Home appliance cooperative operation data analysis method, electronic device, and medium

Abstract

The present disclosure relates to the technical field of smart home and data analysis, and particularly relates to a home electromechanical equipment collaborative operation data analysis method, an electronic device and a storage medium. The method is applied to a distributed system including local and remote nodes, and the method comprises: preprocessing, time window division and feature extraction on multi-source heterogeneous time series data, constructing a feature node set, initializing a local causal graph by using a remote general template in a cold start stage, updating the template based on new data through a federated learning mode, constructing a local individualized causal graph including time lag and intensity when data is sufficient, and supporting incremental update, performing abnormal propagation analysis based on the local causal graph, locating a fault root node, finally performing trend prediction on the root node by multi-model fusion, calculating the remaining available time and giving an early warning. The present disclosure effectively solves the problems of cold start, data heterogeneity, privacy protection and fault root location in collaborative analysis of home electromechanical equipment.

Description

Technical Field

[0001] This disclosure relates to the fields of smart home and data analysis technology, and in particular to a data analysis method, electronic device and storage medium for the coordinated operation of home electromechanical equipment. Background Technology

[0002] With the widespread adoption of IoT and AI technologies, modern homes are seeing an increasing number of smart electromechanical devices deployed, such as central air conditioning, smart washing machines, water heaters, and refrigerators. Over long-term operation, the performance of these devices gradually degrades, and they may malfunction. Collaborative analysis of device operation data to achieve early fault warnings and root cause identification is crucial for improving user experience and reducing maintenance costs.

[0003] However, existing technologies generally face the following challenges when implementing collaborative analysis of home electromechanical devices: The inconsistent data sampling rates and out-of-sync timestamps across different devices, along with missing data and outliers, make it difficult to directly apply traditional causal analysis models, resulting in significant biases in the analysis results.

[0004] Newly connected devices lack historical operational data, making it impossible to immediately build personalized causal models and conduct fault analysis; a long data accumulation period is usually required.

[0005] Centralizing all household data to the cloud to train a global model poses a serious risk of user privacy breaches; while models trained using only local data have insufficient generalization ability due to limited data volume.

[0006] Traditional fault diagnosis often relies on thresholds or single signal analysis, making it difficult to accurately locate the root cause of a fault in complex causal networks, and even more difficult to predict the remaining available time of the equipment from its current state to the occurrence of a fault.

[0007] Therefore, there is an urgent need for a data analysis method for the collaborative operation of home electromechanical equipment that can protect data privacy, effectively utilize cross-device knowledge, and achieve accurate fault root cause location and remaining life prediction. Summary of the Invention

[0008] The purpose of this disclosure is to provide a data analysis method, electronic device and medium for collaborative operation of home electromechanical equipment, in order to solve the technical problems of unreliable causal analysis due to low quality of multi-source heterogeneous time-series data, cold start of new equipment access, contradiction between data privacy and global model sharing, and difficulty in fault root cause location and remaining life prediction in complex causal networks.

[0009] To achieve the above objectives, in a first aspect, this disclosure provides a data analysis method for the collaborative operation of home electromechanical equipment, applied to a distributed system including local computing nodes and remote computing nodes, wherein the remote computing nodes store a pre-trained general causal graph template, comprising the following steps: S1: Data Acquisition and Preprocessing For each data dimension, time-series operational data corresponding to each of the data dimensions of multiple electromechanical devices are acquired; the time-series operational data corresponding to the data dimension of each electromechanical device are preprocessed to obtain a time-series dataset for each data dimension of the electromechanical device. The preprocessing includes at least one of the following: time synchronization processing, resampling processing, missing value imputation processing, and outlier replacement processing. The time-series dataset includes data from each acquisition point under the data dimension of the electromechanical device and the labels of the data. The labels include a global time index, the device identifier of the electromechanical device, and the data dimension identifier; the time-series dataset of the data dimension of each electromechanical device is divided according to a preset fixed time window length to obtain a time-series data subset corresponding to each of the multiple time windows of the data dimension of each electromechanical device; for each electromechanical device, the feature vector of the time-series data subset corresponding to each of the multiple time windows of the data dimension is determined and normalized. The feature vector includes feature values ​​of multiple feature types of the data dimension. Each feature type of each data dimension of each electromechanical device is treated as a feature node to form a feature node set. Each feature node includes the device identifier, data dimension identifier, and feature type identifier of the electromechanical device. Furthermore, the feature values ​​of the same feature node under all time windows constitute the observation sequence of the feature node. S2: Local Causal Graph Construction and Iterative Update If the number of time windows does not reach the second preset threshold, whenever the number of new time windows reaches the first preset threshold, based on the time series data subset of the new time windows, the gradient of the structural information and parameter information of any two feature nodes in the feature node set is determined. The structural information is used to characterize the existence and direction of the causal edge between any two feature nodes in the feature node set, and the parameter information is used to characterize the causal strength parameter and time delay parameter of each existing causal edge in the local causal graph. Differential privacy noise is added to the gradient of the parameter information, and the gradient of the parameter information after adding noise and the structural information are uploaded to the remote computing node. The remote computing node receives gradients from structural and parameter information from multiple local computing nodes, updates the general causal graph template, and pushes the updated general causal graph template to the local computing node. The local computing node receives the updated general causal graph template and uses it to replace the initial local causal graph. The initial local causal graph is generated by matching the feature nodes in the feature node set with the feature nodes in the general causal graph template that matches the multiple electromechanical equipment models obtained from the remote computing node according to a preset matching rule, and based on the matching results. When the number of time windows reaches the second preset threshold, based on the observation sequence of each feature node in the feature node set under all time windows, the feature nodes in each data dimension are filtered according to the preset dimensionality reduction conditions to obtain a low-dimensional feature node set. Based on the observation sequence of each feature node in the low-dimensional feature node set under all time windows, determine the structural information and parameter information of any two feature nodes in the low-dimensional feature node set, construct a local personalized causal graph with time delay order label and causal strength label according to the structural information and the parameter information, and update the local personalized causal graph to the local causal graph. When the number of time windows is greater than the second preset threshold and whenever the number of newly added time windows accumulates to the first preset threshold, based on the time series data subset corresponding to the newly added time window, the gradient of the structure information and parameter information of any two feature nodes in the feature node set is determined. Based on the gradient of the structure information and the gradient of the parameter information, the local causal graph is incrementally updated to obtain the updated local causal graph. Differential privacy noise is added to the gradient of the parameter information, and the gradient of the parameter information with added noise and the structure information are uploaded to the remote computing node. S3: Anomaly Propagation Analysis and Fault Root Cause Location For each feature node, a health baseline is constructed. The health baseline includes the mean and standard deviation of the feature values ​​of the feature node across all time windows. Based on the feature values ​​of the feature node in the current time window and the health baseline, the degree of deviation is determined, and based on the degree of deviation, it is determined whether the feature node is an abnormal node, so as to obtain a set of abnormal nodes. For each abnormal node in the set of abnormal nodes, traverse along the reverse directed edges of the local cause-effect graph starting from the abnormal node to determine the root cause node of the fault. Output a set of fault root cause nodes, which includes fault root cause nodes; S4: Degradation Trend Prediction and Remaining Available Time Calculation For each fault root cause node in the set of fault root cause nodes, perform the following steps: Based on the observation sequence of the root cause node under all time windows, the observation sequence is filtered, differentially processed, and sequence spliced ​​to obtain an enhanced feature sequence. The enhanced feature sequence is predicted using a long short-term memory network and an autoregressive integral moving average model to obtain a first predicted value and a second predicted value. The first predicted value and the second predicted value are weighted and fused according to a preset weight to obtain the feature prediction value of the root cause node in the next time window. The feature prediction value of the next time window is used as historical data, and the above prediction and fusion steps are repeated until the feature prediction value sequence of the root cause node in multiple future time windows is obtained. Based on the health baseline of the fault root cause node and the preset threshold coefficient, a dynamic fault threshold is set; Based on the time when the dynamic fault threshold is first reached in the feature prediction value sequence, the remaining available time of the fault root cause node is determined. When the remaining available time is lower than the preset warning time threshold, a warning message is generated and pushed.

[0010] In a second aspect, this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method described in the first aspect.

[0011] Thirdly, this disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0012] Compared with the prior art, the beneficial effects of this disclosure are: By employing time synchronization, adaptive resampling, segmented missing value imputation, and outlier handling based on operating conditions, the quality of multi-source heterogeneous time-series data has been significantly improved, providing a reliable foundation for subsequent causal analysis.

[0013] Through the federated learning mechanism, new devices can immediately obtain general causal templates from the cloud, enabling preliminary analysis capabilities with zero historical data, thus solving the cold start problem of traditional methods.

[0014] By keeping the raw data locally and only uploading anonymized parameter gradients with added differential privacy noise, user privacy is protected while high-performance, globally applicable templates can be trained using massive amounts of cross-device data through federated aggregation.

[0015] By constructing a directed causal graph with time delay order and causal strength, it is possible to accurately depict physical processes where "cause comes first and effect comes later".

[0016] By employing a fusion model of LSTM and ARIMA to predict the trends of key fault root cause characteristics and combining it with dynamic fault thresholds to calculate the remaining available time, a leap from "fault alarm" to "predictive maintenance" has been achieved. Attached Figure Description

[0017] Figure 1 This is the main flowchart of a data analysis method for collaborative operation of home electromechanical equipment provided in this disclosure; Figure 2 A detailed flowchart of the local causal graph construction and incremental update provided in this disclosure; Figure 3 A schematic diagram illustrating the determination of causal edges between two feature nodes based on conditional independence tests provided in this disclosure; Figure 4 This is a schematic diagram of the path backtracking for locating the root cause node of the fault provided in this disclosure. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by this disclosure to achieve the intended purpose of the invention, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, describes a method for analyzing collaborative operation data of home electromechanical equipment, an electronic device, and a medium according to this disclosure. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used in this disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] It should be noted that all actions involving the acquisition of signals, information, or data in this disclosure are carried out in compliance with data protection laws and policies and with the authorization of the owner.

[0021] The present disclosure will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] The method described in this disclosure can be applied to distributed systems including at least one local computing node and at least one remote computing node. It should be noted that the specific deployment forms of the local and remote computing nodes described below, such as user-side smart home gateways and cloud servers, are merely preferred examples for ease of understanding and do not constitute any limitation on the application scenarios of this disclosure. Those skilled in the art will understand that the functions of the local and remote computing nodes can be flexibly deployed or integrated according to actual computing power requirements, network environment, and privacy protection requirements.

[0023] The local computing node is responsible for the primary tasks of data acquisition, processing, storage, and local computing. In a specific example of this embodiment, the local computing node can be deployed on the user side, such as, but not limited to, a home smart gateway, an edge server, a user terminal device such as a smartphone, tablet, or a local server with computing capabilities. This node has a built-in or accessible fault knowledge base for storing raw operating data, intermediate feature data, local causal graph models, historical warning records, and fault diagnosis-related knowledge from all devices. The local computing node ensures that raw data is processed and analyzed locally, uploading anonymized parameters with privacy protection only when necessary.

[0024] The remote computing node is a node located on the network side, responsible for global model aggregation, storage, and iterative updates. In a specific example of this embodiment, the remote computing node can be deployed in the cloud, such as a cloud server or a central server, to store and iteratively update the universal cause-effect graph template for various device models. This node receives anonymized update parameters from multiple local computing nodes, performs federated aggregation and global model optimization, and pushes the updated universal template back to each local computing node.

[0025] Local and remote computing nodes exchange data via an encrypted communication protocol. The exchange is limited to anonymized parameter updates, which are differentially privacy-enhanced before transmission. Raw runtime data, feature data, and locally generated personalized causal graphs all remain within the local computing node.

[0026] It should be noted that the role division between local and remote nodes in the aforementioned distributed computing system is not a mandatory requirement at the physical hardware level. In an alternative, computationally powerful environment, such as a high-performance home server or industrial control computer, all the functions of the local and remote computing nodes can be integrated into the same computing device. In this case, operations such as "upload" and "push" should be understood as data transfer between different functional modules or different storage areas within the same device. In other words, the method disclosed herein does not depend on specific physical hardware divisions or network topologies; it can be implemented as long as data processing, storage, and data interaction capabilities are available.

[0027] Please refer to Figure 1 The method provided in this embodiment mainly includes the following four stages: S1 data acquisition and preprocessing, S2 local cause-effect graph construction and iterative update, S3 anomaly propagation analysis and fault root cause location, and S4 degradation trend prediction and remaining available time calculation. Each stage is described in detail below.

[0028] S1 Data Acquisition and Preprocessing This step aims to address the technical limitations of existing technologies where multi-source time-series data cannot be directly used for causal analysis due to issues such as inconsistent sampling rates, time asynchrony, missing data, and outliers. Through the following sub-steps, the raw runtime data is processed into a time-aligned, complete, and anomaly-free set of feature nodes, providing input for subsequent causal discovery.

[0029] S1.1 Data Acquisition For each data dimension, the time-series operational data corresponding to each of the multiple electromechanical devices is acquired. For example, in this embodiment, the multiple electromechanical devices include: a central air conditioner, a smart washing machine, a smart water heater, and a refrigerator. The collected data dimensions cover various physical quantities and state quantities, such as: current, voltage, power, vibration acceleration, vibration frequency, device surface temperature, inlet water temperature, outlet water temperature, start / stop status, operating mode, running time, and number of start / stop cycles.

[0030] Each device collects data independently according to its own original sampling rate, and each data entry has a local timestamp of the device, a unique device identifier, and a data dimension identifier.

[0031] S1.2 Data Preprocessing Because local clocks of different devices deviate and raw sampling rates vary, and network fluctuations or occasional sensor malfunctions may lead to missing data or outliers, direct use would severely affect the accuracy of subsequent causal analysis. Therefore, the time-series operational data corresponding to the data dimension of each electromechanical device are preprocessed to obtain a time-series dataset for the data dimension of each electromechanical device. The preprocessing includes at least one of the following: time synchronization processing, resampling processing, missing value imputation processing, and outlier replacement processing.

[0032] Time synchronization processing includes: aligning all time-series running data with the time axis based on the global timestamp of the local computing node, assigning a unique global time index to each data item, and eliminating time deviations caused by independent sampling by devices.

[0033] Resampling processing includes: setting a uniform common sampling rate. For example, in this embodiment, the value is 5Hz, corresponding to a sampling interval. Resampling is performed on the time-aligned timing data: For the original sampling rate is higher than The time-series data is first subjected to anti-aliasing filtering, and then downsampled at equal intervals to make the sampling rate equal to the set sampling rate. To avoid spectral aliasing, the anti-aliasing filtering process specifically involves using an 8th-order Butterworth low-pass filter to filter the time-series data, with the filter cutoff frequency set to half of the common sampling rate. This filtering process is used to suppress spurious frequency components generated by spectral aliasing during downsampling of high-frequency components, ensuring that the downsampled time-series data can truly reflect the low-frequency characteristics of the original data.

[0034] For the original sampling rate is lower than The time-series running data is used to complete the sampling points using cubic spline interpolation, so that the interpolated sampling rate is equal to... This ensures that the interpolated data is smooth and without abrupt changes. Cubic spline interpolation constructs piecewise smooth cubic polynomial curves to ensure that the function values, first derivatives, and second derivatives are continuous at the original data points, thus generating smooth interpolation results without abrupt changes.

[0035] By first performing anti-aliasing filtering on downsampling to suppress high-frequency components before decimation, spurious frequencies are avoided. For upsampling, cubic spline interpolation is used to ensure the continuity of the first and second derivatives, avoiding piecewise linear effects. This resampling method preserves the physical characteristics of the original signal, providing a high-quality, unified time-series data foundation for subsequent causal analysis.

[0036] Missing value imputation processing includes: for consecutive missing points in time-series running data corresponding to the same data dimension, different imputation strategies are adopted according to the missing length. Short-term missing points: When the number of consecutive missing points does not exceed 10 sampling points, linear interpolation is used to fill in the missing points.

[0037] Long-term missing points: When the number of consecutive missing points exceeds 10 sampling points, cubic spline interpolation is used to complete the missing points.

[0038] Outlier replacement processing includes: based on the collected start / stop status data, classifying each sampling point into a steady-state operation phase or a non-stationary transient phase. The specific determination rules are as follows: The start / stop status is a global operational flag for the entire device. In this embodiment, a value of 0 indicates stop, and 1 indicates start. For each sampling point, the device's start / stop status value and the time of the most recent state transition are recorded. When the device is in the start state, if the time interval between the current sampling point and the most recent state transition exceeds a preset transition duration (for example, in this embodiment, the transition duration is set to 1 second, corresponding to 5 sampling points), all data within that sampling point and its preceding and following consecutive intervals are determined to be steady-state operation data; otherwise, they are determined to be non-stationary transient data during the start / stop process. It should be noted that once the device is determined to be in a steady-state operation phase within a certain time period, all sampling points corresponding to all data dimensions of the device within that time period are considered steady-state operation data and are uniformly processed using the following 3σ criterion; similarly, all data dimensions in the non-stationary transient phase are uniformly processed using the sliding window extreme value method.

[0039] For data from the steady-state operation phase, 3 The criteria identify and handle outliers. Specifically, this includes calculating the mean of the time series data for each data dimension. with standard deviation : in, This represents the total number of sampling points in this dimension. For the first The values ​​of each sampling point.

[0040] If sampling points satisfy If an outlier is identified, it is determined to be an outlier. The identified outlier is then replaced with the average of the five valid sampling points before and after the outlier to ensure the continuity of the time series dataset.

[0041] For non-stationary transient data, the sliding window extreme value method is used for anomaly detection: within each sliding window, the difference between the maximum and minimum values ​​of the data within the window is calculated. For example, in this embodiment, the window length is set to 10 sampling points and the sliding step size is 5 sampling points. Only outliers exceeding 3 times the difference are removed, while normal transient impact characteristics are retained to avoid the loss of core fault characteristics.

[0042] After the above processing, a time-series dataset with a unified time axis, consistent sampling intervals, complete data, and no outliers is obtained. The time-series dataset includes data from each collection point under the data dimension of the electromechanical equipment and the labels of the data. The labels include a global time index, the equipment identifier of the electromechanical equipment, and the data dimension identifier.

[0043] S1.3 Time Window Division and Feature Extraction To ensure that the data within each time window reflects the continuous state changes of the equipment within a single operating cycle, and to avoid feature statistical distortion caused by the time window spanning multiple start-stop processes, and for intermittent start-stop equipment, to ensure that each time window falls completely within a single operating cycle or shutdown interval, and to avoid cross-cycle mixed sampling.

[0044] The time-series dataset of each of the electromechanical devices' data dimensions is divided according to a preset fixed time window length, resulting in multiple time-series data subsets corresponding to each time window of each of the electromechanical devices' data dimensions. The length of the fixed time window should be less than the minimum single operating cycle duration of all intermittent start-stop devices. For example, in this embodiment, the fixed time window length is set to 10 minutes, and the sliding step is 5 minutes. This length is less than the minimum single operating cycle duration of all intermittent start-stop devices, thereby ensuring that each time window falls completely within a single operating cycle or shutdown interval. The sliding step of the time window is set to half the window length to achieve overlapping sliding segmentation, thereby enhancing the temporal continuity of the features. After division, the following is obtained: There are 1 time window, and each time window corresponds to a continuous subset of time series data.

[0045] For each of the aforementioned electromechanical devices, feature vectors of time-series data subsets corresponding to multiple time windows of the data dimension are determined and normalized. The feature vectors include feature values ​​of multiple feature types of the data dimension. The feature types of the feature vector include: Time-domain statistical characteristics, such as mean, variance, peak value, peak-to-peak value, kurtosis, skewness, margin factor, and impulse factor, are used to reflect the central tendency and dispersion of a signal.

[0046] Frequency domain features: fundamental frequency amplitude, harmonic proportion, centroid frequency, frequency variance, etc., are extracted based on FFT spectrum and used to characterize the periodic fluctuation components of the signal.

[0047] Time-frequency characteristics: For example, the normalized energy of 16 frequency band nodes after 4-layer db4 wavelet packet decomposition is used to simultaneously characterize the non-stationary characteristics in the time and frequency domains.

[0048] Specific features: defined according to the physical attributes of the data dimensions, such as the peak current ripple and voltage distortion rate of electrical quantities, the resonance peak amplitude of mechanical vibration, and the heating rate of temperature quantities.

[0049] Event characteristics: These include the percentage of runtime within a time window and the number of starts and stops, used to reflect changes in the operating status of the equipment.

[0050] The normalization process includes: for each feature type's feature value sequence, subtracting its mean and dividing by the standard deviation, so that the mean of each feature after normalization is 0 and the standard deviation is 1.

[0051] S1.4 Feature Node Set Construction Each feature type of each data dimension of each electromechanical device is treated as a feature node to form a feature node set. Each of the feature nodes includes an equipment identifier, a data dimension identifier, and a feature type identifier for the electromechanical equipment. Furthermore, the feature values ​​of the same feature node under all time windows constitute the observation sequence of the feature node.

[0052] Through the above data acquisition and preprocessing process, the problems of time alignment, missing anomaly handling, and unit unification of multi-source heterogeneous time series data were effectively solved. The output of a set of feature nodes and their observation sequences with clear structure and reliable quality that can be directly used for causal analysis laid the data foundation for subsequent cross-device causal structure discovery and avoided causal graph bias and fault misjudgment caused by data quality issues.

[0053] S2 Local Causal Graph Construction and Iterative Update Please refer to Figure 2 This step aims to address the "cold start" problem when new user devices are connected, where insufficient local data prevents causal mining. It also enables the gradual evolution of the local causal graph from an initial template to a personalized model. This step is completed collaboratively by local and remote computing nodes, ensuring that the original data never leaves the local machine.

[0054] S2.1 Cold Start Initialization When a new user device connects for the first time or when there is no historical data locally, the local computing node first obtains a generic cause-effect graph template matching the local device model from the remote computing node. This template is pre-trained by the remote computing node based on a large amount of anonymized operational data from devices of the same model and includes two types of parameters: Structural information: The existence and direction of causal edges between each feature node, stored in the form of an adjacency matrix.

[0055] Parameter information: causal strength and time delay order of each causal edge.

[0056] Set of feature nodes The feature nodes in the graph are matched with the feature nodes in the general cause-effect graph template that matches the models of the multiple electromechanical devices obtained from the remote computing node according to the preset matching rules, and an initial local cause-effect graph is generated based on the matching results.

[0057] The preset matching rules include: Exact match: For each causal edge in the general causal graph template, if the feature nodes at both ends of the causal edge are in the feature node set... If the causal edge exists, it is determined to be a complete match, and the causal edge and its corresponding causal strength parameter and time delay parameter are retained; Local missing nodes: For the set of feature nodes that exist in the general cause-effect graph template. For feature nodes that do not exist, delete all causal edges involving those feature nodes; Locally unique nodes: for the set of feature nodes If a feature node exists in the general causal graph template but not in the template, the feature node is treated as an isolated node without causal edges and has no causal edges with other feature nodes.

[0058] By employing a three-level matching rule—retaining edges that match completely, deleting edges corresponding to nodes that exist in the template but not locally, and adding nodes that exist locally but not in the template as isolated nodes—the initial causal graph structure is ensured to be valid and free of invalid edges. Simultaneously, it preserves unique local features for subsequent incremental learning, avoiding manual reconfiguration.

[0059] S2.2 Incremental Updates and Federated Learning Local Computation As the device continues to run, the local computing nodes accumulate a continuous effective time window. A first preset threshold is set. The number of new time windows required for each local incremental update is, for example, taken in this embodiment. One time window. Second preset threshold. : The minimum cumulative time window required to trigger the first full-scale personalized causal graph reconstruction. For example, in this embodiment, we take . A time window.

[0060] Local computing nodes record the total number of currently accumulated valid time windows in real time. .according to Different update strategies are implemented at different stages.

[0061] When the total number of effective time windows Less than the second preset threshold At this time, local data is insufficient to support the construction of stable and reliable personalized causal graphs. Therefore, whenever the number of new time windows reaches... At that time, the local compute node performs the following operations based on this newly added time window subset of time-series data: Determine the gradient of the structural and parameter information of any two feature nodes in the feature node set. The structural information is used to characterize the existence and direction of the causal edge between any two feature nodes in the feature node set, that is, for each pair of feature nodes... Determine whether, within the newly added time window subset of time-series data, there exists a... arrive Or from arrive The causal edges and their directions are defined. The parameter information is used to characterize the causal strength parameter and time delay parameter of each existing causal edge in the local causal graph. The gradient of the parameter information includes the gradient of the causal strength parameter and the gradient of the time delay parameter, which are calculated as follows: for each existing causal edge in the local causal graph, a mean squared error loss function with L2 regularization is constructed based on the newly added time window data, and the partial derivatives of the loss function with respect to the causal strength parameter and the time delay parameter are calculated. The gradient of the integer time delay parameter is approximated using the conventional finite difference method in this field.

[0062] Please refer to Figure 3 Determining the structural information of any two feature nodes includes: For each feature node, the maximum time delay order is determined based on the ratio of the duration of a single complete operation cycle of the electromechanical equipment to which the feature node belongs to the duration of a single time window. The calculation formula is: in, This refers to the duration of a single complete operating cycle for the corresponding equipment. The duration of a single time window, for example, in this embodiment , This is for rounding up. The upper limit is set to 10 to avoid excessive computational load due to excessively large time delay dimensions.

[0063] For each time window Based on the maximum time delay order A time-delay variable set corresponding to the time window is constructed for the feature node. The time-delay variable set includes the feature values ​​of the feature node within the time window. and the time window closest to the aforementioned time window The feature values ​​of each time window .

[0064] For time delay order Feature nodes in time delay order The eigenvalues ​​of all time windows are arranged in chronological order to form a time-delay variable sequence. The insufficient part of the index is filled with 0.

[0065] For any two feature nodes in the feature node set and ,Will Each time-delay variable sequence With another feature node A time-delay variable sequence with a time delay order of 0 Form a pair to be tested The sequence of time-delayed variables, excluding the pair to be tested, i.e., the sequence of time-delayed variables for all other feature nodes, and the feature nodes. and The sequence of other time-delayed variables besides the pair to be tested is used as the condition set. For each pair to be tested, the independence of the execution conditions is tested.

[0066] This embodiment employs a dual-mode conditional independence test: the basic mode combines linear partial correlation coefficient testing with Fisher's Z-transform to detect linear dependencies; the supplementary mode uses kernel partial correlation coefficient testing to capture non-linear dependencies. The specific testing methods are as follows: Linear partial correlation coefficient test: For any two feature nodes in the feature node set and and each time delay order , feature nodes Time-delay variable sequence With feature nodes A variable sequence with a time delay of 0 The pair to be tested is formed. The condition set is composed of all time-delay variable sequences other than the pair to be tested. In order to eliminate condition sets Each time-delay variable sequence pair and The linear effect of the two can be accurately calculated in a given context. The partial correlation coefficients under the values ​​of the variables to be tested are respectively... and As the dependent variable, with the condition set Two ridge regression models are established using all time-delayed variable sequences as independent variables.

[0067] For dependent variable The regression model is: in, : Length is numerical sequences, The number of effective time windows represents the feature nodes. In time delay order The observation sequence.

[0068] : condition set The first in Each of the variables It is a length of A numerical sequence.

[0069] : The number of variables in the condition set.

[0070] : Intercept term, representing the condition when all condition variables are 0. The expected value.

[0071] : No. The regression coefficients of the condition variables, This means that, assuming other condition variables remain unchanged, For every unit change, the dependent variable... The expected change in value.

[0072] : Residual vector, length is ,represent The condition set cannot be used. The part explained by linear combinations.

[0073] To solve for the regression coefficients, ridge regression minimizes the loss function with L2 regularization: in, : Sample number, indicating the first One effective time window.

[0074] : No. In each sample The value of .

[0075] : No. In the nth sample The values ​​that a condition variable can take.

[0076] : Regularization hyperparameter, which controls the strength of the L2 penalty. For example, in this embodiment, we take... This was determined through 5-fold cross-validation.

[0077] The solution to the above optimization problem is: Then the dependent variable The predicted value is: The residual is: Residual sequence Indicates from Middle Removal Condition Set The remaining information after the linear effect.

[0078] For dependent variable Using the exact same ridge regression model: in, A numerical sequence of length N, where N is the number of effective time windows, representing feature nodes. The observation sequence at time delay order 0 is the current window observation sequence.

[0079] : Intercept term.

[0080] : No. The regression coefficients of the condition variables.

[0081] : Residual vector.

[0082] Ridge regression provides coefficient estimates. , The predicted value is: The residual is: , the residual sequence Indicates from Middle Removal Condition Set The remaining information after the linear effect.

[0083] Partial correlation coefficient For the two residual sequences mentioned above and The Pearson correlation coefficient between them is calculated using the following formula: in: Total number of samples, i.e., the number of effective time windows.

[0084] : No. The residual sequence of each sample value, : No. The residual sequence of each sample value.

[0085] : Residual sequence The mean, : Residual sequence The mean.

[0086] Partial correlation coefficient The range of values ​​is The larger its absolute value, the stronger the elimination condition set. After the impact, and The stronger the linear dependency between them.

[0087] To perform a significance test on the partial correlation coefficient, it needs to be converted into a statistic that approximately follows a standard normal distribution. The Fisher Z-transform is used, with the following formula: in, : Natural logarithm function, when hour, Theoretically, it tends towards infinity; in practical calculations, it can be... Perform minor trimming, such as limiting it to ±0.9999.

[0088] Ratio transformation, used to transform... The range of values ​​(-1, 1) is mapped to (0, +∞).

[0089] Construct the test statistic The formula is: in, For condition set The number of variables in the condition set, i.e., the dimension of the condition set.

[0090] The degrees of freedom are used to correct for the influence of sample size and condition set size on the distribution of the statistic.

[0091] For each time delay order Perform the linear partial correlation coefficient test described above, and obtain the corresponding test statistics. A total of Each test corresponds to a pair to be tested. .

[0092] Set significance level The corresponding two-sided critical value for the standard normal distribution is 1.96. If there exists a certain... Make If a significant linear dependency exists between two feature nodes, then we consider that such a dependency exists. Maximum time delay order This serves as the time delay parameter for the causal edge.

[0093] At this point, the existence of a causal edge between the two feature nodes is marked as existing. For two feature nodes whose causal edge is determined to exist through the conditional independence test... and The direction of the causal edge needs to be further determined. In this embodiment, the preset rules for determining the direction of the causal edge include: When the delay order The direction is determined chronologically. Since the cause must occur before the effect, the direction points from the past to the present. exist The state before the time window has an impact. Given the current state, the direction of the causal edge is... .

[0094] When the delay order Since the chronological order cannot be used to distinguish directions, the causal strength in both directions must be calculated separately: by For the reason, As a result, ridge regression was used to obtain the standardized regression system. The absolute value is taken as the causal strength in that direction. .

[0095] by For the reason, Similarly, the strength of causality can be derived from the result. .

[0096] Compare the sizes of the two: If The direction is Otherwise .

[0097] If the causal strength in two directions is equal or the difference is less than a preset threshold, for example, in this embodiment the threshold is 5%, then the causal edge will not be constructed for the time being, and will be determined after more data is collected.

[0098] By comparing the causal strength of two directions and taking the one with greater strength as the direction, this rule is repeatable and can still give a reasonable causal direction even if the sampling frequency is insufficient to distinguish the order, thus filling the technical gap in direction determination in the case of zero time delay.

[0099] The direction of a causal edge is determined according to a preset causal edge direction judgment rule. The feature node at the starting end of the causal edge direction is determined as the cause feature node, and the feature node at the ending end of the direction is determined as the result feature node. The time delay order corresponding to the causal edge is then determined. This serves as the time delay parameter for the causal edge.

[0100] If for all of If the linearity test indicates conditional independence, a kernel partial correlation coefficient test is required.

[0101] Kernel partial correlation coefficient test: For each time delay order ,like and If the linearity test determines that the conditions are independent, then calculate... and If the normalized mutual information (NMI) between them is greater than a preset threshold, for example, in this embodiment the preset threshold is 0.1, then a kernel partial correlation coefficient test needs to be performed to capture possible nonlinear dependencies.

[0102] First, construct the condition set. The joint kernel matrix. Condition set. include There are n time-delay variable sequences, each sequence has The feature values ​​under each time window. Concatenate the feature values ​​of all condition variables under each time window into a vector, obtaining... There are n vectors, each with dimensions of 1. Regarding this The similarity between each pair of vectors is calculated using a radial basis function (RBF) kernel, resulting in a... joint matrix The elements in this matrix reflect the similarity of the overall state of the condition set under the corresponding two time windows.

[0103] After obtaining the joint kernel matrix, it is centered by subtracting the row mean and column mean and adding the overall mean, resulting in a centered kernel matrix. It satisfies the property of zero mean in the feature space, thus eliminating data bias in high-dimensional space.

[0104] With the centralized kernel matrix As independent variables, respectively with and The observed sequence is used as the dependent variable, and kernel ridge regression is performed. Kernel ridge regression learns a nonlinear mapping function by minimizing a loss function with an L2 regularization term, so that the condition set can predict the dependent variable. After regression, for Obtain the predicted value sequence Calculate the residual ,for Obtain the predicted value sequence Calculate the residual These two residual sequences represent the results from... and Middle Removal Condition Set The remaining information after the nonlinear effects.

[0105] After obtaining two residual sequences, they are first centered to obtain centered residual sequences. and Then, the kernel partial correlation coefficient between the two centered residual sequences is calculated. It is obtained by calculating the inner product of the two and dividing by the product of their respective moduli. The value ranges from [0,1]. A larger value indicates that the influence of the condition set has been removed. and The stronger the non-linear dependency between them.

[0106] Since the theoretical distribution of the kernel partial correlation coefficient is unknown, direct parameter testing is not possible. Therefore, the permutation test (PT) is used to evaluate it. The statistical significance. Specifically: maintaining and condition set The order remains unchanged, The observed sequences are randomly rearranged, and the entire process of constructing the joint kernel matrix, centering, kernel ridge regression, and calculating the kernel partial correlation coefficient is repeated. Next, for example, this embodiment takes ,get The rearranged partial correlation coefficient At the same time, it will be based on the unrearranged The partial nuclear correlation coefficient calculated using the exact same steps is denoted as... Then calculate The proportion of these rearrangement coefficients that are greater than or equal to them is denoted as : This represents the probability of observing the current or stronger dependency under the null hypothesis of conditional independence.

[0107] Set significance level .like If the null hypothesis is not met, then the null hypothesis is rejected, and a significant nonlinear dependency is determined to exist between the two time-delayed variable sequences. The null hypothesis is... and Given a set of conditions Conditional independence is assumed. In this case, a causal edge is considered to exist between the two feature nodes. The existence of this causal edge is marked as present. The direction of the causal edge is determined according to a preset causal edge direction judgment rule. The time delay order corresponding to the causal edge is then assigned. The existence and direction of each causal edge are used as the time delay parameter of the causal edge, and are taken as the structural information.

[0108] By explicitly constructing a multi-order time-delay variable sequence for each feature node and independently performing conditional independence tests on each time delay, it is possible to accurately capture time-delayed causal edges where "cause comes first and effect comes later." When the time delay is zero, the direction is determined by comparing the bidirectional causal strength, thus avoiding the shortcomings of traditional methods that cannot handle synchronous causality based solely on time sequence.

[0109] For each pair of feature nodes whose causal edges have been identified and whose directions have been determined through the above steps... To further eliminate spurious correlations caused by common environmental factors or power supply fluctuations, a second test was conducted: First, environmental temperature and humidity feature nodes and power supply parameter feature nodes are obtained. In this embodiment, the environmental temperature and humidity feature nodes include: ambient temperature and ambient humidity; the power supply parameter feature nodes include: RMS value of power supply voltage, RMS value of power supply current, power supply frequency, etc. The observation sequences of the above feature nodes under all time windows are then extracted.

[0110] Secondly, identify the feature nodes at both ends of the causal edge. and Other feature nodes that have significant dependencies on each other, i.e., based on the current local causal graph, find all nodes that simultaneously point to... and The feature nodes. The sequence of time-delay variables with a time delay order of 0 for these common cause feature nodes. Add the condition set. If no such common cause node exists, this part of the condition set is empty. Then, construct the complete condition set. This includes: a sequence of time-delay variables whose time-delay order is 0 for all the aforementioned common cause feature nodes. The observation sequences of environmental temperature and humidity characteristic nodes and the observation sequences of power supply parameter characteristic nodes.

[0111] With condition set ,right A time-delay variable sequence with a time delay order of 0 and A time-delay variable sequence with a time delay order of 0 Independence test of execution conditions. This embodiment uses the same linear partial correlation coefficient test and kernel partial correlation coefficient test as described above.

[0112] If the test result is and Given a set of conditions If the following conditions are independent, then the causal edge is determined. If the correlation is deemed spurious, the causal edge is removed from the structural information and is no longer retained. Conversely, if the test result rejects the conditional independence, the causal edge is retained, and subsequent parameter information calculations continue.

[0113] After initially determining the direction of the causal edge, environmental temperature and humidity, power supply parameters, and common causes pointing to both ends of the node are specifically introduced as a condition set. The conditional independence test is then performed again, which effectively eliminates spurious relationships caused by external interference and significantly improves the authenticity of the causal graph. This is a post-verification mechanism that existing offline causal analysis tools do not have.

[0114] Determining the parameter information of any two feature nodes includes: For two feature nodes that are determined to have a causal edge through the conditional independence test. and Furthermore, the causal edge direction and time delay parameters have been determined. Further calculation of the causal strength parameter of this causal edge is needed to construct complete parameter information.

[0115] Obtain the cause feature node The time delay parameter of the causal edge The corresponding time-delay variable sequence Result Feature Nodes A time-delay variable sequence with a time delay order of 0 and result feature nodes Except for the current cause feature node Other than that and with causal edges pointing to Other feature nodes A time-delay variable sequence with a time delay order of 0 . For feature nodes excluding the current cause Other than that and with causal edges pointing to The number of other feature nodes. If no such other feature nodes exist, the set is empty.

[0116] In order to eliminate right The impact, thus accurately estimating right The direct causal strength is determined by using regression methods to identify causal characteristic nodes. For the result feature nodes The causal effect value is used as the absolute value of the causal effect value as the causal strength parameter of the causal edge, and the time delay parameter of the cause feature node in the causal edge is used as the time delay parameter of the causal edge to generate the parameter information.

[0117] In this embodiment, the same ridge regression method as that used in the aforementioned linear partial correlation coefficient test is employed to construct the following structural equation model: The above model will Represented as intercept term, ,each and residuals A linear combination of these terms. The regression coefficients are solved by minimizing the loss function with L2 regularization: in, The regularization hyperparameters were determined using 5-fold cross-validation. The estimated values ​​of each regression coefficient were then obtained. .

[0118] Focus only ,because This reflects the effect after controlling for the influence of other causal nodes. right The direct causal effect. The absolute value is taken as the causal strength parameter. The time delay parameters are directly adopted from the predetermined ones. Therefore, the parameter information of each causal edge is determined by the causal strength parameter. and time delay parameters Together they constitute.

[0119] For each existing causal edge in the local causal graph, a mean squared error loss function with L2 regularization is constructed based on the newly added time window data. The partial derivatives of the loss function with respect to the causal strength parameter and the time delay parameter are calculated to obtain the gradient of the parameter information. The gradient of the integer time delay parameter is approximated using the conventional finite difference method in this field.

[0120] The local computing node adds differential privacy noise to the gradient of the parameter information, and uploads the gradient of the parameter information with noise and the structure information to the remote computing node.

[0121] The remote computing node receives gradients from structural and parameter information from multiple local computing nodes, updates the general causal graph template, and pushes the updated general causal graph template to the local computing node. The local computing node receives the updated general causal graph template and uses it to replace the initial local causal graph.

[0122] By incorporating all other causal nodes of the outcome node as control variables into the regression model before calculating the causal strength, the confounding causal effect value is obtained, making the causal strength parameter more interpretable and comparable. At the same time, the method of calculating the gradient based on the newly added time window is clarified, which allows the parameters to be updated with incremental data without retraining the full model in each round, greatly reducing the local computational burden.

[0123] The stage of building a personalized cause-effect graph: When the total number of local cumulative effective time windows The second preset threshold is reached for the first time. At this point, the local compute node believes it has sufficient data to support the construction of a personalized causal graph. Instead of directly replacing the federated template, it performs an initial full-scale personalized causal graph reconstruction based on all local historical data. This includes the following sub-steps: Based on the observation sequence of each feature node in the feature node set across all time windows, feature nodes within each data dimension are filtered according to preset dimensionality reduction conditions to obtain a low-dimensional feature node set. This embodiment employs the following two dimensionality reduction conditions: Variance screening: Calculate the variance of the observation sequence for each feature node. If the variance is lower than a preset variance threshold, for example, 0.01 in this embodiment, it indicates that the feature hardly changes during equipment operation and contributes very little to fault diagnosis, so it is discarded.

[0124] Mutual information filtering: For the remaining feature nodes, calculate the normalized mutual information between the observation sequence of each node and the observation sequences of all other nodes. If the mutual information of a node with all other nodes is less than a preset mutual information threshold (for example, 0.1 in this embodiment), it indicates that the node has no significant correlation with other variables and is therefore removed.

[0125] By actively performing variance and mutual information filtering, we eliminate uninformative features with minimal variation and isolated features that are unrelated to all other nodes. This significantly reduces the dimensionality of variables, shortens computation time, and improves the readability and robustness of the causal graph while ensuring effective information.

[0126] Based on the observation sequence of each feature node in the low-dimensional feature node set under all time windows, the structural information and parameter information of any two feature nodes in the low-dimensional feature node set are determined. Based on the structural information and the parameter information, a local personalized causal graph with time delay order label and causal strength label is constructed. The construction of this causal graph is entirely based on local data and does not depend on remote templates.

[0127] Set the local personalized causal graph constructed above as the current local causal graph, replacing the initial local causal graph previously generated by the federated template.

[0128] When the total number of local cumulative effective time windows Greater than the second preset threshold And whenever the cumulative number of newly added time windows reaches the first preset threshold In this case, the local compute node no longer rebuilds the complete personalized causal graph. Instead, it incrementally updates the current local causal graph based on the time-series data subset corresponding to the newly added time window. The specific incremental update process is as follows: Based on the newly added time window The gradients of the structural and parameter information of any two feature nodes in the feature node set are determined using the corresponding time-series data subset.

[0129] The existence and direction of causal edges between each pair of feature nodes obtained based on the newly added data subset are compared with the corresponding structure in the current local causal graph. If the newly added data supports a causal edge that does not exist in the current graph, the existence of that edge is marked as to be added; if the newly added data does not support an existing causal edge in the current graph, that edge is marked as to be deleted; if the direction changes, it is marked as to be reversed. The above differences are considered as gradients of structural information.

[0130] For each causal edge retained in the current local causal graph, a mean squared error loss function with L2 regularization is constructed based on the newly added time window data. The partial derivatives of the loss function with respect to the causal strength parameter and the time delay parameter are calculated to obtain the causal strength gradient. and time delay parameter gradient .

[0131] Based on the gradient described above, perform the following update operation on the current local causal graph: For causal edges marked as "to be added", add the causal edge directly to the current local causal graph, along with the time delay parameter and causal strength parameter calculated based on the new data.

[0132] For causal edges marked as to be deleted: Remove the causal edge and its parameter information from the local causal graph.

[0133] For causal edges marked as to be reversed: first delete the edges in the original direction, then add a new causal edge in the opposite direction, and take the time delay parameter and causal strength parameter as the values ​​calculated from the new data.

[0134] For the retained causal edges, incremental updates are performed using the gradient of the parameter information. The parameter update employs gradient descent; the parameter gradient calculated based on the new data indicates the direction and magnitude of the current parameter adjustment. This is achieved by using the learning rate... The step size is updated along the gradient direction, allowing the updated parameters to better fit the causal dependencies reflected by the newly added data. The formula is: : The causal strength parameter and time delay parameter of the updated causal edge.

[0135] The causal strength parameter and time delay parameter of this causal edge in the current local causal graph.

[0136] Based on new additions The causal intensity gradient and time delay parameter gradient are calculated from data within a time window.

[0137] The learning rate controls the step size for each update. For example, in this embodiment, we take... .

[0138] The updated parameters must meet physical constraints: Causality strength parameter If the calculation result is negative, then take 0; time delay parameter. Rounding is required and the limit is 1 / 3. Within the scope. As the final parameter of the updated local causal graph.

[0139] The local computing node adds differential privacy noise to the gradient of the above parameter information, and uploads the gradient of the parameter information with added noise and the structural information corresponding to this incremental update to the remote computing node, so that the remote computing node can update the general causal graph template.

[0140] S3 Anomaly Propagation Analysis and Fault Root Cause Location Please refer to Figure 4 This step aims to monitor the running status of each feature node in real time based on the constructed local personalized cause-effect graph, identify abnormal nodes, and trace the root cause of the fault along the reverse cause-effect path, so as to provide accurate fault location results for subsequent predictive maintenance.

[0141] S3.1 Health Baseline Construction and Abnormal Node Identification For each feature node Using this node across all time windows, a total of Observation sequence under a time window Construct a health baseline for the feature nodes. The health baseline is used to characterize the statistical distribution characteristics of each feature node under normal operating conditions. In this embodiment, the health baseline includes the mean and standard deviation of the feature values ​​of the feature nodes across all time windows, calculated using the following formula: in, For feature nodes Data center trends under normal operating conditions For feature nodes Data fluctuation range under normal operating conditions The total number of effective time windows used to construct a health baseline.

[0142] Based on the feature values ​​of the current time window of the feature node Determine the degree of deviation from the healthy baseline. This embodiment uses a standardized deviation score, as shown in the following formula: in, For example, in this embodiment, the value is a very small positive number. This is used to avoid division-by-zero errors when the standard deviation is zero. An outlier threshold is set. ,Right now Criteria. If Then determine the feature nodes under the current time window. These are abnormal nodes. All abnormal nodes are aggregated to construct an abnormal node set, defined as follows: .

[0143] S3.2 Fault Root Cause Location Get the set of abnormal nodes Then, for each abnormal node in the set of abnormal nodes, a traversal is performed from the abnormal node along the reverse directed edges of the local causal graph. Specifically, the local causal graph is a directed graph, where each directed edge... Represents feature nodes go through After a time window delay, This creates a causal impact. For each anomalous node, the system traverses the directed edges pointing in the opposite direction to that node in the local causal graph, repeatedly tracing back to nodes without incoming edges. These nodes without incoming edges are the candidate root cause nodes for the anomalous node. After performing the above tracing on all abnormal nodes, the resulting candidate root cause nodes are merged and deduplicated to form the final set of fault root cause nodes. .

[0144] S3.3 Outputs the set of root cause nodes of the fault. Output the set of root cause nodes of the fault Each root node corresponds to a specific device, data dimension, and feature type, such as {"Central Air Conditioner - Current - Fundamental Frequency Amplitude", "Washing Machine - Vibration - Kurtosis"}.

[0145] S4 Degradation Trend Prediction and Remaining Available Time Calculation For each root cause node of the fault Execute the following forecasting and early warning process.

[0146] S4.1 Enhanced Feature Sequence Construction Obtain the root cause node of the fault Observation sequences under all time windows The observed sequence is filtered, differentially processed, and concatenated to obtain an enhanced feature sequence. Specifically: A low-pass filter is used to remove high-frequency noise while preserving low-frequency degradation trends. This embodiment uses a second-order Butterworth low-pass filter with a cutoff frequency set to 0.1 times the sampling frequency. The resulting sequence is obtained after filtering. .

[0147] Perform first-order differencing on the filtered sequence to eliminate non-stationary trends: .

[0148] Original sequence Filter sequence Sum of difference sequences Concatenate columns to form an enhanced feature sequence. Each time window corresponds to a three-dimensional feature vector. This enhanced feature sequence retains both the original absolute value information and the trend change rate information, which helps to improve prediction accuracy.

[0149] S4.2 Multi-model Prediction and Weighted Fusion The enhanced feature sequence was predicted using Long Short-Term Memory (LSTM) network and Autoregressive Integrated Moving Average (ARIMA) model, respectively, to obtain the first and second predicted values.

[0150] LSTM prediction: Train an LSTM model using time window data of enhanced feature sequences. Input the enhanced feature vector of the last time window and output the first predicted value of the next time window. .

[0151] ARIMA prediction: Since ARIMA is a univariate model, the filtered sequence from the enhanced feature sequence is used. As input, an ARIMA model is built, and the model identifies and fits the data to output the second predicted value for the next time window. .

[0152] The first predicted value Second predicted value The feature prediction values ​​for the next time window of the fault root cause node are obtained by weighted fusion according to preset weights. in, For example, in this embodiment, the default weight configuration is as follows: .

[0153] S4.3 Multi-step prediction and fault threshold setting Using the predicted feature value of the next time window as historical data, the above prediction and fusion steps are repeated until a sequence of predicted feature values ​​for multiple future time windows of the fault root cause node is obtained. Specifically, an autoregressive strategy is adopted: Append to the end of the enhanced feature sequence, repeat LSTM and ARIMA prediction and weighted fusion until the future is obtained. Predicted value sequence for each time window This embodiment sets It provides predictions of equipment operating status for the next 5 hours, covering the fault warning cycle for typical household electromechanical equipment.

[0154] Based on the root cause node health baseline parameters , and preset threshold coefficients Set dynamic fault thresholds Dynamic fault thresholds come in two forms: Upper threshold: Applicable when a larger characteristic value indicates more severe equipment degradation, such as vibration amplitude, temperature, and current. In this case, the dynamic fault threshold is defined as: When feature prediction value Greater than or equal to When this happens, the equipment is determined to be in a fault state.

[0155] Lower threshold: Applicable when smaller characteristic values ​​indicate more severe equipment degradation, such as efficiency, health index, and pressure. In this case, the dynamic fault threshold is defined as: When feature prediction value Less than or equal to When this happens, the equipment is determined to be in a fault state.

[0156] S4.4 Remaining Available Time Calculation and Early Warning The remaining available time of the fault root cause node is determined based on the time when the dynamic fault threshold is first reached in the feature prediction value sequence.

[0157] For feature nodes using an upper threshold, in the predicted value sequence In the middle, find the first one that satisfies The time window index is .

[0158] For feature nodes using a lower threshold, in the predicted value sequence In the middle, find the first one that satisfies The time window index is .

[0159] If none of the predicted values ​​trigger the fault threshold, the equipment is deemed to have no fault risk within the prediction period, and the remaining available time is marked as exceeding the prediction period.

[0160] The formula for calculating the remaining available time is: in, The time window sliding step size is set to 5 minutes in this embodiment.

[0161] When the remaining available time is lower than a preset warning time threshold, a warning message is generated and pushed. This embodiment presets a warning threshold. Minutes, when At that time, the local computing node automatically generates standardized fault warning information, including: the device's unique identifier, details of the fault's root cause node, and precise remaining available time. The warning information can be pushed locally to the home control screen and the user's mobile app, or synchronized to a remote operation and maintenance platform as needed, enabling early warning and proactive maintenance of equipment faults.

[0162] The above description is merely a preferred embodiment of this disclosure and is not intended to limit the scope of protection of this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A data analysis method for the collaborative operation of home electromechanical equipment, applied to a distributed system including local computing nodes and remote computing nodes, wherein the remote computing nodes store pre-trained general causal graph templates, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing For each data dimension, obtain the time-series operation data corresponding to each of the data dimensions of multiple electromechanical devices; The time-series operational data corresponding to the data dimension of each electromechanical device are preprocessed to obtain a time-series dataset for the data dimension of each electromechanical device. The preprocessing includes at least one of the following: time synchronization processing, resampling processing, missing value imputation processing, and outlier replacement processing. The time-series dataset includes data from each collection point under the data dimension of the electromechanical device and the labels of the data. The labels include a global time index, the device identifier of the electromechanical device, and the data dimension identifier. The time-series dataset of the data dimension of each electromechanical device is divided according to a preset fixed time window length to obtain a time-series data subset corresponding to each time window of the data dimension of each electromechanical device. For each of the aforementioned electromechanical devices, feature vectors of time-series data subsets corresponding to multiple time windows of the data dimension are determined and normalized. The feature vectors include feature values ​​of multiple feature types of the data dimension. Each feature type of each data dimension of each electromechanical device is treated as a feature node to form a feature node set. Each feature node includes the device identifier, data dimension identifier, and feature type identifier of the electromechanical device. Furthermore, the feature values ​​of the same feature node under all time windows constitute the observation sequence of the feature node. S2: Local Causal Graph Construction and Iterative Update If the number of time windows does not reach the second preset threshold, whenever the number of new time windows reaches the first preset threshold, based on the time series data subset of the new time windows, the gradient of the structural information and parameter information of any two feature nodes in the feature node set is determined. The structural information is used to characterize the existence and direction of the causal edge between any two feature nodes in the feature node set, and the parameter information is used to characterize the causal strength parameter and time delay parameter of each existing causal edge in the local causal graph. Differential privacy noise is added to the gradient of the parameter information, and the gradient of the parameter information after adding noise and the structural information are uploaded to the remote computing node. The remote computing node receives gradients from structural and parameter information from multiple local computing nodes, updates the general causal graph template, and pushes the updated general causal graph template to the local computing node. The local computing node receives the updated general causal graph template and uses it to replace the initial local causal graph. The initial local causal graph is generated by matching the feature nodes in the feature node set with the feature nodes in the general causal graph template that matches the multiple electromechanical equipment models obtained from the remote computing node according to a preset matching rule, and based on the matching results. When the number of time windows reaches the second preset threshold, based on the observation sequence of each feature node in the feature node set under all time windows, the feature nodes in each data dimension are filtered according to the preset dimensionality reduction conditions to obtain a low-dimensional feature node set. Based on the observation sequence of each feature node in the low-dimensional feature node set under all time windows, determine the structural information and parameter information of any two feature nodes in the low-dimensional feature node set, construct a local personalized causal graph with time delay order label and causal strength label according to the structural information and the parameter information, and update the local personalized causal graph to the local causal graph. When the number of time windows is greater than the second preset threshold and whenever the number of newly added time windows accumulates to the first preset threshold, based on the time series data subset corresponding to the newly added time window, the gradient of the structure information and parameter information of any two feature nodes in the feature node set is determined. Based on the gradient of the structure information and the gradient of the parameter information, the local causal graph is incrementally updated to obtain the updated local causal graph. Differential privacy noise is added to the gradient of the parameter information, and the gradient of the parameter information with added noise and the structure information are uploaded to the remote computing node. S3: Anomaly Propagation Analysis and Fault Root Cause Location For each feature node, a health baseline is constructed. The health baseline includes the mean and standard deviation of the feature values ​​of the feature node across all time windows. Based on the feature values ​​of the feature node in the current time window and the health baseline, the degree of deviation is determined, and based on the degree of deviation, it is determined whether the feature node is an abnormal node, so as to obtain a set of abnormal nodes. For each abnormal node in the set of abnormal nodes, traverse along the reverse directed edges of the local cause-effect graph starting from the abnormal node to determine the root cause node of the fault. Output a set of fault root cause nodes, which includes fault root cause nodes; S4: Degradation Trend Prediction and Remaining Available Time Calculation For each fault root cause node in the set of fault root cause nodes, perform the following steps: Based on the observation sequence of the root cause node under all time windows, the observation sequence is filtered, differentially processed, and sequence spliced ​​to obtain an enhanced feature sequence. The enhanced feature sequence is predicted using a long short-term memory network and an autoregressive integral moving average model to obtain a first predicted value and a second predicted value. The first predicted value and the second predicted value are weighted and fused according to a preset weight to obtain the feature prediction value of the root cause node in the next time window. The feature prediction value of the next time window is used as historical data, and the above prediction and fusion steps are repeated until the feature prediction value sequence of the root cause node in multiple future time windows is obtained. Based on the health baseline of the fault root cause node and the preset threshold coefficient, a dynamic fault threshold is set; Based on the time when the dynamic fault threshold is first reached in the feature prediction value sequence, the remaining available time of the fault root cause node is determined. When the remaining available time is lower than the preset warning time threshold, a warning message is generated and pushed.

2. The method according to claim 1, characterized in that, Determining the structural information of any two feature nodes includes: For each feature node, the maximum time delay order is determined based on the ratio of the single complete operation cycle of the electromechanical equipment to which the feature node belongs to the duration of a single time window. For each time window, based on the maximum time delay order, a time delay variable set corresponding to the time window is constructed for the feature node. The time delay variable set includes the feature value of the feature node in the time window and the feature values ​​of the K time windows closest to the time window, where K is the maximum time delay order. For each time delay order from 0 to K, the feature values ​​of the feature node in all time windows under the time delay order constitute a time delay variable sequence. For any two feature nodes, each time-delay variable sequence of one feature node and the time-delay variable sequence of the other feature node with a time-delay order of 0 are combined to form a test pair. All time-delay variable sequences other than the test pair are used as the condition set, and conditional independence tests are performed on each test pair. If there exists a time delay order such that the conditional independence test result indicates a significant dependency between two time-delayed variable sequences, then the existence of a causal edge between the two feature nodes is marked as existing. The direction of the causal edge is determined according to a preset causal edge direction judgment rule. The feature node at the starting end of the direction of the causal edge is determined as the cause feature node, and the feature node at the ending end of the direction is determined as the result feature node. The time delay order corresponding to the causal edge is used as the time delay parameter of the causal edge. The existence and direction of each causal edge are used as the structural information.

3. The method according to claim 2, characterized in that, The determined parameter information includes: for a known causal edge, obtaining the time-delay variable sequence corresponding to the time-delay parameter of the cause feature node on the causal edge, the time-delay variable sequence of the result feature node with a time-delay order of 0, and the time-delay variable sequence of other feature nodes of the result feature node other than the current cause feature node that have a causal edge pointing to the result feature node with a time-delay order of 0; using a regression method to determine the causal effect value of the cause feature node on the result feature node; using the absolute value of the causal effect value as the causal strength parameter of the causal edge; and using the time-delay parameter of the cause feature node on the causal edge as the time-delay parameter of the causal edge to generate the parameter information; The gradient for determining the parameter information includes: based on the time series data subset corresponding to the newly added time window, determining the partial derivative of the loss function with respect to the parameters for the causal strength parameter and time delay parameter of each causal edge, and obtaining the gradient of the parameter information.

4. The method according to claim 2, characterized in that, After determining the direction of the causal edge, the process also includes: For each causal edge with a determined direction, environmental temperature and humidity feature nodes and power supply parameter feature nodes are obtained. The observation sequences of the two feature nodes under all time windows are extracted respectively. The time delay variable sequences of other feature nodes that have a significant dependency relationship with the feature nodes at both ends of the causal edge, the observation sequences of the environmental temperature and humidity feature nodes, and the observation sequences of the power supply parameter feature nodes are used as a condition set. Conditional independence test is performed on the feature nodes at both ends of the causal edge. If the test result is conditional independence, the causal edge is deleted.

5. The method according to claim 1, characterized in that, The preset matching rules include: For each causal edge in the general causal graph template, if the feature nodes at both ends of the causal edge exist in the feature node set, it is determined to be a complete match, and the causal edge and its corresponding causal strength parameter and time delay parameter are retained. For feature nodes that exist in the general causal graph template but not in the feature node set, delete all causal edges involving the feature node. For feature nodes that exist in the feature node set but not in the general causal graph template, treat the feature node as an isolated node without causal edges.

6. The method according to claim 1, characterized in that, The filtering of feature nodes within each data dimension according to preset dimensionality reduction conditions includes: Determine the variance of the observation sequence for each feature node; if the variance is lower than a preset variance threshold, delete the feature node. Determine the normalized mutual information between the observation sequence of each remaining feature node and the observation sequences of all other feature nodes. If the normalized mutual information is less than a preset mutual information threshold, then delete the feature node.

7. The method according to claim 1, characterized in that, The preset causal edge direction determination rule includes: if the time delay order of the causal edge is greater than 0, then the direction is determined according to the time sequence; If the time delay order of the causal edge is equal to 0, determine the causal intensity in the two directions respectively, and take the direction with the greater causal intensity as the direction of the causal edge.

8. The method according to claim 1, characterized in that, The resampling process includes: For time-series running data with an original sampling rate higher than the preset common sampling rate, anti-aliasing filtering is performed first, followed by equal-interval downsampling to make the downsampled sampling rate equal to the common sampling rate. For time-series running data where the original sampling rate is lower than the common sampling rate, cubic spline interpolation is used to complete the sampling points so that the interpolated sampling rate is equal to the common sampling rate.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the data analysis method for the coordinated operation of home electromechanical equipment as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the data analysis method for the coordinated operation of home electromechanical equipment as described in any one of claims 1 to 8.

Images