Energy station equipment whole life cycle monitoring data integration analysis method and system

Abstract

This application discloses a method and system for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment. It relates to the field of energy station equipment monitoring data processing technology. Through unified encoding and timestamp feature generation of multi-source data, standardized access to lifecycle data is achieved. By employing reliability calculation and robust fusion, time alignment, and operating condition compensation normalization, data conflicts and noise interference are effectively suppressed, and real operating conditions are accurately decoupled, avoiding misjudgments of operating condition fluctuations as equipment anomalies. Based on relationship graphs and soft clustering, sub-condition health baselines are constructed to adaptively match the actual operating status of the equipment, improving the accuracy of health assessment and degradation tracking. Combining health indices, degradation indices, and topology propagation relationships, probabilistic reasoning of fault root causes is achieved, with a complete and interpretable chain of evidence. Through stress-corrected remaining life estimation and risk classification, reliable support is provided for preventative maintenance, lifecycle management, and safety early warning of equipment, significantly improving the intelligent operation and maintenance level and operational reliability of energy stations.

Description

Technical Field

[0001] This application relates to the field of energy station equipment monitoring data processing technology, specifically to a method and system for integrating and analyzing monitoring data throughout the entire life cycle of energy station equipment. Background Technology

[0002] Energy stations typically consist of chillers, boilers / heat pumps, heat exchangers, cooling towers, water pumps, valve assemblies, energy storage devices, and their monitoring and control systems. These devices operate via hydraulic or thermal loop coupling. Throughout their entire lifecycle—from design and selection, manufacturing and installation, commissioning and acceptance, long-term operation, maintenance and overhaul, to decommissioning—these devices continuously generate multi-dimensional monitoring data, operating condition data, and electronic data such as alarm, inspection, and maintenance records. Systematic integration and analysis of this data can be used to achieve equipment health assessments, degradation trend tracking, fault precursor identification, remaining life estimation, and operational risk warnings, thereby supporting the safe, stable, and efficient operation of energy stations.

[0003] The current digital construction of energy stations is mostly characterized by platform fragmentation and fragmented data management: the operation side is mainly based on automatic control subsystems, while the maintenance side is mainly based on work order and ledger systems, resulting in limited data interaction and coupled analysis capabilities. For example, Chinese patent CN113269435B discloses a traditional monitoring system that includes a five-prevention system, a SCADA system, a transformer substation monitoring system, and an SVG reactive power compensation device system, but it is difficult to interact with existing monitoring platforms during production and operation, requiring separate monitoring and making coupled monitoring difficult, thus increasing the workload of monitoring personnel. Similarly, Chinese patent CN117910678A mentions that although a BMS undertakes data collection and monitoring, it is constrained by computing power and storage, and alarms are often limited to threshold judgments of single or combined parameters. Energy efficiency indicators may still require manual calculation and analysis, lacking effective automated analysis, anomaly identification, and deviation diagnosis methods, making it difficult to meet higher-level health management needs.

[0004] Due to the lack of a unified data encoding and reliable fusion mechanism covering the entire lifecycle, existing technologies often struggle to achieve stable alignment and decoupling of operating conditions under conditions of multi-source heterogeneity, varying quality, and inconsistent sampling. This can easily lead to misjudging operating condition fluctuations as equipment degradation or masking actual degradation within changes in operating conditions. Furthermore, the lack of correlation modeling between equipment, components, loop topology, and maintenance events results in difficulty tracing the root causes of failures, incomplete evidence chains, and significant fluctuations and insufficient interpretability in remaining life estimates and risk assessments. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution:

[0006] In a first aspect, embodiments of this application provide a method for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment, including:

[0007] Acquire multi-source data on the lifecycle of energy station equipment, and uniformly encode the multi-source data to obtain a feature vector with a timestamp;

[0008] Calculate the data credibility in the feature vector and complete conflict robust fusion, unified time alignment and working condition compensation normalization to obtain the working condition decoupling monitoring vector and uncertainty at each time point;

[0009] Based on the decoupled monitoring vector and uncertainty of the working condition, a relationship graph is constructed and fused to generate a state vector. Soft clustering is performed to obtain the working condition membership degree and a sub-working condition health baseline model is established.

[0010] The equipment health index and degradation index are calculated by combining the state vector and the health baseline model for different working conditions, and the root cause probability is recursively output by combining the propagation relationship.

[0011] Based on the equipment health index, degradation index, and root cause probability estimate, and stress-corrected remaining life, a risk score is calculated and output in a graded manner.

[0012] In one possible implementation, the step of acquiring multi-source data on the lifecycle of energy station equipment and uniformly encoding the multi-source data to obtain a timestamped feature vector includes:

[0013] Obtain multi-source electronic data records corresponding to the target equipment from heterogeneous data sources at the energy station;

[0014] Each electronic data record is objectified as a lifecycle data tuple. ,in: For device instance identification, For lifecycle stage tags, For timestamps, To monitor the main variable vector, For the operating condition context vector, For data source attribute vectors, For version semantic information, Missing mask;

[0015] Regarding the Perform nonlinear embedding respectively to obtain Embedded vectors are generated for stage labels and source identifiers respectively. Then based on the missing mask Gating function right Perform element-wise gating to form the feature vector. .

[0016] In one possible implementation, the term "to the" Perform nonlinear embedding respectively to obtain Embedded vectors are generated for stage labels and source identifiers respectively. Then based on the missing mask Gating function right Perform element-wise gating to form the feature vector. ,include:

[0017] Monitoring branch embedding: ;

[0018] Working condition branch embedding: ;

[0019] Original attribute branch embedding: ;

[0020] Stage and source embedding: ;

[0021] Missing gating fusion and final stitching: ,in: , , , and For matrix parameters, , , , and For bias vectors, It is a non-linear activation function. and This is an embedding mapping from discrete IDs to vectors. For element-wise multiplication, The output vector of the gated function. This involves concatenating vectors.

[0022] In one possible implementation, the step of calculating the data reliability in the feature vector and completing conflict robust fusion, unified time alignment, and operating condition compensation normalization to obtain the operating condition decoupling monitoring vector and uncertainty at each time point includes:

[0023] Based on the above The integrity, consistency, timeliness, source health, drift penalty, calibration aging, and link quality of each data record in the data are used to construct a quality feature vector. And based on historical statistical parameters Computing credible priors Then set the current quality score Posterior with trusted priors The final credibility weight is obtained by fusion:

[0024]

[0025] in: For completeness, For consistency, For timeliness, For the source of health, As a penalty for drifting, To calibrate the aging process, For link quality, This means transposing a column vector into a row vector. Indicates a linear score for the inner product, outputting a scalar;

[0026] A set of conflict records for the same device and the same indicator within the same time window. Robust fusion is performed by minimizing the weighted robust loss to obtain the fused value. ;

[0027] Map the fused asynchronous multisampling rate metrics to a unified time grid. For each indicator Establish a state-space model Write confidence into the observation noise Spatiotemporal alignment values ​​are obtained through Kalman smoothing. ,in: Indicates time The true potential value, This indicates the observations on the aligned grid. Indicates at a point in time The map nodes corresponding to the target device of interest. This represents the random disturbance noise term. Indicates the basic observation noise scale. Indicates in The higher the confidence level, the better. The smaller, To prevent extremely small positive numbers from being divided by zero, Expressing conditional expectation;

[0028] Based on working condition vector Construction index Conditional expectation Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed by using an exponential function; then... Perform working condition compensation normalization And calculate the uncertainty. ,in: To estimate the standard deviation of the variance, This indicates the intensity of uncertainty.

[0029] In one possible implementation, the set of conflict records for the same device and the same indicator within the same time window Robust fusion is performed by minimizing the weighted robust loss to obtain the fused value. ,include:

[0030]

[0031]

[0032] in: This means finding the option that minimizes the objective function. , Indicates the first The first record Individual indicator values, For scaling estimation, the residuals are made dimensionless. To standardize the residuals, for The absolute value, This represents the Huber threshold, which controls when the penalty changes from a second to a first instance.

[0033] In one possible implementation, the condition-based vector... Construction index Conditional expectation Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed through exponential functions, including:

[0034]

[0035]

[0036] in: and For the desired model parameters, The number of dimensions for the operating condition variables. The number of basis functions used for each operating condition variable. Let h be the basis function. It is an exponential function. and These are the parameters of the variance model.

[0037] In one possible implementation, the step of constructing a relationship graph based on the decoupled monitoring vector and uncertainty of the operating conditions, fusing it to generate a state vector, performing soft clustering to obtain the membership degree of the operating conditions, and establishing a sub-operating condition health baseline model includes:

[0038] Constructing heterogeneous temporal relationship graphs Node set Represents devices, components, measurement points, and event entities, and is an edge set. This indicates the relationships between topological connections, containment relationships, influence relationships, and maintenance functions. Represent the set of relation types for edges, and assign relation attribute vectors to the edges. With time difference characteristics ;

[0039] Graph execution relational attention graph fusion: for any side Construct message:

[0040]

[0041] And through the neighborhood Attention weights are obtained by softmax normalization. Based on this aggregation, we obtain: Then through the gating fusion unit Update node representation ,in: Let r be the linear transformation matrix corresponding to relation type r. For nodes u In the l Layer representation vector, For the first The linear mapping matrix of the layer-side edge feature concatenation terms. Represents the edge attribute vector. This means concatenating the edge attributes with the time difference. For nodes In the The neighborhood aggregation vector of the layer, For the first Layers consist of nodes Passed to the node The message vector, For nodes In the The node representation of a layer is a vector. This indicates gating fusion, which merges the old representation with the newly aggregated information according to a gating ratio;

[0042] The normalized monitoring vector and the uncertainty vector are concatenated as follows: and the graph representation of the target device node. splicing to form a fused state vector ,in: for, Let be the uncertainty vector. To monitor the number of dimensions of the main variables;

[0043] right Perform soft clustering with load condition transition priors using Gaussian emission likelihood. With operating condition transition matrix Calculate posterior membership ;

[0044] In the healthy sample set Up, press Calculate the mean health baseline under different working conditions And calculate the shrinkage covariance:

[0045] ,

[0046] in: The shrinkage coefficient is used to prevent... Irreversible, improves stability Operating condition The weighted sample covariance matrix under the following conditions It is the identity matrix. Indicates working conditions The mean of the health baseline.

[0047] In one possible implementation, the pair Perform soft clustering with load condition transition priors using Gaussian emission likelihood. With operating condition transition matrix Calculate posterior membership ,include:

[0048]

[0049]

[0050]

[0051] in: For at any time Observation status From working conditions The generated likelihood value, The center vector of the working condition cluster, Let be the cluster covariance matrix. The square of the Mahalanobis distance. Represents conditional probability. For a moment Implicit operating condition random variables, This is an abbreviation for the observation sequence, indicating... .

[0052] In one possible implementation, the step of calculating the equipment health index and degradation index by combining the state vector and the sub-condition health baseline model, and recursively outputting the root cause probability by combining the propagation relationship, includes:

[0053] The deviation of time k from condition p is calculated based on the baseline of the sub-conditions:

[0054]

[0055] The degenerate residuals are obtained by weighting them according to their membership degree. ,in: For a moment Relative working conditions The degree of deviation, To inverse the covariance, the deviation is made to account for correlation and scale, where P is the total number of case clusters;

[0056] Based on fusion state vector With the output of the prediction model Calculate reconstruction error and the , Energy efficiency deviation Vibration deviation Thermal deviation and a summary of uncertainties The health index obtained through integration:

[0057]

[0058] And according to Degradation index obtained: ,in: The square of the L2 norm. arrive These are the fusion weights for each piece of evidence;

[0059] For the set of failure modes From the characteristics of evidence Calculate the strength of evidence for the log-likelihood ratio for each failure mode: And activated by a threshold Normalization yields anomalous evidence ,in: For the set of evidence features used for fault identification, This is a fault mode. In failure mode The probability model for the occurrence of the following evidence. ReLU-based threshold activation: activation levels below the threshold are not counted as evidence. For the first Evidence trigger threshold for each failure mode For normalization operators;

[0060] Constructing a fault propagation matrix Fault Prior Integrating historical root cause probabilities in the logarithmic field Current evidence Dissemination items Combined with prior terms, we obtain the unnormalized root cause score:

[0061]

[0062] Then normalize the root cause scores to root cause probabilities. ,in: For a moment The root cause probability is an unnormalized intermediate quantity. For the weight of historical memory, As the current weight of evidence, To take the logarithm of each element of the vector, prevent , Indicates propagation gain. For the fault propagation matrix, This indicates that the evidence is disseminated along the propagation relationship to the relevant root causes. Fault prior probability vector This indicates prior addition.

[0063] In one possible implementation, the step of calculating a risk score and outputting it in a graded manner based on the equipment health index, degradation index, and root cause probability estimate and stress-corrected remaining life includes:

[0064] Establish a stochastic evolution model of the degradation index ,in And according to the stress vector Calculate the degradation drift velocity ,in: For random disturbance noise, The variance of the noise during the degradation process. Basic drift term, Based on the drift term, This indicates that the mean is 0 and the variance is 0. Gaussian distribution;

[0065] right Perform exponential smoothing correction: ,in: For smoothing coefficients, The slope window length, In-window weights For weighted slope estimation within the window, The assignment update operator writes the calculation result on the right side back to the variable on the left side.

[0066] Based on failure threshold With correction Calculate the approximate expected remaining lifespan And calculate the lifetime uncertainty. ,in: and In Avoid division by zero or negative drift that could cause anomalies;

[0067] Will Root cause risk items Lifetime mapping risk items Keyness Factors and Integrated computing comprehensive risk score:

[0068]

[0069] in: Original risk score, , , , and All are fusion weights;

[0070] right Risk score obtained by time smoothing Then according to the threshold Will Quantified into risk levels ,in: For the smoothed risk score, This is the risk time smoothing coefficient;

[0071] The , , , , Remaining lifetime estimation and its uncertainty output.

[0072] Secondly, embodiments of this application provide a system for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment, including:

[0073] The acquisition module is used to acquire multi-source data on the life cycle of energy station equipment and to uniformly encode the multi-source data to obtain a feature vector with a timestamp.

[0074] The fusion alignment module is used to calculate the data credibility in the feature vector and complete conflict robust fusion, unified time alignment and working condition compensation normalization to obtain the working condition decoupling monitoring vector and uncertainty at each time point;

[0075] The working condition modeling module is used to construct a relationship graph based on the decoupled monitoring vector and uncertainty of the working condition, fuse and generate a state vector, perform soft clustering to obtain the working condition membership degree, and establish a sub-working condition health baseline model.

[0076] The evaluation and deduction module is used to calculate the equipment health index and degradation index by combining the state vector and the sub-condition health baseline model, and to recursively output the root cause probability by combining the propagation relationship.

[0077] The risk output module is used to calculate a risk score and output it in a graded manner based on the equipment health index, degradation index and root cause probability estimate and stress-corrected remaining life.

[0078] In this embodiment, multi-source data is uniformly encoded and timestamped to break down data barriers between operation and maintenance systems, enabling standardized access to data throughout the entire lifecycle. Reliability calculation and robust fusion, time alignment, and condition compensation normalization are employed to effectively suppress data conflicts and noise interference, accurately decouple real operating conditions, and avoid misjudging condition fluctuations as equipment anomalies. Based on relationship graphs and soft clustering, sub-condition health baselines are constructed, which can adaptively match the actual operating status of equipment, improving the accuracy of health assessment and degradation tracking. Combining health indices, degradation indices, and topology propagation relationships, probabilistic reasoning of fault root causes is achieved, with a complete and interpretable chain of evidence. Stress-corrected remaining life estimation and risk classification provide reliable support for preventative maintenance, lifespan management, and safety early warning, significantly improving the intelligent operation and maintenance level and operational reliability of energy stations. Attached Figure Description

[0079] Figure 1 A schematic diagram of a method for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment, provided in an embodiment of this application;

[0080] Figure 2 A field diagram of the energy station equipment provided in the embodiments of this application;

[0081] Figure 3 This is a schematic diagram of the health index and degradation index curves provided in the embodiments of this application;

[0082] Figure 4 This is a schematic diagram of an energy station equipment full life cycle monitoring data integration and analysis system provided in an embodiment of this application. Detailed Implementation

[0083] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0084] See Figure 1 The energy station equipment full life cycle monitoring data integration and analysis method provided in this embodiment includes:

[0085] S101, acquire multi-source data on the lifecycle of energy station equipment, and uniformly encode the multi-source data to obtain a feature vector with timestamps.

[0086] like Figure 2 The data presented here is a field diagram of the energy station equipment. Monitoring data throughout the entire lifecycle of the equipment originates from various heterogeneous data sources, including sensors, control systems, maintenance records, and operational monitoring platforms. These data sources exhibit significant differences in format, dimension, and sampling frequency, leading to low data utilization efficiency and large biases in analysis results if used directly. Therefore, this step first acquires and standardizes the multi-source data. Through objectification, nonlinear embedding, gating fusion, and vector concatenation, the heterogeneous multi-source data is transformed into a unified format, timestamped feature vector, laying the foundation for subsequent data analysis. Simultaneously, the specific generation methods for each field of the lifecycle tuple, embedding mapping, and gating are clarified.

[0087] Specifically, this embodiment acquires multi-source electronic data records corresponding to the target equipment from heterogeneous data sources in the energy station, covering various types of data such as equipment operation monitoring, operating condition changes, maintenance information, and data source attributes. Each electronic data record is objectified as a lifecycle data tuple. ,in: For device instance identification, For lifecycle stage tags, For timestamps, To monitor the main variable vector, For the operating condition context vector, For data source attribute vectors, For version semantic information, This is a missing mask.

[0088] The unique device code from the master data is used to assign BMS / SCADA measurement points to device instances via the measurement point → device mapping table. Based on the data source type and business events, the entire lifecycle of equipment can be divided into stages, such as BIM corresponding to the design stage, factory inspection data corresponding to the manufacturing stage, acceptance records corresponding to the installation stage, performance test curves corresponding to the commissioning stage, real-time data acquisition corresponding to the operation stage, work order inspection data corresponding to the maintenance stage, replacement and overhaul records corresponding to the overhaul and replacement stage, and scrap assessment data corresponding to the decommissioning stage.

[0089] Prioritize data generation time, such as sensor sampling time, work order creation time, or work time, and then correct it through channel delay during the time alignment process. It consists of real-time values ​​from measurement points and derived features. Real-time values ​​include raw data such as temperature, pressure, flow rate, power, current, valve position, and frequency. Derived features are features calculated from the raw samples, such as sliding window statistics (mean, variance, slope) and spectral features (peak value, bandwidth energy, kurtosis). It consists of variables that affect the monitored quantities but do not represent equipment degradation. The core includes operating mode, start / stop status, load rate, outdoor temperature and humidity, supply and return water temperature difference, set point, valve opening range, loop pressure difference, etc. The data comes from BMS / SCADA system, operation strategy set point and control command log.

[0090] The metadata of the data source is quantified, specifically including dimensions such as the data source system, communication protocol, gateway, sampling period, point type, measurement range / accuracy, upload link quality, and whether it has been filtered or compressed. Data from site migration versions, calibration versions, ledger versions, protocol versions, model versions, etc., are often saved in the form of discrete IDs or timestamps to provide a basis for subsequent data quality assessment and consistency judgment. According to the monitored main variable vector Each dimension is considered missing or invalid and is marked accordingly, including cases where there is no data, the value is out of bounds, there are anomaly markers, or the quality bit fails.

[0091] Regarding the Perform nonlinear embedding respectively to obtain Embedded vectors are generated for stage labels and source identifiers respectively. Then based on the missing mask Gating function right Perform element-wise gating to form the feature vector. .

[0092] For the monitoring master variable vector, operating condition context vector, and data source attribute vector in the data tuple, nonlinear embedding operations are performed to map vectors with different dimensions and physical meanings to a unified feature space. Simultaneously, for discrete data such as lifecycle stage labels and data source identifiers, corresponding embedding vectors are generated through discrete ID-to-vector embedding mapping, achieving vectorized representation of discrete features. The embedding formulas for each branch in this embodiment are as follows:

[0093] Monitoring branch embedding: ;

[0094] Working condition branch embedding: ;

[0095] Original attribute branch embedding: ;

[0096] Stage and source embedding: ;

[0097] Missing gating fusion and final stitching: ,in: , , , and For matrix parameters, , , , and For bias vectors, It is a non-linear activation function. and This is an embedding mapping from discrete IDs to vectors. For element-wise multiplication, The output vector of the gated function. This involves concatenating vectors.

[0098] In this embodiment, stage embeddings, source embeddings, etc., can be obtained offline through learnable lookup tables. If no training is required, a fixed Higgs code can be used for linear mapping. Matrix parameters can be obtained through offline training, for example, by training the encoder with the goal of "reconstructing or predicting the next time step," or by using normalization and linear transformations in engineering implementation. The implementation rule is to set missing dimensions to 0 and valid dimensions to 1, or use... Smoothing avoids hard zeroing, and can also train a small network to... Mapped to gating weights.

[0099] S102, calculate the data credibility in the feature vector and complete conflict robust fusion, unified time alignment and working condition compensation normalization to obtain the working condition decoupling monitoring vector and uncertainty at each time point.

[0100] Even after processing by S101, the feature vectors still suffer from issues such as data reliability discrepancies, time asynchrony, and operating condition coupling. The same indicator for the same equipment may generate conflicting data from multiple sources within the same time window. Different sampling frequencies from different data sources lead to inconsistent data time dimensions, and equipment monitoring indicators are susceptible to changes in operating conditions, making it difficult to reflect the true operating status. This step eliminates interference from data reliability, time, and operating conditions by constructing a data reliability assessment model, robustly fusing conflicting data, aligning time, and performing operating condition compensation normalization. It generates a decoupled monitoring vector for each time point and simultaneously calculates the data uncertainty, quantifying the reliability of the analysis results. Furthermore, it clarifies the calculation and acquisition methods for each core quality feature, parameter, and model.

[0101] Data credibility is a core indicator for evaluating data validity. This embodiment is based on the aforementioned... The integrity, consistency, timeliness, source health, drift penalty, calibration aging, and link quality of each data record in the data are used to construct a quality feature vector. And based on historical statistical parameters Computing credible priors Then set the current quality score Posterior with trusted priors The final credibility weight is obtained by fusion:

[0102]

[0103] in: Completeness is calculated based on the proportion of missing key fields or key measurement points in the record; For consistency, a comprehensive judgment is made based on physical constraints, consistency between adjacent measuring points, and consistency of historical statistics. Physical constraints include rules for checking range, energy conservation, temperature difference sign, and non-negativity of flow rate. Consistency between adjacent points focuses on the logic of related measuring points such as the temperature difference between supply and return water in the same loop, the pressure difference before and after the pump, and the relationship between flow rate. Statistical consistency is the degree of deviation between data under similar operating conditions and historical distribution. Timeliness is determined by the difference between the data entry time and the sampling time, or the difference between the current time and the sampling time. The greater the data delay, the lower the timeliness score. For source health, there are health indicators from the device itself, such as sensor self-test position, fault code, calibration validity period, measurement point quality position, gateway online status, etc. To penalize drift, a rolling window is used to detect data drift. Methods such as CUSUM, EWMA, drift rate, and long-term offset of the difference from the reference measurement point are employed. The more obvious the drift, the larger the penalty value and the lower the reliability. The calibration aging is mapped based on the number of days since the last calibration or replacement. After the recommended calibration cycle of the equipment is exceeded, the calibration aging score gradually decreases. Link quality is calculated using network statistics, including packet loss rate, retransmission rate, offline time, clock drift, etc. The worse the link quality, the lower the data reliability. This means transposing a column vector into a row vector. This represents the linear score of the inner product and outputs a scalar.

[0104] In this embodiment, The cumulative statistics from historical data points or data sources that were judged as valid or invalid, among which... The number of historical valid counts (determined by quality level, manual verification, consistency with reference instruments, etc.). This represents the number of invalid historical occurrences (out of bounds, drift, fault codes, invalid manual annotations, etc.); during the cold start phase, a uniform default value (such as a=b=1) can be used to avoid bias.

[0105] For a set of conflicting records of the same device and the same metric within the same time window, the fusion value is calculated by minimizing the weighted robustness loss based on the credibility weight of each data record. This avoids interference from single outliers in the fusion result and improves the robustness of the data. Specifically, it is implemented using the Huber loss function, which applies a double penalty when the residual is small and a single penalty when the residual exceeds a threshold, balancing the smoothness of squared loss and the robustness of L1 loss. The selection and calculation process of the set and scale required for conflict fusion are as follows:

[0106] A set of conflict records for the same device and the same indicator within the same time window. Robust fusion is performed by minimizing the weighted robust loss to obtain the fused value. Specifically, it includes:

[0107]

[0108]

[0109] in: This means finding the option that minimizes the objective function. , Indicates the first The first record Individual indicator values, For scaling estimation, the residuals are made dimensionless, and robust scaling estimation within the healthy period or the current window is preferred to suppress extreme outliers. To standardize the residuals, for The absolute value, This represents the Huber threshold, which controls when the penalty changes from a secondary to a primary penalty. It can be set as an empirical constant or calibrated using health data quantiles.

[0110] The fused multi-source data still suffers from inconsistent sampling rates, constituting asynchronous data that cannot be directly used for time-series analysis. This embodiment maps the fused asynchronous multi-sampling rate metrics to a unified time grid. For each indicator Establish a state-space model Write confidence into the observation noise Spatiotemporal alignment values ​​are obtained through Kalman smoothing. .

[0111] in: Indicates time The true potential value, This indicates the observations on the aligned grid. Indicates at a point in time The map nodes corresponding to the target device of interest. Process noise : Estimating the variance of the index using the first difference after alignment with the healthy period; the faster the change... The larger. This represents the random disturbance noise term. . This represents the baseline observation noise scale, derived from sensor specification accuracy, resolution, or healthy period observation - slip residual variance estimation, reflecting the noise level of the sensor itself. Indicates in The higher the confidence level, the better. The smaller, To prevent extremely small positive numbers from being divided by zero, Expressing conditional expectation;

[0112] Based on working condition vector Construction index Conditional expectation Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed using an exponential function. Then... Perform working condition compensation normalization And calculate the uncertainty. ,in: To estimate the standard deviation of the variance, Indicates the intensity of uncertainty. Part of it comes from natural fluctuations in operating conditions Another part comes from the uncertainty of alignment estimation. The two superpositions form the total uncertainty of the index at that moment.

[0113] The monitoring indicators of energy station equipment (such as temperature, pressure, and power) are highly coupled with operating conditions (such as load, flow rate, and ambient temperature). Changes in operating conditions can cause fluctuations in these indicators, making it difficult to directly reflect the true health status of the equipment. In this embodiment, operating condition vectors are used... Construction index Conditional expectation Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed through exponential functions, including:

[0114]

[0115]

[0116] in: and For the desired model parameters, The number of dimensions for the operating condition variables. The number of basis functions used for each operating condition variable. Let h be the basis function. It is an exponential function. and These are the parameters of the variance model.

[0117] In this embodiment, the basis function can be any one of cubic B-spline basis functions, piecewise linear basis functions, or Gaussian radial basis functions. In a preferred embodiment, cubic B-spline basis functions are used for each operating condition variable. H nodes are set according to the quantiles of the distribution of healthy samples, and the node sequence is as follows:

[0118]

[0119] in: These are spline nodes, typically set according to the quantiles or equidistant methods of historical samples of operating condition variables.

[0120] The zero-order B-spline basis function is defined as:

[0121]

[0122] For any order The recursive formula for the B-spline basis function is expressed as:

[0123]

[0124] when At that time, the cubic B-spline basis functions are obtained:

[0125]

[0126] The value of H can be selected based on the sample size and fitting complexity. In one implementation:

[0127]

[0128] If the healthy sample size is less than 1000, H can be set to 3-4; if the sample size is large and the operating conditions are significantly nonlinear, H can be set to 5-8. The final value of H can be determined by the Bayesian Information Criterion (BIC) or the principle of minimizing cross-validation error.

[0129] S103, Based on the decoupled monitoring vector and uncertainty of the working condition, construct a relationship graph and fuse it to generate a state vector, perform soft clustering to obtain the working condition membership degree and establish a sub-working condition health baseline model.

[0130] The obtained decoupled monitoring vector only reflects the operating status of a single node of the equipment. Complex topological connections and influence relationships exist between energy station equipment, between equipment and components, between measuring points, and between events. Single-node analysis cannot reflect system-level state correlations. This embodiment constructs a heterogeneous time-series relationship graph to achieve system-level feature fusion and generate a fused state vector. Furthermore, it obtains equipment operating condition membership degrees through soft clustering with operating condition transition priors, and establishes a sub-operating condition health baseline model based on healthy samples, providing a benchmark for subsequent equipment health status assessment. The embodiment also clarifies the implementation methods for core aspects such as graph construction, model parameters, and healthy sample selection.

[0131] Specifically, a relationship graph is constructed based on the physical topology and operational logic of the energy station equipment. Node set Represents devices, components, measurement points, and event entities, and is an edge set. This indicates the relationships between topological connections, containment relationships, influence relationships, and maintenance functions. Represent the set of relation types for edges, and assign relation attribute vectors to the edges. With time difference characteristics In this embodiment, It is obtained by concatenating numerical values ​​based on relationship type, direction, connection strength (such as the influence weight of valve opening and the proportion of flow contribution), and historical co-occurrence strength. Calculated from relevant event timestamps, such as the interval from the last maintenance to the current time, the interval of associated alarm occurrence, the delay from upstream status change to downstream response, etc., can be scalar or segmented encoding.

[0132] To fully explore the correlation features between nodes in the relation graph, relational attention graph fusion is performed using the graph: for any side Construct message:

[0133]

[0134] And through the neighborhood Attention weights are obtained by softmax normalization. Based on this aggregation, we obtain: Then through the gating fusion unit Update node representation .in: Let r be the linear transformation matrix corresponding to relation type r. For nodes u In the l Layer representation vector, For the first The linear mapping matrix of the layer-side edge feature concatenation terms. Represents the edge attribute vector. This means concatenating the edge attributes with the time difference. For nodes In the The neighborhood aggregation vector of the layer, For the first Layers consist of nodes Passed to the node The message vector, For nodes In the The node representation of a layer is a vector. This indicates gating fusion, which merges the old representation with the newly aggregated information according to the gating ratio.

[0135] The fused state vector is formed by fusing single-node monitoring features and system-level map features. Its specific composition comes from the concatenation of a normalized monitoring vector and an uncertainty vector. and the graph representation of the target device node. splicing to form a fused state vector ,in: for, Let be the uncertainty vector. To monitor the number of dimensions of the main variables.

[0136] The operating conditions of energy station equipment exhibit continuity and transition characteristics. The switching between different operating conditions follows certain probabilistic patterns, which traditional hard clustering cannot reflect. This embodiment addresses... Perform soft clustering with load condition transition priors using Gaussian emission likelihood. With operating condition transition matrix Calculate posterior membership Specifically, this includes:

[0137]

[0138]

[0139]

[0140] in: For at any time Observation status From working conditions The generated likelihood value, The center vector of the working condition cluster, Let be the cluster covariance matrix. The square of the Mahalanobis distance. Represents conditional probability. For a moment Implicit operating condition random variables, This is an abbreviation for the observation sequence, indicating... .

[0141] In this embodiment, and Clustering iterative estimation can be performed on health data or full-run data, with initialization using K-means or binning. The transition count or normalization is performed on the historical operating condition sequence; when there is no label, the transition can be inferred from the cluster label sequence, and then Laplace smoothing is added to prevent zero probability. The soft membership degree is obtained by forward-backward recursion calculation and can also be approximated by the emission probability "softmax" in engineering.

[0142] Health sample set throughout the entire life cycle of the device Up, press Calculate the mean health baseline under different working conditions The shrinkage covariance is calculated to improve the stability and invertibility of the covariance matrix, and a health baseline model for different operating conditions is established to provide a benchmark for subsequent equipment health status assessment.

[0143] ,

[0144] in: The shrinkage coefficient is used to prevent... Irreversible, improves stability Operating condition The weighted sample covariance matrix under the following conditions It is the identity matrix. Indicates working conditions The mean of the health baseline.

[0145] In this embodiment, The effectiveness of the baseline model directly determines its accuracy. The core principle of sample selection is to select stable, fault-free operating data and eliminate data from abnormal or fluctuating ranges. Typical sources are stable operating ranges after commissioning and acceptance, and stable operating ranges after major overhauls and replacements. Ranges to be eliminated include known fault windows, alarm-dense windows, sensor faults, calibration anomaly windows, and frequent start-stop transition windows.

[0146] S104, combine the state vector and the sub-condition health baseline model to calculate the equipment health index and degradation index, and recursively output the root cause probability based on the propagation relationship.

[0147] The aforementioned fusion state vector, operating condition membership degree, and sub-operating condition health baseline model are first used in this embodiment to calculate the deviation of time k from operating condition p based on the sub-operating condition baseline:

[0148]

[0149] The degenerate residuals are obtained by weighting them according to their membership degree. ,in: For a moment Relative working conditions The degree of deviation, To inverse the covariance, the deviation is made to take into account correlation and scale, where P is the total number of operating condition clusters.

[0150] The health status of equipment is affected by many factors, and a single indicator cannot fully reflect the true state of the equipment. This embodiment is based on a fused state vector. With the output of the prediction model Calculate reconstruction error and the , Energy efficiency deviation Vibration deviation Thermal deviation and a summary of uncertainties The health index obtained through integration:

[0151]

[0152] And according to Degradation index obtained: This enables a quantitative assessment of the health status and degree of degradation of equipment. Among these: The square of the L2 norm. arrive These are the fusion weights for each piece of evidence. In this embodiment... These are non-negative fusion weights used to characterize the impact of degradation residuals, reconstruction errors, energy efficiency bias, vibration bias, thermal bias, and uncertainty on the health index. In one implementation, each component can first be normalized to 0-1, and then... , The initial values ​​of the fusion weights can be set according to the importance of the project. For example, degradation residuals and energy efficiency deviations have higher weights, followed by vibration deviations and thermal deviations. The uncertainty term is used to penalize low-confidence samples. The initial values ​​of the fusion weights can be determined on a labeled healthy or abnormal sample set, with the goal of minimizing the accuracy of health status identification, AUC, F1 score, or false alarm rate, through grid search, Bayesian optimization, or cross-validation.

[0153] like Figure 3 The graph showing the health index versus degradation index indicates that the health index of the energy station equipment gradually decreases while the degradation index gradually increases over time. However, with maintenance, the health index may rebound, but it never returns to its initial value. This suggests that maintenance can keep the equipment functioning, but as the equipment reaches the end of its lifespan, it will inevitably need to be replaced.

[0154] In this embodiment , and All values ​​are obtained by comparing actual values ​​with health benchmark values. The specific calculation method is as follows:

[0155] First, calculate the equipment's energy efficiency KPIs, such as COP, kW, RT, heat exchange efficiency, and energy consumption per unit flow rate. Then, compare these values ​​with the output of a health model under the same operating conditions to obtain the energy efficiency deviation. The larger the deviation, the worse the equipment's energy efficiency.

[0156] Features such as RMS, peak value, kurtosis, envelope spectrum energy, and characteristic frequency amplitude are extracted from the vibration time or frequency domain. These features are then compared with the feature values ​​of a healthy distribution or baseline model to obtain the vibration deviation. The larger the deviation, the more obvious the equipment vibration anomaly.

[0157] The calculation includes thermal characteristics such as the temperature difference approaching the heat exchanger ends, abnormal supply and return water temperature differences, thermal balance residuals, and temperature differences between the coils and the two ends of the heat exchanger. These are then compared with the healthy operating condition model and the baseline to obtain the thermal deviation. The larger the deviation, the more abnormal the thermal condition of the equipment.

[0158] For the set of failure modes From the characteristics of evidence Calculate the strength of evidence for the log-likelihood ratio for each failure mode: And activated by a threshold Normalization yields anomalous evidence ,in: For the set of evidence features used for fault identification, This is a fault mode. In failure mode The probability model for the occurrence of the following evidence. To ensure that under normal operating conditions, at all times Evidence feature vector The probability of occurrence. ReLU-based threshold activation: activation levels below the threshold are not counted as evidence. For the first Evidence trigger threshold for each failure mode This is the normalization operator.

[0159] In this embodiment, From the fusion state vector Select a subset related to the fault mode, such as differential pressure, flow rate, power, characteristic frequency amplitude, energy efficiency index, residual term, etc. Numerical representations of discrete features such as "alarm count, alarm type, and recent maintenance occurrences" can also be added to improve the accuracy of fault identification.

[0160] This embodiment further constructs a fault propagation matrix to characterize the propagation relationship between fault modes. Combining the prior probability of the fault, it integrates the historical root cause probability, current abnormal evidence, fault propagation term and prior term in the logarithmic domain to avoid the underflow problem caused by probability product. Finally, it outputs the root cause probability vector through normalization to achieve accurate location of the fault root cause.

[0161] Specifically, the propagation matrix Fault Prior Integrating historical root cause probabilities in the logarithmic field Current evidence Dissemination items Combined with prior terms, we obtain the unnormalized root cause score:

[0162]

[0163] Then normalize the root cause scores to root cause probabilities. ,in: For a moment The root cause probability is an unnormalized intermediate quantity. For the weight of historical memory, As the current weight of evidence, To take the logarithm of each element of the vector, prevent , Indicates propagation gain. For the fault propagation matrix, This indicates that the evidence is disseminated along the propagation relationship to the relevant root causes. Fault prior probability vector This indicates prior addition.

[0164] In this embodiment, It can be defined by topology and causal relationships, such as "scaling → decreased heat exchange efficiency → abnormal supply and return water temperature difference". Or it can be obtained from historical co-occurrence statistics or mutual information learning. The more frequently fault evidence co-occurs, the stronger the connection. After construction, it needs to be normalized to make the propagation strength controllable. Statistics are obtained based on equipment type, service life, historical failure frequency, seasonality, etc.; uniform priors can be used during the cold start phase to avoid initial bias.

[0165] By adjusting the validation set, the root cause probability can be made both stable and able to adapt to changes in new evidence. By adjusting parameters based on historical cases, the proportion of influence of propagation factors on the final root cause can be controlled.

[0166] S105, calculate the risk score and output it in a graded manner based on the equipment health index, degradation index and root cause probability estimate and stress-corrected remaining life.

[0167] Based on the equipment health index, degradation index, and root cause probability vector obtained from S104, this step establishes a stochastic evolution model of the degradation index to characterize the equipment degradation pattern. It then calculates the degradation drift velocity using stress vectors and performs index smoothing correction. Based on the failure threshold, it estimates the expected remaining lifespan and lifespan uncertainty. Subsequently, it integrates multi-dimensional indicators such as the degradation index, root cause risk term, lifespan mapping risk term, and criticality factor to calculate a comprehensive risk score. After time smoothing, it completes risk classification according to thresholds, ultimately outputting the core status indicators for the entire equipment lifecycle. This provides accurate basis for energy station equipment operation and maintenance decisions, as detailed below:

[0168] Establish a stochastic evolution model of the degradation index The dynamic changes in the degradation of quantifiable equipment. And according to the stress vector Calculate the degradation drift velocity This reflects the accelerating effect of stress on equipment degradation. Among them: For random disturbance noise, The variance of the noise during the degradation process. Basic drift term, Based on the drift term, This indicates that the mean is 0 and the variance is 0. The Gaussian distribution.

[0169] In this embodiment, the construction The components in the data include: load factor, partial load operation ratio; number of start-stop cycles (e.g., past 1 hour / 1 day), short cycle ratio; cumulative duration of over-temperature / over-pressure, cumulative duration of differential pressure exceeding limits; cumulative duration of high vibration; and cumulative duration of high thermal resistance / low heat exchange efficiency. These can all be obtained from BMS / SCADA and event logs using a sliding window, and can be selected according to the actual situation of the site.

[0170] To reduce the impact of random perturbations on the degradation drift velocity estimation and improve its stability and smoothness, the following measures are taken: Perform exponential smoothing correction: ,in: For smoothing coefficients, The slope window length, In-window weights For weighted slope estimation within the window, The assignment update operator writes the calculation result on the right side back to the variable on the left side.

[0171] Then based on the failure threshold With correction Calculate the approximate expected remaining lifespan And calculate the lifetime uncertainty. ,in: and In Avoid division by zero or negative drift that could cause anomalies.

[0172] In this embodiment, The setting is based on the following: when energy efficiency falls below a certain threshold, vibration reaches an alarm level, or key performance indicators continuously exceed limits. The value is used as a threshold. Historical fault sample statistics: samples from the window before the fault occurred. The quantile is used as the threshold. For fault-free samples, empirical thresholds can be used, such as maintenance trigger points or points where performance is unacceptable.

[0173] Immediately afterwards Root cause risk items Lifetime mapping risk items Keyness Factors and Integrated computing comprehensive risk score:

[0174]

[0175] in: Original risk score, , , , and All are fusion weights;

[0176] right Risk score obtained by time smoothing Then according to the threshold Will Quantified into risk levels ,in: For the smoothed risk score, This is the risk time smoothing coefficient.

[0177] In this embodiment This can be obtained by fusing the following information (and normalizing to 0–1 or 0–10): graph centrality (betweenness, degree, cut vertices), whether it is in the main loop; redundancy: whether there is a backup machine, N+1 configuration; load share: the ratio of cooling / heating handled by this device. Severity of consequences: the scope of power outage caused by the fault, the number of affected users, and the estimated repair time. These are derived from design capacity, operational allocation, topology, and operational experience rules. , , , and Initially, the parameters can be allocated based on experience, for example, the weight of health items and root cause items is greater than that of lifespan items. Then, the parameters can be adjusted by playing back historical alarms to make the false alarm rate or false negative rate meet the target. , , , and All are non-negative risk fusion weights, corresponding to health risk, root cause risk, lifespan risk, criticality risk, and lifespan uncertainty risk, respectively. In one implementation, the following constraints can be imposed:

[0178] ,

[0179] , , , and The initial value can be set according to business preferences; for example, it can be increased in continuous power supply scenarios. and Highly reliable equipment can improve During calibration, historical fault events can be used as positive samples and normal operating ranges as negative samples, with optimization based on a weighted objective of alarm lead time, false alarm rate, and false alarm rate. For image stabilization, select the desired level of smoothness or sensitivity based on the risk warning. It can be set according to the risk distribution quantile during the healthy period, or according to the expected alarm frequency.

[0180] The , , , , Remaining lifetime estimation and its uncertainty output, among which Output can be categorized by level:

[0181]

[0182] In this implementation, satisfy In one embodiment, Alternatively, risk scores can be determined based on quantiles from the distribution of risk scores in a health history sample. For example, the 95th percentile of a normal sample could be used as the quantile. The median of mildly abnormal samples was used as The 25th percentile of the severe fault samples was used as .

[0183] Corresponding to the method for integrating and analyzing monitoring data throughout the entire life cycle of energy station equipment provided in the above embodiments, this application also provides an embodiment of a system for integrating and analyzing monitoring data throughout the entire life cycle of energy station equipment.

[0184] See Figure 4 The energy station equipment full life cycle monitoring data integration and analysis system 20 provided in this embodiment includes:

[0185] The acquisition module 201 is used to acquire multi-source data on the life cycle of energy station equipment and to uniformly encode the multi-source data to obtain a feature vector with a timestamp.

[0186] The fusion alignment module 202 is used to calculate the data credibility in the feature vector and complete conflict robust fusion, unified time alignment and working condition compensation normalization to obtain the working condition decoupling monitoring vector and uncertainty at each time point;

[0187] The working condition modeling module 203 is used to construct a relationship graph based on the working condition decoupled monitoring vector and uncertainty, fuse and generate a state vector, perform soft clustering to obtain the working condition membership degree, and establish a sub-working condition health baseline model.

[0188] The evaluation and deduction module 204 is used to calculate the equipment health index and degradation index by combining the state vector and the sub-condition health baseline model, and to recursively output the root cause probability by combining the propagation relationship.

[0189] The risk output module 205 is used to calculate a risk score and output it in a graded manner based on the equipment health index, degradation index and root cause probability estimate and stress-corrected remaining life.

[0190] The implementation methods of the above modules are the same as those in the method embodiments, and can be referred to in the above method embodiments, so they will not be repeated here.

[0191] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0192] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment, characterized in that, include: Acquire multi-source data on the lifecycle of energy station equipment, and uniformly encode the multi-source data to obtain a feature vector with a timestamp; Calculate the data reliability in the feature vector and complete conflict robust fusion, unified time alignment, and operating condition compensation normalization to obtain the operating condition decoupling monitoring vector and uncertainty at each time point, including: Based on the above The integrity, consistency, timeliness, source health, drift penalty, calibration aging, and link quality of each data record in the data are used to construct a quality feature vector. And based on historical statistical parameters Computing credible priors Then set the current quality score With trusted priors The final credibility weights are obtained by performing posterior fusion: in: For completeness, For consistency, For timeliness, For the source of health, As a penalty for drifting, To calibrate the aging process, For link quality, This means transposing a column vector into a row vector. Indicates a linear score for the inner product, outputting a scalar; Calculated based on the missing percentage of key fields or key measurement points in the records; The determination is based on a comprehensive assessment of physical constraints, consistency between adjacent measuring points, and consistency of historical statistics. Physical constraints include range, energy conservation, temperature difference sign, and flow non-negativity rule verification. Consistency between adjacent points focuses on the relationship between the supply and return water temperature difference in the same loop, the pressure difference before and after the pump, and the flow rate. Statistical consistency is the degree of deviation between the data under similar operating conditions and the historical distribution. The timeliness score is obtained from the difference between the entry time and the sampling time, or the difference between the current time and the sampling time. The greater the data delay, the lower the timeliness score. Information from sensor self-test bit, fault code, calibration validity period, measurement point quality bit, and gateway online status; Data drift was detected using a rolling window, employing CUSUM, EWMA, drift rate, and long-term offset method based on the difference between the reference measurement point and the drift. The more obvious the drift, the larger the penalty value and the lower the reliability. Based on the number of days since the last calibration or replacement, the calibration aging score gradually decreases after the recommended calibration cycle of the equipment has been exceeded. Calculated from network statistics, including packet loss rate, retransmission rate, offline duration, and clock drift, the worse the link quality, the lower the data reliability. Cumulative statistics from historical measurement points or data sources that were judged as valid or invalid, among which... The number of valid historical counts is determined by quality level, manual verification, and consistency with reference instruments. This refers to the number of invalid entries in the past, including: out-of-bounds errors, drifting errors, fault codes, and manually marked invalid entries. A set of conflict records for the same device and the same indicator within the same time window. Robust fusion is performed by minimizing the weighted robust loss to obtain the fused value. ; Map the fused asynchronous multisampling rate metrics to a unified time grid. For each indicator Establish a state-space model Spatiotemporal alignment values ​​are obtained through Kalman smoothing. ,in: Indicates time The true potential value, This indicates the observations on the aligned grid. This represents the random disturbance noise term. Expressing conditional expectation; Based on operating condition vectors Construction index Conditional expectations Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed by using an exponential function; then... Perform working condition compensation normalization And calculate the uncertainty. ,in: To estimate the standard deviation of the variance, Indicates the intensity of uncertainty; The set of conflict records for the same device and the same indicator within the same time window Robust fusion is performed by minimizing the weighted robust loss to obtain the fused value. ,include: in: This means finding the option that minimizes the objective function. , Indicates the first The first record Individual indicator values, For scaling estimation, the residuals are made dimensionless, and robust scaling estimation within the healthy period or the current window is preferred to suppress extreme outliers. The working condition vector-based Construction index Conditional expectations Conditional scale ,in Represented using basis function expansion, Non-negativity is guaranteed through exponential functions, including: in: and For the desired model parameters, The number of dimensions for the operating condition variables. The number of basis functions used for each operating condition variable. Let h be the h-th basis function, and the basis function can be any one of cubic B-spline basis function, piecewise linear basis function, or Gaussian radial basis function; It is an exponential function. and These are the parameters of the variance model; Based on the decoupled monitoring vectors and uncertainties of the aforementioned operating conditions, a relationship graph is constructed and fused to generate a state vector. Soft clustering is then performed to obtain the membership degree of the operating conditions, and a sub-operating condition health baseline model is established, including: Constructing heterogeneous time series relationship graphs Node set Represents devices, components, measurement points, and event entities, and is an edge set. This indicates the relationships between topological connections, containment relationships, influence relationships, and maintenance functions. Represent the set of relation types for edges, and assign relation attribute vectors to the edges. With time difference characteristics ; The It is obtained by concatenating numerical values ​​based on relation type, direction, connection strength, and historical co-occurrence strength. Calculated from relevant event timestamps, including: the interval from the last maintenance to the present, the interval of associated alarm occurrence, and the delay from upstream status change to downstream response, which can be scalar or segmented encoding; Graph execution relational attention graph fusion: for any side Construct message: And through the neighborhood Attention weights are obtained by softmax normalization. Based on this aggregation, we obtain: Then through the gating fusion unit Update node representation ,in: Let r be the linear transformation matrix corresponding to relation type r. For nodes u In the l Layer representation vector, For the first The linear mapping matrix of the layer-side edge feature concatenation terms. Represents the edge attribute vector. This means concatenating the edge attributes with the time difference. For nodes In the The neighborhood aggregation vector of the layer, For the first Layers consist of nodes Passed to the node The message vector, For nodes In the The node representation of a layer is a vector. This indicates gating fusion, which merges the old representation with the newly aggregated information according to a gating ratio; The normalized monitoring vector and the uncertainty vector are concatenated as follows: and the graph representation of the target device node. splicing to form a fused state vector ,in: Let be the uncertainty vector. To monitor the number of dimensions of the main variables; right Perform soft clustering with load condition transition priors using Gaussian emission likelihood. With operating condition transition matrix Calculate posterior membership , The transition count or normalization is performed on the historical operating condition sequence; when there is no label, the transition can be inferred from the cluster label sequence, and then Laplace smoothing is added to prevent zero probability; The pair Perform soft clustering with load condition transition priors using Gaussian emission likelihood. With operating condition transition matrix Calculate posterior membership ,include: in: For at any time Fusion state vector From working conditions The generated likelihood value, The center vector of the working condition cluster, Let be the cluster covariance matrix. The square of the Mahalanobis distance. Represents conditional probability. For a moment Implicit operating condition random variables, This is an abbreviation for the observation sequence, indicating... ; In the health sample index set Up, press Calculate the mean health baseline under different working conditions And calculate the shrinkage covariance: ， in: The shrinkage coefficient is used to prevent... Irreversible, improves stability Operating condition The weighted sample covariance matrix under the following conditions It is the identity matrix. Indicates working conditions The mean of the health baseline; The equipment health index and degradation index are calculated by combining the fused state vector and the sub-condition health baseline model, and the root cause probability is recursively output by combining the propagation relationship, including: The deviation of time k from condition p is calculated based on the baseline of the sub-conditions: The degenerate residuals are obtained by weighting them according to their membership degree. ,in: For a moment Relative working conditions The degree of deviation, To reduce the inverse covariance and make the deviation account for correlation and scale, P is the total number of working condition clusters; Based on fusion state vector With the output of the prediction model Calculate reconstruction error and the , Energy efficiency deviation Vibration deviation Thermal deviation and a summary of uncertainties The integration yields a health index and according to Degradation index obtained: ,in: , and All values ​​are obtained by comparing actual values ​​with health benchmark values. The specific calculation method is as follows: First, calculate the energy efficiency KPIs of the equipment, such as COP, kW, RT, heat exchange efficiency, and energy consumption per unit flow. Then, compare them with the output of the health model under the same operating conditions to obtain the energy efficiency deviation. The larger the deviation, the worse the energy efficiency of the equipment. Features such as RMS, peak value, kurtosis, envelope spectrum energy, and characteristic frequency amplitude are extracted from the vibration time domain or frequency domain and then compared with the feature values ​​of the healthy distribution or baseline model to obtain the vibration deviation. The larger the deviation, the more obvious the equipment vibration abnormality. Calculate the temperature difference approaching the heat exchange end, the abnormal temperature difference between the supply and return water, the thermal balance residual, the temperature difference deviation between the two ends of the coil and the heat exchanger from the thermal characteristics, and then compare it with the healthy operating condition model and the baseline to obtain the thermal deviation. The larger the deviation, the more abnormal the thermal condition of the equipment. For the set of failure modes From the characteristics of evidence Calculate the strength of evidence for the log-likelihood ratio for each failure mode: And activated by a threshold Normalization yields anomalous evidence ,in: For the set of evidence features used for fault identification, From the fusion state vector Select a subset related to the failure mode, including: differential pressure, flow rate, power, characteristic frequency amplitude, energy efficiency index, and residual term. You can also add discrete features: alarm count, alarm type, and numerical representation of the most recent maintenance events. This is a fault mode. In failure mode The probability model for the occurrence of the following evidence. ReLU-based threshold activation: activation levels below the threshold are not counted as evidence. For the first Evidence trigger threshold for each failure mode For normalization operators; Constructing a fault propagation matrix Fault Prior Integrating historical root cause probabilities in the logarithmic field Current evidence Dissemination items Combined with prior terms, we obtain the unnormalized root cause score: Then normalize the root cause scores to root cause probabilities. ,in: For a moment The root cause probability is an unnormalized intermediate quantity. For the weight of historical memory, As the current weight of evidence, To take the logarithm of each element of the vector, prevent , Indicates propagation gain. This is the fault propagation matrix. It can be defined by topology and causality or obtained by historical co-occurrence statistics or mutual information learning. The more frequently fault evidence co-occurs, the stronger the connection. Normalization is required after construction. This indicates that the evidence is disseminated along the propagation relationship to the relevant root causes. Fault prior probability vector Indicates prior addition; Based on the equipment health index, degradation index, and root cause probability estimate, and stress-corrected remaining life, a risk score is calculated and output in a tiered manner, including: Establish a stochastic evolution model of the degradation index ,in And according to the stress vector Calculate the degradation drift velocity ,in: For random disturbance noise, The variance of the noise during the degradation process. Basic drift term, This indicates that the mean is 0 and the variance is 0. Gaussian distribution; construction The components in the data include: load factor, partial load operation ratio, number of start-stop cycles, short cycle ratio, cumulative duration of over-temperature / over-pressure, cumulative duration of differential pressure exceeding limits, cumulative duration of high vibration, and cumulative duration of high thermal resistance / low heat exchange efficiency. right Perform exponential smoothing correction: ,in: For smoothing coefficients, The slope window length, In-window weights For weighted slope estimation within the window, The assignment update operator writes the calculation result on the right side back to the variable on the left side. Based on failure threshold With correction Calculate the approximate expected remaining lifespan And calculate the lifetime uncertainty. ,in: and In Avoid division by zero or negative drift that could cause anomalies; Will Root cause risk items Lifetime mapping risk items Keyness Factors and Integrated computing comprehensive risk score: in: Original risk score, , , , and All are fusion weights; right Risk score obtained by time smoothing Then according to the threshold Will Quantified into risk levels ,in: For the smoothed risk score, This is the risk time smoothing coefficient; The , , , , Remaining lifetime estimation and its uncertainty output.

2. The method for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment according to claim 1, characterized in that, The process of acquiring multi-source data on the lifecycle of energy station equipment and uniformly encoding the multi-source data to obtain a timestamped feature vector includes: Obtain multi-source electronic data records corresponding to the target equipment from heterogeneous data sources at the energy station; Each electronic data record is objectified as a lifecycle data tuple. ,in: For device instance identification, For lifecycle stage tags, For timestamps, To monitor the main variable vector, For the operating condition context vector, For data source attribute vectors, For version semantic information, Versions from site migration, calibration, ledger, protocol, and model are typically saved as discrete IDs or timestamps. Missing mask; Regarding the Perform nonlinear embedding respectively to obtain Embedded vectors are generated for stage labels and source identifiers respectively. Then based on the missing mask Gating function right Perform element-wise gating to form the feature vector. .

3. The method for integrating and analyzing monitoring data throughout the entire lifecycle of energy station equipment according to claim 2, characterized in that, The above Perform nonlinear embedding respectively to obtain Embedded vectors are generated for stage labels and source identifiers respectively. Then based on the missing mask Gating function right Perform element-wise gating to form the feature vector. ,include: Monitoring branch embedding: ; Working condition branch embedding: ; Original attribute branch embedding: ; Stage and source embedding: ; Missing gating fusion and final stitching: ,in: , , , and For matrix parameters, , , , and For bias vectors, It is a non-linear activation function. and This is an embedding mapping from discrete IDs to vectors. For element-wise multiplication, The output vector of the gated function. This involves concatenating vectors.

Images