AI自适应特征映射的多源异构工业设备数据融合方法

Abstract

This invention relates to the field of electronic digital data processing technology and discloses a multi-source heterogeneous industrial equipment data fusion method using AI adaptive feature mapping. The method includes: acquiring heterogeneous original feature streams, monitoring stationarity, and extracting reference anchor points within the physically steady-state interval; constructing an adaptive focusing mapping matrix based on geometric displacement relationships, aligning data in the non-steady-state interval to generate a fused feature vector; constructing a residual distribution tensor through mapping the residual vector, and then calculating its spatial anisotropy index; when the index exceeds a threshold, generating a linear compensation vector based on the centroid offset direction to correct the translation parameters. This invention utilizes the physically steady-state interval as a digital reference system, achieves logical self-healing through secondary mining of the mapping intermediate products, offsets the systematic residuals caused by sensor drift, and improves the physical self-consistency of the fused feature vector.

Description

Technical Field

[0001] This invention relates to a method for fusing multi-source heterogeneous industrial equipment data using AI adaptive feature mapping, belonging to the field of electronic digital data processing technology. Background Technology

[0002] In current industrial equipment operation monitoring, heterogeneous data streams are collected and processed to achieve equipment status identification. Existing practices use interpolation resampling or manifold alignment algorithms to unify high-frequency signals and low-frequency logic data on the time axis and extract feature vectors. Discrete manufacturing conditions involve frequent equipment start-stop switching, and the sensor hardware response time constant has intrinsic differences, causing semantic asymmetry between high sampling rate signals and logic states. Existing technologies increase network depth or increase iteration frequency to approximate the mapping relationship, which introduces processing delays in gateways with limited computing power and makes it difficult to guarantee convergence in noisy environments.

[0003] Sensors are affected by environmental temperature drift or electrical interference, causing the feature mapping relationship to shift over time. Existing discrete calibration schemes only update the reference during stable equipment periods, resulting in observation blind spots between two calibrations. Due to the lack of a mechanism to discriminate the spatial distribution characteristics of the mapping residuals, the system cannot separate random disturbances from systematic biases, leading to non-physical jumps in the fused feature vectors. For example, Chinese invention patent CN120951231A discloses a multi-source heterogeneous data fusion method, device, equipment, and storage medium that uses an external reference system for spatiotemporal benchmark alignment. When the sensor is silently drifting, the algorithm framework lacks sensitivity to the spatial distribution characteristics of the mapping residuals, and cannot identify and correct the structured bias caused by temperature drift. Under dynamic operating conditions, the feature vectors cannot maintain physical self-consistency. This scheme relies on global path distance calculation, which creates a contradiction between computational overhead and real-time response in resource-constrained industrial field gateways, and cannot capture the weak features during transient equipment switching.

[0004] Therefore, how to construct a lightweight mapping mechanism with steady-state constraints and residual feedback, and eliminate the observation blind zone caused by discrete calibration while maintaining physical and logical consistency, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A method for fusing multi-source heterogeneous industrial equipment data using AI adaptive feature mapping, comprising the following steps:

[0006] step Acquire the first and second raw feature streams from heterogeneous data sources;

[0007] step The first original feature stream and the second original feature stream are mapped to the same dimension initial feature space using a preset linear dimensionality reduction matrix to obtain the first feature vector to be fused and the second feature vector to be fused.

[0008] step Calculate the statistical entropy values ​​of the first original feature stream and the second original feature stream within a preset time window, and determine that both the first original feature stream and the second original feature stream have entered the physical steady state region when the statistical entropy value is lower than the preset stability threshold.

[0009] step Calculate the mean centroid of each eigenvector within the physical steady-state interval to extract the first and second reference anchor points;

[0010] step Based on the geometric displacement relationship between the first and second reference anchor points in the initial feature space, an adaptive focus mapping matrix containing translation and rotation parameters is constructed.

[0011] step Using an adaptive focus mapping matrix, spatial alignment is performed on the first and second feature vectors to be fused that are in the non-steady-state region to generate a fused feature vector, and the mapping residual vector generated by the alignment process is extracted.

[0012] step The distribution of the mapped residual vector in the initial feature space within a preset time window is statistically analyzed to construct the residual distribution tensor.

[0013] step Calculate the variance ratio of the residual distribution tensor in each feature dimension to determine the spatial anisotropy index;

[0014] step When the spatial anisotropy index exceeds the preset directional stability threshold, the geometric centroid offset direction of the residual distribution tensor is determined, and a linear compensation vector opposite to the geometric centroid offset direction is generated.

[0015] step The translation parameters in the adaptive focus mapping matrix are updated using the linear compensation vector, and the updated adaptive focus mapping matrix is ​​used to return the execution steps. Alignment processing is used to achieve directional cancellation of systematic residuals.

[0016] Preferred steps The method for determining the preset stability threshold includes: acquiring the original signal of the heterogeneous data source during the constant speed operation phase; calculating the fluctuation variance of the original signal; and using the fluctuation variance as the preset stability threshold to determine the distribution stability of the first original feature stream and the second original feature stream.

[0017] Preferably, in step The following steps are also included: storing the parameter evolution deviation of the adaptive focusing mapping matrix between two adjacent physical steady-state intervals; determining the mapping momentum operator based on the parameter evolution deviation and the corresponding time interval; performing first-order linear prediction compensation on the current adaptive focusing mapping matrix using the mapping momentum operator in the non-steady-state interval to obtain the quasi-transient mapping matrix; and processing the corresponding feature flow using the quasi-transient mapping matrix.

[0018] Preferred steps The method utilizes a pre-set linear dimensionality reduction matrix to map the first and second original feature streams to an initial feature space of the same dimension, including the following steps: calculating the local information entropy density of the first and second original feature streams; determining the mapping dimension ratio of each original feature stream based on the proportional relationship between the local information entropy densities of the first and second original feature streams; and constructing asymmetric mapping matrices according to the mapping dimension ratio to map the first and second original feature streams to an initial feature space with flexible dimension constraints.

[0019] Preferably, the method further includes the following steps: monitoring the signal-to-noise ratio (SNR) of the first and second original feature streams; injecting a standardized detection signal into the adaptive focusing mapping matrix when the SNR is lower than a preset SNR threshold; extracting the transient response sequence generated by the adaptive focusing mapping matrix in response to the detection signal; calculating the coupling sensitivity index of the current feature mapping logic based on the attenuation slope and mapping broadening of the transient response sequence; and adjusting the step size parameter of the adaptive focusing mapping matrix based on the coupling sensitivity index.

[0020] Preferably, the standardized detection signal is defined as a digital logic pulse sequence orthogonal to the main feature frequency, used to maintain the operator activity of the adaptive focus mapping matrix when the first and second original feature streams are in the feature-depleted period.

[0021] Preferably, after generating the fused feature vector, the method further includes the following steps: calculating the magnitude of the mapped residual vector; when the magnitude exceeds a preset residual threshold, performing incremental correction on the fused feature vector using a piecewise linear compensation function; wherein, the piecewise linear compensation function employs... This function is used to correct nonlinear distortions in transient processes.

[0022] Preferred steps The anisotropy index in the medium space is expressed by the formula Calculated; where, The spatial anisotropy index, The first principal eigenvalue of the residual distribution tensor. It is the arithmetic mean of the remaining eigenvalues ​​excluding the first principal eigenvalue.

[0023] Preferably, it further includes the following step: in a length of The circular register stores the mapped residual vector to update the residual distribution tensor; where, The range of values ​​is to .

[0024] Preferred steps The spatial alignment process includes the following steps: using a weighted merging logic based on tensor product to reorganize the projection components of different dimensions into an asymmetric unified state description vector; and outputting the unified state description vector to the industrial control unit for performing equipment operating status identification.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. In multi-source heterogeneous industrial equipment data of AI adaptive feature mapping, the physical steady-state interval of the industrial process is used as the benchmark anchor point to transform the high-dimensional nonlinear spatial alignment process of heterogeneous feature flow into linear focusing processing based on centroid offset vector; the natural stable segment during equipment operation is used as a digital reference system to avoid excessive consumption of computing resources by full manifold iteration, eliminate the pseudo features introduced by traditional interpolation methods due to sampling step size misalignment, and maintain the physical self-consistency of the fused feature vector in reflecting the equipment deterioration trend.

[0027] 2. Based on the interaction between the information entropy density dimensional elastic reduction mechanism and the dynamic decoupling of the mapping operator, the mapping space rank constraint is adjusted in real time according to the contribution weight of each heterogeneous feature stream within the time window; the numerical noise generated by low information intensity data sources during the alignment process is suppressed, and the texture loss of high information intensity features during dimensional compression is prevented, so that the fused state description vector accurately captures the physical information in the asymmetric data source, improving the system's accuracy in representing complex working conditions; the distribution characteristics of the mapping residual generated during the projection process in the feature space are extracted, and the anisotropy index of the residual space is calculated to realize the identification of systematic alignment deviations caused by environmental temperature drift or performance degradation of the front-end sensor; the linear compensation vector generated by the mapping intermediate product is used to perform real-time correction of the translation operator parameters, eliminating the accuracy dead zone caused by physical channel fading, and achieving continuous self-healing of the mapping logic without relying on external anchor points.

[0028] 3. When the feature stream enters the low signal-to-noise ratio range, a standardized logic perturbation signal is injected into the mapping logic and the transient response sequence is extracted to invert the coupling sensitivity index of the feature mapping matrix. When the physical signal enters the feature-poor period, the activity of the mapping operator is maintained to eliminate the initial fusion delay when the equipment enters the dynamic load from the quasi-static state, ensuring that the feature mapping logic is in the optimal focus standby state before the arrival of the real physical signal. The mapping momentum accumulation mechanism works in conjunction with the quasi-transient mapping matrix prediction logic to perform linear prediction compensation using the parameter evolution rate vector between adjacent steady-state calibration cycles. This solves the mapping hysteresis problem caused by discrete trigger calibration during the transient evolution period of the equipment, ensuring the semantic continuity of the feature fusion process at the moment of physical state switching. Attached Figure Description

[0029] Figure 1 This is a flowchart of the adaptive feature mapping method for residual feedback closed loop of the present invention;

[0030] Figure 2 This is a waveform comparison diagram of the spatial alignment and fusion effect of multi-source heterogeneous feature vectors in this invention;

[0031] Figure 3 This is a data fusion system architecture diagram for the present invention, which integrates steady-state anchor points and residual compensation mechanisms. Detailed Implementation

[0032] This specific embodiment is intended to provide a detailed description of the present invention. The following specific embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0033] This invention provides a multi-source heterogeneous industrial equipment data fusion method using AI adaptive feature mapping, comprising five core steps: heterogeneous feature stream access, dimensional elastic reduction, physical steady-state identification, geometric space focusing, and residual feedback closed-loop correction. The processing system synchronously monitors feature streams from different physical channels, utilizing the inherent stable operating phase of the industrial process as a digital benchmark to align high-frequency physical signals and low-frequency logic states on the feature manifold. Furthermore, it corrects systematic deviations caused by sensor environmental temperature drift through secondary mining of residuals accompanying the mapping process. In the data access and preprocessing stage, the system acquires a first and a second original feature stream from heterogeneous data sources. This addresses the issue of different sensor physical dimensions. To address the feature space description inaccuracy problem caused by inconsistent sampling frequencies, the processor calls a preset linear dimensionality reduction matrix to map the first and second original feature streams to an initial feature space of the same dimension, obtaining the first and second feature vectors to be fused. This process involves a dimensionality elastic reduction procedure. The processor calculates the local information entropy density of the first and second original feature streams and determines the mapping dimensionality ratio of each original feature stream based on the proportional relationship between their local information entropy densities. According to the mapping dimensionality ratio, asymmetric mapping matrices are constructed to map heterogeneous signals to an initial feature space with elastic dimensional constraints. In practice, the local information entropy density reflects the heterogeneous signal at the sampling frequency in real time. The logic for obtaining the feature sparsity and mapping dimension ratio within the sample window is as follows: the higher the entropy density, the more basis vectors are allocated to retain detailed textures; conversely, strong compression is performed through an asymmetric matrix to filter out redundant noise, thereby ensuring semantic consistency between high-frequency vibration and low-frequency logic in terms of geometric geodesic distance. In the specific deployment method, the first original feature stream is set as a 64-dimensional vibration signal, and the second original feature stream is set as 128-dimensional process state data. The local information entropy density ratio is calculated to be 2:1, so the mapping dimension ratios allocated to the initial feature space are 24 dimensions and 12 dimensions, respectively. Finally, the two are projected onto a 32-dimensional fusion space to suppress numerical noise introduced by low-information-intensity data sources and construct an asymmetric mapping matrix. At that time, the processor extracts the sliding window sequence of the feature stream, calculates the local information entropy density, determines the mapping weight distribution based on the contribution of each source entropy value, performs feature diagonalization on the original covariance matrix using singular value decomposition (SVD), selects the previous eigenvector with a cumulative contribution rate of more than 90% to form the projection basis, performs linear transformation to achieve dimensionality compression, and uses the magnitude of the singular values ​​of each feature dimension as the physical weight allocation criterion. If the physical component corresponding to the principal eigenvalue changes abruptly, the column vector basis of the asymmetric matrix is ​​induced to undergo orthogonal rotation. The initial coefficients of the mapping matrix are verified by standard orthogonalization to ensure that the local geodesic distance of the physical manifold is maintained in the feature space to be fused after projection, and to avoid phase distortion of high-dimensional nonlinear signals during the compression process.

[0034] To extract an accurate alignment benchmark, the system executes a physical steady-state identification procedure; the processor calculates the statistical entropy values ​​of the first and second original feature streams within a preset time window; when these values ​​are lower than a preset stability threshold, it is determined that both the first and second original feature streams have entered the physical steady-state region; the preset stability threshold is determined by: acquiring the original signals of the heterogeneous data source during the constant-speed operation phase, calculating their fluctuation variance, and using this as the preset stability threshold to determine the stationarity of the distribution; the processor calculates the mean centroid of each feature vector within the physical steady-state region, and extracts the first and second benchmark anchor points; Based on the geometric displacement relationship between the first and second reference anchor points in the initial feature space, an adaptive focusing mapping matrix containing translation and rotation parameters is constructed. In the non-steady-state region, the processor uses the adaptive focusing mapping matrix to perform spatial alignment processing on the first and second feature vectors to be fused, generating a fused feature vector and extracting the mapping residual vector generated by the alignment processing. The spatial alignment processing adopts a weighted merging logic based on tensor product to reorganize the projection components of different dimensions into an asymmetric unified state description vector, which is then output to the industrial control unit for equipment operating status identification.

[0035] To address the silent drift caused by electromagnetic interference or performance aging of the sensor, the system performs residual feedback closed-loop correction; the processor has a length of The circular register stores the mapped residual vector to update the residual distribution tensor, where The value range is selected as 100ms to 500ms. The logic behind setting this time range is that it fully covers at least 5 to 10 standard communication cycles of mainstream industrial bus protocols such as EtherCAT or Profinet. This is sufficient to capture transient residual disturbances from sensors while filtering out random high-frequency pulses caused by electrical interference through time-domain statistical averaging, ensuring the topological stability of the residual distribution tensor. The processor statistically analyzes the distribution of the mapped residual vector in the initial feature space within the preset time window to construct the residual distribution tensor and calculates the variance ratio of this tensor in each feature dimension to determine the spatial anisotropy index. Spatial anisotropy index Through formula Calculated The first principal eigenvalue of the residual distribution tensor. The arithmetic mean of the remaining eigenvalues ​​excluding the first principal eigenvalue; when the spatial anisotropy index When the preset directional stability threshold is exceeded, the geometric centroid offset direction of the residual distribution tensor is determined, and a linear compensation vector opposite to the geometric centroid offset direction is generated. Since a significant increase in the anisotropy of the residual distribution indicates a structural shift in the sensor's physical configuration or environmental temperature drift, the system performs parameter closed-loop correction through changes in physical state. The processor uses this linear compensation vector to update the translation parameters in the adaptive focus mapping matrix and returns to perform alignment processing, achieving directional cancellation of systematic residuals. If the magnitude of the mapped residual vector exceeds the preset residual threshold, the system uses a piecewise linear compensation function to perform incremental correction on the fused feature vector. This piecewise linear compensation function employs... When solving the adaptive focus mapping matrix, the processor establishes a geometric correspondence between the first and second reference anchor points in the physical steady-state interval. It solves the rotation matrix and translation vector by minimizing the objective function, obtains the optimal orthogonal transformation operator based on the Kabsch algorithm, and uses the geometric centroid bias of the residual distribution tensor as input for the linear compensation vector. It performs negative feedback adjustment, where the feedback gain is dynamically calibrated according to the sampling signal-to-noise ratio. If the anisotropy exponent of the residual continues to rise, the translation parameters are compressed and the weights are updated through the logistic step size regression algorithm, forcing the system to return to the physically self-consistent steady state and eliminating the risk of manifold tearing caused by temperature drift.

[0036] To address the perception blind spot generated during state transitions in discrete-triggered calibration, the system introduces a mapping momentum accumulation mechanism. The processor stores the parameter evolution deviation of the adaptive focusing mapping matrix between two adjacent physical steady-state intervals and determines the mapping momentum operator based on this deviation and the corresponding time interval. In the non-steady-state interval, the mapping momentum operator is used to perform first-order linear prediction compensation on the current matrix to obtain a quasi-transient mapping matrix, which is then used to process the corresponding feature stream. When industrial equipment enters a low-load or quasi-static phase, resulting in feature depletion, the system initiates an active detection mechanism. The processor monitors the signal-to-noise ratio (SNR) of the first and second original feature streams. When the SNR falls below a preset SNR threshold, a standardized detection signal, i.e., a signal orthogonal to the main feature frequency, is injected into the adaptive focusing mapping matrix. A sequence of word logic pulses is used to maintain operator activity. In specific implementation, the physical meaning of this probe signal is to act as a digital tracer. By observing the transient response of the matrix to orthogonal pulses, the operator stiffness of the current mapping logic is mapped in real time. The processor extracts the transient response sequence of the adaptive focusing mapping matrix in response to the probe signal, and calculates the coupling sensitivity index of the current feature mapping logic based on its attenuation slope and mapping broadening. Further, the coupling sensitivity index is defined as the reciprocal of the product of the attenuation slope and the mapping broadening in the initial feature space, which is used to quantify the ability of the mapping operator to capture weak shifts in the underlying signal manifold. Based on the coupling sensitivity index, the step size parameter of the adaptive focusing mapping matrix is ​​adjusted to ensure that the system is in a preset standby state when the physical signal is recovered.

[0037] Example 1: The AI ​​adaptive feature mapping multi-source heterogeneous industrial equipment data fusion method provided by this invention is executed under discrete manufacturing conditions. In this application scenario, due to the difference in the response time constants of sensor hardware, during the transient process of frequent equipment start-up and shutdown, high-frequency vibration signals and low-frequency program logic controller state data produce semantic asymmetry on the time axis, causing non-physical jumps in the fused feature vectors. To solve this problem, the system acquires a first original feature stream and a second original feature stream from heterogeneous data sources, and uses a preset linear dimensionality reduction matrix to map the first original feature stream and the second original feature stream to an initial feature space of the same dimension, obtaining a first feature vector to be fused and a second feature vector to be fused. The system calculates the length of the first original feature stream and the second original feature stream in the data fusion method. The statistical entropy value within a preset time window is used. When the statistical entropy value is lower than a preset stability threshold, the first and second original feature streams are determined to have entered the physical steady-state interval. The preset stability threshold is obtained by acquiring the original signals of the heterogeneous data source during the constant-speed operation phase and calculating their fluctuation variance. The processor calculates the mean centroid of each feature vector within the physical steady-state interval, extracts the first and second reference anchor points, and constructs an adaptive focusing mapping matrix containing translation and rotation parameters based on the geometric displacement relationship of the first and second reference anchor points in the initial feature space. Using the adaptive focusing mapping matrix, the first and second feature vectors to be fused within the non-steady-state interval are spatially aligned to generate fused feature vectors, and the mapping residual vector generated by the alignment process is extracted.

[0038] When dealing with mapping offsets caused by environmental temperature drift or sensor aging, the system statistically analyzes the distribution of the mapping residual vector in the initial feature space within a preset time window to construct a residual distribution tensor. It then calculates the variance ratio of the residual distribution tensor across each feature dimension to determine the spatial anisotropy index. Spatial anisotropy index It can be obtained through the following formula: ,in, The spatial anisotropy index, The first principal eigenvalue of the residual distribution tensor. The arithmetic mean of the remaining eigenvalues ​​excluding the first principal eigenvalue; when the spatial anisotropy index When the preset directional stability threshold is exceeded, the geometric centroid offset direction of the residual distribution tensor is determined, and a linear compensation vector opposite to the geometric centroid offset direction is generated to update the translation parameters in the adaptive focusing mapping matrix, thereby achieving directional cancellation of systematic residuals. To eliminate mapping hysteresis within the calibration period, the system stores the parameter evolution deviation of the adaptive focusing mapping matrix between two adjacent physical steady-state intervals, and determines the mapping momentum operator based on the parameter evolution deviation and the corresponding time interval. In the unsteady-state interval, the mapping momentum operator is used to perform first-order linear prediction compensation on the current adaptive focusing mapping matrix to obtain the quasi-transient mapping matrix and process the corresponding features. The system first identifies the source flow; when the signal-to-noise ratio (SNR) of each original feature flow is lower than a preset SNR threshold, a standardized probe signal is injected into the adaptive focusing mapping matrix, defined as a digital logic pulse sequence orthogonal to the main feature frequency; the processor extracts the transient response sequence generated by the adaptive focusing mapping matrix in response to the probe signal, calculates the coupling sensitivity index of the current feature mapping logic based on the attenuation slope and mapping broadening of the transient response sequence, adjusts the step size parameter of the adaptive focusing mapping matrix to ensure that the system is in a preset standby state before the physical signal is recovered, and calculates the magnitude of the mapping residual vector after generating the fused feature vector; when the magnitude exceeds a preset residual threshold, the system uses... The piecewise linear compensation function performs incremental correction on the fused feature vector to handle nonlinear distortions in the transient process; the spatial alignment process reorganizes the projection components of different dimensions into an asymmetric unified state description vector through a weighted merging logic based on tensor product, and outputs it to the industrial control unit for equipment operating status identification.

[0039] Example 2: In the operational status monitoring scenario of a precision bearing machining production line, a high-frequency vibration sensor and a low-frequency programmable logic controller constitute a heterogeneous data environment. Due to electromagnetic environment fluctuations caused by cutting fluid spraying and the thermal expansion and contraction physical effect of the sensor bracket, the first and second original feature flows experience phase drift and spatial configuration shift. The experimental platform is based on a distributed industrial data acquisition architecture, including a maximum sampling rate of... The vibration signal acquisition unit and communication cycle are The logic controller is used to simulate transient conditions during sudden changes in processing load; the data sequence used in the experiment was acquired through a physical experimental platform, whose sensor measurement resolution is [missing information]. The system clock synchronization accuracy is better than In the experimental design, the step size parameter of the adaptive focusing mapping matrix was calibrated. This parameter's setting is constrained by the signal-to-noise ratio (SNR) of the monitored signal and the rate of temperature drift evolution of the sensor. The step size parameter is used to balance the response sensitivity of the mapping logic with its steady-state numerical stability. When the SNR is decreasing, to suppress the interference of random disturbances on the mapping results, the step size parameter tends towards the lower limit of its value range. When the temperature drift evolution rate increases, to maintain the real-time performance of feature focusing, the step size parameter is adjusted towards the upper limit of its value range. When the SNR is... And the thermal displacement rate is Under the operating conditions, the initial value of the step size parameter determined by the above logic is... .

[0040] The experiment involved triggering the data stream access program at the moment the production line started. The processor used a preset linear dimensionality reduction matrix to project the first and second original feature streams to a dimension of [dimensional value missing]. The initial feature space; to simulate a real industrial environment, a set of root mean square amplitudes is actively superimposed at the input. The processor monitors the statistical entropy value within the sliding window, and after identifying the physical steady-state interval of the constant speed processing section, extracts the first and second reference anchor points to solve the adaptive focusing mapping matrix; as the processing load increases, the processor extracts the mapping residual vector and constructs the residual distribution tensor to calculate the spatial anisotropy index. When the spatial anisotropy index When a trend of increase occurs, the system generates a linear compensation vector and performs online correction on the translation parameters in the adaptive focus mapping matrix, so that the error of the fused feature vector converges to within a preset threshold.

[0041] Table 1: Performance Comparison Data of Different Batches under Transient Operating Conditions

[0042]

[0043] According to the data in Table 1, when the system is subjected to electromagnetic interference of the same intensity, the control group lacks a discrimination mechanism for the spatial distribution characteristics of the mapped residual, resulting in a lower spatial anisotropy index. The level is relatively high, resulting in a feature fusion error of [value missing]. The present invention utilizes residual feedback closed-loop correction logic to adjust the spatial anisotropy index. Adjust to Near this point, the feature fusion error drops to [a certain value]. The accuracy of state recognition is from Upgraded to Spatial anisotropy index Through formula Calculated The spatial anisotropy index, The first principal eigenvalue of the residual distribution tensor. The arithmetic mean of all eigenvalues ​​except the first principal eigenvalue is given. Analysis of the partially missing control group data shows that after removing the residual feedback loop, the system cannot offset the structured bias caused by temperature drift, proving a synergistic effect between residual distribution feature mining and the principal mapping logic. Data from the out-of-range control group shows that when the update step size exceeds the upper limit... At that time, due to the response to random noise, the fusion error surges and the recognition accuracy deteriorates.

[0044] Example 3: This example combines Figures 1 to 3 This document describes a method for fusing multi-source heterogeneous industrial equipment data using AI adaptive feature mapping, such as... Figure 1 As shown, the processing flow begins with acquiring the first raw feature stream from heterogeneous data source A and the second raw feature stream from heterogeneous data source B. Step S2 involves initial feature space mapping, generating feature vectors to be fused using a linear dimensionality reduction matrix. Next, step S3 performs statistical entropy monitoring to determine if the value is below a preset stability threshold. If the determination result indicates entry into a steady-state path, step S4 extracts the reference anchor point and calculates the centroid of the mean of each feature vector, then proceeds to step S5 to construct an adaptive focusing mapping matrix and solve for translation and rotation parameters. If the determination result indicates entry into a non-steady-state path, step S6, the non-steady-state region, is executed. The spatial alignment process within the space uses a matrix to align and generate a fused feature vector, which is then output to the industrial control unit. Simultaneously, the mapped residual vector is extracted and fed into a feedback loop. In this loop, step S7 is executed to construct a residual distribution tensor and statistically analyze the distribution state of the mapped residual vector. In step S8, the spatial anisotropy index is calculated, i.e., the variance ratio of each feature dimension is calculated. In step S9, it is determined whether the index exceeds the directional stability threshold. If it exceeds the threshold, step S10 is executed to generate a linear compensation vector and update the translation parameters according to the centroid offset. The updated parameters are used to orient and offset the systematic residuals, and the process returns to step S6 for cyclic processing.

[0045] like Figure 2 As shown, in a Cartesian coordinate system where the horizontal axis represents the feature dimension with values ​​ranging from 0 to 48, and the vertical axis represents the feature vector values ​​with values ​​ranging from 5 to 14, the feature vector waveform curves in three different states are displayed. The solid line represents the first feature vector to be fused before alignment, exhibiting high amplitude fluctuations across multiple dimensions; the dashed line represents the second feature vector to be fused before alignment, with an overall amplitude level lower than the first feature vector and a difference in waveform phase; the dotted line represents the curve after feature vector alignment, falling between the first two, showing the reconstructed feature distribution trend after algorithmic processing. Figure 3As shown, the system architecture revolves around the central adaptive focus mapping core, namely the spatial alignment / feature fusion module. The left input end includes a high-frequency vibration acquisition unit that generates the first original feature stream and a logic state controller that generates the second original feature stream. The two are connected to the central core through a linear dimensionality reduction injection path and an asymmetric matrix projection path, respectively. Above the architecture is a physical steady-state identification anchor point module to provide a reference system and to send reference data to the central core through the centroid extraction path. Below the architecture is a residual anisotropy compensation module responsible for generating linear compensation vectors. It forms a closed-loop control by receiving the mapping residuals from the central core and feeding back parameter correction signals. On the right side of the architecture is an industrial control unit used to receive a unified state description vector, which is the processing result sent by the central core through the fusion feature output path.

[0046] Example 4: In the state monitoring scenario of high-speed collaborative robot welding, the processing system faces a heterogeneous data environment consisting of the angle feature stream fed back by the six-axis joint encoder and the vibration feature stream collected by the end effector accelerometer. Due to the instantaneous strong electromagnetic interference generated during the welding process and the thermal expansion and contraction physical effect caused by the continuous rotation of the robotic arm, there is a spatial topological offset between the first and second original feature streams. The processor acquires the joint data in real time through the industrial bus and calls the preset linear dimensionality reduction matrix to perform projection transformation. The spatial alignment processing is achieved through the following procedure: the processor extracts the aligned first feature vector to be fused. With the second feature vector to be fused The projected components; perform tensor mapping based on Kronecker product to calculate and The tensor product matrix is ​​obtained; a pre-defined weighted coefficient matrix is ​​used to linearly reorganize the tensor product matrix. The logic for obtaining the weighted coefficient matrix is ​​as follows: based on the cross-correlation moment matrix of the first and second feature vectors to be fused in the physical steady-state interval, the principal component contribution rate is extracted in real time and used as the gain weight of the spatial projection, thereby shielding the random noise of the incoherent dimension during the tensor reorganization process; the energy component on the main diagonal is extracted as an asymmetric unified state description vector; when the processor enters the physical steady-state interval, it initiates the calibration procedure of the directional stability threshold, which is based on the statistical analysis of the distribution of the mapping residual vector in the physical steady-state interval; the processor calculates the overall variance of the mapping residual vector sequence in the steady-state interval. And according to the probability distribution The initial values ​​are determined according to principles; the spatial anisotropy index is monitored. The baseline fluctuation range; the directional stability threshold is set to the steady-state condition. 1.5 to 2.0 times the mean; in specific numerical paths, if the overall variance of the mapped residual vector within the steady-state interval is... The calculated value is 0.02. With a steady-state mean of 1.05, the directional stability threshold is calculated by the processor logic as follows: ; The spatial anisotropy index, The overall variance of the mapped residual vector sequence is used; this dynamic calibration procedure based on the background noise level enables the processing system to maintain the adaptability of the judgment logic under different electromagnetic intensity environments.

[0047] In addition, to address the issue of adaptive focusing mapping matrix parameter failure caused by sensor temperature drift evolution, the processing system executes a closed-loop compensation procedure based on feedback gain; when the spatial anisotropy index When the value exceeds the calibrated 1.85, the processor identifies the geometric centroid offset direction of the residual distribution tensor. The processor generates a linear compensation vector with the opposite direction and adjusts the step size parameter according to the following adaptive logic: calculate the rate of change between the magnitude of the mapped residual vector at the current moment and the magnitude at the previous moment; when the rate of change is negative and the absolute value is increasing, maintain the current step size parameter; when the rate of change is positive, determine compensation overload, and compress the step size parameter according to the preset attenuation factor of 0.5. Experimental data shows that, under the condition of an initial step size set to 0.1, due to the intervention of the closed-loop adjustment mechanism, the linear compensation vector guides the feature fusion error to decrease from 0.75 and stabilize at around 0.08 within 5 calculation cycles, avoiding numerical divergence during the compensation process. The experimental environment is set at a welding current of... And the ambient temperature fluctuation reached Under °C conditions, the processing system utilizes a digital reference frame determined by the physical steady-state range, combined with tensor reconstruction logic and residual distribution mining, to monitor the operating status of the robotic arm.

[0048] Example 5: In an industrial robot joint monitoring scenario involving multiple heterogeneous sensors, the processor initiates an initial calibration procedure before performing real-time data fusion processing to determine the linear dimensionality reduction matrix and a preset stability threshold. During the robot's preset trajectory motion phase, the processor simultaneously acquires the first and second original feature streams from the heterogeneous data sources. After performing standardization preprocessing on the original feature streams, the processor extracts the first 32 principal component components through covariance matrix decomposition and calculates the initial coefficient distribution of the linear dimensionality reduction matrix. During the robot's uniform speed running segment, the processor extracts a sliding window with a length of 1000 sampling points, calculates the second moment of the signal amplitude within the window, and obtains the baseline fluctuation variance. And based on the power spectral density of the background electrical noise, Compensation and correction are implemented, and the corrected values ​​are stored in non-volatile memory, which is then used as the preset stability threshold for determining the physical steady-state interval. When the robot enters the load cycle stage of the production site, the pre-debugging procedure is initiated to correct the spatial topological displacement caused by the mechanical installation deviation of the sensors. When the processor captures the first physical steady-state interval, it calculates the statistical moment characteristics of each feature vector in the initial feature space. By solving the centroid position difference between the first and second feature vector sets to be fused, the initial components of the translation parameters in the adaptive focusing mapping matrix are determined. The singular value decomposition method is used to solve for the direction of the principal eigenvectors, and the initial phase of the rotation parameters is determined. The procedure defines the geometric deviation between sensors as the initial state value of the mapping matrix. When the system faces an unsteady acceleration process, it calls the focusing parameters to perform spatial alignment on the first and second feature vectors to be fused.

[0049] In scenarios involving the deployment of multiple heterogeneous sensor clusters, the processor executes a feature space normalization procedure to establish a physical mapping relationship between the first and second original feature streams; the processor simultaneously acquires the original signal sequences of the sensors during the idle operation phase and calculates the sample mean of each sensor channel. with sample standard deviation The first and second original feature streams are converted into dimensionless standardized feature vectors with a mean of 0 and a variance of 1. The processor uses a linear dimensionality reduction matrix to project the standardized feature components onto an initial feature space of dimension 32 to determine the configuration of the first and second feature vectors to be fused on the cross-domain manifold. When the system detects that the signal-to-noise ratio of the first and second original feature streams is lower than a preset signal-to-noise ratio threshold, the processor starts a standardized probe signal injection program. It uses built-in orthogonal pulse synthesis logic to generate a digital logic pulse sequence with a pulse width of 1ms and an amplitude of 5% of the normalized feature vector magnitude, and injects it as a standardized probe signal into the adaptive focus mapping matrix. The transient response sequence generated in response to the probe signal is extracted, and the attenuation slope of the transient response sequence is used as the basis for the extraction. and the mapping broadening in the initial feature space. The coupling sensitivity index of the feature mapping logic is calculated, and the step size parameter of the adaptive focusing mapping matrix is ​​locked in the waiting interval to maintain the semantic continuity of the feature fusion process at the moment of state switching.

[0050] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for fusing multi-source heterogeneous industrial equipment data using AI adaptive feature mapping, characterized in that, Includes the following steps: step Acquire the first and second raw feature streams from heterogeneous data sources; step The first original feature stream and the second original feature stream are mapped to the same dimension initial feature space using a preset linear dimensionality reduction matrix to obtain the first feature vector to be fused and the second feature vector to be fused. step Calculate the statistical entropy values ​​of the first original feature stream and the second original feature stream within a preset time window, and determine that both the first original feature stream and the second original feature stream have entered the physical steady state region when the statistical entropy value is lower than the preset stability threshold. step Calculate the mean centroid of each eigenvector within the physical steady-state interval to extract the first and second reference anchor points; step Based on the geometric displacement relationship between the first and second reference anchor points in the initial feature space, an adaptive focus mapping matrix containing translation and rotation parameters is constructed. step Using an adaptive focus mapping matrix, spatial alignment is performed on the first and second feature vectors to be fused that are in the non-steady-state region to generate a fused feature vector, and the mapping residual vector generated by the alignment process is extracted. step The distribution of the mapped residual vector in the initial feature space within a preset time window is statistically analyzed to construct the residual distribution tensor. step Calculate the variance ratio of the residual distribution tensor in each feature dimension to determine the spatial anisotropy index; step When the spatial anisotropy index exceeds the preset directional stability threshold, the geometric centroid offset direction of the residual distribution tensor is determined, and a linear compensation vector opposite to the geometric centroid offset direction is generated. step The translation parameters in the adaptive focus mapping matrix are updated using the linear compensation vector, and the updated adaptive focus mapping matrix is ​​used to return the execution steps. Alignment processing is performed to achieve targeted cancellation of systematic residuals; And, in the steps The process then includes the following steps: storing the parameter evolution deviation of the adaptive focusing mapping matrix between two adjacent physical steady-state intervals; determining the mapping momentum operator based on the parameter evolution deviation and the corresponding time interval; performing first-order linear prediction compensation on the current adaptive focusing mapping matrix using the mapping momentum operator in the non-steady-state interval to obtain the quasi-transient mapping matrix; and processing the corresponding feature flow using the quasi-transient mapping matrix. Monitor the signal-to-noise ratio (SNR) of the first and second original feature streams; when the SNR is lower than a preset SNR threshold, inject a standardized detection signal into the adaptive focusing mapping matrix; extract the transient response sequence generated by the adaptive focusing mapping matrix in response to the detection signal; calculate the coupling sensitivity index of the current feature mapping logic based on the attenuation slope and mapping broadening of the transient response sequence; and adjust the step size parameter of the adaptive focusing mapping matrix based on the coupling sensitivity index. The standardized detection signal is defined as a sequence of digital logic pulses orthogonal to the main feature frequency, used to maintain the operator activity of the adaptive focus mapping matrix when the first and second original feature streams are in the feature-depleted period.

2. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, step The method for determining the preset stability threshold includes: acquiring the original signal of the heterogeneous data source during the constant speed operation phase; calculating the fluctuation variance of the original signal; and using the fluctuation variance as the preset stability threshold to determine the distribution stability of the first original feature stream and the second original feature stream.

3. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, step The method utilizes a pre-set linear dimensionality reduction matrix to map the first and second original feature streams to an initial feature space of the same dimension, including the following steps: calculating the local information entropy density of the first and second original feature streams; determining the mapping dimension ratio of each original feature stream based on the proportional relationship between the local information entropy densities of the first and second original feature streams; and constructing asymmetric mapping matrices according to the mapping dimension ratio to map the first and second original feature streams to an initial feature space with flexible dimension constraints.

4. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, After generating the fused feature vector, the following steps are also included: calculating the magnitude of the mapped residual vector; when the magnitude exceeds a preset residual threshold, performing incremental correction on the fused feature vector using a piecewise linear compensation function; wherein, the piecewise linear compensation function adopts... This function is used to correct nonlinear distortions in transient processes.

5. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, step The anisotropy index in the medium space is expressed by the formula Calculated; where, The spatial anisotropy index, The first principal eigenvalue of the residual distribution tensor. It is the arithmetic mean of the remaining eigenvalues ​​excluding the first principal eigenvalue.

6. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, It also includes the following steps: In length of The circular register stores the mapped residual vector to update the residual distribution tensor; where, The range of values ​​is to .

7. The method for fusion of multi-source heterogeneous industrial equipment data using AI adaptive feature mapping according to claim 1, characterized in that, step The spatial alignment process includes the following steps: using a weighted merging logic based on tensor product to reorganize the projection components of different dimensions into an asymmetric unified state description vector; and outputting the unified state description vector to the industrial control unit for performing equipment operating status identification.

Images