An electromagnetic clutch response delay evaluation method based on edge computing

Abstract

This invention discloses a method for evaluating the response delay of an electromagnetic clutch based on edge computing, comprising: acquiring multi-source time-series data and preprocessing it to generate a multi-channel sample sequence; calculating the local similarity between channels to establish a time-varying adjacency relation matrix sequence; aggregating the time-varying adjacency relation matrix sequence and weightedly fusing it with a physical topology matrix to form a topological constraint matrix; performing Laplace eigenmaps to generate low-dimensional manifold eigenvectors; mapping the low-dimensional manifold eigenvectors and inputting them into an improved quantum heuristic neural network model to generate a quantum hidden space representation; measuring the quantum hidden space representation to generate a response delay estimate and confidence level; and performing adaptive correction to generate the final delay result. This invention achieves accurate evaluation and dynamic correction of the response delay of an electromagnetic clutch by fusing Laplace eigenmaps and an improved quantum heuristic neural network model.

Description

Technical Field

[0001] This invention relates to the fields of intelligent manufacturing and industrial automation technology, and in particular to a method for evaluating the response delay of an electromagnetic clutch based on edge computing. Background Technology

[0002] Currently, with the rapid development of intelligent manufacturing and equipment control, electromagnetic clutches are increasingly widely used in industrial automation, intelligent transportation, and high-precision control systems. Their performance stability and response speed have become key factors affecting the overall system reliability. Traditional evaluation of electromagnetic clutch response characteristics mainly relies on experimental data analysis and empirical formula fitting, calculating delays by collecting single signals such as current, voltage, and speed. However, under complex operating conditions, the dynamic characteristics of electromagnetic clutches are significantly affected by multiple factors such as temperature, load, and electromagnetic interference. The nonlinear coupling and multi-source heterogeneity of signals make it difficult for traditional methods to accurately characterize the change law of response delay. Furthermore, offline analysis methods lack real-time capability and are highly dependent on manual parameter tuning, failing to meet the needs of intelligent and online system testing.

[0003] At the data analysis and modeling level, existing research has gradually introduced edge computing and intelligent algorithms to achieve real-time perception and delay prediction of electromagnetic clutch states. However, problems still exist, such as shallow data feature extraction, insufficient description of structural relationships, and inadequate model adaptability. While some deep learning-based solutions have improved the fitting ability of nonlinear features, they still have limitations in multi-source data coupling, temporal evolution modeling, and quantum-heuristic structural optimization. Traditional feature mapping and embedding methods struggle to effectively preserve topological structure and temporal coupling information, resulting in insufficient generalization ability and stability of models in dynamic environments.

[0004] Therefore, how to provide a method for evaluating the response delay of an electromagnetic clutch based on edge computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for evaluating the response delay of electromagnetic clutches based on edge computing. This invention utilizes multi-source time-series data fusion, Laplace eigenmaps, and quantum heuristic neural network models to describe in detail the entire process of online evaluation of electromagnetic clutch response delay on edge computing nodes. By constructing a topological constraint matrix reflecting temporal coupling characteristics and structural correlation features, an improved Laplace eigenmap is executed to extract low-dimensional manifold features from multi-source data, and an incremental update mechanism is introduced to adapt to time-varying operating conditions. The low-dimensional feature vectors are mapped to complex quantum state inputs, and an improved quantum heuristic neural network model is used to achieve high-dimensional entanglement and latent space aggregation between features. This invention's method achieves real-time evaluation and adaptive correction of electromagnetic clutch response delay, possessing advantages such as high computational efficiency, strong environmental adaptability, high feature extraction accuracy, and good stability of evaluation results. It can effectively support intelligent diagnostics and equipment self-sensing applications at the edge.

[0006] An edge computing-based method for evaluating the response delay of an electromagnetic clutch, according to an embodiment of the present invention, includes: Multi-source time-series data of the electromagnetic clutch during engagement and release are collected, the multi-source time-series data are preprocessed, a fixed-length sliding time window is constructed based on the action trigger time, and a multi-channel sample sequence arranged in chronological order is generated. Within each sliding time window, local similarity is calculated based on the Euclidean distance and linear correlation between different channels in the multi-channel sample sequence, and a time-varying adjacency matrix sequence is established. Based on the structural layout of the electromagnetic clutch and the signal acquisition circuit, a physical topology matrix is ​​constructed. The comprehensive adjacency matrix formed by weighted aggregation of the time-varying adjacency matrix sequence is then fused with the physical topology matrix according to the weights to obtain a topology constraint matrix containing time coupling information and structural association information. On edge computing nodes, Laplacian eigenmaps are performed on the topological constraint matrix to construct the manifold embedding space of the data samples. Low-dimensional feature vectors are extracted and dynamically corrected through an incremental update mechanism to generate updated low-dimensional manifold feature vectors. The updated low-dimensional manifold eigenvectors are mapped to quantum state inputs, and complex feature representations are formed through amplitude and phase encoding. These complex feature representations are then input into an improved quantum heuristic neural network model to generate a quantum state latent space representation. Perform measurement operations on the hidden space representation of the quantum state to generate response delay estimates and corresponding confidence indices; Adaptive correction is performed on the response delay estimate of the electromagnetic clutch and the corresponding confidence index on the edge computing node to generate the final delay estimate result, and the final delay estimate result is synchronously stored in the edge computing node database.

[0007] Optionally, the multi-source timing data specifically includes electromagnetic coil current data, control terminal voltage data, output shaft speed data, mechanical structure vibration data, and ambient temperature data.

[0008] Optionally, the preprocessing of multi-source time-series data specifically includes denoising, time synchronization, anomaly repair, normalization, and smoothing enhancement.

[0009] Optionally, establishing the time-varying adjacency relation matrix sequence includes: A multi-channel sample sequence is represented as a data set consisting of multiple channel signals changing over time. Each channel corresponds to a time change sequence of a physical quantity, and the sampling time is a continuous time point. Within each fixed-length sliding time window, calculate the sum of squares of the differences between data points of any two channel signals at the same time, and then take the square root of the sum of squares to obtain the Euclidean distance between the two channel signals within the sliding time window. Compare the trends of change of any two channel signals within the same time window and calculate the degree of linear correlation; The Euclidean distance between the two channels and the degree of linear correlation are weighted and fused according to a preset ratio to determine the local similarity between the channels; Based on the local similarity between each channel, establish the adjacency matrix corresponding to the current sliding time window; Arrange all the adjacency matrices generated within the sliding time window in chronological order to form a time-varying adjacency matrix sequence.

[0010] Optionally, obtaining the topological constraint matrix containing temporal coupling information and structural association information includes: Based on the structural layout of the electromagnetic clutch and the signal acquisition line, the channel set is determined and the node and connection relationship is established. A physical topology matrix is ​​constructed. The physical topology matrix is ​​a square matrix, and the number of rows and columns corresponds to the number of multi-source signal channels. Each element in the matrix represents the physical connection strength between the corresponding channels. The time-varying adjacency matrix sequence is numbered in chronological order, with each sliding time window corresponding to an adjacency matrix, and time weight parameters are set for each sliding time window. Based on the time weight parameter, the time-varying adjacency relation matrix sequence is weighted and superimposed to obtain the comprehensive adjacency matrix; The composite adjacency matrix and the physical topology matrix are weighted and fused according to preset fusion coefficients and then normalized to obtain the topology constraint matrix.

[0011] Optionally, generating the updated low-dimensional manifold feature vector includes: Based on the topological constraint matrix, the connection weights of each node are accumulated to calculate the degree matrix. Based on the degree matrix and the topological constraint matrix, the Laplace matrix reflecting the dynamic changes between each channel is calculated. Construct the manifold embedding space of the data samples and establish the corresponding objective function. Introduce temporal smoothing constraints, regularization constraints and structural correlation constraints to keep the interconnected channels as close as possible in the low-dimensional space. Eigenvalue decomposition of the Laplacian matrix is ​​performed in the manifold embedding space to obtain a set of eigenvalues ​​and eigenvectors; from these, the eigenvectors corresponding to several of the smallest non-zero eigenvalues ​​are selected to form an initial set of low-dimensional manifold eigenvectors. The initial low-dimensional manifold eigenvector set is standardized to obtain the initial low-dimensional feature matrix of the current sliding time window; An incremental update operation is performed on the edge computing node. The low-dimensional feature matrix of the previous sliding time window is weighted and fused with the initial low-dimensional feature matrix of the current sliding time window. The update ratio is determined according to the feature change rate to obtain the updated low-dimensional feature matrix.

[0012] Optionally, the generation of the quantum state hidden space representation includes: An improved quantum heuristic neural network model is constructed, comprising a quantum state encoding layer, a gated entanglement processing unit, a complex residual coupling layer, and a quantum hidden space aggregation layer, wherein: The quantum state encoding layer is responsible for mapping the low-dimensional manifold eigenvectors into time-sensitive complex quantum states through amplitude normalization and phase modulation. The gated entanglement processing unit enhances the correlation between features by setting the entanglement strength and the gate coefficient to couple and interfere between multi-channel quantum states. The complex residual coupling layer performs residual fusion in the complex domain to maintain stable feature propagation and prevent information decay. The quantum hidden space aggregation layer performs weighted aggregation based on channel attention weights and introduces temporal smoothing and incremental update mechanisms to continuously evolve and dynamically correct the hidden space representation, thus constructing a quantum state hidden space representation that combines temporal consistency and structural correlation. A phase bias adaptive mechanism is introduced into the quantum state coding layer. The dynamic phase bias is calculated based on the action triggering time, the node correlation strength in the topological constraint matrix, and the signal energy change rate. The updated low-dimensional manifold eigenvector is input into the quantum state coding layer, and amplitude normalization and phase modulation processing are performed to generate a coding matrix with time-dependent characteristics. The encoding matrix is ​​input into the gated entanglement processing unit, and quantum entanglement operations between channels are performed according to the entanglement strength and the gate coefficient. Quantum interference paths between channels are selectively activated through a variable gate structure. When there are strong correlation characteristics between specific channels, the entanglement weight is increased to form a coherent superposition state. Finally, the entanglement matrix reflecting the quantum correlation between channels is output. The entanglement matrix is ​​input into the complex residual coupling layer to perform residual fusion operation based on the complex domain; the amplitude and phase components of the main channel and the sub-channel signals are weighted and superimposed respectively through the residual weight adaptive adjustment mechanism; then the amplitude and phase normalization processing is performed to output the residual matrix that maintains stable propagation of features. The residual matrix is ​​input into the quantum hidden space aggregation layer. The importance of each channel feature is calculated based on the channel attention weight, and a weighted aggregation operation is performed to generate an initial hidden space matrix that reflects the quantum distribution structure of multi-channel features. A temporal smoothing and incremental update mechanism is introduced in the quantum hidden space aggregation layer. The initial hidden space matrices of adjacent sliding time windows are weighted and fused according to the feature change rate. When the feature change between the current sliding time window and the previous sliding time window is small, the high weight is inherited. When the change is large, the weight of the current feature is increased. Finally, the updated hidden space matrix is ​​generated. The updated latent space matrix is ​​subjected to mutation detection and smoothing regularization. When an energy mutation or phase aberration of the channel features is detected, the incremental update weights are dynamically adjusted according to the mutation intensity and regularization constraints are applied. After smoothing and regularization, the final smoothed latent space matrix is ​​generated, and the smoothed latent space matrix is ​​used as the quantum state latent space representation of the final output of the improved quantum heuristic neural network model.

[0013] Optionally, the generation of the response delay estimate and the corresponding confidence index includes: Establish the correspondence between the quantum state hidden space representation and each channel of the multi-channel sample sequence, and define the measurement basis and measurement operator used for measurement; Based on measurement basis and measurement operator, measurement preprocessing is performed on the hidden space representation of quantum state to complete amplitude modulus square and channel normalization, and the observation probability of each channel is obtained. Multiple independent measurements are performed within the same sliding time window, and the observation probabilities of each channel are denoised and weighted to form robust observation results. Based on the calibrated measurement weights and biases, robust observation results are converted into response delay estimates, and a confidence index is calculated that is a combination of measurement dispersion, measurement residuals, and distribution uncertainty. The estimated response delay and its corresponding confidence index are checked against the preset threshold range and physical constraints. If the check passes, the estimated response delay and its corresponding confidence index are output. If the check fails, the parameters are updated and remeasured.

[0014] Optionally, generating the final delay estimation result includes: Based on the estimated response delay and the corresponding confidence index, combined with the operating status of the edge nodes and the on-site working conditions, the correction constraints are determined. The physical condition correction is calculated based on changes in ambient temperature, power supply voltage, and load. The data correction is calculated based on the confidence index, residuals, and trend terms. The two types of corrections are weighted and combined to form a total correction. The phase offset parameters and measurement layer weights are updated to obtain the corrected response delay estimate and confidence index. Based on the corrected response delay estimate and confidence level, an adaptive strategy is executed sequentially to generate a stable response delay estimate and confidence level as the final delay estimation result. The final delay estimation results, correction parameters, time stamps, and operating condition information are written into the edge computing node database.

[0015] The beneficial effects of this invention are: This invention achieves high-precision online evaluation and adaptive optimization of electromagnetic clutch response delay by integrating advanced technologies such as multi-source time-series data, Laplace eigenmaps, and improved quantum-inspired neural network models. Compared with traditional methods that rely on single-signal monitoring and static threshold judgment, this invention comprehensively utilizes multi-source heterogeneous data such as current, voltage, speed, vibration, and temperature to construct a topological constraint model reflecting temporal coupling and structural correlation characteristics, enabling full-domain feature characterization of the clutch engagement and release dynamic process under complex operating conditions. By performing Laplace eigenmaps and incremental manifold updates on edge computing nodes, this invention achieves low-dimensional embedding representation and continuous evolution of multi-channel time-series features, effectively enhancing the sensitivity and stability to changes in dynamic response delay.

[0016] In intelligent modeling, this invention proposes an improved quantum-inspired neural network structure, introducing mechanisms such as quantum state encoding, gated entanglement, and latent space aggregation to achieve high-dimensional interaction and enhanced global correlation of multi-channel complex features. Combined with the distributed processing of an edge computing architecture, this invention enables real-time deployment and adaptive correction of the model at the field equipment, significantly reducing evaluation latency and improving robustness. Through quantum state latent space measurement and dynamic confidence assessment, this invention can output response latency estimates and their confidence indices, providing a highly reliable basis for monitoring equipment performance degradation and predicting health status. This invention overcomes the problems of low data coupling, weak dynamic feature extraction capabilities, and insufficient evaluation timeliness in existing technologies, providing an innovative solution for intelligent monitoring, online diagnosis, and adaptive control of electromagnetic clutches. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an electromagnetic clutch response delay evaluation method based on edge computing proposed in this invention; Figure 2 This is a schematic diagram of the improved quantum heuristic neural network model structure for an edge computing-based electromagnetic clutch response delay evaluation method proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 and Figure 2 An edge computing-based method for evaluating the response delay of an electromagnetic clutch includes: Multi-source time-series data of the electromagnetic clutch during engagement and release are collected, the multi-source time-series data are preprocessed, a fixed-length sliding time window is constructed based on the action trigger time, and a multi-channel sample sequence arranged in chronological order is generated. Within each sliding time window, local similarity is calculated based on the Euclidean distance and linear correlation between different channels in the multi-channel sample sequence, and a time-varying adjacency matrix sequence is established. Based on the structural layout of the electromagnetic clutch and the signal acquisition circuit, a physical topology matrix is ​​constructed. The comprehensive adjacency matrix formed by weighted aggregation of the time-varying adjacency matrix sequence is then fused with the physical topology matrix according to the weights to obtain a topology constraint matrix containing time coupling information and structural association information. On edge computing nodes, Laplacian eigenmaps are performed on the topological constraint matrix to construct the manifold embedding space of the data samples. Low-dimensional feature vectors are extracted and dynamically corrected through an incremental update mechanism to generate updated low-dimensional manifold feature vectors. The updated low-dimensional manifold eigenvectors are mapped to quantum state inputs, and complex feature representations are formed through amplitude and phase encoding. These complex feature representations are then input into an improved quantum heuristic neural network model to generate a quantum state latent space representation. Perform measurement operations on the hidden space representation of the quantum state to generate response delay estimates and corresponding confidence indices; Adaptive correction is performed on the response delay estimate of the electromagnetic clutch and the corresponding confidence index on the edge computing node to generate the final delay estimate result, and the final delay estimate result is synchronously stored in the edge computing node database.

[0020] In this embodiment, the multi-source timing data specifically includes electromagnetic coil current data, control terminal voltage data, output shaft speed data, mechanical structure vibration data, and ambient temperature data.

[0021] In this embodiment, the preprocessing of multi-source time-series data specifically includes denoising, time synchronization, anomaly repair, normalization, and smoothing enhancement.

[0022] In this embodiment, establishing the time-varying adjacency relation matrix sequence includes: A multi-channel sample sequence is represented as a data set consisting of multiple channel signals changing over time. Each channel corresponds to a time change sequence of a physical quantity, and the sampling time is a continuous time point. Within each fixed-length sliding time window, calculate the sum of squares of the differences between data points of any two channel signals at the same time, and then take the square root of the sum of squares to obtain the Euclidean distance between the two channel signals within the sliding time window. The trends of any two channel signals within the same sliding time window are compared, and the degree of linear correlation is calculated. Specifically, the calculation of the degree of linear correlation involves: Within the same sliding time window, the product of the corresponding sampling points of any two channel signals is taken and the average value is calculated to obtain the co-variance. The squared average of the differences between the sampling points and the average value of the two channel signals is calculated to obtain the fluctuation intensity of the two channel signals. The co-variance is divided by the product of the fluctuation intensity of the two channel signals to obtain the linear correlation degree between negative one and one. The Euclidean distance between the two channels and their linear correlation are weighted and fused according to a preset ratio to determine the local similarity between the channels. Specifically, determining the local similarity between the channels involves: Based on the sensitivity of the influence of Euclidean distance and linear correlation on the delay estimation results in historical samples, corresponding weight coefficients are set. Within the current sliding time window, the Euclidean distance is normalized and then inverted to represent the similarity direction. It is then weighted and summed with the linear correlation according to the weight coefficients to obtain a local similarity that comprehensively reflects the amplitude difference and trend consistency. Based on the local similarity between channels, an adjacency matrix corresponding to the current sliding time window is established. Specifically, establishing the adjacency matrix corresponding to the current sliding time window involves: Each channel signal is used as the row and column index of the matrix, and the local similarity between any two channels is filled into the corresponding matrix position to form a symmetrical adjacency matrix. Arrange all the adjacency matrices generated within the sliding time window in chronological order to form a time-varying adjacency matrix sequence.

[0023] In this embodiment, obtaining the topological constraint matrix containing temporal coupling information and structural association information includes: Based on the structural layout of the electromagnetic clutch and the signal acquisition line, the channel set is determined and the node and connection relationship is established. A physical topology matrix is ​​constructed. The physical topology matrix is ​​a square matrix, and the number of rows and columns corresponds to the number of multi-source signal channels. Each element in the matrix represents the physical connection strength between the corresponding channels. The time-varying adjacency relation matrix sequence is numbered chronologically, with each sliding time window corresponding to one adjacency relation matrix. A time weight parameter is set for each sliding time window. Specifically, setting the time weight parameter for each sliding time window involves: For each sliding time window, the adjacency matrix is ​​assigned a time weight parameter according to the time sequence, and the time weight parameter decreases with the window number. Based on the time weight parameter, the time-varying adjacency relation matrix sequence is weighted and superimposed to obtain a comprehensive adjacency matrix. Specifically, the comprehensive adjacency matrix is ​​obtained as follows: Each adjacency matrix is ​​multiplied element by element with its corresponding time weight, and then summed and normalized to obtain a comprehensive adjacency matrix that reflects the average correlation strength between channels over the entire time period. The composite adjacency matrix and the physical topology matrix are weighted and fused according to a preset fusion coefficient and then normalized to obtain a topology constraint matrix. Specifically, the obtained topology constraint matrix is ​​as follows: The total connection weight of each node is calculated row by row in the merged matrix, and each element is divided by the total connection weight to obtain a row-normalized topological weight matrix. Global range constraints are applied to the row-normalized topological weight matrix, and all elements in the matrix are linearly mapped according to the maximum and minimum values, so that the overall weight distribution is limited to between zero and one. Boundary corrections are performed on possible numerical overflows and minima in the matrix to form a numerically stable and structurally consistent topological constraint matrix.

[0024] In this embodiment, generating the updated low-dimensional manifold feature vector includes: Based on the topological constraint matrix, the connection weights of each node are accumulated to calculate the degree matrix. Specifically, the calculation of the degree matrix is ​​as follows: The total connection weight of a node is obtained by summing all the connection weights in the corresponding row of each node in the topological constraint matrix. The total connection weights of each node are then filled into the diagonal positions to construct the degree matrix. Based on the degree matrix and the topological constraint matrix, the Laplace matrix reflecting the dynamic changes between each channel is calculated. Specifically, the calculated Laplace matrix is ​​as follows: The topological constraint matrix is ​​symmetricized and nonnegativity checked to remove diagonal self-loop weights while maintaining topological constraint symmetry. The total connection weights of corresponding nodes are taken at the diagonal positions of the topological constraint matrix, and the negative values ​​of the connection weights between corresponding channels are taken at the off-diagonal positions, forming a matrix structure where the total weights are at the diagonal and the negative connection weights are at the off-diagonal positions, resulting in the updated topological constraint matrix. The updated topological constraint matrix is ​​then normalized, and boundary corrections are performed on minimal values ​​to obtain a positive semi-definite and numerically stable Laplace matrix. Construct the manifold embedding space of the data samples and establish the corresponding objective function. Introduce temporal smoothing constraints, regularization constraints, and structural correlation constraints to keep the interconnected channels as close as possible in the low-dimensional space, where: Temporal smoothing constraints compare the changes in low-dimensional feature vectors in adjacent sliding time windows, constraining the difference in low-dimensional feature vectors between adjacent sliding time windows to not exceed a set threshold, thus maintaining the continuity and stability of low-dimensional feature vectors in the time series. Regularization constraints constrain the overall energy level of low-dimensional eigenvectors, resulting in low-dimensional eigenvectors that maintain balance on a numerical scale. Structural correlation constraints apply strong coupling constraints to the correlation channels based on the connection strength between nodes in the Laplacian matrix, keeping the embedding positions of the correlation channels close in the low-dimensional space, and preserving the temporal continuity and structural coupling relationship of the signal in the dimensionality reduction mapping. The establishment of the objective function is specifically as follows: A structure preservation term is established based on the connection weights between nodes in the topological constraint matrix. A temporal smoothing term is introduced, and the feature with excessively large changes is constrained by calculating the difference in changes of the low-dimensional feature matrix within adjacent sliding time windows. Finally, a regularization term is added to perform energy normalization and scale balance on the overall feature vector. By weighting and minimizing the structure preservation term, temporal smoothing term, and regularization term, an objective function that both preserves structural correlation and has temporal smoothness is obtained. The Laplacian matrix is ​​eigenvalued and decomposed in the manifold embedding space to obtain a set of eigenvalues ​​and eigenvectors. Several eigenvectors corresponding to the smallest non-zero eigenvalues ​​are selected to form an initial set of low-dimensional manifold eigenvectors. Specifically, this initial set of low-dimensional manifold eigenvectors is constructed as follows: The Laplacian matrix is ​​mapped to the manifold embedding space to obtain a normalized Laplacian matrix that preserves the geometric relationships between samples. A target dimension is set in the manifold embedding space, and eigenvalues ​​and eigenvectors of the normalized Laplacian matrix are obtained through eigenvalue decomposition. During the decomposition process, iterative optimization and orthogonal constraints ensure the stability and discriminability of the eigenvectors. After the eigenvalue decomposition algorithm converges, vectors corresponding to zero eigenvalues ​​are removed, retaining only the eigenvectors corresponding to the smallest non-zero eigenvalues. The retained eigenvectors are combined sequentially to construct an initial set of low-dimensional manifold eigenvectors reflecting the manifold structure of the samples. The initial low-dimensional manifold eigenvector set is standardized to obtain the initial low-dimensional feature matrix of the current sliding time window. Specifically, the standardization process for the initial low-dimensional manifold eigenvector set involves: The mean and variance of each initial low-dimensional manifold feature vector within the current sliding time window are calculated; zero-mean processing is performed on each initial low-dimensional manifold feature vector and scaling is performed according to the variance; amplitude limiting and outlier clipping are performed, and consistency correction is performed on the same feature direction of adjacent sliding time windows; the corrected initial low-dimensional manifold feature vectors are combined column by column to form the initial low-dimensional feature matrix of the current sliding time window, and the standardization process ends. An incremental update operation is performed on the edge computing node. The low-dimensional feature matrix of the previous sliding time window is weighted and fused with the initial low-dimensional feature matrix of the current sliding time window. The update ratio is determined according to the feature change rate to obtain the updated low-dimensional feature matrix.

[0025] In this embodiment, the generation of the quantum state hidden space representation includes: An improved quantum heuristic neural network model is constructed, comprising a quantum state encoding layer, a gated entanglement processing unit, a complex residual coupling layer, and a quantum hidden space aggregation layer, wherein: The quantum state encoding layer is responsible for mapping the low-dimensional manifold eigenvectors into time-sensitive complex quantum states through amplitude normalization and phase modulation. The gated entanglement processing unit enhances the correlation between features by setting the entanglement strength and the gate coefficient to couple and interfere between multi-channel quantum states. The complex residual coupling layer performs residual fusion in the complex domain to maintain stable feature propagation and prevent information decay. The quantum hidden space aggregation layer performs weighted aggregation based on channel attention weights and introduces temporal smoothing and incremental update mechanisms to continuously evolve and dynamically correct the hidden space representation, thus constructing a quantum state hidden space representation that combines temporal consistency and structural correlation. The construction of the improved quantum heuristic neural network model specifically involves: A quantum state encoding layer is added to the quantum state input layer of the quantum heuristic neural network model to form a quantum state encoding layer. Amplitude normalization and phase modulation are used to transform the low-dimensional manifold feature vectors into a time-dependent encoding matrix. A gated entanglement processing unit is inserted after the quantum state encoding layer to enhance the feature correlation between different channels by setting the entanglement strength and gate coefficient. A complex residual coupling layer is inserted after the gated entanglement processing unit to perform residual fusion in the complex domain and output the residual matrix. A quantum latent space aggregation layer is inserted after the complex residual coupling layer to perform weighted aggregation of multi-channel features according to channel attention weights and introduce time-series smoothing and incremental update mechanisms to generate an updated latent space matrix. An adaptive phase bias mechanism is introduced into the quantum state encoding layer. The dynamic phase bias is calculated based on the action trigger time, the node correlation strength in the topological constraint matrix, and the signal energy change rate. The updated low-dimensional manifold eigenvector is then input into the quantum state encoding layer, where amplitude normalization and phase modulation are performed to generate an encoding matrix with time-dependent characteristics. The phase offset adaptive mechanism calculates the dynamic phase offset based on the action trigger time, the correlation strength between nodes in the topological constraint matrix, and the rate of change of channel signal energy. When the correlation between nodes is enhanced, the phase offset adaptive mechanism automatically adjusts the phase offset; when the signal changes steadily, the phase offset tends to converge. The calculation of the dynamic phase offset is specifically as follows: The reference phase benchmark for the current sliding time window is determined based on the action trigger time; the correlation strength of each node in the topological constraint matrix is ​​extracted, and the energy change rate of each channel signal within the current sliding time window is calculated; the correlation strength and energy change rate are weighted and fused according to preset weights to obtain the timing response coefficients at the node level; the reference phase benchmark is used as the starting point, and the phase is dynamically adjusted according to the magnitude of the timing response coefficients to generate the corresponding dynamic phase offset. The encoding matrix is ​​input into the gated entanglement processing unit, and quantum entanglement operations between channels are performed based on the entanglement strength and the gate coefficient. A variable gate structure is used to selectively activate quantum interference paths between channels. When strong correlations exist between specific channels, entanglement weights are increased to form a coherent superposition state. The final output is an entanglement matrix reflecting the quantum correlation between channels. Specifically, the output entanglement matrix reflecting the quantum correlation between channels is as follows: The initial entanglement strength parameters are determined based on the similarity and correlation between the features of each channel in the encoding matrix. A gating coefficient is introduced to dynamically control the entanglement activation state between different channels according to the importance of the channel and the rate of feature change. When the gating coefficient exceeds a set threshold, the quantum interference path between the corresponding channels is opened, and the superposition amplitude and phase relationship are adjusted according to the entanglement strength. This process is repeated between multiple channels to form a multi-level quantum superposition structure. The entanglement matrix reflecting the quantum correlation and interference relationship between channels is obtained through normalization. The entanglement matrix is ​​input into the complex residual coupling layer, and a residual fusion operation based on the complex domain is performed. Through an adaptive adjustment mechanism for residual weights, the amplitude and phase components of the main channel and branch channel signals are weighted and superimposed respectively. Subsequently, amplitude and phase normalization processing is performed, and a residual matrix that maintains stable feature propagation is output. Specifically, the output residual matrix that maintains stable feature propagation is as follows: The signals of each channel in the entanglement matrix are decomposed into amplitude components and phase components, representing signal strength and phase characteristics, respectively. The signal strength of the main channel is used as the benchmark, and an adaptive adjustment mechanism for residual weights is introduced to calculate the residual based on the phase and amplitude components between the branch channels and the main channel. The residual is applied to the amplitude and phase components respectively, and the complex domain of the features is fused by weighted superposition. During the fusion process, the amplitude and phase of the superposition result are normalized, and the residual matrix that maintains stable propagation of the features is output. The residual matrix is ​​input into the quantum latent space aggregation layer. The importance of each channel feature is calculated based on the channel attention weights, and a weighted aggregation operation is performed to generate an initial latent space matrix reflecting the quantum distribution structure of the multi-channel features. Specifically, generating the initial latent space matrix reflecting the quantum distribution structure of the multi-channel features involves: Based on the residual matrix, the amplitude and phase components of each channel are extracted as importance indicators of channel characteristics. The channel attention weights are calculated based on the indicators, and then the complex quantum states of each channel are superimposed according to the attention weights through weighted aggregation operations. During the aggregation process, phase consistency and energy conservation are maintained at the same time. Finally, an initial quantum hidden space matrix that can comprehensively characterize the distribution relationship of multi-channel features is generated. A temporal smoothing and incremental update mechanism is introduced within the quantum hidden space aggregation layer. The initial hidden space matrices of adjacent sliding time windows are weighted and fused according to their feature change rates. When the feature changes between the current and previous sliding time windows are small, high weights are maintained; when the changes are large, the weights of the current features are increased. Finally, an updated hidden space matrix is ​​generated, where: The temporal smoothing and incremental update mechanism maintains the continuity and stability of the initial latent space matrix in the time series, while also having the ability to respond quickly to changes in new features. By comparing the feature change rates between adjacent sliding time windows, it adaptively adjusts the fusion ratio of historical information and current features. When the change is small, it smoothly inherits the features, and when the change is large, it increases the weight of the current features. The updated latent space matrix undergoes mutation detection and smoothing regularization. When an energy mutation or phase shift in the channel features is detected, the incremental update weights are dynamically adjusted based on the mutation intensity, and regularization constraints are applied. After smoothing and regularization, the final smoothed latent space matrix is ​​generated. This smoothed latent space matrix is ​​used as the quantum state latent space representation of the final output of the improved quantum heuristic neural network model. Specifically, the generation of the final smoothed latent space matrix involves: Abrupt changes are detected based on the energy changes and phase shifts of channel features between adjacent sliding time windows. When a sudden increase in energy or an abnormal phase shift is detected, the incremental update weights are dynamically adjusted according to the intensity of the abrupt change to weaken the influence of abnormal channels in the overall matrix. Smoothing constraints are applied to the feature components of adjacent channels and the time dimension in the updated latent space matrix. Then, through regularization, excessive shifts in feature amplitude and phase are limited, resulting in a smooth latent space matrix.

[0026] In this embodiment, the generation of the response delay estimate and the corresponding confidence index includes: Establish the correspondence between the quantum state hidden space representation and each channel of the multi-channel sample sequence, and define the measurement basis and measurement operator used for measurement; Based on measurement basis and measurement operator, measurement preprocessing is performed on the quantum state hidden space representation, completing amplitude modulus square and channel-wise normalization to obtain the observation probability of each channel. Specifically, the observation probability of each channel is obtained as follows: Based on measurement basis and measurement operator, a projection operation is performed on the complex features in the quantum hidden space representation to map the complex features from the quantum hidden space to the observable space. The real and imaginary parts of the complex feature amplitude represent the coherent components and phase shifts of the quantum hidden space representation, respectively. The mapped complex features are then subjected to a modulo-square operation to obtain the amplitude energy intensity of each channel, which serves as the initial observation probability. The observation probabilities of all channels are then normalized to obtain stable and comparable observation probabilities for each channel. Multiple independent measurements are performed within the same sliding time window, and the observation probabilities of each channel are denoised and weighted to form robust observation results. Based on the calibrated measurement weights and biases, robust observation results are converted into response delay estimates, and a confidence index is calculated, which is a comprehensive result of measurement dispersion, measurement residuals, and distribution uncertainty. The calculation of the confidence index is as follows: The dispersion of response delay estimates obtained from multiple measurements is calculated. Measurement residuals are calculated based on the differences between predicted values ​​and historical calibration values. The probability distribution of measurement results from different channels is statistically analyzed to calculate the distribution uncertainty. The measurement dispersion, measurement residuals, and distribution uncertainty are then weighted and fused. The dispersion weight reflects data stability, the residual weight reflects model accuracy, and the uncertainty weight reflects the impact of environmental interference. Finally, the comprehensive weighted result is normalized to generate a confidence index. The estimated response delay and its corresponding confidence index are checked against the preset threshold range and physical constraints. If the check passes, the estimated response delay and its corresponding confidence index are output. If the check fails, the parameters are updated and remeasured.

[0027] In this embodiment, generating the final delay estimation result includes: Based on the estimated response delay and the corresponding confidence index, combined with the operating status of the edge nodes and the on-site working conditions, the correction constraints are determined. The physical condition correction is calculated based on changes in ambient temperature, power supply voltage, and load. The data correction is calculated based on the confidence index, residuals, and trend terms. The two types of corrections are weighted and combined to form a total correction. The phase offset parameters and measurement layer weights are updated to obtain the corrected response delay estimate and confidence index. Based on the corrected response delay estimate and confidence level, an adaptive strategy is executed sequentially to generate a stable response delay estimate and confidence level as the final delay estimation result. The final delay estimation results, correction parameters, time stamps, and operating condition information are written into the edge computing node database.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-time evaluation scenario of the response performance of electromagnetic clutches in a new energy vehicle assembly line in a certain city. Ten prototype vehicles were selected as test subjects on the assembly line. During engagement and disengagement, multi-source time-series data, including electromagnetic coil current, control terminal voltage, output shaft speed, structural vibration, and ambient temperature, were simultaneously collected. The monitoring nodes were deployed on an edge computing terminal (ARM64 quad-core processor, 2GB memory). Online processing of the data stream was achieved through local sliding time window segmentation and topological constraint fusion calculations. All calculation results were transmitted back to the Manufacturing Execution System (MES) monitoring platform in real time.

[0029] The test period is from April to September 2025, covering 10 typical operating conditions, including steady-state operation at room temperature, high temperature and high load, low temperature start-up, high frequency start-stop, and extreme combined conditions. The sampling time for each condition varies from 20 to 720 minutes. The ambient temperature range is 10℃ to 80℃, and the load variation is within ±20%. During operation, the edge nodes continuously complete data acquisition, topology modeling, and response prediction with a delay of 0.2 seconds, and can output the estimated engagement delay and confidence interval of the electromagnetic clutch in real time.

[0030] The method of this invention maintains stable output under various complex environments. The delay confidence interval is consistently maintained within ±2 milliseconds, indicating high reliability of the prediction results. The average false alarm rate is 3.6%, and the false negative rate is 2.5%. The delay estimate fluctuates little under different operating conditions, demonstrating the algorithm's robustness to noise and nonlinear disturbances.

[0031] Compared to traditional detection methods based on empirical thresholds and offline analysis, the edge computing architecture of this invention achieves a localized closed loop for data processing and feature extraction, avoiding cloud latency and data loss issues, and significantly improving the system's real-time performance and reliability. Application results show that this invention can not only accurately reflect the dynamic response characteristics of electromagnetic clutches under different operating conditions, but also achieve reliable assessment, providing data support for equipment life prediction and intelligent maintenance.

[0032] Table 1 Evaluation Results of Electromagnetic Clutch Response Delay under Different Operating Conditions

[0033] As shown in Table 1, the method of this invention conducted multiple rounds of response delay evaluation experiments on electromagnetic clutches under different operating conditions from April to September 2025. These experiments covered various operating states, including steady-state operation at room temperature, low-temperature start-up, high-temperature and high-load operation, high-frequency start-stop, and extreme combined operating conditions, with a total data sample exceeding 1800 minutes. The experimental results demonstrate that, supported by multi-source time-series data fusion and edge computing architecture, the method of this invention can stably output the estimated response delay and its confidence interval during the engagement and release processes of the electromagnetic clutch, exhibiting high detection accuracy and real-time response capability.

[0034] Overall, under normal temperature and light load fluctuation conditions, the average response latency remained between 84 and 86 milliseconds, with a latency confidence interval of approximately ±2 milliseconds, indicating stable and reliable detection results. Under high temperature, high load, and extreme combined conditions, the response latency increased slightly but remained within 96 milliseconds, demonstrating that the method of this invention can accurately characterize the dynamic response features under different environmental and load conditions. In all test conditions, the average false alarm rate was 3.6%, and the false negative rate was 2.5%, significantly better than traditional threshold-based schemes, demonstrating the robustness of the algorithm in complex disturbance and noise environments. Regarding real-time performance, the average detection time of the edge nodes deployed in this invention after completing data acquisition, feature mapping, and quantum state inference is less than 0.2 seconds, achieving near real-time latency assessment and state judgment.

[0035] The distribution trend of the results shows that as the sampling time and operating environment change, the delay estimation output remains continuous and smooth, with no significant fluctuations in the confidence interval. This indicates that the fusion of Laplace eigenmaps and topological constraint modeling effectively maintains temporal consistency and structural correlation. Especially during long-term steady-state durability operation, the model can still maintain a confidence boundary of ±2 milliseconds and a false alarm rate of less than 4%, demonstrating the stability and adaptability of the method in long-term operation. This fully verifies the high accuracy, high real-time performance, and early warning capability of the method under various operating conditions, providing reliable technical support for the intelligent detection and health management of electromagnetic clutches.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the response delay of an electromagnetic clutch based on edge computing, characterized in that, include: Multi-source time-series data of the electromagnetic clutch during engagement and release are collected, the multi-source time-series data are preprocessed, a fixed-length sliding time window is constructed based on the action trigger time, and a multi-channel sample sequence arranged in chronological order is generated. Within each sliding time window, local similarity is calculated based on the Euclidean distance and linear correlation between different channels in the multi-channel sample sequence, and a time-varying adjacency matrix sequence is established. Based on the structural layout of the electromagnetic clutch and the signal acquisition circuit, a physical topology matrix is ​​constructed. The comprehensive adjacency matrix formed by weighted aggregation of the time-varying adjacency matrix sequence is then fused with the physical topology matrix according to the weights to obtain a topology constraint matrix containing time coupling information and structural association information. On edge computing nodes, Laplacian eigenmaps are performed on the topological constraint matrix to construct the manifold embedding space of the data samples. Low-dimensional feature vectors are extracted and dynamically corrected through an incremental update mechanism to generate updated low-dimensional manifold feature vectors. The updated low-dimensional manifold eigenvectors are mapped to quantum state inputs, and complex feature representations are formed through amplitude and phase encoding. These complex feature representations are then input into an improved quantum heuristic neural network model to generate a quantum state latent space representation. Perform measurement operations on the hidden space representation of the quantum state to generate response delay estimates and corresponding confidence indices; Adaptive correction is performed on the response delay estimate of the electromagnetic clutch and the corresponding confidence index on the edge computing node to generate the final delay estimate result, and the final delay estimate result is synchronously stored in the edge computing node database.

2. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The multi-source timing data specifically includes electromagnetic coil current data, control terminal voltage data, output shaft speed data, mechanical structure vibration data, and ambient temperature data.

3. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The preprocessing of multi-source time-series data specifically includes denoising, time synchronization, anomaly repair, normalization, and smoothing enhancement.

4. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The establishment of the time-varying adjacency relation matrix sequence includes: A multi-channel sample sequence is represented as a data set consisting of multiple channel signals changing over time. Each channel corresponds to a time change sequence of a physical quantity, and the sampling time is a continuous time point. Within each fixed-length sliding time window, calculate the sum of squares of the differences between data points of any two channel signals at the same time, and then take the square root of the sum of squares to obtain the Euclidean distance between the two channel signals within the sliding time window. Compare the trends of change of any two channel signals within the same sliding time window and calculate the degree of linear correlation. The Euclidean distance between the two channels and the degree of linear correlation are weighted and fused according to a preset ratio to determine the local similarity between the channels; Based on the local similarity between each channel, establish the adjacency matrix corresponding to the current sliding time window; Arrange all the adjacency matrices generated within the sliding time window in chronological order to form a time-varying adjacency matrix sequence.

5. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The process of obtaining the topological constraint matrix, which includes temporal coupling information and structural association information, includes: Based on the structural layout of the electromagnetic clutch and the signal acquisition line, the channel set is determined and the node and connection relationship is established. A physical topology matrix is ​​constructed. The physical topology matrix is ​​a square matrix, and the number of rows and columns corresponds to the number of multi-source signal channels. Each element in the matrix represents the physical connection strength between the corresponding channels. The time-varying adjacency matrix sequence is numbered in chronological order, with each sliding time window corresponding to an adjacency matrix, and time weight parameters are set for each sliding time window. Based on the time weight parameter, the time-varying adjacency relation matrix sequence is weighted and superimposed to obtain the comprehensive adjacency matrix; The composite adjacency matrix and the physical topology matrix are weighted and fused according to preset fusion coefficients and then normalized to obtain the topology constraint matrix.

6. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The generation of the updated low-dimensional manifold feature vector includes: Based on the topological constraint matrix, the connection weights of each node are accumulated to calculate the degree matrix. Based on the degree matrix and the topological constraint matrix, the Laplace matrix reflecting the dynamic changes between each channel is calculated. Construct the manifold embedding space of the data samples and establish the corresponding objective function. Introduce temporal smoothing constraints, regularization constraints and structural correlation constraints to keep the interconnected channels as close as possible in the low-dimensional space. Eigenvalue decomposition of the Laplacian matrix is ​​performed in the manifold embedding space to obtain a set of eigenvalues ​​and eigenvectors; from these, the eigenvectors corresponding to the smallest non-zero eigenvalues ​​are selected to form an initial set of low-dimensional manifold eigenvectors. The initial low-dimensional manifold eigenvector set is standardized to obtain the initial low-dimensional feature matrix of the current sliding time window; An incremental update operation is performed on the edge computing node. The low-dimensional feature matrix of the previous sliding time window is weighted and fused with the initial low-dimensional feature matrix of the current sliding time window. The update ratio is determined according to the feature change rate to obtain the updated low-dimensional feature matrix.

7. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The generated quantum state hidden space representation includes: An improved quantum heuristic neural network model is constructed, comprising a quantum state encoding layer, a gated entanglement processing unit, a complex residual coupling layer, and a quantum hidden space aggregation layer, wherein: The quantum state encoding layer is responsible for mapping the low-dimensional manifold eigenvectors into time-sensitive complex quantum states through amplitude normalization and phase modulation. The gated entanglement processing unit enhances the correlation between features by setting the entanglement strength and the gate coefficient to couple and interfere between multi-channel quantum states. The complex residual coupling layer performs residual fusion in the complex domain to maintain stable feature propagation and prevent information decay. The quantum hidden space aggregation layer performs weighted aggregation based on channel attention weights and introduces temporal smoothing and incremental update mechanisms to continuously evolve and dynamically correct the hidden space representation, thus constructing a quantum state hidden space representation that combines temporal consistency and structural correlation. A phase bias adaptive mechanism is introduced into the quantum state coding layer. The dynamic phase bias is calculated based on the action triggering time, the node correlation strength in the topological constraint matrix, and the signal energy change rate. The updated low-dimensional manifold eigenvector is input into the quantum state coding layer, and amplitude normalization and phase modulation processing are performed to generate a coding matrix with time-dependent characteristics. The encoding matrix is ​​input into the gated entanglement processing unit, and quantum entanglement operations between channels are performed according to the entanglement strength and the gate coefficient. Quantum interference paths between channels are selectively activated through a variable gate structure. When there are strong correlation characteristics between specific channels, the entanglement weight is increased to form a coherent superposition state. Finally, the entanglement matrix reflecting the quantum correlation between channels is output. The entanglement matrix is ​​input into the complex residual coupling layer to perform residual fusion operation based on the complex domain; the amplitude and phase components of the main channel and the sub-channel signals are weighted and superimposed respectively through the residual weight adaptive adjustment mechanism; then the amplitude and phase normalization processing is performed to output the residual matrix that maintains stable propagation of features. The residual matrix is ​​input into the quantum hidden space aggregation layer. The importance of each channel feature is calculated based on the channel attention weight, and a weighted aggregation operation is performed to generate an initial hidden space matrix that reflects the quantum distribution structure of multi-channel features. A temporal smoothing and incremental update mechanism is introduced in the quantum hidden space aggregation layer. The initial hidden space matrices of adjacent sliding time windows are weighted and fused according to the feature change rate. When the feature change between the current sliding time window and the previous sliding time window is small, the high weight is inherited. When the change is large, the weight of the current feature is increased. Finally, the updated hidden space matrix is ​​generated. The updated latent space matrix is ​​subjected to mutation detection and smoothing regularization. When an energy mutation or phase aberration of the channel features is detected, the incremental update weights are dynamically adjusted according to the mutation intensity and regularization constraints are applied. After smoothing and regularization, the final smoothed latent space matrix is ​​generated, and the smoothed latent space matrix is ​​used as the quantum state latent space representation of the final output of the improved quantum heuristic neural network model.

8. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The generation of response delay estimates and corresponding confidence indices include: Establish the correspondence between the quantum state hidden space representation and each channel of the multi-channel sample sequence, and define the measurement basis and measurement operator used for measurement; Based on measurement basis and measurement operator, measurement preprocessing is performed on the hidden space representation of quantum state to complete amplitude modulus square and channel normalization, and the observation probability of each channel is obtained. Multiple independent measurements are performed within the same sliding time window, and the observation probabilities of each channel are denoised and weighted to form robust observation results. Based on the calibrated measurement weights and biases, robust observation results are converted into response delay estimates, and a confidence index is calculated that is a combination of measurement dispersion, measurement residuals, and distribution uncertainty. The estimated response delay and its corresponding confidence index are checked against the preset threshold range and physical constraints. If the check passes, the estimated response delay and its corresponding confidence index are output. If the check fails, the parameters are updated and remeasured.

9. The method for evaluating the response delay of an electromagnetic clutch based on edge computing according to claim 1, characterized in that, The generation of the final delay estimation result includes: Based on the estimated response delay and the corresponding confidence index, combined with the operating status of the edge nodes and the on-site working conditions, the correction constraints are determined. The physical condition correction is calculated based on changes in ambient temperature, power supply voltage, and load. The data correction is calculated based on the confidence index, residuals, and trend terms. The two types of corrections are weighted and combined to form a total correction. The phase offset parameters and measurement layer weights are updated to obtain the corrected response delay estimate and confidence index. Based on the corrected response delay estimate and confidence level, an adaptive strategy is executed sequentially to generate a stable response delay estimate and confidence level as the final delay estimation result. The final delay estimation results, correction parameters, time stamps, and operating condition information are written into the edge computing node database.

Images