenvectors with a cumulative contribution rate of more than 95%, constructs the initial coupled feature matrix, and writes the principal component eigenvectors into the read-only storage area of ​​the feature storage module to establish the fixed geometric reference system corresponding to the orthogonal projection transformation.

[0068] Based on a fixed geometric reference frame, the data analysis module continuously calculates the orthogonal residual vector set generated by real-time multi-source heterogeneous digital signal mapping during the subsequent 72-hour physical break-in operation cycle. The data analysis module extracts the maximum and average values ​​of the Euclidean modulus in the orthogonal residual vector set, and multiplies the absolute value of the difference between the maximum and average values ​​by a control coefficient. The product results are then averaged to calculate a preset judgment threshold that adapts to the current wind turbine mechanical topology, where the control coefficient is... As a dimensionless quantity with values ​​limited to the range of 1.5 to 2.0, the aforementioned extremely narrow closed-loop constraint engineering safety numerical range of 1.5 to 2.0 is not derived from subjective constant patching, but rather from the analytical unenvelope domain extracted by the digital twin dynamic simulation system of a megawatt-level offshore ultra-large transmission chain wind turbine equipment, constructed by combining environmental meteorological and fluid dynamic variables, after 300,000 supercomputing Monte Carlo sensitivity impact experiments. The simulation pressure detection results show that if this W value falls below the critical line of 1.5, even when the controlled unit is dealing with common normal coastal gust wind direction change conditions, the flexible structure turbulent bouncing vibration energy it generates will be significantly reduced. A surge in volume can instantly break through the residual threshold, triggering a complete false alarm mechanism that paralyzes and shuts down the system. Conversely, if the value is set too loosely and exceeds the 2.0 upper limit, the huge tolerance black hole will swallow and erase the tiny ripples of fragile topological feature variation signals caused by the micron-level precursor roller surface micro-detachment evolution of the inner and outer rings of the transmission bearing, which are deeply buried in the low-energy, low-frequency weak signal band. This precisely confirms that the limited range is the physical critical equilibrium state of tolerance for accurately suppressing and calming the interference of violent wind load fluctuations while simultaneously maintaining the sensitivity peak of the irreversible fatigue failure detection of the underlying material. The pre-baseline calibration process compensates for the initial modulus offset introduced by the mechanical tolerance of different equipment and outputs the dynamic judgment boundary of specific physical nodes.

[0069] Example 5: In the edge computing node's feature subspace extraction scenario, the data analysis module receives 512 consecutive real-time observation vectors from the time series and arranges them column-by-column to construct a feature matrix. The data analysis module applies a singular value decomposition algorithm to the feature matrix to extract an orthogonal matrix composed of left singular vectors and a diagonal matrix containing singular values ​​arranged in descending order. The data analysis module calculates the square of each singular value and divides it by the sum of the squares of all singular values ​​to obtain the energy contribution rate of a single dimension. The data analysis module accumulates the energy contribution rate item by item in descending order of singular values. When the accumulated value first reaches a 95% threshold, the corresponding first... A manifold basis vector is constructed using left singular vectors, where... Let the dimensionless positive integer representing the dimension of the subspace be the first... The space spanned by a left singular vector defines the associated topological boundary of the controlled object under stable operating conditions.

[0070] Based on the manifold basis vectors, the data analysis module multiplies the subsequently received real-time observation vectors with the transpose matrix corresponding to the manifold basis vectors to calculate the orthogonal projection coordinates of the real-time observation vectors in the stable manifold space. The data analysis module calculates the product of the manifold basis vectors and the orthogonal projection coordinates to reconstruct the filtered observation vectors, and calculates the numerical difference between the corresponding dimensions of the real-time observation vectors and the filtered observation vectors to generate orthogonal residual vectors. The data analysis module calculates the Euclidean magnitude of the orthogonal residual vectors and outputs the Euclidean magnitude to the preset judgment threshold comparison stage. The feature subspace extraction and dimensionality reduction projection logic transforms the correlation attributes of multi-source sensing signals into matrix order reduction operation steps, and outputs quantized spatial topological distance metric parameters to the dynamic residual discrimination mechanism.

[0071] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. An anomaly detection system based on edge computing sensor data fusion, characterized in that, include: Data acquisition module, feature storage module, and data analysis module; The data analysis module is connected to both the data acquisition module and the feature storage module. The data analysis module detects the controlled object through the following processing steps: Step S1: Obtain N multi-source heterogeneous digital signals representing the real-time operating state of the controlled object, and perform dimensional normalization processing on the N multi-source heterogeneous digital signals to construct a real-time observation vector mapped to the N-dimensional feature space. Step S2: Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors; The coupling feature matrix is ​​generated by singular value decomposition of historical sample data of the controlled object under stable operating conditions. Step S3: Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition. Step S4: Calculate the magnitude of the orthogonal residual vector and compare the magnitude with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold. Step S5: When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module. The data analysis module is also used to adjust the parallel granularity of processing tasks according to the available computing power resources on the edge side; the data analysis module adjusts the iteration step size of the orthogonal projection transformation according to the value of available computing power resources, and maintains a dynamic balance between computing power allocation and anomaly detection accuracy while maintaining a processing latency of less than 10ms; and the data acquisition module acquires a total of 12 multi-source heterogeneous digital signals in parallel, including vibration, current and temperature, which characterize the real-time operating status of the spindle. The feature storage module is also used to receive the global manifold update strategy from the cloud platform; the feature storage module performs online incremental correction on the pre-stored coupling feature matrix according to the aging model of the controlled object to compensate for the drift of the baseline manifold caused by the evolution of the physical properties of the controlled object. The preset judgment threshold is set as an adaptive threshold that dynamically shifts with the complexity of the environment; the data analysis module dynamically sets the judgment boundary of the preset judgment threshold based on the variance envelope of the orthogonal residual vector within the historical time window. It also includes a feedback verification module for verifying abnormal warning commands; the feedback verification module extracts sampling slices before and after the output time of the abnormal warning command, and uploads the sampling slices to the server for failure mode analysis, while the command data analysis module adjusts the sampling frequency parameters.

2. The edge computing-based sensor data fusion anomaly detection system according to claim 1, characterized in that, It also includes a logic compensation module to maintain the continuity of the evolution of the N-dimensional feature space on the time axis. The logic compensation module completes the data of the real-time observation vector by refining the following sub-steps of step S1: Step S11, monitor the real-time data integrity of the data acquisition module, and extract the correlation prediction value of the previous sampling period when a single data packet loss is detected in the N-channel multi-source heterogeneous digital signals; Step S12, fill in the missing signal bits using a linear interpolation algorithm to correct the topological offset of the real-time observation vector.

3. The edge computing-based sensor data fusion anomaly detection system according to claim 1, characterized in that, The data analysis module is also used to run the residual discrimination mechanism; the residual discrimination mechanism uses orthogonal residual vectors to shield the linear common-mode component interference in the N-channel multi-source heterogeneous digital signals, so as to resist the level drift caused by ambient temperature fluctuations in the N-channel multi-source heterogeneous digital signals.

4. The edge computing-based sensor data fusion anomaly detection system according to claim 1, characterized in that, The data acquisition module includes an asynchronous sampling calibration unit, used to align timestamps for the sampling frequency differences of N multi-source heterogeneous digital signals; The asynchronous sampling calibration unit uses the high-frequency signal as the sampling reference and performs sampling value fitting on the low-frequency signal to ensure the consistency of each component in the time domain in the real-time observation vector.

5. The edge computing-based sensor data fusion anomaly detection system according to claim 1, characterized in that, The data analysis module is also used to perform multi-dimensional information entropy measurement; the data analysis module calculates the information distribution entropy value of the orthogonal residual vector in different component dimensions, and locates the fault signal path based on the information distribution entropy value.

6. An anomaly detection method implemented in an edge computing-based sensor data fusion anomaly detection system, used to implement the edge computing-based sensor data fusion anomaly detection system as described in claim 1, characterized in that, Includes the following steps: N multi-source heterogeneous digital signals representing the real-time operating status of the controlled object are acquired, and the N multi-source heterogeneous digital signals are normalized to construct a real-time observation vector mapped to an N-dimensional feature space. Retrieve the coupled feature matrix composed of manifold basis vectors pre-stored in the feature storage module, and perform an orthogonal projection transformation on the real-time observation vector to the stable manifold space spanned by the manifold basis vectors; Extract the normal projection component of the real-time observation vector relative to the stable manifold space to generate an orthogonal residual vector characterizing the degree to which the real-time observation vector deviates from the physical constraints of the stable operating condition; Calculate the magnitude of the orthogonal residual vector and compare it with a preset judgment threshold to determine the duration of the magnitude exceeding the preset judgment threshold; When the duration reaches the preset time window threshold, it is determined that the controlled object has undergone nonlinear structural decoupling, and an abnormal warning command is output to the external execution terminal module.

Images