Electric fireplace fault self-diagnosis method and system

Abstract

The invention relates to the technical field of intelligent household equipment and Internet of Things operation and maintenance, and discloses an electric fireplace fault self-diagnosis method and system, and the method comprises the steps: obtaining multi-dimensional sensing data and historical records; performing feature extraction and clustering calculation according to the multi-dimensional sensing data to obtain a difference feature set; inputting the difference feature set into a pre-constructed causal model for deduction, and extracting to obtain a potential abnormal link; according to the potential abnormal links and the historical records, fusing to obtain a state matrix, constructing a difference directed graph based on the state matrix, and positioning an abnormal source; and inputting the state vector corresponding to the abnormal source into a pre-constructed Bayesian network for deduction to determine a fault starting point. According to the method, the fault root cause of the electric fireplace can be accurately positioned.

Description

Technical Field

[0001] The present invention relates to the technical fields of smart home devices and Internet of Things operation and maintenance technologies, and particularly relates to a method and system for self-diagnosing faults of an electric fireplace. Background Art

[0002] In the field of modern smart home devices, as a comprehensive device integrating heating and decorative lighting effects, the internal operation of an electric fireplace involves complex interactions of multiple physical fields such as electric energy conversion, heat transfer, and light effect display. With the increasing high integration and intelligence of device functions, the reliable monitoring of the device operation state is gradually evolving towards fault prediction and health management, aiming to achieve precise abnormal monitoring and maintenance before the device suffers serious damage or poses safety hazards.

[0003] Currently, for the fault diagnosis of electric fireplaces and similar devices, in an existing technology, it usually relies on a hard threshold judgment mechanism of a single sensor (such as over-temperature alarm or over-current power-off), or simple comparison is performed using cross-sectional data based on static rules. However, during the actual dynamic operation of an electric fireplace, there are high dynamic coupling and significant time response differences among multiple parameters such as electric energy input, heat output, and light effect presentation. For example, a minor abnormality in the control circuit may first cause a hidden fluctuation in electric energy input, and after a certain thermal inertia delay, it may lead to an offset in heat output, and then trigger a chain instability in the light effect module. The existing technology only relies on isolated static thresholds or single-dimensional surface data, which breaks the dynamic interaction law and deep causal topological relationship of multi-dimensional parameters in time series. When a fault occurs in the device and causes multi-parameter linkage anomalies, the existing diagnostic system often can only capture the most obvious surface features at the end of the propagation chain, and cannot clarify the time sequence and spatial propagation context of parameter anomalies.

[0004] Therefore, there is a technical problem in the existing technology that it is difficult to accurately trace the path of abnormal chain reactions and effectively locate the initial source of faults. Summary of the Invention

[0005] The present invention provides a method and system for self-diagnosing faults of an electric fireplace to solve the technical problem in the existing technology that it is difficult to accurately trace the path of abnormal chain reactions and effectively locate the initial source of faults.

[0006] In a first aspect, to solve the above technical problem, the present invention provides a method for self-diagnosing faults of an electric fireplace, including:

[0007] Obtain multi-dimensional sensing data, and perform feature extraction on the multi-dimensional sensing data to obtain a response sequence set;

[0008] Perform clustering analysis according to the response sequence set to obtain a set of differential features;

[0009] The differential feature set is input into a pre-built causal model to obtain an offset sequence. Based on the offset sequence, deviation analysis is performed to obtain potential abnormal links.

[0010] Obtain historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set with the historical records according to the offset triggering order to obtain a state matrix, and construct a difference directed graph based on the state matrix;

[0011] The directed edges with weights exceeding a preset weight threshold in the differential directed graph are extracted and spliced ​​together to obtain the reaction path. The in-degree of the nodes in the reaction path is calculated to obtain the in-degree value. The nodes with an in-degree value of zero are located as the source of the anomaly.

[0012] Extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the response path to obtain a consistency score, and determine the anomaly source with the consistency score higher than a preset consistency threshold as the fault starting point.

[0013] Secondly, the present invention provides a self-diagnostic system for electric fireplace malfunctions, comprising:

[0014] The data processing module is used to acquire multidimensional sensing data and extract features from the multidimensional sensing data to obtain a response sequence set.

[0015] The feature difference module is used to perform cluster analysis based on the response sequence set to obtain a set of difference features;

[0016] The offset detection module is used to input the differential feature set into a pre-built causal model to obtain an offset sequence, and perform deviation analysis based on the offset sequence to obtain potential abnormal links;

[0017] The graph construction module is used to acquire historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set with the historical records according to the offset triggering order to obtain a state matrix, and construct a difference directed graph based on the state matrix.

[0018] The source localization module is used to extract the directed edges in the differential directed graph whose weights exceed a preset weight threshold, splice them to obtain the reaction path, calculate the in-degree of the nodes in the reaction path to obtain the in-degree value, and locate the nodes with an in-degree value of zero as the source of the anomaly.

[0019] The traceability output module is used to extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the response path to obtain a consistency score, and determine the anomaly source with the consistency score higher than a preset consistency threshold as the fault starting point.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) This invention constructs an extended matrix by acquiring multidimensional sensing data and calculating the time delay correlation values ​​between response sequence sets, and compares historical delay sequences with dynamic time warping algorithm to extract offset triggering order. It effectively overcomes the defect that traditional single-point static threshold cannot capture the time delay effect between parameters, deeply explores the implicit coupling and time-series linkage law of multidimensional parameters such as power, heat and light effect in the dynamic operation of electric fireplace, significantly improves the comprehensiveness and robustness of fault feature extraction in complex and variable environment, and effectively avoids misjudgment and omission caused by local surface data fluctuation.

[0022] (2) This invention extracts potential abnormal links by inputting the differential feature set into the causal model, and further combines the in-degree analysis of the differential directed graph with the Bayesian network to perform posterior probability inference, compares the response path to obtain a consistency score, realizes the transformation of the diagnostic paradigm from superficial feature matching to deep causal tracing, can accurately strip away the complex downstream chain reaction interference, directly lock the initial abnormal source with zero in-degree, greatly improve the accuracy and computational efficiency of fault root cause location, and greatly shorten the troubleshooting and maintenance cycle of equipment.

[0023] (3) This invention combines the fault origin and the state of its downstream nodes associated in the reaction path to perform time-series mapping, and deduce the predicted trajectory of each node. Based on the node linkage relationship, the propagation branch is reconstructed to output a complete traceability chain. This not only clearly and intuitively restores the fault occurrence time and spatial propagation path, but also gives the system the ability to dynamically predict the subsequent deterioration trend of the equipment. It successfully upgrades the traditional passive post-event maintenance to proactive fault prediction and health management, and comprehensively ensures the stability and safety of electric fireplaces in long-term operation. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the self-diagnosis method for electric fireplace faults provided in the first embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the structure of the electric fireplace fault self-diagnosis system provided in the second embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Reference Figure 1 The first embodiment of the present invention provides a self-diagnosis method for electric fireplace faults, including the following steps:

[0028] S11, acquire multidimensional sensing data, and perform feature extraction on the multidimensional sensing data to obtain a response sequence set;

[0029] S12, perform cluster analysis based on the response sequence set to obtain a set of differential features;

[0030] S13, input the differential feature set into the pre-constructed causal model to obtain the offset sequence, and perform deviation analysis based on the offset sequence to obtain potential abnormal links;

[0031] S14, Obtain historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set and the historical records according to the offset triggering order to obtain a state matrix, and construct a differential directed graph based on the state matrix;

[0032] S15, extract the directed edges in the differential directed graph whose weights exceed a preset weight threshold and splice them to obtain the reaction path, calculate the in-degree of the nodes in the reaction path to obtain the in-degree value, and locate the nodes with an in-degree value of zero as the source of the anomaly.

[0033] S16, extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the reaction path to obtain a consistency score, and determine the anomaly source with the consistency score higher than the preset consistency threshold as the fault starting point.

[0034] In step S11, multidimensional sensing data is acquired, and feature extraction is performed on the multidimensional sensing data to obtain a response sequence set.

[0035] The process of extracting features from the multidimensional sensing data to obtain a response sequence set includes:

[0036] The multidimensional sensing data is aligned using timestamps to obtain fused data; the multidimensional sensing data includes temperature data, current data, and voltage data.

[0037] The features of the fused data are extracted to obtain an initial feature set, and the initial feature set is completed using interpolation to obtain a complete feature set;

[0038] A state mapping matrix is ​​obtained by performing time-series analysis on the complete feature set;

[0039] The response sequence set is obtained by calculating the state switching delay time based on the state mapping matrix.

[0040] In one implementation, this embodiment acquires multi-dimensional sensing data through a sensor network deployed inside the electric fireplace. The sensor network includes temperature sensors, current sensors, and voltage sensors. Each sensor independently collects physical quantity data according to a preset sampling frequency, and an absolute millisecond-level timestamp is appended to each collected data point in the underlying hardware driver.

[0041] It should be noted that the preset sampling frequency is determined based on the Nyquist sampling theorem for changes in the physical field of the device. In this embodiment, the highest frequency of change in the temperature physical field within the operating environment of the electric fireplace and the highest frequency of change in the control circuit signal are obtained respectively. The maximum value among all frequencies is extracted, and this maximum value is multiplied by a coefficient of two to five. The calculation result is set as the preset sampling frequency, for example, 10 Hz. This embodiment extracts millisecond-level timestamps from the multidimensional sensing data and uses a sliding time window algorithm for time alignment. This embodiment sets a time tolerance window, the size of which is equal to the reciprocal of the preset sampling frequency. Using the scale of the system reference clock as the main axis, this embodiment concatenates the temperature, current, and voltage values ​​collected within the same time tolerance window into a one-dimensional parallel data record, and arranges all parallel data records under all time windows in sequence to generate the fused data.

[0042] In one implementation, this embodiment extracts features from the fused data to obtain an initial feature set. Specifically, this embodiment sets a fixed-length feature extraction time span, and calculates statistical indicators for various types of sensor data within each time span. The features extracted in this embodiment include the arithmetic mean and first-order difference rate of change of temperature data, as well as the extreme value difference and fluctuation frequency of current and voltage data. This embodiment combines the calculated feature values ​​into a multi-dimensional feature vector, which is then concatenated according to the time series to form the initial feature set.

[0043] It should be noted that due to sensor hardware transmission jitter, some dimensions of the initial feature set may be missing data. This embodiment uses interpolation to complete the initial feature set to obtain a complete feature set. It is worth noting that the interpolation method used is Lagrange polynomial interpolation. This embodiment traverses the initial feature set to identify missing nodes with missing values. This embodiment extracts a preset number of valid feature values ​​adjacent to the missing node on the time axis as reference nodes. Preferably, the preset number is set to 2 or 3, which is based on ensuring that the local interpolation curve smoothly approximates the actual physical values ​​while avoiding the Runge phenomenon, which causes severe oscillations at the edges of the interpolation interval when using higher-order polynomials. A Lagrange basis function is constructed using the timestamps of the reference nodes and the feature values. The timestamps corresponding to the missing nodes are substituted into the constructed polynomial to calculate the feature fitting value for that dimension. This embodiment uses the feature fitting value to replace the corresponding missing data, generating the complete feature set without missing values.

[0044] In one implementation, a state mapping matrix is ​​obtained by performing time-series analysis on the complete feature set. A shortest-distance classification algorithm based on Mahalanobis distance is then used to map continuous feature vectors to discrete physical operating states.

[0045] It should be noted that this embodiment pre-constructs a standard state database. The database construction process involves injecting standard operating commands into the test electric fireplace under controlled laboratory physical conditions, causing the equipment to perform standby, heating, power-off, and overheat protection actions. The system synchronously collects complete feature set data under the corresponding actions, and labels each segment of feature set data with known physical operating states. It then calculates and extracts the covariance matrix and class center coordinates of various state features, stores them as feature interval templates, and inputs them into the standard state database.

[0046] In one implementation, the feature vector of each time point in the complete feature set to be analyzed is compared with each feature interval template in the standard state database. Combining the covariance matrix of the templates, the Mahalanobis distance between the feature vector to be analyzed and the class center of each feature interval template is calculated. The physical working state corresponding to the template with the smallest distance value is selected as the mapping state for that time point. Finally, the time axis coordinates, multidimensional feature vectors, and inferred mapping states are combined accordingly to construct the two-dimensional state mapping matrix.

[0047] For example, suppose the state mapping matrix records the internal state transition process of an electric fireplace within a specific time period. The first time node recorded by the matrix corresponds to the mapping state of heating element off, at which time the first-order differential rate of change of temperature is continuously negative and the current value is close to 0 amperes; the mapping state corresponding to the second time node recorded by the matrix changes to normal heating, at which time the first-order differential rate of change of temperature changes from negative to positive and the current value jumps to 5.2 amperes. In this embodiment, the adjacent nodes in the state mapping matrix that have undergone state transitions are retrieved, and the end timestamp of the interval before the state change and the start timestamp of the interval after the state change are extracted. The state switching delay time is calculated by subtracting the end timestamp from the start timestamp. The entire mapping matrix is ​​traversed, and all state switching delay times generated when various sensor parameters participate in state switching are spliced ​​and serialized in chronological order to generate the response sequence set.

[0048] In step S12, cluster analysis is performed based on the response sequence set to obtain a set of differential features, including:

[0049] The response sequence set is dimensionless to obtain a standardized sequence set, and the correlation coefficient between each pair of sequences in the standardized sequence set is calculated using a cross-correlation function to obtain the intensity value.

[0050] The interaction set is obtained by aggregating the sequence pairs whose intensity values ​​exceed a preset association threshold;

[0051] The time delay difference of each sequence pair in the interaction set is calculated as the time delay correlation value, and the extended interaction matrix is ​​constructed based on the time delay correlation value;

[0052] Clustering algorithms are used to group sequences with similar time-delay correlation values ​​in the extended interaction matrix to obtain differential response clusters;

[0053] Extract the deviation vectors of the differential response clusters, and merge the deviation vectors to obtain the differential feature set.

[0054] In one implementation, the sequences in the response sequence set are dimensionless to obtain a standardized sequence set. This embodiment uses the Z-Score normalization algorithm to calculate the arithmetic mean and standard deviation of all values ​​within a single sequence. The arithmetic mean is subtracted from each original value in the sequence, and the difference is divided by the standard deviation to generate a standard numerical sequence with a mean of zero and a variance of one. Then, this normalization calculation is performed on all sequences in the response sequence set, and all standard numerical sequences are aggregated to generate the standardized sequence set. Next, the correlation coefficient between each pair of sequences in the standardized sequence set is calculated using a cross-correlation function to obtain the strength value. Specifically, the two standard numerical sequences are shifted point-by-point relative to each other on the time axis, the discrete inner product is calculated at different time shift steps, and a normalization operation is performed. The global maximum value in the normalized discrete inner product function sequence is extracted and determined as the strength value characterizing the degree of correlation between the two sequences.

[0055] It should be noted that in this embodiment, the sequence pairs with intensity values ​​exceeding a preset association threshold are aggregated to obtain the interaction set. The preset association threshold is determined using a statistical analysis method based on a Gaussian mixture model. Association strength data of all sequence pairs generated during the device's historical normal operation cycle are collected. The expectation-maximization algorithm is used to fit a bimodal Gaussian distribution density function containing strong and weak association components. The x-coordinate value corresponding to the intersection point of the two Gaussian component probability density curves is calculated and set as the preset association threshold. All two sequences with intensity values ​​greater than this threshold are bound together as a sequence pair, and all sequence pairs satisfying the condition are aggregated to construct the interaction set.

[0056] In one implementation, the time delay difference between each sequence pair in the interaction set is calculated as the time delay correlation value, and the extended interaction matrix is ​​constructed based on the time delay correlation value. For each sequence pair in the interaction set, the time shift step corresponding to the global maximum value generated during the aforementioned cross-correlation function calculation is extracted. The product of this time shift step and the data sampling period is calculated, and the absolute value of this product is taken to obtain the time delay correlation value of the two sequences in the physical response. In this embodiment, the parameter type of each sequence is used as the row and column indices of the matrix. The calculated time delay correlation value is filled into the corresponding matrix element positions, and all zeros are filled into the main diagonal positions of the matrix to construct the diagonally symmetric extended interaction matrix.

[0057] Furthermore, this embodiment utilizes a clustering algorithm to group sequences with similar time-delay correlation values ​​in the extended interaction matrix to obtain differential response clusters. A density-based spatial clustering algorithm (DBSCAN) is employed. The preset neighborhood radius parameter in this DBSCAN algorithm is determined based on the inflection point analysis of a k-distance graph, and the preset minimum number of contained points parameter is determined by rounding down the natural logarithm of the total dimension of the sequences. Using each parameter in the extended interaction matrix as a data point and the time-delay correlation value as a distance metric, data points within the same neighborhood radius density connectivity range are grouped into the same category, outputting multiple independent differential response clusters.

[0058] In this embodiment, the deviation vectors of the differential response clusters are extracted, and the deviation vectors are merged to obtain the differential feature set. For each differential response cluster, the arithmetic mean of the time-delay correlation values ​​of all sequence pairs within the cluster is calculated. This arithmetic mean is then subtracted from each individual time-delay correlation value within the cluster to generate the deviation vector composed of multiple discrete differences. Following the index order of the original matrix, the deviation vectors of all differential response clusters are row-wise concatenated to generate the differential feature set.

[0059] In step S13, the differential feature set is input into a pre-constructed causal model to obtain an offset sequence, and deviation analysis is performed based on the offset sequence to obtain potential abnormal links.

[0060] Specifically, deviation analysis is performed based on the offset sequence to identify potential abnormal links, including:

[0061] The offset sequence is analyzed to obtain the delay time, the delay time is compared with a preset reference delay interval to obtain the excess portion, and the ratio of the excess portion to the upper limit of the reference delay interval is calculated to obtain the deviation.

[0062] The deviation is compared with a preset deviation threshold, and parameters with values ​​greater than the preset deviation threshold are extracted to construct a potential fault set. The parameters in the potential fault set are then identified as the potential abnormal links.

[0063] In one implementation, the differential feature set is input into a pre-constructed causal model to obtain the offset sequence. The pre-constructed causal model is a spatiotemporal causal inference model jointly constructed based on a directed acyclic graph and a graph convolutional neural network. The construction and training process of this model involves collecting multi-dimensional sensor data sequences of electric fireplaces during historical operation, including various standard operating conditions and known fault offset conditions. The Granger causality test algorithm is used to calculate the P-value and F-statistic between the time series of each parameter. Parameter pairs with P-values ​​less than a preset significance level (e.g., 0.05, representing the statistical rejection of the null hypothesis of no causal relationship) are extracted to construct directed topological edges, and the corresponding F-statistic values ​​are assigned to these directed edges as the initial weights for causal connections, thereby constructing an objective causal graph adjacency matrix without business experience intervention. The network structure of the causal model sequentially includes an input layer, a three-layer graph convolutional operation layer, and a time series inference output layer composed of a multilayer perceptron regression head. The mean squared error is used as the loss function, and the adaptive moment estimation (Adam) algorithm is used to iteratively update the neuron weight parameters of each network layer along the time axis. In this embodiment, the difference feature set is used as the initial node feature input to the trained model. Spatial topological features are extracted through a three-layer graph convolution operation layer. The spatial topological features are flattened and mapped into a one-dimensional time delay scalar using the multilayer perceptron regression head, generating the offset sequence representing the response time drift of each parameter.

[0064] In one implementation, parsing the offset sequence to obtain the delay time includes extracting the scalar time difference value corresponding to each parameter in the offset sequence, performing an absolute value operation on it, and determining the absolute value value as the delay time of the corresponding parameter.

[0065] In one implementation, the deviation is obtained by comparing the delay time with a preset reference delay interval to obtain the excess portion, and by calculating the ratio of the excess portion to the upper limit of the reference delay interval.

[0066] It is worth noting that the preset baseline delay interval is determined based on the normal test dataset of the electric fireplace's entire life cycle. In this embodiment, the physical response time required for various parameters to undergo state switching under fault-free conditions is statistically analyzed, and the first quartile and the third quartile of the sample set are calculated. The first quartile is set as the lower limit, and the third quartile is set as the upper limit. The baseline delay interval is constructed using these two points.

[0067] In one implementation, it is determined whether the delay time is greater than the upper limit of the baseline delay interval. If it is determined to be greater, the upper limit of the baseline delay interval is subtracted from the delay time to calculate the physically excessive lag time, which is defined as the excess portion. Subsequently, the excess portion is used as the dividend, and the upper limit of the baseline delay interval is used as the divisor to perform a division operation. The resulting ratio is the deviation. It should be clearly stated that the deviation here represents the relative lag degree in the time dimension, and has different physical dimensions and mathematical meanings from the deviation concept involving physical numerical fluctuations in subsequent steps.

[0068] In one implementation, the deviation is compared with a preset deviation threshold, parameters with values ​​greater than the preset deviation threshold are extracted to construct a potential fault set, and the parameters in the potential fault set are located as the potential abnormal links.

[0069] It should be noted that the preset deviation threshold is determined based on the receiver operating characteristic (ROC) curve analysis rule. In this embodiment, based on historical fault repair records and a sensor data sample library, delayed deviation samples that cause actual physical faults are marked as positive, while normal deviation samples caused only by environmental noise are marked as negative. An ROC curve is constructed by traversing all possible deviation values, and the Youden index (i.e., the sum of sensitivity and specificity minus one) is calculated for each value. The deviation cutoff value corresponding to the maximum value of the Youden index is selected and set as the preset deviation threshold. Next, numerical comparisons are performed to filter out all parameters whose deviation is greater than the preset deviation threshold, constructing the potential fault set. Each parameter in the potential fault set is then positioned as a potential abnormal link for downstream logic analysis.

[0070] For example, suppose that the delay time of the light effect presentation parameter is 2.8 seconds obtained by parsing the offset sequence output by the causal model. The preset baseline delay range for this light effect presentation parameter, determined based on historical statistics, is [0.2 seconds, 1.0 seconds]. In this embodiment, a comparison reveals that 2.8 seconds is greater than the upper limit of 1.0 seconds. Subtracting 1.0 from 2.8 yields an excess of 1.8 seconds. Next, this embodiment divides the excess of 1.8 by the upper limit of 1.0 to calculate the deviation as 1.8. The preset deviation threshold calculated based on the ROC curve is 0.45. Finally, comparing the deviation of 1.8 with the threshold of 0.45, it is determined that the deviation is greater than the threshold, thus identifying the light effect presentation parameter as a potential abnormality.

[0071] In step S14, historical records are obtained, and the offset triggering order is obtained by comparing the historical records with the potential abnormal links. The response sequence set and the historical records are fused according to the offset triggering order to obtain a state matrix. A differential directed graph is constructed based on the state matrix.

[0072] The process of obtaining the offset triggering order by comparing the historical records with the potential abnormal links includes:

[0073] The historical records are analyzed to extract the delayed feature segments of the potential abnormal links, and the delayed feature segments are matched with preset historical abnormal segments to obtain matching segments;

[0074] The optimal alignment path is calculated by comparing the matching segment with the response sequence set using a dynamic time warping algorithm.

[0075] The time mapping relationship in the optimal alignment path is analyzed, the time offset of the response sequence set relative to the matching segment is extracted, and the offset triggering order is obtained by sorting the time offsets.

[0076] In one implementation, the historical records are parsed to extract delay feature segments of the potential abnormal links. Using the potential abnormal links generated in the above steps as search keywords, the historical records containing the device's past operating logs are used to extract sequence data of a preset span length before and after the state switching time point of the potential abnormal link, generating the delay feature segments. The preset span length is set based on the historical average time span of the electric fireplace completing one complete physical response cycle (e.g., from start-up to reaching steady-state temperature), for example, set to 50 sampling points before and after. The delay feature segments are matched with preset historical abnormal segments, and the Euclidean distance between them in the same time dimension is calculated. The historical abnormal segment with the smallest Euclidean distance is selected as the matching segment.

[0077] It should be noted that the preset historical anomaly segment is objectively determined based on an unsupervised anomaly detection algorithm. This embodiment collects all delayed sequence data within the historical lifecycle of the electric fireplace, calculates the anomaly score for each sequence using the Isolation Forest algorithm, and plots a score curve by arranging all scores in descending order. The second derivative of this score curve is calculated using an inflection point detection algorithm, and the anomaly score corresponding to the point where the second derivative reaches its maximum value (i.e., the curve declines most sharply, and a significant discontinuity appears between the abnormal and normal data) is extracted as the cutoff threshold. Sequences with anomaly scores higher than this cutoff threshold are extracted and stored in the benchmark database as the preset historical anomaly segment.

[0078] In one implementation, a dynamic time warping algorithm is used to compare the matching segment with the response sequence set to calculate the optimal alignment path. The specific execution process of this dynamic time warping algorithm is as follows: A two-dimensional distance matrix is ​​constructed, where the number of rows equals the length of the response sequence set and the number of columns equals the length of the matching segment. The absolute difference between each point in the response sequence set and each point in the matching segment is calculated and filled into the corresponding positions in the two-dimensional distance matrix. Subsequently, a dynamic programming algorithm is used to find a continuous path from the upper left corner to the lower right corner in the two-dimensional distance matrix. During the pathfinding process, the accumulated distance value for each step is calculated, and the path that minimizes the total accumulated distance value is selected and output as the optimal alignment path representing the alignment relationship between the two sequences.

[0079] This embodiment analyzes the time mapping relationship in the optimal alignment path and extracts the time offset of the response sequence set relative to the matching segment. Specifically, it extracts the index coordinates of the matching point pairs in the optimal alignment path, subtracts the corresponding time index of the matching segment from the time index of the response sequence set, and calculates a scalar difference. This scalar difference, multiplied by the data sampling period, is the time offset. The time offsets of all acquired parameters are sorted in ascending order, and the sorted parameter sequence structure is used to generate the offset triggering order.

[0080] Specifically, the process involves fusing the response sequence set and the historical records according to the offset triggering order to obtain a state matrix, and constructing a differential directed graph based on the state matrix, including:

[0081] Extract the response sequence set and the historical record within the time window corresponding to the offset triggering order, and arrange the response sequence set and the historical record according to the parameter type to obtain the state matrix;

[0082] The deviation between real-time data and historical data in the state matrix is ​​calculated. The parameter nodes are used as graph vertices, the temporal correlation directions between parameters are used as directed edges, and the deviation values ​​are used as directed edge weights to construct the difference directed graph.

[0083] In one implementation, the response sequence set and the historical record are extracted within the time window corresponding to the offset triggering order. Specifically, based on the offset triggering order generated in the preceding steps, the timestamp of the first sorting parameter is extracted as the start time, and the timestamp of the last sorting parameter is extracted as the end time, thus forming the time window. Within this time window, real-time sensing sequence segments from the response sequence set and corresponding sequence segments from the historical record are extracted. According to parameter types such as current, voltage, temperature, and photosensitivity values, the extracted response sequence set and the historical record are arranged row-wise to correspond, constructing a two-dimensional state matrix.

[0084] In one implementation, the deviation between real-time data and historical data in the state matrix is ​​calculated. Specifically, this involves traversing each row of the state matrix, subtracting the historical data value at the corresponding position from the real-time data value, calculating the absolute difference, and determining this absolute difference as the deviation value.

[0085] It is worth noting that this embodiment uses graph theory topology to generate the differential directed graph, instantiating each parameter node as a graph vertex. Based on the sequential relationship of parameter actions in the offset triggering sequence, this embodiment draws directed edges from the parameter node whose action occurred first to the parameter node whose action occurred later, serving as the temporal correlation direction between parameters. The calculated deviation value is assigned to the corresponding directed edge as its weight attribute. By integrating all vertices, directed edges, and weight values, a directed weighted graph structure is constructed, namely the differential directed graph.

[0086] In step S15, directed edges with weights exceeding a preset weight threshold in the differential directed graph are extracted and concatenated to obtain a reaction path. The in-degree of nodes in the reaction path is calculated to obtain the in-degree value. Nodes with an in-degree value of zero are identified as anomaly sources, including:

[0087] Traverse all directed edges in the differential directed graph, filter out edges with weights lower than a preset weight threshold, and extract the remaining directed edges as high-risk propagation links.

[0088] The high-risk transmission links are spliced ​​together end to end in chronological order to obtain the reaction path;

[0089] The in-degree value is obtained by counting the number of preceding associated edges of each node in the reaction path, and the starting node with an in-degree value of zero is extracted and marked as the source of the anomaly.

[0090] In one implementation, all directed edges in the differential directed graph are traversed, and edges with weights lower than a preset weight threshold are filtered out. It should be noted that the preset weight threshold is determined using the Gaussian distribution benchmark in statistics. The deviation values ​​between all parameter nodes generated by the data acquisition device under fault-free baseline operating conditions are weighted, and the arithmetic mean and standard deviation of these normal weight data are calculated. The arithmetic mean is added to three times the standard deviation (3-sigma criterion), and the sum is set as the preset weight threshold. The weight of each directed edge in the differential directed graph is compared with the preset weight threshold. Edges with weight values ​​less than or equal to the threshold and their isolated nodes are directly deleted, and the remaining directed edges with weight values ​​greater than the threshold are extracted and defined as the high-risk propagation links.

[0091] In one implementation, the high-risk propagation links are concatenated end-to-end in chronological order to obtain the response path. Timestamps corresponding to all high-risk propagation links are extracted, and these links are sorted in ascending order of time. The target endpoint vertex of the previous high-risk propagation link is found, and it is matched with the starting source vertex of the next high-risk propagation link that is immediately adjacent in time. This completes the end-to-end concatenation operation of the entire topology, generating one or more continuous directed acyclic path sequences, i.e., the response path.

[0092] It is worth noting that in this embodiment, the in-degree value is obtained by counting the number of preceding associated edges of each node in the reaction path. Each graph vertex in the reaction path is traversed, and the total number of directed edges pointing to that vertex is calculated. This total number of directed edges is the in-degree value of that vertex in graph theory. Starting nodes with an in-degree value of zero are extracted from the calculation results. Since an in-degree value of zero indicates that the node has no preceding triggering cause in the extracted propagation network, the node possesses independence from initial physical perturbations.

[0093] It should be further explained that if multiple nodes with an in-degree of zero are detected, the sum of the weights of all out-degree directed edges emanating from these nodes in the differential directed graph is calculated. The system extracts the node with the largest sum of weights (i.e., the node with the greatest physical damage or impact on downstream areas), uniquely marks it, and locates it as the source of the anomaly.

[0094] In step S16, the state vector corresponding to the anomaly source is extracted, and the state vector is input into a pre-constructed Bayesian network to deduce the probability path. The probability path is compared with the reaction path to obtain a consistency score, and the anomaly source with the consistency score higher than the preset consistency threshold is determined as the fault starting point.

[0095] Specifically, the state vector is input into a pre-constructed Bayesian network to deduce a probabilistic path. A consistency score is obtained by comparing the probabilistic path with the response path. Anomalies whose consistency scores exceed a preset consistency threshold are identified as fault origins. This includes:

[0096] The state vector is used as the evidence node input to the pre-constructed Bayesian network graph to perform probability inference and update the node state, thereby obtaining the posterior probability distribution of each associated node.

[0097] Calculate the conditional dependency probability of each node in the posterior probability distribution, and extract the branch with the highest cumulative probability to obtain the probability path;

[0098] The consistency score is obtained by calculating the node overlap rate between the probability path and the reaction path, and the abnormal source that exceeds the preset consistency threshold is extracted and determined as the fault starting point.

[0099] In one implementation, this embodiment uses the state vector as the input of the evidence node into a pre-constructed Bayesian network graph, performs probability inference to update the node state, and obtains the posterior probability distribution of each associated node.

[0100] It should be noted that the pre-constructed Bayesian network graph is a probabilistic graphical model built based on the device's historical operation dataset. This embodiment obtains a historical operation dataset containing all sensor parameter dimensions, uses the PC structure learning algorithm to calculate the conditional independence between each parameter node, and establishes a directed acyclic graph structure between nodes. The maximum likelihood estimation algorithm is used to calculate the conditional probability table for each node in the directed topology, and the directed acyclic graph structure and the conditional probability table are combined to generate the pre-constructed Bayesian network graph. Next, the state vector corresponding to the anomaly source is extracted, and the deviation values ​​of each term in the state vector are set as the observed evidence variables for the corresponding nodes in the Bayesian network graph. The joint tree exact inference algorithm is used to perform message passing and probability inference updates on the entire network, calculating the marginal posterior probability set of each associated node in the network, i.e., the posterior probability distribution, given the observed evidence variables.

[0101] In one implementation, the conditional dependency probability of each node in the posterior probability distribution is calculated, and the branch with the highest cumulative probability is extracted to obtain the probability path. Starting from the anomaly source, all downstream connected branches of the Bayesian network graph are traversed along the directed edges. Using the chain rule of probability theory, based on the topological directed dependencies of the Bayesian network graph, the conditional probability values ​​in the conditional probability table corresponding to each node in the connected branch are multiplied together to calculate the posterior joint probability distribution of the joint occurrence of the anomaly state sequences of each node in the branch, which is then used as the conditional dependency probability of each complete connected branch. Finally, the connected branch with the highest value is extracted, and the sequence of nodes and edges contained in this connected branch is determined as the probability path.

[0102] It is worth noting that, in this embodiment, the consistency score is obtained by calculating the node overlap rate between the probability path and the response path. Specifically, this includes extracting all first node sets in the probability path and all second node sets in the response path; calculating the number of intersection elements between the first node sets and the second node sets, and calculating the number of union elements between the first node sets and the second node sets; dividing the number of intersection elements by the number of union elements to calculate the Jacarbe similarity index, and determining the Jacarbe similarity index as the consistency score.

[0103] It should be noted that the preset consistency threshold is determined based on objective analysis of the Receiver Operating Characteristic (ROC) curve. In this embodiment, past maintenance records are collected, and similarity samples of actual fault propagation paths confirmed by manual investigation are marked as positive, while similarity samples of false alarm paths caused by occasional environmental disturbances are marked as negative. ROC curves are then plotted. Subsequently, the similarity value corresponding to the maximization of the Youden index is selected as the preset consistency threshold. When the consistency score is greater than the preset consistency threshold, the deduced high-probability propagation pattern is confirmed to match the actually observed abnormal physical diffusion process, and the corresponding abnormal source is extracted and identified as the fault origin.

[0104] In another implementation, after identifying the source of the anomaly with a consistency score higher than a preset consistency threshold as the starting point of the fault, the method further includes:

[0105] Extract the linkage relationships of nodes downstream of the fault origin in the reaction path;

[0106] The state parameters of the fault initiation point are smoothed to generate a reference fluctuation sequence. The reference fluctuation sequence is then mapped and extrapolated to the downstream nodes along the node linkage relationship using a preset adjustment vector to obtain the predicted trajectory of each node.

[0107] Extract the propagation branches in the predicted trajectory whose parameter deviation is greater than a preset linkage threshold, merge the propagation branches with the reaction path to perform topology reconstruction to generate a reconstruction network, and output the reconstruction network as the tracing chain.

[0108] In one implementation, the state parameters of the fault initiation point are smoothed to generate a reference fluctuation sequence. The original sensing time series collected within a predetermined number of consecutive time windows (e.g., 3 to 5) from the fault initiation point are extracted as the state parameters. A simple moving average filtering algorithm is used, where the length of the moving window is objectively determined based on the cutoff frequency requirement of the low-pass filter. This window is set to effectively filter out high-frequency electromagnetic noise or sensor glitches higher than the system control frequency. The arithmetic mean of the state parameter values ​​within this window is calculated, and the original high-frequency jitter values ​​are replaced with this arithmetic mean in a time-sliding order to generate the reference fluctuation sequence.

[0109] It should be noted that in this embodiment, a preset adjustment vector is used to map and extrapolate the benchmark fluctuation sequence along the node linkage relationship to the downstream nodes, thereby obtaining the predicted trajectory of each node. The preset adjustment vector is a transfer coefficient matrix determined based on ordinary least squares regression fitting. Matrix multiplication is performed between the matrix of the benchmark fluctuation sequence and the preset adjustment vector to calculate the theoretical extrapolated value of the node at the endpoint of each directed edge within a future time window, generating the predicted trajectory.

[0110] In one implementation, propagation branches with parameter deviations greater than a preset linkage threshold are extracted from the predicted trajectory, the propagation branches are merged with the reaction path to perform topology reconstruction to generate a reconstruction network, and the reconstruction network is output as the tracing chain.

[0111] It should be noted that in this embodiment, the absolute difference between the predicted values ​​of each item in the predicted trajectory and the normal operating condition baseline is calculated, and the absolute difference is divided by the sum of the normal operating condition baseline value and a preset small smoothing constant (e.g., set to 10 to the power of -5) to calculate the parameter deviation. The introduction of this extremely small constant is intended to prevent the electric fireplace from undergoing division-by-zero operations under extreme physical conditions such as when the baseline value is zero, such as when it is turned off or in standby mode.

[0112] It is worth noting that the preset linkage threshold is determined based on the three sigma criterion of Gaussian distribution. The parameter deviation of each inference branch is compared with the preset linkage threshold, and directed connections exceeding the threshold are retained as propagation branches. A union operation in graph theory is performed, and the vertices and directed edges of the newly extracted propagation branches are topologically merged with the vertices and directed edges of the original reaction paths to generate the structurally complete reconstructed network. The graph structure data of the reconstructed network is then output as the tracing chain.

[0113] In summary, this invention acquires multi-dimensional sensor data and calculates the time-delay correlation between response sequence sets. By combining a dynamic time warping algorithm to construct a differential directed graph containing time-series constraints, it achieves precise quantification of the dynamic coupling laws of multiple physical fields such as electrical energy, heat, and light efficiency inside an electric fireplace. This effectively overcomes the shortcomings of traditional single-point static thresholds, which cannot capture parameter time-delay effects. By introducing a pre-constructed spatiotemporal causal inference model and a Bayesian network graph, and utilizing posterior probability inference based on the chain rule and graph theory topology analysis, it directly identifies the initial anomaly source with zero in-degree. This realizes a paradigm shift from superficial feature alarms to deep causal tracing, accurately stripping away the complex downstream chain reaction interference. Furthermore, by combining the state parameters of the fault starting point for time-series mapping inference and topology reconstruction to output a complete tracing chain, it not only clearly restores the spatial propagation of the fault but also endows the system with the ability to dynamically predict the subsequent deterioration trend of the equipment. This enables high-precision fault self-diagnosis and root cause localization of electric fireplaces in complex and variable operating environments, successfully upgrading the traditional passive post-fault troubleshooting to proactive fault prediction and health management (PHM), and comprehensively ensuring the long-term operational stability and safety of the equipment.

[0114] Reference Figure 2 The second embodiment of the present invention provides a self-diagnostic system for electric fireplace faults, comprising:

[0115] The data processing module is used to acquire multidimensional sensing data and extract features from the multidimensional sensing data to obtain a response sequence set.

[0116] The feature difference module is used to perform cluster analysis based on the response sequence set to obtain a set of difference features;

[0117] The offset detection module is used to input the differential feature set into a pre-built causal model to obtain an offset sequence, and perform deviation analysis based on the offset sequence to obtain potential abnormal links;

[0118] The graph construction module is used to acquire historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set with the historical records according to the offset triggering order to obtain a state matrix, and construct a difference directed graph based on the state matrix.

[0119] The source localization module is used to extract the directed edges in the differential directed graph whose weights exceed a preset weight threshold, splice them to obtain the reaction path, calculate the in-degree of the nodes in the reaction path to obtain the in-degree value, and locate the nodes with an in-degree value of zero as the source of the anomaly.

[0120] The traceability output module is used to extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the response path to obtain a consistency score, and determine the anomaly source with the consistency score higher than a preset consistency threshold as the fault starting point.

[0121] It should be noted that the electric fireplace fault self-diagnosis system provided in this embodiment of the invention is used to execute all the process steps of the electric fireplace fault self-diagnosis method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0122] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0123] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A self-diagnostic method for electric fireplace faults, characterized in that, include: Acquire multidimensional sensing data, and extract features from the multidimensional sensing data to obtain a response sequence set; Cluster analysis is performed on the response sequence set to obtain a set of differential features; The differential feature set is input into a pre-built causal model to obtain an offset sequence. Based on the offset sequence, deviation analysis is performed to obtain potential abnormal links. Obtain historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set with the historical records according to the offset triggering order to obtain a state matrix, and construct a difference directed graph based on the state matrix; The directed edges with weights exceeding a preset weight threshold in the differential directed graph are extracted and spliced ​​together to obtain the reaction path. The in-degree of the nodes in the reaction path is calculated to obtain the in-degree value. The nodes with an in-degree value of zero are located as the source of the anomaly. Extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the response path to obtain a consistency score, and determine the anomaly source with the consistency score higher than a preset consistency threshold as the fault starting point.

2. The self-diagnosis method for electric fireplace faults according to claim 1, characterized in that, The step of extracting features from the multidimensional sensing data to obtain a response sequence set includes: The multidimensional sensing data is aligned using timestamps to obtain fused data; the multidimensional sensing data includes temperature data, current data, and voltage data. The features of the fused data are extracted to obtain an initial feature set, and the initial feature set is completed using interpolation to obtain a complete feature set; A state mapping matrix is ​​obtained by performing time-series analysis on the complete feature set; The response sequence set is obtained by calculating the state switching delay time based on the state mapping matrix.

3. The self-diagnosis method for electric fireplace faults according to claim 1, characterized in that, The step of performing cluster analysis based on the response sequence set to obtain a set of differential features includes: The response sequence set is dimensionless to obtain a standardized sequence set. The correlation coefficient between each pair of sequences in the standardized sequence set is calculated using the cross-correlation function to obtain the intensity value. The sequence pairs whose intensity values ​​exceed a preset association threshold are extracted and aggregated to obtain an interaction set; The time delay difference of each sequence pair in the interaction set is calculated as the time delay correlation value, and an extended interaction matrix is ​​constructed based on the time delay correlation value; Clustering algorithms are used to group sequences with similar time-delay correlation values ​​in the extended interaction matrix to obtain differential response clusters; Extract the deviation vectors of the differential response clusters, and merge the deviation vectors to obtain the differential feature set.

4. The self-diagnosis method for electric fireplace faults according to claim 1, characterized in that, The step of performing deviation analysis based on the offset sequence to obtain potential abnormal links includes: The offset sequence is analyzed to obtain the delay time, the delay time is compared with a preset reference delay interval to obtain the excess portion, and the ratio of the excess portion to the upper limit of the reference delay interval is calculated to obtain the deviation. The deviation is compared with a preset deviation threshold, and parameters with values ​​greater than the preset deviation threshold are extracted to construct a potential fault set. The parameters in the potential fault set are then identified as potential abnormal links.

5. A self-diagnostic method for electric fireplace faults according to claim 1, characterized in that, The step of obtaining the offset triggering order by comparing the historical records with the potential abnormal links includes: The historical records are analyzed to extract the delayed feature segments of the potential abnormal links, and the delayed feature segments are matched with preset historical abnormal segments to obtain matching segments; The optimal alignment path is calculated by comparing the matching segment with the response sequence set using a dynamic time warping algorithm. The time mapping relationship in the optimal alignment path is analyzed, the time offset of the response sequence set relative to the matching segment is extracted, and the offset triggering order is obtained by sorting the time offsets.

6. The self-diagnosis method for electric fireplace faults according to claim 1, characterized in that, The step of fusing the response sequence set and the historical records according to the offset triggering order to obtain a state matrix, and constructing a differential directed graph based on the state matrix, includes: Extract the response sequence set and the historical record within the time window corresponding to the offset triggering order, and arrange the response sequence set and the historical record according to the parameter type to obtain the state matrix; The deviation between real-time data and historical data in the state matrix is ​​calculated. The parameter nodes are used as graph vertices, the temporal correlation directions between parameters are used as directed edges, and the deviation values ​​are used as directed edge weights to construct a difference directed graph.

7. The self-diagnosis method for electric fireplace faults according to claim 1, characterized in that, The process of extracting directed edges from the differential directed graph whose weights exceed a preset weight threshold and concatenating them to obtain a reaction path, calculating the in-degree of nodes in the reaction path to obtain in-degree values, and locating nodes with in-degree values ​​of zero as anomaly sources includes: Traverse all directed edges in the differential directed graph, filter out edges with weights lower than a preset weight threshold, and extract the remaining directed edges as high-risk propagation links. The high-risk transmission links are spliced ​​together end to end in chronological order to obtain the response path; The in-degree value is obtained by counting the number of preceding associated edges of each node in the reaction path, and the starting node with an in-degree value of zero is extracted and marked as the source of the anomaly.

8. A self-diagnostic method for electric fireplace faults according to claim 1, characterized in that, The process of inputting the state vector into a pre-constructed Bayesian network to deduce a probabilistic path, comparing the probabilistic path with the response path to obtain a consistency score, and identifying anomaly sources with consistency scores higher than a preset consistency threshold as fault initiation points includes: The state vector is used as the evidence node input to the pre-constructed Bayesian network graph to perform probability inference and update the node state, thereby obtaining the posterior probability distribution of each associated node. Calculate the conditional dependency probability of each node in the posterior probability distribution, and extract the branch with the highest cumulative probability to obtain the probability path; The consistency score is obtained by calculating the node overlap rate between the probability path and the reaction path, and the abnormal source that exceeds the preset consistency threshold is extracted and determined as the fault starting point.

9. A self-diagnostic method for electric fireplace faults according to claim 1, characterized in that, After identifying the anomaly source with a consistency score higher than a preset consistency threshold as the fault initiation point, the method further includes: Extract the linkage relationship of downstream nodes located at the fault starting point in the reaction path; The state parameters of the fault initiation point are smoothed to generate a reference fluctuation sequence. The reference fluctuation sequence is then mapped and extrapolated to the downstream nodes along the node linkage relationship using a preset adjustment vector to obtain the predicted trajectory of each node. Extract the propagation branches in the predicted trajectory whose parameter deviation is greater than a preset linkage threshold, merge the propagation branches with the reaction path to perform topology reconstruction to generate a reconstruction network, and output the reconstruction network as a tracing chain.

10. A self-diagnostic system for electric fireplace faults, characterized in that, include: The data processing module is used to acquire multidimensional sensing data and extract features from the multidimensional sensing data to obtain a response sequence set. The feature difference module is used to perform cluster analysis based on the response sequence set to obtain a set of difference features; The offset detection module is used to input the differential feature set into a pre-built causal model to obtain an offset sequence, and perform deviation analysis based on the offset sequence to obtain potential abnormal links; The graph construction module is used to acquire historical records, compare the delay between the historical records and the potential abnormal links to obtain the offset triggering order, fuse the response sequence set with the historical records according to the offset triggering order to obtain a state matrix, and construct a difference directed graph based on the state matrix. The source localization module is used to extract the directed edges in the differential directed graph whose weights exceed a preset weight threshold, splice them to obtain the reaction path, calculate the in-degree of the nodes in the reaction path to obtain the in-degree value, and locate the nodes with an in-degree value of zero as the source of the anomaly. The traceability output module is used to extract the state vector corresponding to the anomaly source, input the state vector into a pre-constructed Bayesian network to deduce the probability path, compare the probability path with the response path to obtain a consistency score, and determine the anomaly source with the consistency score higher than a preset consistency threshold as the fault starting point.

Images