Embedded device online test and remote maintenance method and system

Abstract

The invention provides an embedded device online test and remote maintenance method and system, and relates to the technical field of device maintenance, and the method comprises the steps: collecting multi-source heterogeneous data, selecting a sampling period based on Shannon entropy, employing wavelet transform and fractal theory to extract features, inputting an anomaly detection model to generate an anomaly degree vector, and carrying out the detection of the anomaly degree vector; extracting features by using a capsule neural network and performing fault type probability distribution prediction; calculating system state indexes by combining a phase-space reconstruction technology; inputting a causal reasoning network to generate a fault diagnosis report; retrieving a maintenance scheme based on the fault diagnosis report; and the maintenance effect is evaluated through the twin neural network, and the maintenance instruction is iteratively optimized according to the effect until the maintenance effect reaches the standard.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment maintenance, and in particular to an online testing and remote maintenance method and system for embedded equipment. Background Art

[0002] With the rapid development of the Internet of Things and edge computing, more and more embedded devices are deployed in various application scenarios, such as industrial control, smart homes, medical devices, etc. Embedded devices usually need to run continuously for a long time. Therefore, real-time online testing and remote maintenance of embedded devices are becoming more and more important.

[0003] Traditional embedded device testing and maintenance methods usually rely on manual regular inspections and offline diagnosis, which are inefficient, unable to detect and handle equipment failures in a timely manner, and are costly. They require a lot of manpower and material resources and are difficult to cope with complex and changing operating environments.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiment of the present invention provides an embedded device online testing and remote maintenance method and system, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides an online testing and remote maintenance method for an embedded device, comprising:

[0007] Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, calculate the information entropy value based on the Shannon entropy principle, select a sampling period based on a preset information entropy threshold, perform wavelet transform decomposition on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence according to the selected sampling period, and obtain a first characteristic matrix. Use fractal theory to calculate the Hausdorff dimension of storage performance data and the fractal dimension of network status data to obtain a second characteristic matrix representing timing complexity characteristics. Input the first characteristic matrix and the second characteristic matrix into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector.

[0008] The abnormality degree vector is input into a capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by a hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than a preset health threshold, a maintenance process is triggered.

[0009] According to the fault diagnosis report, similar fault cases are retrieved in the maintenance knowledge base, and the related maintenance plans are extracted, and the success rate of each maintenance plan in the historical records is calculated. The maintenance plan weight is generated and the maintenance instruction sequence is planned in combination with the Monte Carlo tree search algorithm. The maintenance instruction sequence is executed and the execution status is collected according to the preset monitoring nodes. An effect evaluation model based on the twin neural network is constructed, and the similarity between the current state characteristics of the system and the target state characteristics is calculated to obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, the maintenance instruction sequence is adjusted based on the historical maintenance data through the policy gradient optimization algorithm. The maintenance and adjustment are repeated until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, and the maintenance is terminated and the maintenance process is stored in the maintenance knowledge base.

[0010] In an optional embodiment,

[0011] Edge computing nodes are deployed in embedded devices to collect multi-source heterogeneous data sequences. For each data sequence, the information entropy value is calculated based on the Shannon entropy principle. The sampling period is selected based on a preset information entropy threshold. The processor power consumption data and memory usage data in the multi-source heterogeneous data sequence are decomposed by wavelet transform according to the selected sampling period to obtain a first feature matrix. The Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics. The first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector including:

[0012] Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data corresponding to the embedded devices, and generate multi-source heterogeneous data sequences by combining timestamps, wherein the multi-source heterogeneous data include processor power consumption data, memory usage data, storage performance data, and network status data of the embedded devices;

[0013] For each data sequence in the multi-source heterogeneous data sequence, multiple time windows are divided based on the Shannon entropy principle, the data probability distribution in each time window is calculated, the data probability distribution is substituted into the Shannon entropy formula, and the information entropy value of the time window is calculated. If the information entropy value is greater than a preset information entropy threshold, the sampling period is set to the first sampling period, otherwise, it is set to the second sampling period, wherein the first sampling period is less than the second sampling period, and the set sampling period is fed back to the edge computing node;

[0014] Based on the selected sampling period, performing multi-level wavelet decomposition on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence by using the Dauby wavelet, obtaining corresponding approximate coefficients and detail coefficients at each layer, arranging the approximate coefficients and the detail coefficients in order from low frequency to high frequency, and obtaining the first feature matrix;

[0015] For the storage performance data, a point set consisting of storage data throughput samples within a unit time range is statistically analyzed based on fractal theory, and the space where the point set is located is gridded in combination with the box dimension method, and the number of grids containing the points is counted and the side length of each grid is changed to obtain multiple storage performance data pairs, and linear fitting is performed, and the slope of the obtained fitting straight line is used as the Hausdorff dimension. For the network status data, the correlation integral under different time lags is calculated, and the optimal embedding dimension is selected in combination with the Takkens embedding theorem to reconstruct the phase space to obtain a trajectory matrix, and the fractal dimension is calculated based on the trajectory matrix, and the Hausdorff dimension and the fractal dimension are arranged in chronological order to construct a second characteristic matrix;

[0016] The first feature matrix and the second feature matrix are normalized to generate a normalized feature matrix. The normalized feature matrix is ​​transformed by combining an anomaly detection model constructed based on fifth-order generalized Chebyshev polynomials to obtain Chebyshev coefficients. The Chebyshev coefficients are constructed as an abnormal feature vector. The distance from the abnormal feature vector to a preset normal sample boundary is calculated and normalized to obtain the abnormal degree vector.

[0017] In an optional embodiment,

[0018] Substituting the data probability distribution into the Shannon entropy formula, the information entropy value of the time window is calculated as shown in the following formula:

[0019] ;

[0020] Among them, H(X t ) represents the time window X t The information entropy value, λ trepresents the time-varying forgetting factor, H(X t-1 ) represents the information entropy value of time window t-1, w i (t) represents the weight coefficient of the ith event, p i Represents the time window X t The probability of the i-th event occurring in represents the correction term, n represents the total number of times, β represents the KL divergence regularization term weight factor, and K(p||q) represents the KL divergence between the data distribution p and the prior distribution q.

[0021] In an optional embodiment,

[0022] The abnormality degree vector is input into a capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and the correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by a hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than a preset health threshold, the maintenance process is triggered, including:

[0023] Adding the abnormality degree vector to a pre-trained capsule neural network, wherein the structure of the capsule neural network corresponds to the system architecture of the embedded device, each capsule unit corresponds to a system component in the embedded device, wherein the system components include a processor, memory, storage, and network, and the capsule neural network performs feature extraction on the abnormality degree vector to obtain an abnormal feature vector;

[0024] The capsule neural network learns the spatial relationship and constraint relationship between different components according to the dynamic routing algorithm, adaptively adjusts the routing weights between different capsule layers in combination with the mutual information of the component states corresponding to the embedded device, and modifies the message transmission strength between different components based on the adjusted routing weights;

[0025] Perform Gaussian mixture clustering on the abnormal feature vectors corresponding to the capsule neural network, iteratively estimate the parameters of the preset Gaussian mixture model through the expectation maximization algorithm, calculate the posterior probability that each feature vector in the abnormal feature vector belongs to a different Gaussian distribution, and update the mean, covariance and mixing coefficient of each Gaussian distribution based on the posterior probability to obtain the probability distribution of the fault type;

[0026] According to the probability distribution of the fault type, a nonlinear prediction model is constructed based on the phase space reconstruction technology, the system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, the Lyapunov index and the correlation dimension of the system state trajectory are calculated, the probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a pre-constructed causal reasoning network, the causal dependency relationship is modeled through a directed acyclic graph, the hidden causal structure is learned by iteratively updating the edge weights, a fault diagnosis report is obtained and a state transfer matrix is ​​constructed, and the hidden system state is determined based on the state transfer matrix;

[0027] The hidden system state is used as the hidden state space of the hidden Markov model, and the fault diagnosis report is used as the observed state space of the hidden Markov model. The state transfer matrix and the emission probability matrix predetermined from historical fault data are combined, and the system state evolution path is predicted through a forward-backward algorithm and the health status of the current equipment is calculated to obtain the equipment health score. If the equipment health score is lower than the preset health threshold, the maintenance process is triggered.

[0028] In an optional embodiment,

[0029] A nonlinear prediction model is constructed based on phase space reconstruction technology, a system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, and the Lyapunov index and correlation dimension of the system state trajectory are calculated, including:

[0030] Reconstruct the state sequence of the embedded device into a phase space trajectory, wherein the state sequence includes state values ​​at multiple time points, and reconstruct the state sequence into a phase space vector according to the selected embedding dimension and reconstruction delay, wherein the phase space vector is composed of the state value at the current moment and the state values ​​at multiple previous time points, wherein the embedding dimension is the minimum integer greater than the fractal dimension of the system attractor, and the reconstruction delay is determined by the first zero point of the average mutual information;

[0031] A nonlinear prediction model is constructed based on the phase space trajectory. For each state vector to be predicted, a plurality of nearest state vectors are selected in the neighborhood of the current state vector to be predicted, and the plurality of nearest state vectors are linearly combined to obtain a predicted value of the state vector to be predicted. The coefficients of the linear combination are obtained by minimizing the reconstruction error.

[0032] Gaussian mixture clustering is used to obtain a priori probability distribution of fault types, the priori probability distribution is mapped to the phase space trajectory, a weight coefficient is assigned to a state vector, and weighted processing is performed on the nonlinear prediction model according to the weight coefficient;

[0033] Based on the phase space trajectory after weighted processing, the Lyapunov exponent spectrum of the system evolution is calculated, and the largest Lyapunov exponent is selected from the Lyapunov exponent spectrum for output;

[0034] The correlation dimension is calculated by using a correlation dimension algorithm, and the correlation dimension is obtained by counting the number of state pairs whose distance in the phase space is less than a given scale and calculating the slope of the double logarithmic curve.

[0035] In an optional embodiment,

[0036] Retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan maintenance instruction sequences in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, adjust the maintenance instruction sequence based on the historical maintenance data through the policy gradient optimization algorithm, repeat maintenance and adjustment until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, end maintenance and store the maintenance process in the maintenance knowledge base, including:

[0037] Inputting the fault diagnosis report of the embedded device into a pre-built maintenance knowledge base, the maintenance knowledge base stores a plurality of historical fault cases, each of which includes fault diagnosis information, maintenance plan and effect feedback information, the fault diagnosis information includes CPU occupancy data, memory usage data, and system resource status data, and calculating the semantic similarity between the fault diagnosis report and each historical fault case based on the fault feature vector in the fault diagnosis report, and retrieving the historical fault cases with the high similarity as candidate historical fault cases;

[0038] Calculate the success rate of each maintenance solution in the candidate historical fault cases, where the success rate is obtained by dividing the number of successful repairs of the maintenance solution in the maintenance knowledge base by the total number of executions, and set a corresponding weight coefficient according to the success rate of each maintenance solution, where the weight coefficient increases as the success rate increases;

[0039] A maintenance decision tree is constructed based on a Monte Carlo tree search algorithm, wherein the nodes of the maintenance decision tree include a process restart operation node, a memory cleanup operation node, a system update operation node, and a data backup operation node. Based on the weight coefficient, the Monte Carlo tree search algorithm is used to perform multiple rounds of iterative sampling on the maintenance decision tree. In each round of iteration, the number of visits and the cumulative reward value of each node are recorded, and a maintenance instruction sequence is generated according to the number of visits and the cumulative reward value.

[0040] A monitoring probe is set on the embedded device to collect execution status data during the execution of the maintenance instruction sequence, wherein the execution status data includes CPU occupancy data, memory usage data, storage capacity data, network traffic data and system log data, and the execution status data is input into a twin neural network, wherein the twin neural network includes two parameter-sharing feature extraction subnetworks, and the feature extraction subnetworks are used to process the pre-maintenance status data and the post-maintenance status data respectively to obtain a pre-maintenance feature vector and a post-maintenance feature vector, and the cosine similarity between the post-maintenance feature vector and the preset target feature vector is calculated to obtain a maintenance effect score;

[0041] When the maintenance effect score is less than a preset threshold, the maintenance instruction sequence is adjusted using a policy gradient optimization algorithm, the selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, the reward signal value is calculated based on the difference in the maintenance effect scores of two adjacent maintenance operations, the parameters of the probability distribution function are updated using the reward signal value, an optimized maintenance sequence is generated, and the optimized maintenance sequence is repeatedly executed until the maintenance effect scores within a preset number of consecutive cycles are all greater than the preset threshold;

[0042] The fault diagnosis report, the maintenance instruction sequence, the optimized maintenance sequence and the maintenance effect score are combined into a new historical fault case and stored in the maintenance knowledge base. All historical fault cases in the maintenance knowledge base are clustered and sorted to construct an equipment operation and maintenance knowledge graph.

[0043] In an optional embodiment,

[0044] The selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, and the reward signal value is calculated based on the difference in maintenance effect scores between two adjacent maintenance operations. The reward signal value is used to update the parameters of the probability distribution function as shown in the following formula:

[0045] ;

[0046] Among them, σ' represents the updated parameter, σ represents the current parameter, α represents the learning rate, F represents the Fisher information matrix, and F -1 represents the inverse of the Fisher information matrix, E πRepresents the strategy distribution Expectations, γ r represents the discount factor at time step r, δ r represents the temporal difference error at time step r, which is solved by the immediate reward signal and the baseline. Represents the policy function The logarithmic gradient, s r represents the system state at time step r, a r represents the action taken at time step r and system state s.

[0047] A second aspect of an embodiment of the present invention provides an embedded device online testing and remote maintenance system, comprising:

[0048] The first unit is used to deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, an information entropy value is calculated based on the Shannon entropy principle, a sampling period is selected based on a preset information entropy threshold, and a wavelet transform is performed on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence according to the selected sampling period to obtain a first feature matrix, and the Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics, and the first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector;

[0049] The second unit is used to input the abnormality degree vector into the capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and the features output by the capsule network are subjected to Gaussian mixture clustering to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology, and the Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into the causal reasoning network, a fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by the hidden Markov model, and the equipment health score is calculated. If the equipment health score is lower than a preset health threshold, a maintenance process is triggered;

[0050] The third unit is used to retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan the maintenance instruction sequence in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, the maintenance instruction sequence is adjusted based on the historical maintenance data through the policy gradient optimization algorithm, and the maintenance and adjustment are repeated until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, and the maintenance is terminated and the maintenance process is stored in the maintenance knowledge base.

[0051] According to a third aspect of the embodiments of the present invention,

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the embodiments of the present invention is:

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] In the present invention, a variety of advanced technologies are combined to extract more comprehensive features from multi-source heterogeneous data, more accurately identify fault types, and predict the system state evolution path, thereby improving the accuracy of fault diagnosis. Based on the fault diagnosis results, similar cases are automatically retrieved from the maintenance knowledge base and maintenance plans are generated. The Monte Carlo tree search algorithm is combined to plan the maintenance instruction sequence to achieve automated remote maintenance without human intervention, reducing maintenance costs and time. The effect evaluation model based on the twin neural network and the policy gradient optimization algorithm are adopted, which can adjust the maintenance instruction sequence in real time according to the maintenance effect, and improve the maintenance effect through continuous learning and optimization. The maintenance process is stored in the maintenance knowledge base, forming a virtuous circle and continuously improving the reliability and stability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a flow chart of an online testing and remote maintenance method for an embedded device according to an embodiment of the present invention;

[0060] Figure 2 It is a structural diagram of an online testing and remote maintenance system for embedded devices according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0063] Figure 1 FIG. 1 is a flow chart of an online testing and remote maintenance method for an embedded device according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0064] S1. Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, calculate the information entropy value based on the Shannon entropy principle, select a sampling period based on a preset information entropy threshold, perform wavelet transform decomposition on the processor power consumption data and memory usage data in the multi-source heterogeneous data sequence according to the selected sampling period, obtain a first feature matrix, calculate the Hausdorff dimension of the storage performance data and the fractal dimension of the network status data using fractal theory, obtain a second feature matrix representing the timing complexity characteristics, input the first feature matrix and the second feature matrix into an anomaly detection model constructed based on generalized Chebyshev polynomials, and generate an anomaly degree vector;

[0065] The edge computing node refers to a computing resource node deployed at the edge of the network, which can process data and provide computing services, reduce data transmission delays, and improve real-time processing capabilities. The Shannon entropy principle is a concept in information theory, which is used to measure the degree of disorder or uncertainty of information. The larger the entropy, the more complex and difficult to predict the information. The information entropy threshold refers to the limit set according to the calculation of information entropy. When the information entropy is less than the threshold, the complexity of the data is considered to be low and may be ignored or considered useless. The wavelet transform decomposition is a mathematical method that decomposes a signal into multiple frequency components. It can be converted between the time domain and the frequency domain, which helps to analyze the local characteristics of the signal. The fractal theory is a mathematical theory that studies geometric shapes with self-similar properties. These shapes have similar structural characteristics at different scales. The Hausdorff dimension is used to describe the dimension of fractals with complex geometric structures. It measures the complexity of fractals and is usually used to represent the details of fractal graphics at different scales. The fractal dimension is an important parameter in fractal theory and is used to quantitatively describe the complexity of fractal graphics. The fractal dimension is usually a non-integer value, reflecting the self-similarity of fractal graphics. The generalized Chebyshev polynomials are an extended form of Chebyshev polynomials, which are widely used in approximation and optimization problems and can adapt to more complex constraints and requirements.

[0066] In an optional embodiment,

[0067] Edge computing nodes are deployed in embedded devices to collect multi-source heterogeneous data sequences. For each data sequence, the information entropy value is calculated based on the Shannon entropy principle. The sampling period is selected based on a preset information entropy threshold. The processor power consumption data and memory usage data in the multi-source heterogeneous data sequence are decomposed by wavelet transform according to the selected sampling period to obtain a first feature matrix. The Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics. The first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector including:

[0068] Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data corresponding to the embedded devices, and generate multi-source heterogeneous data sequences by combining timestamps, wherein the multi-source heterogeneous data include processor power consumption data, memory usage data, storage performance data, and network status data of the embedded devices;

[0069] For each data sequence in the multi-source heterogeneous data sequence, multiple time windows are divided based on the Shannon entropy principle, the data probability distribution in each time window is calculated, the data probability distribution is substituted into the Shannon entropy formula, and the information entropy value of the time window is calculated. If the information entropy value is greater than a preset information entropy threshold, the sampling period is set to the first sampling period, otherwise, it is set to the second sampling period, wherein the first sampling period is less than the second sampling period, and the set sampling period is fed back to the edge computing node;

[0070] Based on the selected sampling period, performing multi-level wavelet decomposition on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence by using the Dauby wavelet, obtaining corresponding approximate coefficients and detail coefficients at each layer, arranging the approximate coefficients and the detail coefficients in order from low frequency to high frequency, and obtaining the first feature matrix;

[0071] For the storage performance data, a point set consisting of storage data throughput samples within a unit time range is statistically analyzed based on fractal theory, and the space where the point set is located is gridded in combination with the box dimension method, and the number of grids containing the points is counted and the side length of each grid is changed to obtain multiple storage performance data pairs, and linear fitting is performed, and the slope of the obtained fitting straight line is used as the Hausdorff dimension. For the network status data, the correlation integral under different time lags is calculated, and the optimal embedding dimension is selected in combination with the Takkens embedding theorem to reconstruct the phase space to obtain a trajectory matrix, and the fractal dimension is calculated based on the trajectory matrix, and the Hausdorff dimension and the fractal dimension are arranged in chronological order to construct a second characteristic matrix;

[0072] The first feature matrix and the second feature matrix are normalized to generate a normalized feature matrix. The normalized feature matrix is ​​transformed by combining an anomaly detection model constructed based on fifth-order generalized Chebyshev polynomials to obtain Chebyshev coefficients. The Chebyshev coefficients are constructed as an abnormal feature vector. The distance from the abnormal feature vector to a preset normal sample boundary is calculated and normalized to obtain the abnormal degree vector.

[0073] The multi-level wavelet decomposition refers to decomposing a signal into components of multiple frequency bands through a multi-level decomposition method, and each layer provides information at a different level, which is helpful for signal denoising and feature extraction. The storage data throughput sample refers to the throughput data sample obtained by collecting the rate and time of data transmission in the storage system, which is used to analyze and optimize the performance of the storage system. The fitting straight line is an optimal straight line found by methods such as the least squares method, so that the error between the data point and the straight line is minimized. It is often used in regression analysis and data modeling. The Takens embedding theorem is a theorem in time series data analysis, which shows that by properly selecting delays and dimensions, the time series of nonlinear systems can be converted into multidimensional phase space, thereby facilitating the analysis of the dynamic behavior of the system. The phase space reconstruction is a technology that maps time series data to a high-dimensional space, which helps to reveal the laws and behaviors of dynamic systems and is often used to analyze complex nonlinear systems. The Chebyshev coefficient is the coefficient representing each term in the Chebyshev polynomial, which is used to describe the error characteristics of the polynomial when approximating certain functions, and is widely used in numerical analysis and approximation theory.

[0074] Edge computing nodes are deployed in embedded devices to collect multi-source heterogeneous data of the devices, including processor power consumption, memory usage, storage performance, and network status data. The collection process is combined with timestamps to form multiple time series, each of which corresponds to a data type.

[0075] Information entropy is calculated for each data sequence, and the data sequence is divided into multiple time windows. The distribution of statistical data values ​​in each window is calculated, and the probability distribution of the data in the window is calculated. The information entropy value of the time window is calculated based on the probability distribution. An information entropy threshold is set in advance. If the calculated information entropy value is greater than the threshold, the sampling period of the data sequence is set to a shorter first sampling period, such as 1 millisecond; otherwise, it is set to a longer second sampling period, such as 10 milliseconds. A shorter sampling period is used to capture data with faster information changes, and a longer sampling period is used for data with slower information changes. The set sampling period is fed back to the edge computing node to guide subsequent data collection. For example, the probability distribution of processor power consumption data in a certain time window is {0.1, 0.2, 0.3, 0.4}, the calculated information entropy value is 1.846, and the preset information entropy threshold is 1.5, then the sampling period of the processor power consumption data sequence is set to 1 millisecond.

[0076] Extract features from the data according to the selected sampling period. Perform wavelet transform decomposition on the processor power consumption data and memory usage data. Use, for example, Daubey wavelet for multi-level decomposition, and obtain approximate coefficients and detail coefficients at each level of decomposition. Arrange the approximate coefficients and detail coefficients of all levels in order from low frequency to high frequency to form the first feature matrix. For example, perform three-level wavelet decomposition on a certain section of processor power consumption data to obtain the approximate coefficient A3 and detail coefficients D3, D2, and D1, then the arrangement order of the first feature matrix is ​​A3, D3, D2, and D1.

[0077] The fractal dimension of storage performance data and network status data is calculated. For storage performance data, the point set consisting of the storage data throughput samples per unit time is counted. The space where the point set is located is gridded using the box dimension method, and the number of grids containing the points is counted. The side length of the grid is continuously changed to obtain multiple sets of storage performance data pairs. These data pairs are linearly fitted, and the slope of the fitted line is the Hausdorff dimension. For example, the storage data throughput sample point set obtained by statistics is {(1, 2), (2, 4), (3, 6)}, and the slope obtained after linear fitting is 2, then the Hausdorff dimension is 2. For network status data, the correlation integrals under different time delays are calculated. The optimal embedding dimension is selected according to the Takkens embedding theorem, and the phase space is reconstructed to obtain the trajectory matrix. The fractal dimension is calculated based on the trajectory matrix. The calculated Hausdorff dimension and fractal dimension are arranged in chronological order to form the second characteristic matrix.

[0078] The first feature matrix and the second feature matrix are normalized, for example, the value of each element is scaled to between 0 and 1 to generate a normalized feature matrix. Using a pre-trained anomaly detection model based on fifth-order generalized Chebyshev polynomials, the normalized feature matrix is ​​transformed to obtain Chebyshev coefficients. The Chebyshev coefficients are constructed into an abnormal feature vector. The distance from the abnormal feature vector to the pre-set normal sample boundary is calculated, and normalized to obtain an abnormal degree vector. For example, the calculated abnormal feature vector is (0.2, 0.3, 0.5), the distance to the normal sample boundary is 0.1, and the abnormal degree vector obtained after normalization is 0.1.

[0079] In this embodiment, by combining multi-source heterogeneous data and time series characteristics, the operating status of embedded devices can be more comprehensively characterized, thereby improving the accuracy of anomaly detection and reducing false alarms and missed alarms. The sampling period is dynamically adjusted based on information entropy. While ensuring detection accuracy, the overhead of data collection and processing can be reduced, the battery life of embedded devices can be extended, and data processing can be performed on edge computing nodes, which reduces data transmission delays and bandwidth consumption. It is suitable for resource-constrained embedded devices.

[0080] In an optional embodiment,

[0081] Substituting the data probability distribution into the Shannon entropy formula, the information entropy value of the time window is calculated as shown in the following formula:

[0082] ;

[0083] Among them, H(X t ) represents the time window X t The information entropy value, λ t represents the time-varying forgetting factor, H(X t-1 ) represents the information entropy value of time window t-1, w i (t) represents the weight coefficient of the ith event, p i Represents the time window X t The probability of the i-th event occurring in represents the correction term, n represents the total number of times, β represents the KL divergence regularization term weight factor, and K(p||q) represents the KL divergence between the data distribution p and the prior distribution q.

[0084] In this embodiment, by introducing a time-varying forgetting factor, the model can better capture the dynamic changes of data distribution, thereby more accurately identifying abnormal data. The introduction of the KL divergence regularization term can effectively avoid the information entropy calculation deviation caused by data sparsity, thereby enhancing the robustness of the model. The information entropy value is calculated in a recursive manner, avoiding repeated calculations, thereby reducing the computational complexity.

[0085] S2. Input the abnormality degree vector into the capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into the causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by the hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than the preset health threshold, the maintenance process is triggered.

[0086] The capsule neural network is a deep learning model for processing image and video data. It encapsulates local features through capsule units and maintains spatial and posture information through a routing mechanism, thereby improving the robustness of the model to spatial transformation and rotation. The capsule unit is a basic computing unit in the capsule neural network, usually containing a small neural network, which is used to detect features and represent them in the form of vectors, while saving information such as the posture and direction of the features. The mutual information is a concept in information theory, which is used to measure the degree of dependence between two variables. The greater the mutual information, the closer the relationship between the two variables. The routing weight is the weight of information transmitted between capsule units through a dynamic routing process in the capsule network, which is used to determine the connection strength between different capsules and affect the effect of information transmission. The Gaussian mixture clustering is a clustering method based on a probability model, which regards data as consisting of multiple Gaussian distributions and realizes data clustering by estimating the parameters of the Gaussian distribution. The Lyapunov index is used to measure the sensitivity of the trajectory in a dynamic system and indicates the degree of dependence of the system on the initial conditions. The larger the Lyapunov index is, the more sensitive the system is to the initial conditions and the more unstable the system is. The correlation dimension is an indicator to measure the dimensional characteristics of a complex system or fractal graph. It is usually used to describe the chaos and nonlinear characteristics of the system and reflect the self-similarity of the system at different scales. The causal inference network is a model used to infer the causal relationship between variables. It is usually constructed by constructing a graphical model to represent and analyze the causal dependency between variables for prediction and inference. The system state evolution path describes the state change trajectory of a dynamic system at different time points, showing the time evolution process of the system and the law of state transition.

[0087] In an optional embodiment,

[0088] The abnormality degree vector is input into a capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and the correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by a hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than a preset health threshold, the maintenance process is triggered, including:

[0089] Adding the abnormality degree vector to a pre-trained capsule neural network, wherein the structure of the capsule neural network corresponds to the system architecture of the embedded device, each capsule unit corresponds to a system component in the embedded device, wherein the system components include a processor, memory, storage, and network, and the capsule neural network performs feature extraction on the abnormality degree vector to obtain an abnormal feature vector;

[0090] The capsule neural network learns the spatial relationship and constraint relationship between different components according to the dynamic routing algorithm, adaptively adjusts the routing weights between different capsule layers in combination with the mutual information of the component states corresponding to the embedded device, and modifies the message transmission strength between different components based on the adjusted routing weights;

[0091] Perform Gaussian mixture clustering on the abnormal feature vectors corresponding to the capsule neural network, iteratively estimate the parameters of the preset Gaussian mixture model through the expectation maximization algorithm, calculate the posterior probability that each feature vector in the abnormal feature vector belongs to a different Gaussian distribution, and update the mean, covariance and mixing coefficient of each Gaussian distribution based on the posterior probability to obtain the probability distribution of the fault type;

[0092] According to the probability distribution of the fault type, a nonlinear prediction model is constructed based on the phase space reconstruction technology, the system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, the Lyapunov index and the correlation dimension of the system state trajectory are calculated, the probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a pre-constructed causal reasoning network, the causal dependency relationship is modeled through a directed acyclic graph, the hidden causal structure is learned by iteratively updating the edge weights, a fault diagnosis report is obtained and a state transfer matrix is ​​constructed, and the hidden system state is determined based on the state transfer matrix;

[0093] The hidden system state is used as the hidden state space of the hidden Markov model, and the fault diagnosis report is used as the observed state space of the hidden Markov model. The state transfer matrix and the emission probability matrix predetermined from historical fault data are combined, and the system state evolution path is predicted through a forward-backward algorithm and the health status of the current equipment is calculated to obtain the equipment health score. If the equipment health score is lower than the preset health threshold, the maintenance process is triggered.

[0094] The system components are the basic modules or parts that constitute a system, usually with specific functions and interactions. The combination and collaboration of system components determine the performance and behavior of the entire system. The dynamic routing algorithm is an algorithm in the capsule network, which is used to dynamically adjust the connection strength between capsules according to the input information, so that information can be efficiently and accurately transmitted from low levels to high levels. The mixing coefficient refers to the weight coefficient of each component or distribution in a mixed model (such as a Gaussian mixture model), which determines the importance or contribution of different components in the data generation process. The hidden causal structure refers to the potential that cannot be directly observed by observing variables in the causal inference model. Causal relationships or factors, these hidden causal relationships have an important impact on the behavior of the system. The observation state space refers to the set of all system states that can be obtained through observation in a dynamic system, representing all possible observation states of the system at a given time. The emission probability matrix is ​​an important parameter in the hidden Markov model (HMM), which represents the probability of generating a specific observation under a certain state, and is used to describe the relationship between hidden states and observations. The forward-backward algorithm is a dynamic programming algorithm in the hidden Markov model, which is used to calculate the probability of an observation sequence and perform parameter estimation of the model, and calculate the state probability at each moment through forward and backward steps.

[0095] Collect the operating data of embedded devices and extract indicators related to the health status of the devices, such as processor temperature, memory usage, storage space occupancy, and network latency, quantify and standardize them to form an abnormality degree vector. For example, assuming that the collected processor temperature is 80°C, the memory usage is 90%, the storage space occupancy is 95%, and the network latency is 100ms, after standardization, the abnormality degree vector [0.8, 0.9, 0.95, 1.0] is formed.

[0096] The abnormality degree vector is input into the pre-trained capsule neural network. The structure of the capsule neural network corresponds to the system architecture of the embedded device. Each capsule unit corresponds to a system component in the device, such as processor, memory, storage, and network. The capsule neural network extracts features from the input abnormality degree vector to obtain an abnormal feature vector. For example, suppose that after feature extraction by the capsule neural network, the abnormal feature vector [0.7, 0.8, 0.9, 0.6] is obtained.

[0097] The capsule neural network uses a dynamic routing algorithm to learn the spatial and constraint relationships between different components. At the same time, the routing weights between different capsule layers are adaptively adjusted according to the mutual information of the component states corresponding to the embedded device. For example, if the mutual information of the processor temperature and memory usage is high, the routing weights between the corresponding capsule layers will also be high, thereby enhancing the strength of message transmission between the processor and memory. Assume that the routing weight between the processor and memory is 0.8, and the routing weight between storage and network is 0.5.

[0098] Gaussian mixture clustering is performed on the abnormal feature vectors output by the capsule neural network. The expectation maximization algorithm is used to iteratively estimate the parameters of the Gaussian mixture model and calculate the posterior probability that each feature vector belongs to a different Gaussian distribution. The mean, covariance, and mixing coefficient of each Gaussian distribution are updated based on the posterior probability to obtain the probability distribution of the fault type. For example, suppose that the probability distribution of two fault types is finally obtained, with the probability of fault type A being 0.7 and the probability of fault type B being 0.3.

[0099] According to the obtained probability distribution of fault types, a nonlinear prediction model is constructed using phase space reconstruction technology. The system state sequence constructed according to the probability distribution of fault types is input into the nonlinear prediction model to calculate the Lyapunov index and correlation dimension of the system state trajectory. For example, assume that the calculated Lyapunov index is 0.5 and the correlation dimension is 2.5.

[0100] The probability distribution of the fault type, Lyapunov index, and correlation dimension are input into the pre-built causal reasoning network. The causal reasoning network models the causal dependency through a directed acyclic graph, and learns the hidden causal structure by iteratively updating the edge weights to obtain the fault diagnosis report and construct the state transition matrix. For example, the fault diagnosis report indicates that the probability of fault type A is high and is related to processor overheating. The state transition matrix describes the probability of the system transitioning from one state to another.

[0101] The hidden system state is used as the hidden state space of the hidden Markov model, the fault diagnosis report is used as the observed state space, and the forward-backward algorithm is used to calculate the health score of the current device by combining the state transition matrix and the emission probability matrix predetermined from the historical fault data. For example, assume that the calculated device health score is 0.6.

[0102] If the device health score is lower than a preset health threshold, such as 0.7, a maintenance process is triggered. The maintenance process may include sending alarm notifications, generating maintenance recommendations, and automatically executing maintenance operations.

[0103] In this embodiment, capsule neural networks, phase space reconstruction, and causal reasoning technologies are combined to more accurately identify fault types and root causes. By predicting the health status of equipment through a hidden Markov model, potential faults can be discovered in advance, unexpected equipment downtime can be avoided, and maintenance costs can be reduced. By timely triggering the maintenance process, faults can be effectively prevented from occurring, thereby improving the reliability and stability of the system.

[0104] In an optional embodiment,

[0105] A nonlinear prediction model is constructed based on phase space reconstruction technology, a system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, and the Lyapunov index and correlation dimension of the system state trajectory are calculated, including:

[0106] Reconstruct the state sequence of the embedded device into a phase space trajectory, wherein the state sequence includes state values ​​at multiple time points, and reconstruct the state sequence into a phase space vector according to the selected embedding dimension and reconstruction delay, wherein the phase space vector is composed of the state value at the current moment and the state values ​​at multiple previous time points, wherein the embedding dimension is the minimum integer greater than the fractal dimension of the system attractor, and the reconstruction delay is determined by the first zero point of the average mutual information;

[0107] A nonlinear prediction model is constructed based on the phase space trajectory. For each state vector to be predicted, a plurality of nearest state vectors are selected in the neighborhood of the current state vector to be predicted, and the plurality of nearest state vectors are linearly combined to obtain a predicted value of the state vector to be predicted. The coefficients of the linear combination are obtained by minimizing the reconstruction error.

[0108] Gaussian mixture clustering is used to obtain a priori probability distribution of fault types, the priori probability distribution is mapped to the phase space trajectory, a weight coefficient is assigned to a state vector, and weighted processing is performed on the nonlinear prediction model according to the weight coefficient;

[0109] Based on the phase space trajectory after weighted processing, the Lyapunov exponent spectrum of the system evolution is calculated, and the largest Lyapunov exponent is selected from the Lyapunov exponent spectrum for output;

[0110] The correlation dimension is calculated by using a correlation dimension algorithm, and the correlation dimension is obtained by counting the number of state pairs whose distance in the phase space is less than a given scale and calculating the slope of the double logarithmic curve.

[0111] The phase space trajectory is a trajectory obtained by reconstructing the phase space, which represents the evolution process of the system in the multidimensional state space, and is usually used to study the long-term behavior and structural characteristics of dynamic systems. The fractal dimension of the system attractor is a parameter used to describe the complexity of the shape of the attractor of the dynamic system. An attractor with a higher fractal dimension represents a system with higher complexity and chaos. The first zero point refers to the point in a mathematical function where the function value reaches zero for the first time, and is usually used to analyze and find the critical point or important turning point of the system. The double logarithmic curve refers to a curve drawn when the coordinate axis adopts a logarithmic scale, which is often used to display phenomena with exponential growth or power-law distribution, and can effectively display changes of different magnitudes.

[0112] Collect multiple state variables of the equipment, such as temperature, pressure, vibration, etc., to form a time series, and map the time series to the phase space. The mapping process requires determining two key parameters: embedding dimension and reconstruction delay. The embedding dimension refers to the number of state variables that constitute the phase space vector, which should be greater than the minimum integer of the system attractor fractal dimension to ensure that the phase space can fully reflect the dynamic characteristics of the system. The reconstruction delay refers to the time interval between adjacent state variables, which can be determined by the first zero point of the average mutual information to ensure that there is sufficient independence between adjacent state variables. For example, suppose that the time series data of the device temperature is collected once per second, and a total of 1000 data points are collected. By calculating the average mutual information, it is found that the first zero point appears at a delay of 5 seconds. If it is judged empirically that the system attractor fractal dimension is less than 3, the embedding dimension can be selected as 4. The original one-dimensional time series is converted into a four-dimensional phase space trajectory, where each phase space vector consists of the temperature value at the current moment and the temperature values ​​of the previous three moments (delayed by 5 seconds, 10 seconds and 15 seconds respectively).

[0113] Based on the constructed phase space trajectory, a nonlinear prediction model is established. For each future state vector that needs to be predicted, several historical state vectors closest to it are found in the phase space. The linear combination of historical state vectors is used to predict the value of the future state vector. The coefficients of the linear combination are determined by minimizing the reconstruction error to ensure the accuracy of the prediction results. For example, to predict the temperature at a certain moment in the future, find the five closest historical state vectors in the phase space. By adjusting the linear combination coefficients of these five vectors so that the difference between the combination result and the actual temperature value is minimized, the predicted value can be obtained.

[0114] In order to improve the accuracy of prediction, the prediction model is weighted in combination with the prior probability distribution of the fault type. The historical data is analyzed using methods such as Gaussian mixture clustering to obtain the prior probability distribution of different fault types, map the probability distribution to the phase space trajectory, and assign a weight coefficient to each state vector. The size of the weight coefficient reflects the probability of the fault type to which the state vector belongs. Finally, the nonlinear prediction model is weighted according to the weight coefficient, so that the model pays more attention to those state vectors related to high-probability fault types. For example, suppose that through Gaussian mixture clustering analysis, it is found that there are two fault types of the equipment, A and B, whose prior probabilities are 0.8 and 0.2 respectively. If a state vector in the phase space is judged to belong to fault type A, it is assigned a larger weight coefficient, such as 0.8; if it is judged to belong to fault type B, it is assigned a smaller weight coefficient, such as 0.2.

[0115] Based on the weighted phase space trajectory, the Lyapunov index spectrum and correlation dimension of the system are calculated. The Lyapunov index reflects the convergence or divergence speed of the system state trajectory. The maximum positive Lyapunov index indicates that the system is in a chaotic state. The larger the value, the higher the instability of the system. The correlation dimension reflects the fractal dimension of the system attractor and is used to characterize the complexity of the system. For example, if the maximum Lyapunov index is 0.5 and the correlation dimension is 2.5, it means that the system is in a chaotic state and has a certain degree of complexity.

[0116] In this embodiment, by weighting the prediction model in combination with the prior probability distribution of the fault type, the prediction accuracy can be effectively improved, thereby more accurately predicting the future state of the equipment. Through the phase space reconstruction technology, the nonlinear characteristics in the equipment state time series can be effectively extracted, making the prediction model more robust to noise and interference. Not only can the future state of the equipment be predicted, but also the health status of the equipment can be evaluated through chaotic characteristic quantities such as the Lyapunov exponent and correlation dimension, providing more comprehensive fault information, thereby better guiding the maintenance and maintenance of the equipment.

[0117] S3. Retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan the maintenance instruction sequence in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, adjust the maintenance instruction sequence based on the historical maintenance data through the policy gradient optimization algorithm, repeat maintenance and adjustment until the maintenance effect score is greater than the maintenance effect threshold for 20 consecutive cycles, end the maintenance and store the maintenance process in the maintenance knowledge base.

[0118] The Monte Carlo tree search algorithm is a decision tree search method based on random sampling, which is used to solve large-scale problems with decision space, and is particularly widely used in games and reinforcement learning. The twin neural network is a special neural network structure, which usually includes two identical sub-networks that share the same weights and parameters. The maintenance of the knowledge base refers to the system's management and update process of information in the knowledge base, which usually involves adding new knowledge, deleting outdated information, correcting erroneous data, and integrating knowledge from multiple sources.

[0119] In an optional embodiment,

[0120] Retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan maintenance instruction sequences in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, adjust the maintenance instruction sequence based on the historical maintenance data through the policy gradient optimization algorithm, repeat maintenance and adjustment until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, end maintenance and store the maintenance process in the maintenance knowledge base, including:

[0121] Inputting the fault diagnosis report of the embedded device into a pre-built maintenance knowledge base, the maintenance knowledge base stores a plurality of historical fault cases, each of which includes fault diagnosis information, maintenance plan and effect feedback information, the fault diagnosis information includes CPU occupancy data, memory usage data, and system resource status data, and calculating the semantic similarity between the fault diagnosis report and each historical fault case based on the fault feature vector in the fault diagnosis report, and retrieving the historical fault cases with the high similarity as candidate historical fault cases;

[0122] Calculate the success rate of each maintenance solution in the candidate historical fault cases, where the success rate is obtained by dividing the number of successful repairs of the maintenance solution in the maintenance knowledge base by the total number of executions, and set a corresponding weight coefficient according to the success rate of each maintenance solution, where the weight coefficient increases as the success rate increases;

[0123] A maintenance decision tree is constructed based on a Monte Carlo tree search algorithm, wherein the nodes of the maintenance decision tree include a process restart operation node, a memory cleanup operation node, a system update operation node, and a data backup operation node. Based on the weight coefficient, the Monte Carlo tree search algorithm is used to perform multiple rounds of iterative sampling on the maintenance decision tree. In each round of iteration, the number of visits and the cumulative reward value of each node are recorded, and a maintenance instruction sequence is generated according to the number of visits and the cumulative reward value.

[0124] A monitoring probe is set on the embedded device to collect execution status data during the execution of the maintenance instruction sequence, wherein the execution status data includes CPU occupancy data, memory usage data, storage capacity data, network traffic data and system log data, and the execution status data is input into a twin neural network, wherein the twin neural network includes two parameter-sharing feature extraction subnetworks, and the feature extraction subnetworks are used to process the pre-maintenance status data and the post-maintenance status data respectively to obtain a pre-maintenance feature vector and a post-maintenance feature vector, and the cosine similarity between the post-maintenance feature vector and the preset target feature vector is calculated to obtain a maintenance effect score;

[0125] When the maintenance effect score is less than a preset threshold, the maintenance instruction sequence is adjusted using a policy gradient optimization algorithm, the selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, the reward signal value is calculated based on the difference in the maintenance effect scores of two adjacent maintenance operations, the parameters of the probability distribution function are updated using the reward signal value, an optimized maintenance sequence is generated, and the optimized maintenance sequence is repeatedly executed until the maintenance effect scores within a preset number of consecutive cycles are all greater than the preset threshold;

[0126] The fault diagnosis report, the maintenance instruction sequence, the optimized maintenance sequence and the maintenance effect score are combined into a new historical fault case and stored in the maintenance knowledge base. All historical fault cases in the maintenance knowledge base are clustered and sorted to construct an equipment operation and maintenance knowledge graph.

[0127] Build a maintenance knowledge base that stores a large number of historical fault cases. Each case contains fault diagnosis information (such as CPU usage, memory usage, system resource status, etc.), corresponding maintenance plan, and feedback information after maintenance. In order to facilitate retrieval, the information is stored in a structured manner, such as using a database or knowledge graph.

[0128] When an embedded device fails and generates a fault diagnosis report, it is input into the system. The system extracts fault features in the report, such as high CPU usage, memory leak, etc. The system calculates the similarity between the fault features and the fault features of each historical case in the maintenance knowledge base, such as calculating text similarity or vector cosine similarity. Several historical cases with the highest similarity are selected as candidate cases.

[0129] The system analyzes the historical success rate of each maintenance solution in the candidate case. The success rate is calculated by counting the number of times the solution successfully solved similar faults in the past and dividing it by the total number of executions of the solution. For example, if a solution successfully solved 8 faults in the past 10 executions, its success rate is 80%. The system assigns a corresponding weight to each solution based on its success rate. The higher the success rate, the greater the weight. For example, the success rate can be used directly as the weight, or some nonlinear function can be used to map the success rate to the weight.

[0130] The system builds a maintenance decision tree to plan the maintenance instruction sequence. The nodes of the decision tree represent different maintenance operations, such as restarting the process, cleaning memory, system update, data backup, etc. The system uses the Monte Carlo tree search algorithm to simulate the decision tree multiple times according to the previously calculated solution weights. In each round of simulation, the algorithm explores down the path of the decision tree and records the number of visits and cumulative reward values ​​of each node (such as the maintenance effect score). Based on the number of visits and cumulative reward values ​​of the nodes, the path with the most visits and the highest reward value is selected to generate the maintenance instruction sequence. For example, the final generated sequence may be "Restart process->Clean memory->System update".

[0131] Before executing the generated maintenance instruction sequence, it is necessary to set up monitoring probes on the embedded device to collect system status data during the execution process, such as CPU occupancy, memory usage, storage capacity, network traffic, and system logs.

[0132] The collected state data before maintenance and state data after maintenance are input into a twin neural network respectively. The twin neural network contains two sub-networks with the same structure for feature extraction. The two sub-networks share the same parameters, map the state data before and after maintenance to the same feature space, calculate the similarity (e.g. cosine similarity) between the feature vector of the state after maintenance and the feature vector of the preset target state, and obtain the maintenance effect score. The target state feature vector represents the ideal system state.

[0133] If the maintenance effect score is lower than the preset threshold, the maintenance instruction sequence needs to be optimized. The system uses a policy gradient optimization algorithm to represent the selection strategy of the maintenance instruction sequence as a probability distribution function. The reward signal is calculated based on the difference in the maintenance effect scores of two adjacent maintenance operations, and the parameters of the probability distribution function are updated using the reward signal to generate a new maintenance instruction sequence. Repeat the optimized maintenance sequence until the maintenance effect score for 20 consecutive cycles is greater than the preset threshold.

[0134] The fault diagnosis report, the executed maintenance instruction sequence, the optimized maintenance instruction sequence, and the maintenance effect score are recorded in the maintenance knowledge base as a new historical case. All historical cases in the maintenance knowledge base are regularly clustered and sorted to build a knowledge graph of equipment operation and maintenance to better support future maintenance work. For example, the fault diagnosis information of a case is "CPU occupancy rate 95%", the maintenance plan is "restart process", and the effect feedback is "CPU occupancy rate reduced to 10%".

[0135] For example, a fault diagnosis report of an embedded device shows that the CPU usage is too high, reaching 90%. The system searches the knowledge base and finds three similar cases, whose maintenance solutions are "restart process", "clean memory" and "system update", with success rates of 70%, 60% and 50% respectively. The system assigns weights of 0.7, 0.6 and 0.5 to these three solutions respectively. After Monte Carlo tree search, the maintenance instruction sequence finally generated is "restart process". After execution, the CPU usage is reduced to 20%, and the maintenance effect score is high, so no optimization is required. The system records this maintenance information in the knowledge base.

[0136] In this embodiment, by automatically generating and optimizing the maintenance instruction sequence, human intervention is reduced, troubleshooting time is shortened, and maintenance efficiency is improved. By utilizing historical maintenance data and machine learning algorithms, the cause of the fault can be diagnosed more accurately and the most effective maintenance plan can be selected, thereby improving maintenance accuracy. By continuously updating the maintenance knowledge base and optimization algorithms, the system can continuously learn new failure modes and maintenance strategies, thereby enhancing the system's adaptability and improving the system's robustness.

[0137] In an optional embodiment,

[0138] The selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, and the reward signal value is calculated based on the difference in maintenance effect scores between two adjacent maintenance operations. The reward signal value is used to update the parameters of the probability distribution function as shown in the following formula:

[0139] ;

[0140] Among them, σ' represents the updated parameter, σ represents the current parameter, α represents the learning rate, F represents the Fisher information matrix, and F -1 represents the inverse of the Fisher information matrix, E π Represents the strategy distribution Expectations, γ r represents the discount factor at time step r, δ r represents the temporal difference error at time step r, which is solved by the immediate reward signal and the baseline. Represents the policy function The logarithmic gradient, s r represents the system state at time step r, a r represents the action taken at time step r and system state s.

[0141] For example, assume that when an embedded device fails, there are three maintenance plans: Plan A executes restart process and clears memory, Plan B executes system update and data backup, and Plan C executes resetting configuration and rebuilding index.

[0142] The processing method of the prior art adopts a fixed order of execution A→B→C. The maintenance effect scores after each scheme is executed are: Scheme A scores 0.45, Scheme B scores 0.55, and Scheme C scores 0.65. The total maintenance time is 180 minutes, and the average CPU occupancy is 85%.

[0143] The policy gradient optimization process of the present invention is as follows: In the initial state, the initial values ​​of the probability distribution σ of the three schemes are all 0.33, the learning rate α is set to 0.01, and the discount factor γ is set to 0.95. The score of the first round of iterative execution scheme C is 0.65, and the calculated time difference error δ1 is equal to 0.15 (0.65 minus the baseline 0.50), and the updated probability distribution becomes [0.25, 0.30, 0.45]. The score of the second round of iterative execution scheme B is 0.55, and the calculated time difference error δ2 is equal to -0.10 (0.55 minus 0.65), and the updated probability distribution becomes [0.20, 0.25, 0.55]. The score of the third round of iterative execution scheme C is 0.70, and the calculated time difference error δ3 is equal to 0.15 (0.70 minus 0.55), and the final probability distribution is updated to [0.15, 0.20, 0.65].

[0144] After optimization, the execution order becomes C→B→C. The maintenance effect scores after the execution of each plan are: the first execution of C scores 0.70, the execution of B scores 0.75, and the second execution of C scores 0.85. The total maintenance time is shortened to 120 minutes, and the average CPU usage is reduced to 65%.

[0145] The comparative analysis results show that: in terms of maintenance effect, the final score of the existing technology is 0.65, and the final score of the present invention is 0.85, with an improvement of 30.8%; in terms of time efficiency, the existing technology requires 180 minutes, while the present invention only requires 120 minutes, saving 33.3% of time; in terms of resource utilization, the CPU occupancy of the existing technology is 85%, while the CPU occupancy of the present invention is 65%, saving 23.5% of resources.

[0146] From this example, we can see that the policy gradient optimization algorithm of the present invention improves the maintenance effect score from 0.65 to 0.85 by dynamically adjusting the maintenance sequence; reduces maintenance time by 33.3%; saves 23.5% of computing resources; adaptively finds the optimal maintenance path, and identifies solution C as the most effective. These data directly confirm the technical effects of the present invention in improving maintenance efficiency, saving resources, and dynamically optimizing. At the same time, through the dynamic adjustment of probability distribution, the system can allocate more resources to maintenance solutions with better effects, such as increasing the probability of solution C from 0.33 to 0.65, achieving optimal resource allocation.

[0147] In this embodiment, by learning the optimal maintenance instruction sequence, the performance and reliability of the equipment can be maximized, thereby reducing maintenance time and cost. By dynamically adjusting the probability distribution function, limited maintenance resources can be concentrated on the equipment and components that need maintenance most, avoiding waste of resources. According to the actual operating status of the equipment and feedback on the maintenance effect, the maintenance strategy can be dynamically adjusted to adapt to different equipment and environments.

[0148] Figure 2 FIG. 1 is a schematic diagram of the structure of an online testing and remote maintenance system for embedded devices according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0149] The first unit is used to deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, an information entropy value is calculated based on the Shannon entropy principle, a sampling period is selected based on a preset information entropy threshold, and a wavelet transform is performed on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence according to the selected sampling period to obtain a first feature matrix, and the Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics, and the first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector;

[0150] The second unit is used to input the abnormality degree vector into the capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and the features output by the capsule network are subjected to Gaussian mixture clustering to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology, and the Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into the causal reasoning network, a fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by the hidden Markov model, and the equipment health score is calculated. If the equipment health score is lower than a preset health threshold, a maintenance process is triggered;

[0151] The third unit is used to retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan the maintenance instruction sequence in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, the maintenance instruction sequence is adjusted based on the historical maintenance data through the policy gradient optimization algorithm, and the maintenance and adjustment are repeated until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, and the maintenance is terminated and the maintenance process is stored in the maintenance knowledge base.

[0152] According to a third aspect of the embodiments of the present invention,

[0153] An electronic device is provided, comprising:

[0154] processor;

[0155] a memory for storing processor-executable instructions;

[0156] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the embodiments of the present invention is:

[0158] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0159] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An online testing and remote maintenance method for embedded devices, characterized in that: include: Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, calculate the information entropy value based on the Shannon entropy principle, select a sampling period based on a preset information entropy threshold, perform wavelet transform decomposition on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence according to the selected sampling period, and obtain a first characteristic matrix. Use fractal theory to calculate the Hausdorff dimension of storage performance data and the fractal dimension of network status data to obtain a second characteristic matrix representing timing complexity characteristics. Input the first characteristic matrix and the second characteristic matrix into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector. The abnormality degree vector is input into a capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by a hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than a preset health threshold, a maintenance process is triggered. According to the fault diagnosis report, similar fault cases are retrieved in the maintenance knowledge base, and the related maintenance plans are extracted, and the success rate of each maintenance plan in the historical records is calculated. The maintenance plan weight is generated and the maintenance instruction sequence is planned in combination with the Monte Carlo tree search algorithm. The maintenance instruction sequence is executed and the execution status is collected according to the preset monitoring nodes. An effect evaluation model based on the twin neural network is constructed, and the similarity between the current state characteristics of the system and the target state characteristics is calculated to obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, the maintenance instruction sequence is adjusted based on the historical maintenance data through the policy gradient optimization algorithm. The maintenance and adjustment are repeated until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, and the maintenance is terminated and the maintenance process is stored in the maintenance knowledge base.

2. The method according to claim 1, characterized in that Edge computing nodes are deployed in embedded devices to collect multi-source heterogeneous data sequences. For each data sequence, the information entropy value is calculated based on the Shannon entropy principle. The sampling period is selected based on a preset information entropy threshold. The processor power consumption data and memory usage data in the multi-source heterogeneous data sequence are decomposed by wavelet transform according to the selected sampling period to obtain a first feature matrix. The Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics. The first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector including: Deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data corresponding to the embedded devices, and generate multi-source heterogeneous data sequences by combining timestamps, wherein the multi-source heterogeneous data include processor power consumption data, memory usage data, storage performance data, and network status data of the embedded devices; For each data sequence in the multi-source heterogeneous data sequence, multiple time windows are divided based on the Shannon entropy principle, the data probability distribution in each time window is calculated, the data probability distribution is substituted into the Shannon entropy formula, and the information entropy value of the time window is calculated. If the information entropy value is greater than a preset information entropy threshold, the sampling period is set to the first sampling period, otherwise, it is set to the second sampling period, wherein the first sampling period is less than the second sampling period, and the set sampling period is fed back to the edge computing node; Based on the selected sampling period, performing multi-level wavelet decomposition on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence by using the Dauby wavelet, obtaining corresponding approximate coefficients and detail coefficients at each layer, arranging the approximate coefficients and the detail coefficients in order from low frequency to high frequency, and obtaining the first feature matrix; For the storage performance data, a point set consisting of storage data throughput samples within a unit time range is statistically analyzed based on fractal theory, and the space where the point set is located is gridded in combination with the box dimension method, and the number of grids containing the points is counted and the side length of each grid is changed to obtain multiple storage performance data pairs, and linear fitting is performed, and the slope of the obtained fitting straight line is used as the Hausdorff dimension. For the network status data, the correlation integral under different time lags is calculated, and the optimal embedding dimension is selected in combination with the Takkens embedding theorem to reconstruct the phase space to obtain a trajectory matrix, and the fractal dimension is calculated based on the trajectory matrix, and the Hausdorff dimension and the fractal dimension are arranged in chronological order to construct a second characteristic matrix; The first feature matrix and the second feature matrix are normalized to generate a normalized feature matrix. The normalized feature matrix is ​​transformed by combining an anomaly detection model constructed based on fifth-order generalized Chebyshev polynomials to obtain Chebyshev coefficients. The Chebyshev coefficients are constructed as an abnormal feature vector. The distance from the abnormal feature vector to a preset normal sample boundary is calculated and normalized to obtain the abnormal degree vector.

3. The method according to claim 2, characterized in that Substituting the data probability distribution into the Shannon entropy formula, the information entropy value of the time window is calculated as shown in the following formula: ； Among them, H(X t ) represents the time window X t The information entropy value, λ t represents the time-varying forgetting factor, H(X t-1 ) represents the information entropy value of time window t-1, w i (t) represents the weight coefficient of the ith event, p i Represents the time window X t The probability of the i-th event occurring in represents the correction term, n represents the total number of times, β represents the KL divergence regularization term weight factor, and K(p||q) represents the KL divergence between the data distribution p and the prior distribution q.

4. The method according to claim 1, characterized in that: The abnormality degree vector is input into a capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and Gaussian mixture clustering is performed on the features output by the capsule network to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology. The Lyapunov index and the correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a causal reasoning network. A fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by a hidden Markov model. The equipment health score is calculated. If the equipment health score is lower than a preset health threshold, the maintenance process is triggered, including: Adding the abnormality degree vector to a pre-trained capsule neural network, wherein the structure of the capsule neural network corresponds to the system architecture of the embedded device, each capsule unit corresponds to a system component in the embedded device, wherein the system components include a processor, memory, storage, and network, and the capsule neural network performs feature extraction on the abnormality degree vector to obtain an abnormal feature vector; The capsule neural network learns the spatial relationship and constraint relationship between different components according to the dynamic routing algorithm, adaptively adjusts the routing weights between different capsule layers in combination with the mutual information of the component states corresponding to the embedded device, and modifies the message transmission strength between different components based on the adjusted routing weights; Perform Gaussian mixture clustering on the abnormal feature vectors corresponding to the capsule neural network, iteratively estimate the parameters of the preset Gaussian mixture model through the expectation maximization algorithm, calculate the posterior probability that each feature vector in the abnormal feature vector belongs to a different Gaussian distribution, and update the mean, covariance and mixing coefficient of each Gaussian distribution based on the posterior probability to obtain the probability distribution of the fault type; According to the probability distribution of the fault type, a nonlinear prediction model is constructed based on the phase space reconstruction technology, the system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, the Lyapunov index and the correlation dimension of the system state trajectory are calculated, the probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into a pre-constructed causal reasoning network, the causal dependency relationship is modeled through a directed acyclic graph, the hidden causal structure is learned by iteratively updating the edge weights, a fault diagnosis report is obtained and a state transfer matrix is ​​constructed, and the hidden system state is determined based on the state transfer matrix; The hidden system state is used as the hidden state space of the hidden Markov model, and the fault diagnosis report is used as the observed state space of the hidden Markov model. The state transfer matrix and the emission probability matrix predetermined from historical fault data are combined, and the system state evolution path is predicted through a forward-backward algorithm and the health status of the current equipment is calculated to obtain the equipment health score. If the equipment health score is lower than the preset health threshold, the maintenance process is triggered.

5. The method according to claim 4, characterized in that A nonlinear prediction model is constructed based on phase space reconstruction technology, a system state sequence constructed according to the probability distribution of the fault type is added to the nonlinear prediction model, and the Lyapunov index and correlation dimension of the system state trajectory are calculated, including: Reconstruct the state sequence of the embedded device into a phase space trajectory, wherein the state sequence includes state values ​​at multiple time points, and reconstruct the state sequence into a phase space vector according to the selected embedding dimension and reconstruction delay, wherein the phase space vector is composed of the state value at the current moment and the state values ​​at multiple previous time points, wherein the embedding dimension is the minimum integer greater than the fractal dimension of the system attractor, and the reconstruction delay is determined by the first zero point of the average mutual information; A nonlinear prediction model is constructed based on the phase space trajectory. For each state vector to be predicted, a plurality of nearest state vectors are selected in the neighborhood of the current state vector to be predicted, and the plurality of nearest state vectors are linearly combined to obtain a predicted value of the state vector to be predicted. The coefficients of the linear combination are obtained by minimizing the reconstruction error. Gaussian mixture clustering is used to obtain a priori probability distribution of fault types, the priori probability distribution is mapped to the phase space trajectory, a weight coefficient is assigned to a state vector, and weighted processing is performed on the nonlinear prediction model according to the weight coefficient; Based on the phase space trajectory after weighted processing, the Lyapunov exponent spectrum of the system evolution is calculated, and the largest Lyapunov exponent is selected from the Lyapunov exponent spectrum for output; The correlation dimension is calculated by using a correlation dimension algorithm, and the correlation dimension is obtained by counting the number of state pairs whose distance in the phase space is less than a given scale and calculating the slope of the double logarithmic curve.

6. The method according to claim 1, characterized in that Retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan maintenance instruction sequences in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, adjust the maintenance instruction sequence based on the historical maintenance data through the policy gradient optimization algorithm, repeat maintenance and adjustment until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, end maintenance and store the maintenance process in the maintenance knowledge base, including: Inputting the fault diagnosis report of the embedded device into a pre-built maintenance knowledge base, the maintenance knowledge base stores a plurality of historical fault cases, each of which includes fault diagnosis information, maintenance plan and effect feedback information, the fault diagnosis information includes CPU occupancy data, memory usage data, and system resource status data, and calculating the semantic similarity between the fault diagnosis report and each historical fault case based on the fault feature vector in the fault diagnosis report, and retrieving the historical fault cases with the high similarity as candidate historical fault cases; Calculate the success rate of each maintenance solution in the candidate historical fault cases, where the success rate is obtained by dividing the number of successful repairs of the maintenance solution in the maintenance knowledge base by the total number of executions, and set a corresponding weight coefficient according to the success rate of each maintenance solution, where the weight coefficient increases as the success rate increases; A maintenance decision tree is constructed based on a Monte Carlo tree search algorithm, wherein the nodes of the maintenance decision tree include a process restart operation node, a memory cleanup operation node, a system update operation node, and a data backup operation node. Based on the weight coefficient, the Monte Carlo tree search algorithm is used to perform multiple rounds of iterative sampling on the maintenance decision tree. In each round of iteration, the number of visits and the cumulative reward value of each node are recorded, and a maintenance instruction sequence is generated according to the number of visits and the cumulative reward value. A monitoring probe is set on the embedded device to collect execution status data during the execution of the maintenance instruction sequence, wherein the execution status data includes CPU occupancy data, memory usage data, storage capacity data, network traffic data and system log data, and the execution status data is input into a twin neural network, wherein the twin neural network includes two parameter-sharing feature extraction subnetworks, and the feature extraction subnetworks are used to process the pre-maintenance status data and the post-maintenance status data respectively to obtain a pre-maintenance feature vector and a post-maintenance feature vector, and the cosine similarity between the post-maintenance feature vector and the preset target feature vector is calculated to obtain a maintenance effect score; When the maintenance effect score is less than a preset threshold, the maintenance instruction sequence is adjusted using a policy gradient optimization algorithm, the selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, the reward signal value is calculated based on the difference in the maintenance effect scores of two adjacent maintenance operations, the parameters of the probability distribution function are updated using the reward signal value, an optimized maintenance sequence is generated, and the optimized maintenance sequence is repeatedly executed until the maintenance effect scores within a preset number of consecutive cycles are all greater than the preset threshold; The fault diagnosis report, the maintenance instruction sequence, the optimized maintenance sequence and the maintenance effect score are combined into a new historical fault case and stored in the maintenance knowledge base. All historical fault cases in the maintenance knowledge base are clustered and sorted to construct an equipment operation and maintenance knowledge graph.

7. The method according to claim 6, characterized in that The selection strategy of the maintenance instruction sequence is constructed as a probability distribution function, and the reward signal value is calculated based on the difference in maintenance effect scores between two adjacent maintenance operations. The reward signal value is used to update the parameters of the probability distribution function as shown in the following formula: ； Among them, σ' represents the updated parameter, σ represents the current parameter, α represents the learning rate, F represents the Fisher information matrix, and F -1 represents the inverse of the Fisher information matrix, E π Represents the strategy distribution Expectations, γ r represents the discount factor at time step r, δ r represents the temporal difference error at time step r, which is solved by the immediate reward signal and the baseline. Represents the policy function The logarithmic gradient, s r represents the system state at time step r, a r represents the action taken at time step r and system state s.

8. An online testing and remote maintenance system for embedded devices, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy edge computing nodes in embedded devices and collect multi-source heterogeneous data sequences. For each data sequence, an information entropy value is calculated based on the Shannon entropy principle, a sampling period is selected based on a preset information entropy threshold, and a wavelet transform is performed on the processor power consumption data and memory occupancy data in the multi-source heterogeneous data sequence according to the selected sampling period to obtain a first feature matrix, and the Hausdorff dimension of the storage performance data and the fractal dimension of the network status data are calculated using fractal theory to obtain a second feature matrix representing the time series complexity characteristics, and the first feature matrix and the second feature matrix are input into an anomaly detection model constructed based on generalized Chebyshev polynomials to generate an anomaly degree vector; The second unit is used to input the abnormality degree vector into the capsule neural network for feature extraction. Each capsule unit of the capsule neural network corresponds to a system component. The routing weights between capsule layers are dynamically adjusted according to the mutual information of the component status, and the features output by the capsule network are subjected to Gaussian mixture clustering to obtain the probability distribution of the fault type. A nonlinear prediction model is established based on the phase space reconstruction technology, and the Lyapunov index and correlation dimension of the system state trajectory are calculated. The probability distribution of the fault type, the Lyapunov index and the correlation dimension are input into the causal reasoning network, a fault diagnosis report is generated and a state transfer matrix is ​​constructed. The system state evolution path is predicted by the hidden Markov model, and the equipment health score is calculated. If the equipment health score is lower than a preset health threshold, a maintenance process is triggered; The third unit is used to retrieve similar fault cases in the maintenance knowledge base according to the fault diagnosis report, extract related maintenance plans, calculate the success rate of each maintenance plan in the historical records, generate maintenance plan weights and plan the maintenance instruction sequence in combination with the Monte Carlo tree search algorithm, execute the maintenance instruction sequence and collect the execution status according to the preset monitoring nodes, build an effect evaluation model based on the twin neural network, calculate the similarity between the current state characteristics of the system and the target state characteristics, and obtain a maintenance effect score. If the maintenance effect score is lower than the preset maintenance effect threshold, the maintenance instruction sequence is adjusted based on the historical maintenance data through the policy gradient optimization algorithm, and the maintenance and adjustment are repeated until the maintenance effect score is greater than the maintenance effect threshold in 20 consecutive cycles, and the maintenance is terminated and the maintenance process is stored in the maintenance knowledge base.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images