A layered fault diagnosis method, system and device for a precision analytical instrument

Abstract

The application provides a layered fault diagnosis method, system and equipment of a precision analytical instrument, relates to the field of equipment fault diagnosis, and sets physical components, sensing features and prior knowledge of the diagnosed equipment as different nodes in a heterogeneous dynamic graph, and records the relationship between the nodes by physical coupling edges, data statistical edges and knowledge rule edges. Local sensitive sub-graphs are calculated during normal monitoring; when the abnormal triggering condition is reached, the complete heterogeneous dynamic graph is switched to. The fault probability and prediction uncertainty obtained by reasoning are calibrated by the sensing reliability and the physical feasible region, and the explicit fault, the boundary state low confidence or the normal result is output. The fault samples confirmed by the repair record or the artificial review are saved as evidence chains, and enter the sample enhancement and knowledge updating process after the mechanism consistency verification passes. The component misjudgment caused by the sensing anomaly and the graph calculation amount in the normal monitoring stage are reduced.

Description

Technical Field

[0001] This invention relates to the field of instrument fault diagnosis, and in particular to a hierarchical fault diagnosis method, system and equipment for precision analytical instruments. Background Technology

[0002] Mass spectrometers, chromatographs, electron microscopes, and other precision analytical instruments are widely used in scientific research, industrial testing, environmental monitoring, life sciences, and energy and chemical engineering. These instruments typically consist of multiple functional components, sensing units, actuation units, and control units, performing measurement, analysis, or control tasks under various operating conditions, long cycles, or continuous operation. With increasing instrument integration and automation, problems such as component degradation, abnormal structural connections, blockages or leaks in transmission pathways, abnormal power supply control, and sensor drift often manifest as slow drift, subtle anomalies, or combined shifts in multiple parameters. If these issues are not identified and addressed early, they can lead to inaccurate instrument output, performance degradation, or unplanned downtime.

[0003] Existing fault diagnosis technologies have the following technical shortcomings: First, the operational information of complex instruments comes from dispersed sources. This includes equipment output data such as measurement results, response characteristics, and performance indicators; monitoring parameters reflecting the status of core functional components, such as electrical parameters, temperature, pressure, flow rate, rotational speed, and vacuum; and information such as environmental conditions, load conditions, control commands, alarms, and maintenance records. These data are physically coupled, statistically correlated, and linked by fault rules. However, existing diagnostic methods often treat these as independent feature inputs, lacking a unified, structured expression of physical mechanisms, data correlations, and prior knowledge. This makes it difficult to reliably utilize the relationships between component degradation, environmental disturbances, and control changes.

[0004] Secondly, obtaining real-world fault samples is difficult, especially for early weak faults, boundary state faults, newly emerging fault modes, and composite fault samples. Existing data augmentation methods are mostly based on pure statistical distributions to expand samples. The generated data often lacks constraints from the physical operating laws of equipment or fault evolution mechanisms, which limits the improvement of the model's generalization ability and may introduce inaccurate fault boundaries.

[0005] Third, online monitoring systems are mostly deployed on embedded computing units or edge computing platforms, which are limited by computing resources, storage space, and energy consumption. Deep learning models typically have significant computational overhead, and if the complete model is continuously invoked during continuous monitoring, it may affect the equipment's data acquisition, measurement analysis, or real-time control tasks.

[0006] Fourth, existing diagnostic models typically output fault categories or probabilities, but they lack sufficient records of the correspondence between abnormal features, related components, and fault modes. They also lack a process for updating the model and knowledge base using new samples, manual verification results, and historical maintenance knowledge. After long-term operation, component states, degradation thresholds, operating condition distributions, and fault modes may all change. If the diagnostic model and knowledge base remain unchanged, the diagnostic results are likely to gradually deviate from the actual equipment state.

[0007] Therefore, there is a need to provide a hierarchical fault diagnosis method, system, and equipment for precision analytical instruments to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to provide a hierarchical fault diagnosis method, system, and device for precision analytical instruments. It organizes physical mechanisms, data association, and prior knowledge in the same diagnostic process, first determines whether the sensing parameters can participate in component state reasoning, then selects a local sensitive subgraph or a complete heterogeneous dynamic graph based on the abnormal triggering conditions, and calibrates the diagnostic results with the physical feasible domain, thus solving the problems existing in the background technology.

[0009] To achieve the above objectives, the present invention provides a hierarchical fault diagnosis method for precision analytical instruments, comprising the following steps: S1: Collect parameters and data from precision analytical instruments, perform time alignment and sliding window segmentation; obtain a heterogeneous dynamic graph that updates with the sliding window; S2: Perform a sensing reliability assessment on the sensing unit; S3: Extract local sensitive subgraphs from the heterogeneous dynamic graph, use a local graph neural network to infer the local sensitive subgraphs, obtain the fault probability and prediction uncertainty, and determine whether the local inference result reaches the preset anomaly threshold; if it does not reach the preset anomaly threshold, do not perform full graph inference, and use the local fault probability and local prediction uncertainty as the inference result of the current sliding window; if it reaches the preset anomaly threshold, switch to full graph inference mode, use a heterogeneous graph neural network that distinguishes different edge types to perform full graph inference on the heterogeneous dynamic graph, obtain the full graph fault probability and full graph prediction uncertainty, and use the full graph fault probability and full graph prediction uncertainty as the inference result of the current sliding window; S4: Based on the physical feasible region constraint, perform reliable fault determination, and perform consistency calibration on the fault probability, prediction uncertainty, sensing reliability, and physical feasible region output by the current sliding window in S3; when S3 does not trigger full graph inference, the fault probability and prediction uncertainty are the local fault probability and local prediction uncertainty obtained by local inference; when S3 triggers full graph inference, the fault probability and prediction uncertainty are the full graph fault probability and full graph prediction uncertainty obtained by full graph inference; if the verification is consistent, output the fault determination result of the current sliding window as the diagnostic result; S5: If the verification is inconsistent, mark it as a potential new fault mode or a sample to be reviewed, and proceed to the remote review or expert analysis process. S6: For potential new failure modes or samples to be reviewed, small sample augmentation is performed using covariance-related perturbations constrained by the physical feasible region. After augmentation, the data statistical edges are recalculated, and samples that do not meet the failure rules, operating condition consistency, or time series continuity are deleted before returning to S3.

[0010] Preferably, the parameters in S1 include component status parameters and operating condition parameters, and the data includes control event data and output response data. Time alignment and sliding window segmentation are performed according to the equipment instance and operating condition. Window statistical features are calculated within each sliding window to form sensing feature nodes. For the first Parameters of continuous components in equipment instances Operating parameters Sampled values according to Perform robust normalization; in, This is the median of historical normal operating data. This represents the median absolute deviation. A positive number not less than the corresponding sensor resolution; each sliding window contains There are sampling points, and the window step size is . Each sampling point has window statistical characteristics including window median, window median absolute deviation, normalized change slope, out-of-limit ratio, and missing ratio. Using functional components and sensing units as physical component nodes, and fault modes and maintenance rules as prior knowledge nodes, physical coupling edges, data statistics edges, and knowledge rule edges are established to obtain a heterogeneous dynamic graph that updates with the sliding window. The weights of the physical coupling edges are represented as follows: ; in, For correlation strength, This is a correction factor for operating conditions; The weights of the data statistics edges are represented as follows: ; in, The normalized feature sequence corresponding to the i-th continuous parameter and the normalized feature sequence corresponding to the j-th continuous parameter are in the optimal time delay The correlation coefficient at the location, For the maximum search latency, To ensure consistency in common variations, clip[·,0,1] indicates that the value is truncated to 0 to 1; The weights of knowledge rule edges are represented as follows: ; in, The initial confidence level for the rule. and These represent the number of confirmed hits and the number of mismatches, respectively.

[0011] Preferably, in S2, the five-dimensional node attributes consisting of the window center, the absolute deviation of the window center, the normalized change slope, the over-limit ratio, and the missing ratio are input into the local graph neural network, and the normal state probability and the sensor fault state probabilities corresponding to the bias, drift, missing, spike, jamming, and accuracy degradation of the sensing parameters are output in sequence, and the sensing reliability is determined based on the normal state probability. When the sensing confidence level is not lower than 0.80, the corresponding sensing parameter is determined as the original confidence parameter; when the sensing confidence level is lower than 0.80, the output state is a correctable bias fault or drift fault, and the corrected correlation estimation residual does not exceed 3 times the robust scale of the normal residual, the corrected sensing parameter is determined as the corrected confidence parameter; when the sensing confidence level is lower than 0.80 and the correction condition is not met, the corresponding sensing parameter is determined as an invalid parameter, and a missing mask is set for the invalid parameter to prevent the invalid parameter from being the sole criterion for the degradation of the corresponding physical component.

[0012] Preferably, in S3, prior knowledge nodes are used as seed nodes, and local sensitive subgraphs are extracted from the heterogeneous dynamic graph based on edge weights and node contributions; the local sensitive subgraphs corresponding to the current working condition or target fault type are called, and a lightweight model is used for feature aggregation to output fault probability and uncertainty estimates; When any of the following factors—failure probability, prediction uncertainty, node deviation, or data statistics edge change—reaches the corresponding trigger threshold, a heterogeneous graph neural network that distinguishes different edge types is used to perform full graph reasoning on the heterogeneous dynamic graph.

[0013] Preferably, one-hop and two-hop adjacent nodes with an edge weight of not less than 0.35 are retained; when the number of nodes exceeds 32, the node contribution is calculated according to the node deviation, the product of the path edge weights from the seed node to the candidate node and the historical rule hit rate, and the nodes are pruned in descending order of their contribution. The sensor local graph neural network sequentially maps the node attribute vector formed by window statistical features into 16-dimensional and 8-dimensional hidden states, and reads the second-layer hidden state of the seed sensor node for classification. The component state local graph neural network uses graph-level average pooling and a classification layer to output the fault probability.

[0014] Preferably, if the failure probability output by the local sensitive subgraph is not less than 0.40, the prediction uncertainty reaches the uncertainty trigger threshold, the absolute deviation of the sensing feature node is not less than 3, or the weight change of any data statistical edge relative to the previous sliding window is not less than 0.20, then switch to full graph inference. The uncertainty trigger threshold is the larger of the 95th percentile of the prediction uncertainty on the normal validation set and 0.15. The heterogeneous graph neural network performs adjacency propagation on physical coupling edges, data statistics edges and knowledge rule edges respectively, and then performs feature transformation and graph-level pooling after concatenating the propagation results of different edge types.

[0015] Preferably, the training phase and online inference phase of the lightweight model in S3 are executed separately. The training set is divided into device instances and continuous running segments. Each continuous running segment is divided into training segment, validation segment and test segment in chronological order. The same sliding window does not cross different datasets. Sensor local graph neural network, component state local graph neural network and relation type perception heterogeneous graph neural network for full graph inference are trained separately. The model weights, node order, category order, relation matrix, normal operating domain version and model summary are saved.

[0016] Preferably, during the online inference phase, the frozen model weights are invoked, and at least three samples are obtained for the same fault mode, which are confirmed by maintenance records or manual review. After the samples come from at least two independent running segments, an independent model update and verification process is initiated.

[0017] A hierarchical fault diagnosis system for a precision analytical instrument includes a data processing module, a heterogeneous dynamic graph construction module, a sensor reliability assessment module, a hierarchical reasoning module, a reliability determination module, a model version management module, a knowledge iteration module, and an interactive output module. The output of the data processing module is connected to the heterogeneous dynamic graph construction module; the output of the heterogeneous dynamic graph construction module is connected to the sensing reliability assessment module and the hierarchical inference module; the sensing reliability assessment module sends the original reliability parameters, corrected reliability parameters, or invalid parameters to the hierarchical inference module; the output of the hierarchical inference module is connected to the reliability determination module; the reliability determination module sends explicit fault samples and boundary state low-confidence samples to the knowledge iteration module; the model version management module provides a consistent configuration version to the heterogeneous dynamic graph construction module, the sensing reliability assessment module, the hierarchical inference module, and the reliability determination module; the knowledge iteration module only updates the corresponding version after independent verification is passed; the interactive output module receives the reliability determination results and diagnostic evidence, which are used to implement the steps of a hierarchical fault diagnosis method for a precision analytical instrument.

[0018] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a hierarchical fault diagnosis method for a precision analytical instrument.

[0019] Therefore, the present invention employs the above-described hierarchical fault diagnosis method, system, and equipment for precision analytical instruments, and the beneficial technical effects are as follows: (1) The heterogeneous dynamic graph constructed by the present invention includes physical component nodes, sensing feature nodes and prior knowledge nodes, and records the association between different information sources with physical coupling edges, data statistics edges and knowledge rule edges, so that the operating data, physical structure and fault knowledge of complex instruments can participate in reasoning in a unified structure.

[0020] (2) The present invention introduces a layered switching mechanism. Under normal conditions, only the local sensitive subgraph is calculated, reducing the number of nodes and edges involved in reasoning. When an anomaly occurs, the complete heterogeneous dynamic graph is called to supplement cross-component, cross-feature and knowledge rule association information, which is suitable for resource-constrained online diagnostic scenarios such as embedded and edge computing.

[0021] (3) In this invention, the physical feasible domain constraint and the probability-uncertainty dual threshold judgment are applied to the model output together, and the samples are divided into three categories: clear fault, low confidence of boundary state and normal, which reduces the possibility of misjudgment caused by physically infeasible abnormal modes.

[0022] (4) In this invention, the fault evidence chain, mechanism consistency verification, small sample enhancement and knowledge base incremental update are formed into a write-back process after verification, so that the diagnostic model and knowledge base can be updated as new degradation states, new fault modes and operating condition distributions occur during the service of the equipment.

[0023] (5) The present invention places the sensor reliability assessment before the component status diagnosis. When the sensor is abnormal, a correction value or invalid mark is used, and the correction uncertainty is passed to the component diagnosis, which can reduce the situation of interpreting the measurement link failure as the degradation of the tested component. Attached Figure Description

[0024] Figure 1 This is an overall framework diagram of a hierarchical fault diagnosis system for a precision analytical instrument according to the present invention; Figure 2 This is a flowchart of a hierarchical fault diagnosis system for a precision analytical instrument according to the present invention; Figure 3 This is a graph showing the convergence result of training the lightweight model in an embodiment of the present invention; Figure 4 This is a diagram showing the classification results of the offline fault injection test set in an embodiment of the present invention; Figure 5 The above are comparison diagrams of sensor trusted gating ablation and hierarchical inference latency in embodiments of the present invention, wherein (a) is the gating ablation result diagram under offline fault injection conditions, and (b) is the inference latency result diagram under a specified operating environment. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0027] Example 1 like Figure 2 As shown, this invention provides a hierarchical fault diagnosis method for precision analytical instruments. Before component state reasoning, a sensor reliability assessment is performed. Based on sensor reliability, calibration flags, and calibration uncertainties, sensor parameters are determined to be in an original reliable, calibrated reliable, or invalid state. During the normal monitoring phase, only the local sensitive subgraph corresponding to the target fault mode is calculated. The complete heterogeneous dynamic graph is calculated only after the triggering condition is met. The graph model output also needs to be calibrated for consistency with the physical feasible domain corresponding to the equipment instance and operating conditions. Knowledge rules and models are updated only using samples confirmed by maintenance records or manual review. The above steps sequentially define the sensor input state, graph reasoning range, fault output conditions, and update sample source.

[0028] Specifically, the following steps are included: S1: Collect parameters and data from precision analytical instruments, perform time alignment and sliding window segmentation; obtain a heterogeneous dynamic graph that updates with the sliding window; In this embodiment, a precision analytical instrument refers to an instrument or device that includes multiple functional units such as sample introduction, signal excitation or conversion, transmission or separation, detection, environmental maintenance, power supply and control, and is capable of outputting analytical results, component status parameters, environmental operating condition parameters and control event data.

[0029] Multi-source operational data of the device under diagnosis is collected according to a preset sampling period. This multi-source operational data includes: device functional output data, such as measurement results, response characteristics, and performance indicators; core component status data, such as monitoring parameters reflecting mechanical, thermal, electrical, fluid, vibroacoustic, kinematic, or material degradation states; environmental and operating condition data, such as environmental parameters, load parameters, operating modes, and operating condition markers; and control and event information, such as control commands, feedback signals, alarm records, maintenance records, and status switching records. The collected data must include at least sensitive parameters related to the preset target failure mode, critical failure mode, or critical performance degradation state.

[0030] Data quality marking, time alignment, resampling, delay compensation, sliding window segmentation, state marking, and robust normalization are performed on the raw data to obtain the input features of the heterogeneous dynamic graph. For missing, spiked, biased, or drifting data, only anomaly candidates and valid masks are recorded before sensing reliability assessment; automatic completion or correction that would eliminate the corresponding fault representation is not performed. For space principal component analyzers or trace gas analyzers, the 96 types of operational parameters collected can be functionally divided into eight groups: The first group consists of ion source and filament parameters, including filament current, filament emission current, filament cumulative working time, electron energy, anode voltage, and ion source operating status; the second group consists of vacuum parameters, including ion pump voltage, ion pump current, vacuum chamber pressure, vacuum ready flag, and vacuum abnormality flag; the third group consists of ion optics parameters, including lens voltage, deflection or focusing voltage, and corresponding control status; the fourth group consists of power supply parameters, including power supply voltage, current, enable status, and power fault codes for each circuit; the fifth group consists of temperature parameters, including ion source temperature, mass spectrometer temperature, electronics temperature, and ambient temperature; the sixth group consists of gas path and valve parameters, including gas supply pressure, flow rate, valve on / off status, and sample injection status; the seventh group consists of control and communication parameters, including operating stage, control commands, feedback status, communication status, and system error codes; and the eighth group consists of output response parameters, including mass channel response, peak intensity, baseline, noise, and measurement results. Unconfigured parameters for different models are represented by a missing mask, and it is not required that every device has the same number of fields.

[0031] Continuous parameters form five-dimensional sensing feature nodes; state variables form five-dimensional state feature nodes consisting of the current state code, state duration, number of switches within the window, time since the most recent switch, and proportion of invalid states. Parameter groups, field names, units, ranges, resolutions, sampling periods, associated device instances, and valid operating conditions are written into the field configuration table, enabling the mapping from original fields to graph nodes to be reproduced.

[0032] To eliminate the differences in dimensions and benchmarks between different models, different equipment instances, and different operational phases, the first... Parameters of continuous components in equipment instances Operating parameters Sampled values according to Perform robust normalization; in, This is the median of historical normal operating data. This represents the median absolute deviation. It is a positive number not less than the corresponding sensor resolution; when the MAD of historical data is zero, it utilizes... To avoid zero denominators, baselines are saved separately for each equipment instance and operating condition. Recalibration is performed after sensor replacement, maintenance, or changes in operating condition distribution. Each sliding window contains... There are sampling points, and the window step size is . For each continuous parameter, calculate at least the window center value, absolute deviation of the window center value, least squares change slope, proportion of sampling points outside the normal operating range, and proportion of missing sampling points to form a five-dimensional node attribute vector. For switch quantities and status codes, calculate the duration of each state, the number of switching cycles, and the time interval from the current window center time.

[0033] Data quality screening is performed in the following order: First, delete records where the number of fields does not match the header; then mark invalid characters and values ​​exceeding the sensor's design range as missing; if there are no more than 3 consecutive missing values, use linear interpolation; if there are more than 3 consecutive missing values, retain the missing value marker; timestamps are resampled according to 2 seconds or the actual sampling cycle of the equipment, and the operating conditions are divided by the moment of control state change. When different field codes are used in different sites, first delete DCBH, DCBW and their following parameter numbers and "engineering value" suffixes, then map them to a unified field according to the Chinese parameter names.

[0034] Functional components and sensing units are used as physical component nodes. Each physical component node corresponds to a core component, functional unit, connection interface, execution unit, or sensing unit in the device being diagnosed. Its attributes include rated operating parameters, design boundaries, cumulative operating time, current working status, health status, degradation stage, and maintenance status.

[0035] The sensing feature nodes correspond to preprocessed features such as time domain, frequency domain, time-frequency, trend, statistics, control, and output response. Their attributes include real-time feature values, sliding window statistics, rate of change, normal fluctuation range, degradation critical range, and fault interval.

[0036] Using failure modes and maintenance rules as prior knowledge nodes, corresponding to existing failure modes, fault tree rules, failure mode and impact analysis results, historical failure cases, degradation thresholds, handling strategies and maintenance knowledge, etc., their attributes include failure name, scope of impact, triggering rules, hazard level, handling strategy, knowledge source and version information, etc.

[0037] Define the edge types and weight update rules for the heterogeneous dynamic graph. Edge types include physically coupled edges, data statistical edges, and knowledge rule edges, resulting in a heterogeneous dynamic graph that updates with the sliding window. Physical coupling edges are determined based on equipment structure, functional relationships, energy or matter transfer relationships, and control relationships. They encompass the relationships between components, functional units, and components and sensing features. The weights of physical coupling edges are updated when the instrument hardware configuration or operating mode changes, and are represented as follows: ; in, The correlation strength is represented by a value between 0 and 1. This is the operating condition correction factor; it is 0 when the relevant component is not in operation under the current operating condition, 0.5 when it is in operation but in the transition phase, and 1 when it is in stable operation.

[0038] The data statistics edge is established based on the correlation, time delay consistency or mutual information of different sensing features within the sliding window. Its weight is recalculated each time the sliding window is updated, and can be determined by indicators such as Pearson correlation coefficient, Spearman correlation coefficient, mutual information, dynamic time warping distance or covariance.

[0039] The data statistics are updated within each sliding window. First in - to Calculate the time delay correlation coefficient between two feature sequences within the specified range, and take the one with the largest absolute value. and corresponding delay ; then calculate the consistency of common changes ,in and This is a robust metric for the difference sequence. The weights of the data statistics edges are represented as follows: ; in, The normalized feature sequence corresponding to the i-th continuous parameter and the normalized feature sequence corresponding to the j-th continuous parameter are in the optimal time delay The correlation coefficient at the location, For the maximum search latency, To ensure consistency across variations, clip[·,0,1] indicates that the value is truncated to between 0 and 1; in this embodiment, only the value is retained. Edges with a minimum strength of 0.35.

[0040] Knowledge rule edges are established based on causal relationships defined in fault tree rules, expert knowledge, historical cases, or maintenance knowledge. Connections can be established between fault mode nodes and corresponding sensitive component nodes and abnormal feature nodes. Their weights are adjusted when fault modes are added or updated in the knowledge base.

[0041] The initial weights of the knowledge rule edges are given by fault tree analysis, failure mode and effects analysis, or maintenance procedures. After the rules are put into operation, they are assigned weights according to... The update uses a one-plus smoothing method to avoid abrupt changes in edge weights caused by a small number of samples.

[0042] in, The initial confidence level for the rule. and These represent the number of confirmed hits and the number of mismatches, respectively.

[0043] In addition, when the working condition changes or abnormal triggering conditions are met, the weights of all edges are re-evaluated and updated.

[0044] S2: To avoid misjudging sensor malfunctions as degradation of the tested component, six types of sensor fault characteristics are extracted before component condition diagnosis. Sensing reliability assessment is performed on the sensing unit; the five-dimensional node attributes, consisting of the window midpoint, window midpoint absolute deviation, normalized change slope, out-of-limit ratio, and missing ratio, are input into a local graph neural network. The network outputs the probability of a normal state and the corresponding sensor fault state probabilities for bias, drift, missing, spike, stagnation, and accuracy degradation of the sensing parameters. Sensing reliability is determined based on the probability of a normal state. Let z(t) be the normalized sequence of the current window, and let z(t) be the robust scale of the normal window. Window length is Bias fault is represented as Drift fault is represented as Missing faults are represented as Or it may continue to approach invalid coding; spike faults are represented as ,in Large values ​​are only observed at a few sampling times; jamming faults are characterized by continuous sampling value changes that are less than the sensor resolution; accuracy degradation is characterized by a continuous increase in the residual variance of the sensing value relative to the estimated value of the associated parameter.

[0045] Bias score Depend on The drift score is obtained by multiplying the smaller value of 1 by the low slope indicator. Depend on The smaller value of 1 is obtained; missing score Depend on The smaller value of 1 is obtained; peak score The lag score is obtained by normalizing the proportion of sampling points within the window that satisfy |z| > 5 and whose adjacent points recover to |z| < 3 by a factor of 0.05. The accuracy degradation score is set to 1 if the window range is no greater than twice the sensor resolution and the absolute deviation of at least one associated parameter is greater than 1; otherwise, it is set to 0. It is obtained by normalizing the ratio of the robust scale of the current correlation estimation residual to the robust scale of the normal residual. It is determined by the 99.5th percentile of the absolute value of the slope of the normal validation set.

[0046] In a sensor-embedded local graph neural network, the model outputs the probability of a normal state. and the probability of six types of fault states, and according to Determine sensor reliability; in a degraded implementation where model files are unavailable, according to... The reliability of the calculation rules is assessed and a manual review is requested. The component failure is not automatically confirmed based on the downgrade result. Use the original parameters when the value is not lower than 0.80; Use the corrected confidence parameter when the value is below 0.80, the model output bias or drifts, and the corrected association estimate residuals do not exceed three times the robust scale of the normal residuals; otherwise... For values ​​below 0.80, set invalid parameters. For bias and drift, respectively... and Correction: For short-term missing or isolated spikes, substitute them with robust regression values ​​of related nodes under the same operating conditions; retain invalid parameters when there are continuous missing, stuck, or substitution residuals exceed limits. Set a missing parameter mask to prevent invalid parameters from being the sole criterion for determining the degradation of the corresponding physical component. The above thresholds are the initial implementation parameters and will be recalibrated by the validation set after obtaining confirmed samples.

[0047] This embodiment pre-configures local sensitive subgraphs for typical fault modes, including component performance degradation, structural or connection anomalies, transmission path anomalies, blockages or leaks, power supply or control anomalies, sensor anomalies, output response anomalies, and communication anomalies. The local sensitive subgraphs retain only nodes and edges that are physically coupled, statistically correlated, or associated with specific fault modes. Their size is smaller than the complete heterogeneous dynamic graph, and they can be expanded, pruned, or reconfigured based on operating conditions, node states, or edge weight changes.

[0048] The local sensitive subgraph is generated according to the following steps: using the prior knowledge node corresponding to the target failure mode as the seed, search for one-hop and two-hop adjacent nodes; delete edges with a weight lower than 0.35 that are not mandatory physical connections; calculate the contribution score based on the node deviation, the product of the path edge weights from the seed node to the candidate node, and the historical rule hit rate; when there are more than 32 nodes, retain them in descending order of contribution score. The resulting local sensitive subgraph remains consistent when the input graph and parameter configuration remain unchanged.

[0049] When the absolute deviation of any node not included in the subgraph reaches 3, or the change of the data statistics edge connected to that node relative to the previous window reaches 0.20, the node and its one-hop adjacent node are temporarily incorporated into the local sensitive subgraph; the temporary node is removed after 5 consecutive windows have returned to normal.

[0050] For filament-related diagnostics, two levels of local sensitivity sub-graphs are set up. The first level is the local sensitivity sub-graph of the filament current sensor, whose nodes include at least the filament current sensor, the five-dimensional characteristics of the filament current, the filament power supply, the filament enable state, the electron energy, the anode voltage, the filament emission current, and knowledge nodes of bias, drift, missing, spike, jamming, and accuracy degradation. The output vector corresponds to the normal and the above six types of sensor states respectively.

[0051] The second level is a local sensitivity subgraph of filament state, whose nodes include at least filament components, filament current characteristics assessed with confidence, filament emission current characteristics, ion source temperature characteristics, filament cumulative operating time, ion pump voltage, ion pump current, and filament aging knowledge nodes. Only when the output of the first level is reliable, or when the correction value given by the first level meets the correction uncertainty requirement, will the filament current node participate in the second-level inference as a valid state; otherwise, its missing mask is set to 1, and the other related nodes provide low-confidence candidate conclusions.

[0052] S3: Extract local sensitive subgraphs from the heterogeneous dynamic graph, use a local graph neural network to reason about the local sensitive subgraphs to obtain the fault probability and prediction uncertainty, and determine whether the local reasoning result reaches a preset anomaly threshold; if it does not reach the preset anomaly threshold, do not perform full graph reasoning, and use the local fault probability and local prediction uncertainty as the reasoning result of the current sliding window; if it reaches the preset anomaly threshold, switch to full graph reasoning mode, use a heterogeneous graph neural network that distinguishes different edge types to perform full graph reasoning on the heterogeneous dynamic graph to obtain the full graph fault probability and full graph prediction uncertainty, and use the full graph fault probability and full graph prediction uncertainty as the reasoning result of the current sliding window; Under normal monitoring conditions, the system invokes local sensitive subgraphs corresponding to the current operating condition or target fault type, and uses a lightweight graph neural network model for feature aggregation, outputting fault probability and uncertainty estimates. The lightweight model includes graph convolutional networks or graph attention networks, and uncertainty is assessed through model ensemble or the degree of dispersion of the predicted probability distribution.

[0053] The lightweight model employs a two-layer graph convolutional network. Let the adjacency matrix after adding self-loops be... The degree matrix is The node attribute matrix is Then the first layer output The second layer is obtained according to the same normalized adjacency relation. ;right After graph-level average pooling, the probabilities of each failure mode are obtained through a fully connected layer and a Softmax function. An executable configuration is as follows: input dimension 5, first hidden layer 16 dimensions, second hidden layer 8 dimensions, and random inactivation rate 0.20.

[0054] The complete heterogeneous dynamic graph employs a relation-type-aware heterogeneous graph neural network. Normalized adjacency matrices are constructed for physically coupled edges, data statistical edges, and knowledge rule edges, respectively. Each relation matrix performs adjacency propagation on the hidden states of nodes in the previous layer. The propagation results of the three types are concatenated along the feature dimension, and a new hidden state of the node is obtained through a trainable transformation matrix (excluding relation sharing) and nonlinear activation. Finally, graph-level average pooling and a classification layer output the fault probability. This structure preserves the source differences of the three types of edges and does not require unconnected nodes to occupy dense relation edges.

[0055] The model training set is divided by device instance and continuous runtime segments. Each continuous runtime segment forms a training segment, a validation segment, and a test segment in chronological order. The sliding window is generated only within the corresponding segment, and sampling points from the same window must not enter different datasets. The normal operation domain, data statistics edges, and normalization parameters are calculated only for the training segment. The joint loss includes cross-entropy loss, physical consistency loss, Brier calibration loss, and L2 regularization term. An executable configuration uses a learning rate of 0.001, a batch size of 32, and a maximum of 120 training epochs. Training stops and the parameters of the lowest-period validation loss are restored when the validation loss does not decrease for 15 consecutive epochs.

[0056] The model training phase and the online inference phase are executed separately. During the training phase, window samples containing device instance identifiers, operating condition identifiers, continuous running segment identifiers, and confirmation labels are read to construct node feature tensors. Adjacency tensor and node effective mask , where B is the batch size; the complete heterogeneous dynamic graph uses an adjacency list stored by edge type to avoid allocating dense matrices to unconnected nodes.

[0057] The training output should include at least the model weights, field configuration versions, normal operating domain versions, physically feasible domain versions, threshold configurations, and the training data time range. During the online inference phase, only the above frozen versions are read; model weights are not modified during a single inference iteration. Newly confirmed samples are first added to the update library, and the model version is only switched after a comparison before and after updating is completed on an independent validation set.

[0058] The sensor local subgraph and the component state local subgraph use independent parameters. The former is trained with labels of normal, bias, drift, missing, spike, hysteresis, and accuracy degradation, while the latter is trained with labels of normal, performance degradation, and boundary state. The output of the sensor local subgraph is passed to the component state local subgraph through three input states: original reliable, corrected reliable, or invalid. The two subgraphs do not share the final classification layer.

[0059] The uncertainty prediction is calculated using stochastic deactivation inference. During the inference phase, the stochastic deactivation layer remains active, and the calculation is performed on the same graph. Next forward computation, The range is 10 to 50; the variance of each failure probability is used to characterize the model uncertainty, and the normalized entropy of the average prediction distribution is used as the data uncertainty. The two are weighted to obtain the prediction uncertainty between 0 and 1.

[0060] The two-level local sensitivity subgraphs can each employ a parameter-independent two-layer graph convolutional network. The softmax output dimension of the sensor subgraph is 7, and the specific parameters for each running configuration are shown in Table 1. The output order is normal, bias, drift, missing, spike, hysteresis, and accuracy degradation. The filament state subgraph outputs at least three probabilities: normal, filament performance degradation, and boundary state. Both models simultaneously output the predicted class, maximum class probability, prediction uncertainty, the three nodes with the highest contribution, and their corresponding key edges for use in the complete heterogeneous dynamic graph and for manual review.

[0061] Table 1 Model and runtime configuration parameters

[0062] In S3, prior knowledge nodes are used as seed nodes. Local sensitive subgraphs are extracted from the heterogeneous dynamic graph based on edge weights and node contributions. Local sensitive subgraphs corresponding to the current operating condition or target fault type are called, and a lightweight model is used for feature aggregation to output fault probability and uncertainty estimates. One-hop and two-hop adjacent nodes with edge weights of not less than 0.35 are retained. When the number of nodes exceeds 32, node contributions are calculated based on node deviation, the product of path edge weights from the seed node to the candidate node, and the historical rule hit rate. Nodes are then pruned in descending order of contribution to obtain local sensitive subgraphs.

[0063] When any of the following factors—failure probability, prediction uncertainty, node deviation, or statistical edge change—reaches its corresponding trigger threshold, a heterogeneous graph neural network that distinguishes different edge types is used to perform full-graph inference on the heterogeneous dynamic graph. The system automatically switches to full-graph inference mode when any of the following abnormal trigger conditions are met: (1) The failure probability of the local sensitive subgraph output reaches or exceeds the preset trigger threshold; (2) The uncertainty of the output of the local sensitive subgraph reaches or exceeds the preset uncertainty trigger threshold; (3) The attribute of at least one physical component node or sensing feature node exceeds the normal operating range or the trend of change exceeds the limit; (4) The weight, weight change rate, or time delay relationship of at least one data statistical edge exceeds the normal range; (5) The relationships between different nodes exhibit a preset abnormal pattern; (6) The equipment experiences a change in operating conditions, a sudden change in load, a change in control status, or an abnormality in communication quality.

[0064] The aforementioned trigger thresholds are calibrated through fault injection experiments, simulation data, historical operating data, equipment design parameters, or expert knowledge. Upon entering the full-graph fine-grained reasoning mode, the system employs a deep graph neural network model to aggregate features from the complete heterogeneous dynamic graph and outputs the fault prediction results after full-graph aggregation.

[0065] When a field is missing, a timestamp is interrupted, or a communication anomaly causes the proportion of valid nodes in the current window to fall below the configured value, the system does not output a clear component failure, but instead outputs a data quality anomaly and retains the diagnostic results of the previous trusted window. When a local sensitive subgraph becomes disconnected due to the deletion of data statistics edges, the system retains the preset physical coupling edges and node self-loops, so that the local graph can still perform inference.

[0066] When the complete heterogeneous dynamic graph inference fails to complete within the preset maximum response time, or when the model file or field configuration version is inconsistent with the normal operating domain version, the system suspends automatic fault confirmation for that window, retains the output of the local sensitive subgraph and increases prediction uncertainty, and records timeout or version inconsistency events. This degradation path prevents computational or configuration anomalies from being interpreted as equipment component failures.

[0067] In this embodiment, the sensor local graph neural network sequentially maps the node attribute vector formed by window statistical features into 16-dimensional and 8-dimensional hidden states, and reads the second-layer hidden state of the seed sensor node for classification. The component state local graph neural network uses graph-level average pooling and a classification layer to output the fault probability.

[0068] If the failure probability output by the local sensitive subgraph is not less than 0.40, the prediction uncertainty reaches the uncertainty trigger threshold, the absolute deviation of the sensing feature node is not less than 3, or the weight change of any data statistical edge relative to the previous sliding window is not less than 0.20, switch to full graph inference. The uncertainty trigger threshold is the larger of the 95th percentile of the prediction uncertainty on the normal validation set and 0.15. The heterogeneous graph neural network performs adjacency propagation on physical coupling edges, data statistics edges and knowledge rule edges respectively, and then performs feature transformation and graph-level pooling after concatenating the propagation results of different edge types.

[0069] In S3, the training phase and online inference phase of the lightweight model are executed separately. The training set is divided into device instances and continuous running segments. Each continuous running segment is divided into training segment, validation segment and test segment in chronological order. The same sliding window does not cross different datasets. Sensor local graph neural network, component state local graph neural network and relation type perception heterogeneous graph neural network for full graph inference are trained separately. The model weights, node order, category order, relation matrix, normal operating domain version and model summary are saved.

[0070] During the online inference phase, the frozen model weights are invoked. For the same fault mode, at least three samples are obtained that have been verified by maintenance records or manual review. After the samples come from at least two independent running segments, an independent model update and verification process is initiated.

[0071] S4: Reliable fault determination is based on physical feasible region constraints. Through fault injection experiments, semi-physical simulation, historical operating data, equipment design parameters, and expert knowledge, the physical feasible region of each node's attributes and edge weight relationships is pre-defined and divided into normal operation intervals, degradation critical intervals, and fault intervals according to operating status. During inference, abnormal features identified by the model are compared with the physical feasible region: if node attributes or relationships deviate from the physical feasible region, the output uncertainty is increased according to the degree of deviation to reduce misjudgments caused by physically infeasible abnormal patterns; if abnormal features fall within the degradation critical interval or fault interval, and the relationships conform to preset physical mechanisms or knowledge rules, the output uncertainty is maintained or reduced.

[0072] Taking filament aging as an example, the physical consistency score The weighted average is obtained under the following conditions: the filament current continuously increases relative to the baseline of this equipment and operating condition; the filament emission current continuously decreases relative to the baseline; the ion source temperature increases relative to the baseline; and the statistical weights or time delay relationships between the filament current and the filament emission current continuously change. The four weights can be set to 0.30, 0.30, 0.20, and 0.20 respectively. Here, "increase" or "decrease" refers to longitudinal deviation within the equipment, and absolute value comparisons between different equipment are not used.

[0073] Let the fault probability output by the graph neural network be... Prediction uncertainty is Then you can press Calibration is performed. This formula suppresses high-probability outputs lacking mechanistic support; for predictions conforming to fault rules, and The basic consistency is maintained. Consistency conditions and weights are set separately for different failure modes.

[0074] The decision threshold is primarily determined by the normal validation set and confirmed fault samples. In the initial deployment phase when samples are insufficient, the fault probability trigger threshold, explicit fault threshold, and uncertainty threshold can be set to 0.40, 0.80, and 0.15, respectively. After obtaining confirmed samples, the initial values ​​are replaced with validation set quantiles that meet the preset false alarm limit.

[0075] The physical consistency score was obtained by weighting the four filament aging conditions at 0.30, 0.30, 0.20, and 0.20. And combine effective sensing reliability to calibrate the failure probability. and forecast uncertainty In the initial parameter configuration, a definite fault is defined when the probability after calibration is not lower than 0.80 and the uncertainty is not higher than 0.15; a low-confidence boundary state sample is defined when the probability is between 0.40 and 0.80, the prediction uncertainty is higher than 0.15, or the sensing parameter is in the boundary state; the rest are defined as normal samples.

[0076] The fault probability, prediction uncertainty, sensing reliability, and physical feasible region obtained by full graph reasoning or local sensitive subgraph reasoning are consistent. If the verification is consistent, the fault judgment result of the current sliding window is output as the diagnosis result. For clearly defined fault samples and low-confidence boundary state samples, a fault evidence chain is extracted. The evidence chain can be obtained through methods such as attention weights, node contributions, edge contributions, or fault rule matching, and is used to record the correspondence between abnormal features, related components, and fault modes. The extracted evidence chain is then compared and verified with knowledge graphs, fault tree rules, historical cases, failure mode and effect analysis results, semi-physical simulation data, or expert knowledge.

[0077] In this embodiment, for filament condition diagnosis, the filament current sensor status is first output using the local sensitivity subgraph of the filament current sensor. Then, the original or corrected reliable parameters are input into the local sensitivity subgraph of the filament status. The physical consistency score of filament aging is calculated based on the increase of filament current relative to the baseline of this device, the decrease of filament emission current relative to the baseline of this device, the increase of ion source temperature, and the statistical edge change between filament current and filament emission current. When the filament current is an invalid parameter, the predictive uncertainty of filament condition diagnosis is increased, and the output of a clear conclusion of filament performance degradation supported only by the filament current is prohibited.

[0078] S5: If the verification is inconsistent, mark it as a potential new fault mode or a sample to be reviewed, and proceed to the remote review or expert analysis process. S6: For potential new failure modes or samples to be reviewed, small sample augmentation is performed using covariance-related perturbations constrained by the physical feasible region. After augmentation, the data statistical edges are recalculated, and samples that do not meet the failure rules, operating condition consistency, or time series continuity are deleted before returning to S3.

[0079] Small sample augmentation employs constrained perturbations, rather than generating data without restrictions. For a confirmed sample x, the feature covariance matrix under the same operating condition is used. Mid-sampling perturbation and order ,in Take a value between 0.02 and 0.10. This represents the physical feasible region corresponding to the failure mode. Recalculate the data statistics edges, retaining only samples with unchanged operating condition labels, no sudden jumps in the time series, and a physical consistency score of not less than 0.8.

[0080] Knowledge updates require a confirmation threshold. At least three samples, verified by maintenance records or manual review, must be obtained for the same new fault mode, and these samples must come from at least two independent runtime segments before adjustments to knowledge rule edges and decision thresholds are allowed. Unconfirmed low-confidence samples are only added to the review pool. Model version, parameter summary, training data time range, and rule change records are saved before and after updates for rollback and traceability.

[0081] like Figure 1 As shown, a hierarchical fault diagnosis system for a precision analytical instrument includes a data processing module, a heterogeneous dynamic graph construction module, a sensor reliability assessment module, a hierarchical reasoning module, a reliability determination module, a model version management module, a knowledge iteration module, and an interactive output module. The output of the data processing module is connected to the heterogeneous dynamic graph construction module; the output of the heterogeneous dynamic graph construction module is connected to the sensing reliability assessment module and the hierarchical inference module; the sensing reliability assessment module sends the original reliability parameters, corrected reliability parameters, or invalid parameters to the hierarchical inference module; the output of the hierarchical inference module is connected to the reliability determination module; the reliability determination module sends explicit fault samples and boundary state low-confidence samples to the knowledge iteration module; the model version management module provides a consistent configuration version to the heterogeneous dynamic graph construction module, the sensing reliability assessment module, the hierarchical inference module, and the reliability determination module; the knowledge iteration module only updates the corresponding version after independent verification is passed; the interactive output module receives the reliability determination results and diagnostic evidence, which are used to implement the steps of a hierarchical fault diagnosis method for a precision analytical instrument.

[0082] The data processing module is used to collect multi-source operational data, perform time alignment, sliding window segmentation, and window statistical feature calculation to form sensor feature nodes.

[0083] The heterogeneous dynamic graph construction module is used to create physical component nodes, sensor feature nodes, and prior knowledge nodes, configure physical coupling edges, data statistics edges, and knowledge rule edges, and update the dynamics with the sliding window.

[0084] The sensor reliability assessment module is used to output the probabilities of normal, bias, drift, missing, spike, jamming and accuracy degradation, calculate the sensor reliability, and determine the sensor parameters as original reliable parameters, corrected reliable parameters or invalid parameters.

[0085] The hierarchical reasoning module is used to extract local sensitive subgraphs and switch between local sensitive subgraph reasoning and full heterogeneous dynamic graph reasoning, outputting the failure probability and prediction uncertainty.

[0086] The reliability determination module is used to calibrate the failure probability and prediction uncertainty based on the sensor reliability and the physical feasible region to obtain the failure determination result.

[0087] The model version management module is used to save model weights, node order, category order, relation matrix, normal operating domain version and model summary, and to stop automatic fault confirmation when versions are inconsistent.

[0088] The knowledge iteration module is used to receive maintenance records or manual review results, and initiate constrained sample augmentation, model update and independent validation after the confirmation sample conditions are met.

[0089] The interactive output module is used to output fault type, sensor reliability, maximum category probability, prediction uncertainty, associated components, abnormal characteristics, evidence chain and handling suggestions, and to interact with local monitoring terminals, remote management platforms or manual review terminals.

[0090] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a hierarchical fault diagnosis method for a precision analytical instrument.

[0091] Example 2 This embodiment focuses on the mass spectrometry unit in a trace gas analyzer. This unit includes a filament, ion source, vacuum system, detector, sample introduction system, and power supply system. The diagnostic program is deployed in an edge computing unit connected to the instrument controller and executes the following steps: Step 1: Collect operational data and construct a heterogeneous dynamic graph. In this embodiment, data such as filament current, filament emission current, cumulative filament operating time, ion source temperature, mass spectrometer tube temperature, ion pump voltage, ion pump current, electron energy, anode voltage, filament enable, vacuum ready flag, operating stage, and system error status are collected according to the equipment instance and operating conditions. Sampling records are written to a timestamped circular buffer. The field mapping unit standardizes field names, units, and status codes. Transitional segments after startup, shutdown, range switching, and maintenance are only used for status recording and are not used to update the normal operating domain.

[0092] Taking a sampling period of approximately 2 seconds as an example, a 120-second window is formed by 60 sampling points, and the online window step size is 10 sampling points, meaning the diagnosis is updated every 20 seconds. The median and median absolute deviation of each continuous parameter are calculated from historical normal segments of the same equipment under the same stable operating conditions, and robust normalization is performed; the window node attributes are taken in the following order: window median, window median absolute deviation, normalization change slope, out-of-limit ratio, and missing ratio.

[0093] ; Using functional components and sensing units as physical component nodes, each physical component node includes at least the filament, ion source, vacuum system, and power supply system. Using fault modes and maintenance rules as prior knowledge nodes, each prior knowledge node includes at least filament aging, ion source contamination, vacuum leakage, power supply anomaly, and communication anomaly. Physical coupling edges are established based on component connections and energy or matter transfer relationships. Data statistics edges are updated based on the time delay correlation and common change consistency of the current window. Knowledge rule edges are established based on the fault tree and maintenance rules. Using the filament aging knowledge node as a seed, one-hop and two-hop nodes with edge weights not less than 0.35 are retained. When the number of nodes exceeds 32, the contribution is calculated based on node deviation, the product of path edge weights, and historical rule hit rate, and then pruned to obtain a local sensitivity subgraph of the filament state.

[0094] The data statistics edges are recalculated within each window. For two continuous features, the position with the highest absolute correlation coefficient within the preset maximum search delay is selected, and then combined with the delay attenuation term and the consistency of common changes to form an edge weight between 0 and 1. If the missing proportion of either of the two features involved in the calculation exceeds 0.20, the corresponding edge weight is not updated in this window and an invalid flag is recorded; if the number of valid samples is less than half the window length, the edge structure of the previous reliable window is used, but the change of this edge does not trigger full graph inference. Discrete control events only establish directed edges with continuous features before and after the event, avoiding the use of long-term unchanged status codes as highly correlated continuous parameters.

[0095] Step Two: Perform a sensor reliability assessment, first determining the independent state of the filament current measurement point. The sensor's local graph neural network takes the filament current and its associated nodes (power supply, enable, electron energy, anode voltage, and filament emission current) as inputs, and outputs seven probabilities: normal, bias, drift, missing, spike, jammed, and accuracy degradation. The normal state probability is then... Determined as sensor reliability .

[0096] When the sensing confidence level is not lower than 0.80, the filament current is used as the original confidence parameter. When the sensing confidence level is lower than 0.80, the output state is biased or drifting, and the corrected correlation estimation residual does not exceed three times the robust scale of the normal residual, the filament current is used as the corrected confidence parameter. In other cases below 0.80, a missing mask is set, and the parameter is treated as invalid. The bias and drift slope are estimated by the robust regression relationship between the last confidence window before the anomaly occurs and the filament emission current, electron energy, and anode voltage. For short-term missing values ​​or isolated spikes not exceeding three sampling periods, the correlation estimate can be used as a substitute; automatic substitution is not performed when there are continuous missing values, lag, or when the substitution residual exceeds the limit.

[0097] ; When the sensor model file is unavailable, a degradation reference value is calculated based on the scores of six rule categories: bias, drift, missing values, spikes, jamming, and accuracy degradation. This reference value is only used to trigger manual review and is not used to automatically confirm component failures; filament current, even in a degraded state, must not be used as the sole criterion for determining filament performance degradation.

[0098] Step 3: Perform local-priority hierarchical graph inference. Under normal monitoring conditions, the filament current sensor subgraph is first processed by a two-layer sensor local graph neural network, and then the original or corrected confidence parameters are fed into the filament state local sensitive subgraph. The node input dimension of both local models is 5, and the hidden layer dimensions are 16 and 8 respectively, with a random deactivation rate of 0.20. The component state local model outputs the probabilities of normal, filament performance degradation, and boundary states after graph-level average pooling. Random deactivation inference is executed 20 times, and the prediction uncertainty is formed by the class probability variance and the average prediction entropy.

[0099] When the filament performance degradation probability is not less than 0.40, the prediction uncertainty reaches the larger of the 95th percentile of the normal validation set and 0.15, the absolute deviation of any node is not less than 3, or the weight change of any data statistics edge relative to the previous window is not less than 0.20, the system switches to the complete heterogeneous dynamic graph. The complete graph model performs adjacency propagation on physically coupled edges, data statistics edges, and knowledge rule edges respectively. After concatenating the three types of propagation results, feature transformation and graph-level pooling are performed. If the triggering conditions are not met, the local sensitive subgraph continues to be used.

[0100] Step 4: Based on the reliable fault determination output from the physical feasible domain, and assuming the filament current is valid, sequentially determine the following: filament current increases relative to the baseline of this device; filament emission current decreases relative to the baseline of this device; ion source temperature increases; and the statistical edge between the filament current and filament emission current changes. These four conditions are weighted at 0.30, 0.30, 0.20, and 0.20 to form a filament aging physical consistency score, which is then combined with the effective sensing reliability to calibrate the model probability and prediction uncertainty.

[0101] ; When the post-calibration probability is not lower than 0.80 and the post-calibration uncertainty is not higher than 0.15, a clear filament performance degradation is output; when the probability is between 0.40 and 0.80, the post-calibration uncertainty is higher than 0.15, or any critical parameter is in a boundary state, a low-confidence boundary state sample is output; otherwise, normal outputs are provided. When the filament current is marked as invalid, the component diagnostic uncertainty is increased, and a clear degradation conclusion supported solely by the filament current is prohibited from being output. A persistent anomaly is only confirmed when the same anomaly condition is met for three consecutive windows.

[0102] S5. Record diagnostic evidence and perform controlled updates. For each diagnosis, save the device instance, operating conditions, window start and end times, field configuration version, model version, normal operating domain version, local subgraph nodes, full graph triggering reasons, sensor input status, probabilities before and after calibration, prediction uncertainty, and evidence chain. After maintenance, sensor replacement, or hardware configuration switching, freeze the original normal operating domain. After obtaining 30 to 100 stable, alarm-free, and reliable sensing windows, generate a new operating domain. During recalibration, only output the post-maintenance pending calibration status.

[0103] Only samples confirmed through maintenance records or manual verification are allowed to be written into the knowledge base. An independent model update and verification process is initiated only after at least three confirmed samples for the same fault mode are obtained, and these samples come from at least two independent runtime segments. If the performance of the verification set degrades or the false alarm rate exceeds that of the original model, the original model and the original knowledge base are retained. After communication interruption is recovered, the counting restarts from the first complete window, without splicing discontinuous data before and after the interruption.

[0104] Example 3 This embodiment uses historical operation files from four field trace gas analyzers to illustrate the training interface and verification process. The raw data includes 153 CSV files with 345,385 rows of records, exhibiting two field structures: 204 columns and 278 columns. The processing program eliminates differences in field prefixes and engineering value suffixes based on Chinese parameter names, mapping filament current, filament emission current, ion source temperature, mass spectrometer tube temperature, ion pump voltage, ion pump current, electron energy, anode voltage, cumulative filament operating time, and status flags to unified fields. Field files primarily use GB18030 encoding; when a read fails, the filename and error location are recorded, and missing files are not included in the training data.

[0105] Training data is organized by device instance and continuous runtime segments. Each segment is divided chronologically into a 60% training segment, a 20% validation segment, and a 20% test segment. Windows must not cross segment boundaries, and adjacent windows within the same continuous segment must not be assigned to different datasets. The offline training window length is 60 sampling points, and the step size is 30 sampling points. Online inference still uses an update step size of 10 sampling points. Device baseline, robustness scale, normal operating domain, and data statistics edges are calculated only from the training segment.

[0106] The sensor local model contains 7 parameter nodes and employs a two-layer graph convolutional network with an input dimension of 5 and hidden layers of 16 and 8. It reads the second hidden state of the filament current seed node to output seven types of sensing state probabilities. The component state local model uses the same hidden dimension and outputs normal, filament performance degradation, and boundary state probabilities through graph-level average pooling. The complete graph model contains 10 parameter nodes, propagating for the three types of relationships separately, with hidden layer dimensions of 24 and 12. All three models employ a Brier calibration loss with a random inactivation rate of 0.20 and a weight of 0.10, and an L2 regularization term with a weight of 0.0001; the two component state models also employ a physical consistency loss with a weight of 0.20.

[0107] ; Training uses a fixed random seed and saves model weights, node order, class order, three-class relation matrices, device baseline, normal operating domain version, and SHA-256 digests of model files. Online inference only loads frozen models; automatic fault acknowledgment stops if any model file, node order, or relation matrix version is inconsistent. The convergence process of the validation set loss for the three models is described in [link to documentation]. Figure 3 .

[0108] The existing field records do not simultaneously provide the start and end times and type of the fault confirmed by maintenance personnel; therefore, the operating status code is not directly used as the ground truth for supervised learning. Field data is only used to verify field access, time segmentation, equipment baseline, and graph relationship calculation; classification performance is evaluated using offline fault injection samples with fixed random seeds. The injection basis is limited to stable operating candidate windows with an absolute normalized deviation of no more than 2 for the median, an absolute value of the normalized slope no more than 1.5, a missing proportion no more than 0.05, and an absolute deviation of no more than 3 for the median of the normalized window. This selection does not indicate that the corresponding window has been confirmed as healthy by maintenance personnel.

[0109] Sensor fault injection includes: applying a bias of 4 to 7 times the robust scale to the filament current; applying a linear drift with a final value of 5 to 9 times the robust scale within a window; continuously setting 18 to 31 missing points; randomly setting 5 isolated spikes with amplitudes of 8 to 12 times the robust scale; fixing the window value to the median to create hysteresis; and superimposing random perturbations with a standard deviation of 3.5 times the robust scale to create accuracy degradation. The filament performance degradation sample simultaneously increases the filament current, decreases the filament emission current, and increases the ion source temperature; the boundary states use perturbation amplitudes smaller than those of clearly degraded samples.

[0110] The test results are shown in Table 2 and Figure 4As shown. For sensor seven-class classification, 3115 injected samples were used, achieving an accuracy of 0.9368 and a macro-average F1 score of 0.9345. After 1000 grouped self-sampling iterations based on field or equipment instances, the 95% range of the macro-average F1 score was 0.8705 to 0.9983. For component three-class classification, 744 injected samples were used, achieving both an accuracy and a macro-average F1 score of 0.9462, corresponding to a 95% range of 0.9145 to 0.9523.

[0111] During evaluation, precision, recall, and F1 scores for each category are calculated separately within each device instance, and then the arithmetic mean of the categories is taken to obtain the macro-average F1. Confidence intervals are calculated using the field or device instance as the resampling unit, not using overlapping sliding windows. Each resampling completely preserves the test window of the selected device, and after 1000 resampling iterations, the 2.5% and 97.5% quantiles are taken. The gated ablation group and the always-executed complete graph group use the same data partitioning, random seed, and number of training rounds as the method in this embodiment.

[0112] Table 2 Offline Reproduction Results and Evidence Attributes

[0113] In 2670 sensor fault injection windows, without sensor reliability gating, the proportion of sensor anomalies judged as filament performance degradation was 44.27%; after gating, it was 17.30%. See the comparison results below. Figure 5 (a) shows that determining the sensor status before injection can reduce misjudgments of component degradation caused by abnormal measurement points.

[0114] Inference latency was measured under an environment with an Intel Core i5-12400F processor, Windows 10, Python 3.12.13, and NumPy 2.3.5, with a fixed batch size of 128 samples repeated 20 times. The average latency per sample for the local graph was 0.002132 milliseconds, and for the complete graph it was 0.011378 milliseconds. The complete graph call rate for the normal window was 31.05%, and the average latency for the hierarchical structure was 0.005665 milliseconds, representing a reduction of approximately 50.2% compared to always executing the complete graph. Figure 5 As shown in (b) above. This latency varies with hardware, batch size, and software version; a record of the operating environment should be maintained during deployment.

[0115] After inputting an unmodified on-site CSV file into the frozen model, the program sequentially outputs the sensor category probability, sensor reliability, local component probability, prediction uncertainty, full graph triggering cause, and final component probability. Upon reaching the triggering condition, it calls the full graph model, completing the processing flow specified by the online inference interface. This window lacks a maintenance confirmation label; therefore, it does not indicate a genuine sensor fault or filament degradation in the corresponding instrument. The aforementioned offline indicators only reflect the software execution results and classification separability under specified injection conditions and do not replace the accuracy, recall, or early warning time of actual faults.

[0116] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0117] Therefore, the present invention employs the aforementioned hierarchical fault diagnosis method, system, and equipment for a precision analytical instrument, which collects data and constructs a heterogeneous dynamic graph containing physical components, sensing features, and prior knowledge nodes; assesses sensing reliability, corrects or invalidates abnormal parameters; performs local sensitive subgraph reasoning, switches to full graph reasoning when conditions are met, and reduces component misjudgment caused by sensing anomalies and graph computation during normal monitoring phases based on equipment baseline and physical feasible domain calibration results.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A hierarchical fault diagnosis method for a precision analytical instrument, characterized in that, Includes the following steps: S1: Collect parameters and data from precision analytical instruments, perform time alignment and sliding window segmentation; obtain a heterogeneous dynamic graph that updates with the sliding window; S2: Perform a sensing reliability assessment on the sensing unit; S3: Extract local sensitive subgraphs from the heterogeneous dynamic graph, use a local graph neural network to reason about the local sensitive subgraphs to obtain the fault probability and prediction uncertainty, and determine whether the local reasoning result reaches the preset anomaly threshold; if it does not reach the preset anomaly threshold, do not perform full graph reasoning, and use the local fault probability and local prediction uncertainty as the reasoning result of the current sliding window; If the preset abnormal threshold is reached, switch to full graph reasoning mode, use a heterogeneous graph neural network that distinguishes different edge types to perform full graph reasoning on the heterogeneous dynamic graph, obtain the full graph failure probability and full graph prediction uncertainty, and use the full graph failure probability and full graph prediction uncertainty as the reasoning result of the current sliding window; S4: Based on the physical feasible region constraint, perform reliable fault determination and perform consistency calibration on the fault probability, prediction uncertainty, sensing reliability and physical feasible region output by the current sliding window in S3. If the verification is consistent, output the fault determination result of the current sliding window as the diagnosis result; S5: If the verification is inconsistent, mark it as a potential new fault mode or a sample to be reviewed, and proceed to the remote review or expert analysis process. S6: For potential new failure modes or samples to be reviewed, small sample augmentation is performed using covariance-related perturbations constrained by the physical feasible region. After augmentation, the data statistical edges are recalculated, and samples that do not meet the failure rules, operating condition consistency, or time series continuity are deleted before returning to S3.

2. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 1, characterized in that, S1 contains parameters including component status parameters and operating condition parameters, and data including control event data and output response data. It is time-aligned and segmented into sliding windows according to equipment instance and operating condition. Calculate window statistical features within each sliding window to form sensing feature nodes; For the Parameters of continuous components in equipment instances Operating parameters Sampled values according to Perform robust normalization; in, This is the median of historical normal operating data. This represents the median absolute deviation. A positive number not less than the corresponding sensor resolution; each sliding window contains There are sampling points, and the window step size is . Each sampling point has window statistical characteristics including window median, window median absolute deviation, normalized change slope, out-of-limit ratio, and missing ratio. Using functional components and sensing units as physical component nodes, and fault modes and maintenance rules as prior knowledge nodes, physical coupling edges, data statistics edges, and knowledge rule edges are established to obtain a heterogeneous dynamic graph that updates with the sliding window. The weights of the physical coupling edges are represented as follows: ; in, For correlation strength, This is a correction factor for operating conditions; The weights of the data statistics edges are represented as follows: ； in, The normalized feature sequence corresponding to the i-th continuous parameter and the normalized feature sequence corresponding to the j-th continuous parameter are in the optimal time delay The correlation coefficient at the location, For the maximum search latency, To ensure consistency in common variations, clip[·,0,1] indicates that the value is truncated to 0 to 1; The weights of knowledge rule edges are represented as follows: ; in, The initial confidence level for the rule. and These represent the number of confirmed hits and the number of mismatches, respectively.

3. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 2, characterized in that, In S2, the five-dimensional node attributes consisting of the window center value, the absolute deviation of the window center value, the normalized change slope, the over-limit ratio, and the missing ratio are input into the local graph neural network. The output is the normal state probability and the sensor fault state probability corresponding to the bias, drift, missing, spike, stagnation, and accuracy degradation of the sensing parameters in sequence. The sensing reliability is determined based on the normal state probability. When the sensing confidence level is not lower than 0.80, the corresponding sensing parameter is determined as the original confidence parameter; when the sensing confidence level is lower than 0.80, the output state is a correctable bias fault or drift fault, and the corrected correlation estimation residual does not exceed 3 times the robust scale of the normal residual, the corrected sensing parameter is determined as the corrected confidence parameter; when the sensing confidence level is lower than 0.80 and the correction condition is not met, the corresponding sensing parameter is determined as an invalid parameter, and a missing mask is set for the invalid parameter to prevent the invalid parameter from being the sole criterion for the degradation of the corresponding physical component.

4. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 2, characterized in that, In S3, prior knowledge nodes are used as seed nodes. Local sensitive subgraphs are extracted from the heterogeneous dynamic graph based on edge weights and node contribution. Local sensitive subgraphs corresponding to the current working condition or target fault type are called. A lightweight model is used to perform feature aggregation and output fault probability and uncertainty estimate. When any of the following factors—failure probability, prediction uncertainty, node deviation, or data statistics edge change—reaches the corresponding trigger threshold, a heterogeneous graph neural network that distinguishes different edge types is used to perform full graph reasoning on the heterogeneous dynamic graph.

5. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 4, characterized in that, One-hop and two-hop adjacent nodes with an edge weight of not less than 0.35 are retained; when the number of nodes exceeds 32, the node contribution is calculated according to the node deviation, the product of the path edge weights from the seed node to the candidate node, and the historical rule hit rate, and the nodes are pruned in descending order of their contribution. The sensor local graph neural network sequentially maps the node attribute vector formed by window statistical features into 16-dimensional and 8-dimensional hidden states, and reads the second-layer hidden state of the seed sensor node for classification. The component state local graph neural network uses graph-level average pooling and a classification layer to output the fault probability.

6. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 4, characterized in that, If the failure probability output by the local sensitive subgraph is not less than 0.40, the prediction uncertainty reaches the uncertainty trigger threshold, the absolute deviation of the sensing feature node is not less than 3, or the weight change of any data statistical edge relative to the previous sliding window is not less than 0.20, switch to full graph inference. The uncertainty trigger threshold is the larger of the 95th percentile of the prediction uncertainty on the normal validation set and 0.

15. The heterogeneous graph neural network performs adjacency propagation on physical coupling edges, data statistics edges and knowledge rule edges respectively, and then performs feature transformation and graph-level pooling after concatenating the propagation results of different edge types.

7. The hierarchical fault diagnosis method for a precision analytical instrument according to claim 6, characterized in that, In S3, the training phase and online inference phase of the lightweight model are executed separately. The training set is divided into device instances and continuous running segments. Each continuous running segment is divided into training segment, validation segment and test segment in chronological order. The same sliding window does not cross different datasets. Sensor local graph neural network, component state local graph neural network and relation type perception heterogeneous graph neural network for full graph inference are trained separately. The model weights, node order, category order, relation matrix, normal operating domain version and model summary are saved.

8. The method for stratified fault diagnosis of a precision analytical instrument according to claim 7, characterized in that, During the online inference phase, the frozen model weights are invoked. For the same fault mode, at least three samples are obtained that have been verified by maintenance records or manual review. After the samples come from at least two independent running segments, an independent model update and verification process is initiated.

9. A hierarchical fault diagnosis system for a precision analytical instrument, characterized in that, It includes a data processing module, a heterogeneous dynamic graph construction module, a sensor reliability assessment module, a hierarchical reasoning module, a reliability determination module, a model version management module, a knowledge iteration module, and an interactive output module; The output of the data processing module is connected to the heterogeneous dynamic graph construction module; the output of the heterogeneous dynamic graph construction module is connected to the sensing reliability assessment module and the hierarchical inference module. The sensing reliability assessment module sends the original reliability parameters, corrected reliability parameters, or invalid parameters to the hierarchical inference module; The output of the hierarchical reasoning module is connected to the credibility determination module; The confidence determination module will send the identified fault samples and low-confidence boundary state samples to the knowledge iteration module; The model version management module provides a consistent configuration version to the heterogeneous dynamic graph construction module, the sensor credibility assessment module, the hierarchical inference module, and the credibility determination module. The knowledge iteration module updates the corresponding version only after independent verification is passed. The interactive output module receives credible judgment results and diagnostic evidence to implement the steps of the hierarchical fault diagnosis method for a precision analytical instrument as described in any one of claims 1-8.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes a computer program, it implements the steps of the hierarchical fault diagnosis method for a precision analytical instrument as described in any one of claims 1 to 8.

Images