Method for automated testing and failure prediction of variable frequency devices

Abstract

The application provides a method for automatic testing and fault prediction of a variable frequency device, comprising: performing initialization and loop self-checking on a test resource; collecting associated data to construct a test snapshot sequence and a device health state graph; performing three operations of multi-modal coding, space-time joint health modeling and mixed expert model fusion on the test snapshot sequence and the device health state graph to obtain a fault prediction result; generating a retest action and collecting a retest test snapshot according to the fault prediction result; performing fault positioning diagnosis combined with the retest test snapshot, and performing incremental learning based on a diagnosis result. The application has the beneficial effects of realizing unified organization of measurement data, device state, control events, environmental data and link topology relationship, and closed-loop processing of system-level fault prediction, retest action generation, component-level fault positioning diagnosis and incremental learning.

Description

Technical Field

[0001] This invention belongs to the field of automated testing and intelligent diagnosis, and in particular relates to automated testing and fault prediction methods for variable frequency equipment. Background Technology

[0002] Testing of frequency converter components, frequency synthesizer components and related RF links typically relies on signal sources, spectrum analyzers, power meters, switch matrices and control software.

[0003] Most existing automated testing solutions measure indicators such as gain, output power, spurious emissions, phase noise, and frequency stability according to preset procedures, and give a pass or fail conclusion based on threshold rules.

[0004] The existing technology has the following problems: First, existing solutions typically use single measurement results as the criterion, making it difficult to uniformly organize and jointly analyze historical test data, equipment status data, control event data, environmental data, and link topology relationships; Secondly, when test results are abnormal, existing solutions can usually only pinpoint the error in a specific test item, making it difficult to further distinguish whether the abnormality originates from the module under test, the switch channel, the RF adapter card, the semi-steel cable, or the test instrument itself. Third, existing testing procedures are mostly static, making it difficult to dynamically insert supplementary testing actions based on the current failure probability distribution and diagnostic uncertainty; Fourth, the existing solution lacks a continuous learning mechanism for field applications, making it difficult to use expert confirmation results, maintenance results, and retest results to update the model in a closed loop.

[0005] Therefore, a new automated testing and intelligent fault prediction method is needed to solve the above problems. Summary of the Invention

[0006] In view of this, the present invention aims to propose an automated testing and fault prediction method for frequency converters, in order to solve at least one of the aforementioned technical problems.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows: Automated testing and fault prediction methods for variable frequency drives (VFDs) include: Perform initialization and loopback self-check on the test resources; Collect correlated data to construct a test snapshot sequence and device health status graph; Three operations—multimodal coding, spatiotemporal joint health modeling, and hybrid expert model fusion—are performed on the test snapshot sequence and equipment health status diagram to obtain fault prediction results. Based on the fault prediction results, generate supplementary test actions and collect supplementary test snapshots; Fault location diagnosis is performed by combining supplementary test snapshots, and incremental learning is performed based on the diagnosis results.

[0008] Furthermore, the process of initializing and performing a loopback self-test on the test resources includes: Detect the communication status of the test resources and load the corresponding instrument address configuration, test process template, routing table, device health status diagram, historical baseline database, and model threshold configuration; The control signal source outputs a standard signal, which forms a loop path through a switch matrix, and the loop data is collected by a spectrum analyzer and a power meter. The loopback data is constructed into a test snapshot and compared with the historical baseline database. Based on the comparison results, it is determined whether the loopback path meets the preset baseline conditions, and the loopback self-test results are output. The test resources include signal sources, spectrum analyzers, power meters, switch matrices, and the module under test.

[0009] Furthermore, the following operations are performed before collecting the associated data: The test process template is parsed into a sequence of executable actions; The six-throw single-pole channel and the two-throw single-pole channel in the switch matrix are abstracted into a routing resource pool, and a mapping relationship between the logical routing table and the actual switch action is established. Execute path connection instructions, path disconnect instructions, and path query instructions according to the executable action sequence to establish the current routing path; The working mode and parameter configuration are sent to the module under test based on the executable action sequence.

[0010] Furthermore, the process of constructing the test snapshot sequence includes: Related data is collected via a unified data bus and aligned to a unified timeline; A test snapshot is constructed based on identification information, path configuration information, measurement result information, and status monitoring information; Combine consecutive test snapshots into a test snapshot sequence.

[0011] Furthermore, the process of constructing the device health status diagram includes: Treat the test system entity and the object under test as nodes; The relationships between entities, including radio frequency signal connection, control, power supply influence, clock synchronization, and thermal coupling, are treated as edges. The test system entity includes test instruments, routing switching units, transmission interconnection units, and environmental conditioning units, while the test object entity includes the test module and its internal functional sub-units.

[0012] Furthermore, the process of performing multimodal encoding on the test snapshot sequence and device health status graph includes: The features within the test snapshot sequence and device health status map are divided into four categories: structured numerical features, spectrum and scan curve features, event flow features, and topological features. Different neural networks are used to process the four features respectively to obtain a unified state representation containing time information, frequency domain information, behavioral information, and structural information. Structured numerical features are encoded using a temporal convolutional network; The spectral and scanning curve features are encoded using a frequency domain transformer network; The event stream features are encoded using an event sequence encoding method; Topological features are encoded using a graph attention network.

[0013] Furthermore, the process of performing spatiotemporal joint health modeling and hybrid expert model fusion on the test snapshot sequence and device health status map includes: Based on a unified state representation, the state of each node and each edge is updated according to the test time to obtain the spatiotemporal joint health characteristics. The spatiotemporal joint health features are input into a hybrid expert model; The gating network assigns weights to each expert based on the current test context and merges the outputs of each expert.

[0014] Furthermore, the process of generating supplementary testing actions based on the fault prediction results includes: The results from all experts will be used as the fault prediction result. When the fault prediction result meets the preset triggering conditions, the active testing strategy network is invoked. The proactive testing strategy network receives the current diagnostic decision status information and outputs supplementary testing actions. It inserts the supplementary testing actions into the executable action sequence and executes them to collect supplementary test snapshots. The current diagnostic decision status information includes at least one of the following: fault status information, diagnostic uncertainty status information, test execution status information, and candidate supplementary test value information.

[0015] Furthermore, the process of performing fault location diagnosis by combining supplementary test snapshots includes: Based on diagnostic correlation data, correlation calculations are performed, and graph attention mechanism is used to establish the correlation between anomalies and nodes and edges, and diagnostic results are output. The diagnostic correlation data is a dataset used to characterize test anomalies, retest feedback, historical baselines, link context, and model inference results. The diagnostic results include fault categories, fault location probability distributions, confidence scores, and remaining life estimates of the target component.

[0016] Furthermore, the incremental learning process based on the diagnostic results includes: Perform model training and updates based on model training data; The model training and update include self-supervised pre-training, supervised fine-tuning, and incremental learning; the model training data includes normal sample data, simulated fault sample data, and labeled feedback data. Physical consistency constraints are introduced during the model training and update process.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. By testing the combined organization of snapshot sequences and device health status graphs, a unified representation of multi-source heterogeneous data was achieved.

[0018] 2. By adopting a hierarchical approach of system-level fault prediction and component-level fault location diagnosis, a progressive analysis from risk assessment to fault location is achieved.

[0019] 3. By generating supplementary testing actions through an active testing strategy network, the testing process can be dynamically adjusted.

[0020] 4. By integrating hybrid expert models and physical consistency constraints, the consistency between the fault diagnosis process and the system structure is improved.

[0021] 5. By having experts confirm the data and maintenance results and participate in incremental learning, a closed loop for model updates was achieved. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the workflow of the automated testing and fault prediction method described in an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application. In the absence of conflict, the embodiments and features in the embodiments of this invention can be combined with each other. In order to make the purpose, technical solution and advantages of this application clearer, the following describes this application in further detail with reference to the accompanying drawings and embodiments.

[0025] like Figure 1As shown, the automated testing and fault prediction method for frequency converters includes: S1. Perform initialization and loopback self-check on the test resources; S2. Collect related data to construct a test snapshot sequence and device health status map; S3. Perform three operations on the test snapshot sequence and equipment health status diagram: multimodal coding, spatiotemporal joint health modeling, and hybrid expert model fusion to obtain the fault prediction results; S4. Generate supplementary test actions based on the fault prediction results and collect supplementary test snapshots; S5. Combine the supplementary test snapshots to perform fault location diagnosis, and perform incremental learning based on the diagnosis results.

[0026] Step S1 describes the process of initializing and performing a loopback self-test on the test resources, which includes: The test resources include a signal source, spectrum analyzer, power meter, switch matrix, and module under test; Detect the communication status of the test resources and load the corresponding instrument address configuration, test process template, routing table, device health status diagram, historical baseline database, and model threshold configuration; The signal source is controlled to output a standard signal, which forms a loop path through the switch matrix, and the loop data is collected by the spectrum analyzer and power meter. The loopback data is constructed into a test snapshot and compared with the historical baseline database. Based on the comparison result, it is determined whether the loopback path meets the preset baseline conditions, and the loopback self-test result is output.

[0027] Instrument address configuration: This is the connection and addressing information required for communication control of each test resource. It is used to define the communication interface, communication address, port parameters, and communication timing parameters of each test resource.

[0028] Test process templates are predefined process rules for test tasks before execution. They are used to describe test items, test steps, step order, parameter configuration, judgment conditions, and exception handling logic.

[0029] The routing table is a mapping between logical connections and actual switching actions in the test link. It is used to define the mapping between logical test ports, the current routing path, and the actual actions of the switching matrix.

[0030] A device health status diagram is a data model that organizes the devices, channels, cables, sub-units, and their connections in a test system into a graphical structure. It represents the composition, connection relationships, and health status characteristics of test resources and test links, including: The set of nodes (e.g., signal source, spectrum analyzer, power meter, switch matrix, switch channel, RF adapter card, semi-steel cable, module under test and its internal sub-units), the set of edges (e.g., RF signal connection relationship, control relationship, power supply influence relationship, clock synchronization relationship, thermal coupling relationship), node attributes, and edge attributes.

[0031] The historical baseline database is used to store test results, status data, and allowable fluctuation ranges under normal operating conditions.

[0032] Model threshold configuration is a set of threshold parameters used by the intelligent analysis model and decision logic. It is used to define various decision thresholds used in the process of loopback self-test, fault prediction, supplementary test triggering and fault location.

[0033] Before collecting the associated data in step S2, the following operations are performed: The test process template is parsed into a sequence of executable actions; The six-throw single-pole channel and the two-throw single-pole channel in the switch matrix are abstracted into a routing resource pool, and a mapping relationship between the logical routing table and the actual switch action is established. Execute path connection instructions, path disconnect instructions, and path query instructions according to the executable action sequence to establish the current routing path; The working mode and parameter configuration are sent to the module under test based on the executable action sequence.

[0034] The process of constructing the test snapshot sequence in step S2 includes: Related data is collected via a unified data bus and aligned to a unified timeline; A test snapshot is constructed based on identification information, path configuration information, measurement result information, and status monitoring information; Combine consecutive test snapshots into a test snapshot sequence.

[0035] In this embodiment: Related data includes test index data, spectrum curve data, device status snapshots, event logs, switch action records, ambient temperature and humidity data, module register information, self-test information, maintenance records, and link topology relationships; The identification information includes the test task number, module number, and test step number; Path configuration information includes the current route path; Measurement results include instrument readings and spectrum curves; Status monitoring information includes current device status, current environment status, and current module status.

[0036] Step S2, the process of constructing the device health status diagram, includes: Treat the test system entity and the object under test as nodes; The relationships between entities, including radio frequency signal connection, control, power supply influence, clock synchronization, and thermal coupling, are treated as edges. The test system entity includes test instruments, routing switching units, transmission interconnection units, and environmental conditioning units, while the test object entity includes the test module and its internal functional sub-units.

[0037] Step S3, which involves multimodal encoding of the test snapshot sequence and the device health status graph, includes: The features within the test snapshot sequence and device health status map are divided into four categories: structured numerical features, spectrum and scan curve features, event flow features, and topological features. Different neural networks are used to process the four features respectively to obtain a unified state representation containing time information, frequency domain information, behavioral information, and structural information. Structured numerical features are encoded using a temporal convolutional network; The spectral and scanning curve features are encoded using a frequency domain transformer network; The event stream features are encoded using an event sequence encoding method; Topological features are encoded using a graph attention network.

[0038] The aforementioned temporal convolutional network is responsible for understanding the evolution of structured numerical features over time; the frequency domain transformer network is responsible for understanding the overall structure of the spectrum and scan curve on the frequency axis and the long-distance frequency band correlation; the event sequence encoding is responsible for understanding the occurrence order, interval relationship and behavioral chain of discrete events; and the graph attention network is responsible for understanding the internal topological connections of the system and the propagation of influence between key nodes.

[0039] In some embodiments, the process of multimodal encoding of the test snapshot sequence and the device health status graph is as follows: Structured numerical characteristics include at least the following: gain, output power, 1 dB compression point, spurious suppression, phase noise, frequency stability, voltage, current, temperature, phase-locked loop lockout status, switch action count, route switching settling time, instrument self-test status, and maintenance records.

[0040] Step 1: Multi-source heterogeneous data acquisition: Collect multi-source heterogeneous data generated by the target equipment during operation, testing, maintenance or inspection.

[0041] Among them, the structured numerical features come from the equipment telemetry interface, the test instrument output interface, the control system, the environmental monitoring system, and the maintenance management system. This type of data is mainly manifested as continuous values ​​that change over time, discrete state quantities, count quantities, and maintenance record-derived state quantities. The spectrum and sweep curve characteristics are derived from spectrum analyzers, vector network analyzers, sweep frequency testing devices, signal detection modules, or built-in diagnostic units, including but not limited to spectrum curves, power sweep curves, gain sweep curves, spurious distribution curves, amplitude frequency response curves, and phase frequency response curves; Event flow characteristics are derived from equipment operation logs, alarm logs, control operation logs, self-test logs, maintenance logs, and fault record logs. Event flow characteristics include at least event type, event occurrence time, event source, event level, and additional event attributes. Topology features are derived from system configuration files, device connection tables, routing tables, control link relationships, power supply relationships, clock reference relationships, switching matrix relationships, or module dependencies, and are used to characterize the connection structure and interaction relationships between nodes in the system.

[0042] Step 2: Unified Time Alignment and Sample Construction: Since data from different sources differ in sampling frequency, triggering mechanism, update time, and expression form, a unified time alignment and sample construction are performed on all types of data before feature encoding. Determine the analysis time window and construct samples in units of the analysis time window. The analysis time window can be a fixed-length time window or an event-triggered time window constructed around a predetermined key event.

[0043] For structured numerical features, resampling, interpolation, preservation, or aggregation are performed according to a unified time base to form a multivariate time series arranged in chronological order. For the characteristics of the spectrum and scanning curve, select the test results or observation results corresponding to the analysis time window, and ensure that the frequency axis, scanning direction and sampling point sequence are consistent; For event stream characteristics, the original order of events and timestamps are preserved, and the event time is converted into a relative time representation relative to the start time of the analysis time window; For topological features, extract and analyze the corresponding network connection state, node state and edge relationship state within the time window to construct graph structure samples at the current time or under the current time window; During the sample construction process, outliers, missing values, invalid values ​​and obvious collection errors can be identified and processed, and the units, dimensions, naming and status representation of data from different sources can be standardized.

[0044] Step 3: Temporal convolutional network encoding of structured numerical features: After obtaining time series samples of structured numerical features, they are input into a temporal convolutional network for encoding in order to extract the temporal evolution patterns in the structured numerical features.

[0045] First, the structured numerical features corresponding to each moment within the analysis time window are arranged in chronological order to form a multivariate time series input; Each time sampling point corresponds to a feature vector, which includes the gain, output power, 1 dB compression point, spurious suppression, phase noise, frequency stability, voltage, current, temperature, phase-locked loop lockout status, switch action count, route switching stabilization time, instrument self-test status, and status representation obtained from maintenance records at that moment.

[0046] To avoid the adverse effects of differences in the units of different features on the coding results, continuous numerical values ​​are first normalized or standardized, state data are numerically represented, count features are scaled uniformly, maintenance records are structured and mapped, and the locations of missing values ​​are explicitly marked.

[0047] Subsequently, the processed multivariate time series is input into a temporal convolutional network; Temporal convolutional networks perform one-dimensional convolution operations along the time axis on multivariate inputs to extract collaborative change patterns among features within a local temporal neighborhood. By stacking convolutional layers layer by layer, they can identify features such as short-term fluctuations, sudden anomalies, state transitions, and coordinated changes among multiple indicators.

[0048] Temporal convolutional networks use dilated convolutions to gradually expand the temporal receptive field, enabling the network to not only capture short-period local changes, but also learn long-range dependency patterns such as trend drift, slow degradation, periodic oscillations, and precursors to fault formation over longer time periods.

[0049] After multi-layer convolutional feature extraction, a high-level representation of structured numerical features corresponding to the current analysis time window is obtained. This high-level representation reflects the dynamic characteristics, trend characteristics, and abnormal evolution characteristics of the equipment's operating status in the time dimension.

[0050] Step 4: Frequency Domain Transformer Network Coding of Spectrum and Scan Curve Characteristics: For the spectral and scanning curve features, this embodiment uses a frequency domain converter network for encoding to extract frequency domain structural features and cross-frequency band correlations.

[0051] The spectral curves or scan curves are organized into a one-dimensional frequency domain sequence according to the frequency axis. Each frequency sampling point or scan sampling point corresponds to at least one observation value. The observation value can be a power value, amplitude value, gain value, phase value, noise value, or response value. In the case of multiple correlation curves, a multi-channel form can also be used to represent them together. Before inputting into the frequency domain transformer network, the raw frequency domain data is preprocessed, including frequency axis unification, sampling point alignment, amplitude normalization, baseline correction, noise floor correction, and test condition unification. For spikes, spurs, local distortions, or stray structures that can reflect abnormal signs, it is preferable to retain their original morphological information. The frequency domain sequence is divided into several continuous frequency domain segments to reduce the processing complexity of long sequences and enhance the ability to express local structures. Each frequency domain segment is used as a frequency domain marker unit and is attached with corresponding frequency position information. The frequency position information is used to characterize the relative position, center frequency or frequency band offset of the segment in the overall frequency range.

[0052] After the above frequency domain segment sequence is input into the frequency domain transformer network, the network establishes the correlation between each frequency domain segment through the attention mechanism. Thus, it can not only extract local modes such as local peaks, dips, sidebands, spurious noise and noise floor rise, but also learn the global coupling relationship between the main peak and harmonics, in-band and out-of-band, and local distortion and overall spectral shape.

[0053] After processing by a multi-layer frequency domain converter network, the high-level characterization of the spectrum and scan curve corresponding to the current analysis time window is output. This characterization comprehensively reflects the overall frequency domain morphology, key frequency band anomalies, cross-frequency band structural relationships, and scan response change patterns.

[0054] Step 5: Event sequence encoding of event stream features: For event flow characteristics, this embodiment uses event sequence encoding to characterize them in order to extract the event occurrence order, time interval, event propagation chain, and behavioral evolution pattern.

[0055] First, extract and analyze all events that occur within the time window in chronological order to form an event sequence. Each event includes at least the event type, the time of occurrence, the source of the event, and additional attributes of the event. To ensure that events from different sources are comparable, event types are standardized and mapped, event attributes are normalized and represented, and a unified event vocabulary is established. Event types can include phase-locking success, phase-locking failure, switching action, route switching, self-test start, self-test pass, self-test failure, maintenance execution, alarm occurrence, alarm recovery, module restart, communication interruption, and communication recovery.

[0056] Subsequently, a corresponding event semantic representation is generated for each event. The event semantic representation includes at least an event type representation, an event time representation, an event source representation, and an event attribute representation.

[0057] Event time indicates at least the time offset of the event relative to the start of the analysis time window, as well as the time interval between adjacent events; event source indicates the module, channel, node, or subsystem to which the event belongs; event attributes indicate additional information such as the severity of the event, duration, whether it has been recovered, and whether it was manually triggered.

[0058] After obtaining the event-by-event representation, a complete event sequence representation is constructed according to the actual occurrence order of the events. By sequentially encoding the event sequence, the sequential dependencies, trigger chain relationships, cluster occurrence patterns, recurrence patterns, and procedural features such as alarm-maintenance-recovery between events can be extracted.

[0059] For time windows with a small number of events, no event identifier or blank padding can be set; for time windows with a large number of events, key event sequences can be retained based on the importance of events, time proximity, or abnormal correlation to ensure that the sample length is manageable and consistent.

[0060] Ultimately, a high-level representation of the event flow under the current analysis time window is obtained, which reflects the temporal logic characteristics of the device behavior process, the anomaly propagation process, and the operation response process.

[0061] Step 6: Graph Attention Network Encoding of Topological Structure Features: For topological features, this embodiment uses graph attention networks for encoding to extract internal connection relationships, dependencies, and anomaly propagation structure information of the device system.

[0062] A graph structure is constructed based on the physical connection relationships, signal flow relationships, control relationships, routing relationships, power supply relationships, reference source relationships, or module dependencies of the frequency converter equipment. Nodes in the graph are used to represent devices, modules, ports, channels, functional units, or subsystems, and edges in the graph are used to represent the connection relationships, control relationships, dependencies, or transmission relationships between nodes.

[0063] After constructing the graph structure, node attribute features are attached to each node. Node attribute features may include the node's output power, current, temperature, phase-locked state, self-test state, alarm count, switching count, maintenance status, and other operating parameters related to the node. If necessary, edge attributes can be added, such as connection type, direction, attenuation, latency, primary / backup relationship, allowed switching relationship, or priority relationship.

[0064] After the constructed graph structure is input into the graph attention network, the network uses the node as the center and weighted aggregates the information of its neighboring nodes; Graph attention networks do not assign the same weight to all neighboring nodes. Instead, they automatically learn the importance weights of different neighboring nodes in representing the current node's state, based on the correlation, connection, and state characteristics between the current node and its neighbors.

[0065] Through the attention aggregation process described above, each node, when updating its own representation, not only retains its own attribute information, but also integrates the state information of key neighboring nodes and local structural environment information, thereby obtaining a node representation that can reflect the local network influence relationship.

[0066] Through multi-layer graph attention network propagation, node representations can gradually absorb information from nodes at greater distances, enabling the modeling of influence relationships within first-level neighbors, second-level neighbors, and even larger topologies, thereby capturing fault propagation paths, critical dependency links, and abnormal clustering areas.

[0067] After updating the node representations, the entire graph is aggregated to obtain a high-level representation of the topology corresponding to the current analysis time window. This representation can summarize the overall connection structure of the system, the influence of key nodes, the spread of local anomalies, and the characteristics of global topological state changes.

[0068] Step 7: Fusion of multimodal coding results: After obtaining the structured numerical feature coding results, spectrum and scan curve coding results, event stream coding results, and topology coding results respectively, the four types of coding results are fused to form a unified state representation.

[0069] First, semantic space alignment is performed on the four types of encoding results to make the high-level representations from different modalities comparable and fusionable in a unified feature space; Subsequently, the results of each modality coding are jointly modeled to establish the correspondence and complementarity between temporal dynamic features, frequency domain structural features, event behavior features, and topological dependency features.

[0070] Through the above fusion process, the following information can be represented simultaneously: The changing trends of equipment operating parameters in the time dimension, the structural changes of spectrum and scan curve in the frequency dimension, the process evolution law in the event log, and the propagation relationship of abnormal influences in the system structure; The unified state representation obtained in this way can reflect the actual operating state of the equipment more comprehensively and stably than single modal features.

[0071] Step S3, which involves spatiotemporal joint health modeling and hybrid expert model fusion of the test snapshot sequence and device health status diagram, includes: Based on the unified state representation obtained from multimodal coding operations, the state of each node and each edge is updated according to the test time to obtain the spatiotemporal joint health feature; The spatiotemporal joint health features are input into a hybrid expert model; The gating network assigns weights to each expert based on the current test context and merges the outputs of each expert.

[0072] In a hybrid expert model, each expert corresponds to a different health mechanism or source of abnormality. Among them, the RF link degradation expert is used to characterize the performance degradation of the RF transmission link; the power supply and bias expert is used to characterize power supply and bias stability anomalies; the local oscillator and phase-locked loop expert is used to characterize the reference frequency link and locking stability anomalies; the switch and mechanical life expert is used to characterize the life degradation of switching components and mechanical actuators; the environmental stress expert is used to characterize the impact of environmental stress factors such as temperature on the health status of the equipment; and the test system self-anomaly expert is used to characterize the interference of test system anomalies on the test results.

[0073] Each expert performs parallel modeling based on the same spatiotemporal joint health characteristics, and outputs health representation, abnormality score or fault judgment results under the corresponding mechanism.

[0074] The gating network generates corresponding weights for each expert based on the current test context, spatiotemporal joint health features, and data credibility information. These weights are then used as fusion coefficients output by each expert to obtain comprehensive health modeling results.

[0075] When the weight or anomaly score of the anomaly expert in the test system exceeds the preset conditions, the conclusion of the equipment failure can also be suppressed, corrected, or retested.

[0076] In some embodiments, the process of performing spatiotemporal joint health modeling and hybrid expert model fusion on the test snapshot sequence and the device health status graph is as follows.

[0077] I. Spatiotemporal Joint Health Modeling: Obtain the test snapshot sequence corresponding to the target device at multiple test moments, and construct the device health status map. The test snapshot sequence is used to characterize the state evolution information of the device during continuous testing. Each test snapshot includes at least the structured numerical features, spectrum or scan curve features, event features, and auxiliary context information related to the device operating status corresponding to that test moment. The device health status diagram is used to characterize the functional nodes inside the device and their interconnections. Nodes represent devices, modules, channels, ports or functional units, while edges represent signal transmission relationships, power supply relationships, reference source relationships, control relationships, routing relationships or other dependencies between nodes.

[0078] After obtaining the test snapshot sequence and device health status graph, the device health status graph is updated in time sequence according to the test time.

[0079] Using each test moment in the test snapshot sequence as a time reference, the test result corresponding to that test moment is mapped to the corresponding node and edge in the device health status graph; For node status updates, the output power, gain, voltage, current, temperature, phase-locked state, self-test state, alarm state, maintenance state, or other health parameters corresponding to the node can be written into the corresponding node attributes. For edge state updates, the insertion loss change, on / off state, switching state, propagation stability, reference coupling state, power supply branch state or other link state corresponding to the connection relationship between nodes can be written into the corresponding edge attributes; By repeatedly performing the above mapping and update process at each test moment, a graph state sequence that evolves over time is formed, thereby obtaining a spatiotemporal joint health feature that simultaneously contains information on time-dimensional changes and topological-dimensional dependencies. The spatiotemporal joint health feature not only reflects the local health status of the device at a single moment, but also reflects the state evolution trend between different test moments and the propagation relationship of anomalies in the device topology.

[0080] II. Hybrid Expert Model Fusion: The hybrid expert model includes RF link degradation experts, power supply and bias experts, local oscillator and phase-locked loop experts, switch and mechanical life experts, environmental stress experts, and test system self-abnormality experts; The aforementioned experts are all specialized discriminant branches or modeling units constructed for different health mechanisms or sources of abnormality. Their inputs are all spatiotemporal joint health features, but the features they focus on, the interpretive paths of their health mechanisms, and their outputs differ. Specifically: (1) RF link degradation experts are used to analyze the health status of equipment related to RF signal transmission, amplification, filtering, mixing and link loss; Based on spatiotemporal joint health characteristics, the expert focused on extracting node state changes and edge state changes related to the radio frequency transmission path, analyzing whether features such as gain decrease, output power attenuation, advance of the one-decibel compression point, spurious enhancement, passband distortion, increase in local insertion loss, or frequency response distortion continue to evolve along the radio frequency path, and combined with the topological dependency relationship between upstream and downstream nodes of the path to determine whether the anomaly is caused by local device degradation or by the cumulative deterioration of cascaded links. The expert outputs health characteristics, degradation scores, and information on suspected abnormal paths or nodes related to RF link degradation.

[0081] (2) Power and bias experts are used to analyze anomalies in equipment related to power supply stability, bias operating point and power supply branch health status; Based on the time-dependent changes in voltage, current, temperature, output capacity, and related events, and combined with the power supply connection relationships, shared power supply relationships, and bias dependence relationships in the equipment health status diagram, the expert judges whether there are phenomena such as undervoltage, overcurrent, bias drift, branch instability, load coupling imbalance, or power degradation. The expert further distinguished between single-node power supply anomalies and multi-node common-cause anomalies caused by common power supply branches, and output corresponding power health scores, bias stability scores, and information on suspected abnormal power supply branches.

[0082] (3) Local oscillator and phase-locked loop experts are used to analyze the health status of the reference source, local oscillator module, frequency synthesizer, phase-locked loop and frequency stability; Based on the phase-locked loop status, frequency stability, phase noise, loss-of-lock events, relocking behavior, and related spectral characteristics, the expert determined whether there was instability in the reference source, phase-locked loop jitter, local oscillator link degradation, abnormal frequency synthesis, or abnormal clock allocation. The expert combined the dependency path between the reference source and downstream nodes in the equipment health status diagram to identify whether the anomaly was concentrated in a local phase-locked unit or propagated from a common reference link to multiple dependent nodes.

[0083] The expert outputs the local oscillator link health score, lock stability score, frequency stability anomaly risk, and information on suspected abnormal reference nodes or phase-locked nodes.

[0084] (4) Switch and mechanical life experts are used to analyze the health status of equipment related to RF switches, relays, routing units, mechanical actuators and their lifespan consumption; Based on switch action counts, route switching stabilization time, switching event sequences, and performance changes before and after switching, the expert analyzed whether there were phenomena such as contact performance degradation, switching jitter, mechanical jamming, action lag, contact life decay, or abnormal recovery after route switching. The expert also combined the edge status and path status related to routing switching in the equipment health status diagram to determine whether the anomaly was strongly related to a specific switch node or a specific switching edge. The expert outputs the switch health score, mechanical lifespan consumption level, switching risk level, and information on suspected abnormal switch nodes or paths.

[0085] (5) Environmental stress experts are used to analyze the coupling relationship between equipment health status and environmental factors such as temperature; Based on the temperature evolution information in the test snapshot sequence and the correlation between temperature and parameters such as output power, gain, current, phase noise, and frequency stability, the expert determined whether the equipment was in a cold start, temperature rise, thermal stability, thermal cycle, or abnormal overheating state, and further analyzed whether the abnormality was a transient shift caused by environmental changes or a performance degradation caused by long-term accumulation of thermal stress. By combining the status of thermally sensitive nodes and their adjacent coupled nodes in the equipment health status diagram, the expert identified the local concentration or diffusion of thermal effects and output environmental stress impact scores, thermal instability risk levels, and risk information of thermally sensitive nodes.

[0086] (6) The test system's self-anomaly expert is used to identify whether the current anomaly is introduced by the test system itself, rather than by the device under test; Based on the instrument's self-test status, calibration status, test link status, acquisition consistency, control events, test logs, and the consistency between multiple test results, the expert determined whether there was instrument drift, abnormal test link contact, triggering and synchronization abnormalities, abnormal range settings, abnormal acquisition noise, or abnormal control process. The expert can also combine the equipment health status diagram to determine whether the current anomaly is consistent with the actual topology propagation logic of the tested equipment; if the anomaly does not conform to the internal propagation law of the equipment, or multiple unrelated objects show the same anomaly pattern at the same time, it is first determined to be an anomaly of the test system itself. The expert outputs the probability of system anomalies, the reliability of the current test results, and a judgment on whether to recommend retesting, recalibrating, or switching test links.

[0087] III. Integration Process: After obtaining the parallel outputs of each expert, the spatiotemporal joint health features are input into the gating network to dynamically assign weights to each expert. In addition to spatiotemporal joint health features, the input to the gating network may also include the current test context information. The test context information includes at least the test item type, working mode, working frequency band, activation path, output level, environmental range, recent maintenance status, instrument self-test status, and whether it is in a switched state. Based on the characteristics of the current sample and the test context, the gating network calculates the applicability scores of the current sample for the RF link degradation expert, power supply and bias expert, local oscillator and phase-locked loop expert, switch and mechanical life expert, environmental stress expert and test system self-anomaly expert. The applicability scores were normalized to obtain the final weights for each expert. Preferably, each expert's weight is a non-negative value, and the sum of all expert weights is one, to represent the relative contribution of each expert to the current comprehensive health judgment.

[0088] It is worth noting that the principle for allocating weights to experts in a gating network is as follows: When a certain type of health mechanism has a stronger explanatory power in the current sample, the corresponding expert receives a higher weight; when a certain type of mechanism has a low relevance to the current test context, or when the data relied upon by the expert is incomplete or has low reliability, the corresponding expert's weight decreases. For example, when phase noise increases, frequency stability decreases, and phase-locked loop (PLL) jitter exists, the weight of the local oscillator and PLL experts increases; when multiple shared power supply branch nodes simultaneously experience decreased output capacity, voltage fluctuations, and current anomalies, the weight of the power supply and bias experts increases; when the settling time after switching increases significantly and the anomalies mainly occur on specific routing paths, the weight of the switch and mechanical life experts increases; when performance deteriorates under high temperature conditions but partially recovers after cooling, the weight of the environmental stress expert increases; when the instrument self-test fails and multiple irrelevant objects simultaneously exhibit the same measurement distortion, the weight of the test system self-anomaly expert increases.

[0089] After the weight allocation is completed, the outputs of each expert are merged to obtain a comprehensive health modeling result; Fusion can be achieved through feature-level fusion, fraction-level fusion, or a hybrid fusion approach combining both. In the feature-level fusion approach, the health representation vectors output by each expert are weighted according to their corresponding weights to form a unified comprehensive health representation. In the score-based fusion approach, the abnormal scores, failure probabilities, health scores, or risk scores output by each expert are weighted and summed according to their corresponding weights to obtain the final judgment result. In the hybrid fusion approach, the health characteristics of each expert are first weighted and fused, and then combined with the expert scores, confidence levels and test credibility information for secondary synthesis. By using the above fusion method, the final output can simultaneously reflect the contribution of different health mechanisms to the current state of the device.

[0090] Preferably, when performing expert fusion, in addition to considering the expert weights generated by the gating network, the confidence level of the expert's own output, the quality of the corresponding input data, and the consistency between the expert's conclusions and the propagation relationship of the device topology can also be considered. When an expert has high relevance in the current test context, but their output is unstable, the corresponding input quality is low, or it is significantly inconsistent with the device state propagation logic, the expert's actual fusion contribution can be suppressed or corrected, thereby improving the robustness and credibility of the comprehensive health modeling results.

[0091] In other embodiments, the test system's self-anomaly expert also has a result correction function; When the weight or anomaly score of the test system's anomaly expert exceeds the preset threshold, the output confidence of the equipment's fault determination can be reduced, or the current test result can be marked as a suspicious result, triggering a retest, recalibration, switching of test links, or manual review process. By introducing an anomaly expert into the test system, measurement artifacts introduced by the test system can be avoided from being misjudged as equipment failures, thereby improving the accuracy of health assessment and fault diagnosis results.

[0092] Step S4, which generates supplementary testing actions based on the fault prediction results, includes: The results from all experts will be used as the fault prediction result. When the fault prediction result meets the preset triggering conditions, the active testing strategy network is invoked. The proactive testing strategy network receives the current diagnostic decision status information and outputs supplementary testing actions. It inserts the supplementary testing actions into the executable action sequence and executes them to collect supplementary test snapshots. The current diagnostic decision status information includes at least one of the following: fault status information, diagnostic uncertainty status information, test execution status information, and candidate supplementary test value information.

[0093] The proactive testing strategy network is used to select the optimal supplementary testing action from the candidate supplementary testing action set when the current diagnostic conclusion is insufficient, the diagnostic uncertainty is high, multiple fault mechanisms are difficult to distinguish, or the current test results need further verification. This allows for prioritizing the supplementary testing operations that are most valuable to the current diagnosis under limited testing time, testing cost, and available resources.

[0094] Supplementary testing actions include, but are not limited to: Retest current abnormal frequency points or bands; expand the spectrum scanning range or improve the spectrum resolution; perform retesting after route switching on specific RF paths; perform local power, voltage, current or temperature retesting on specific nodes; add phase noise, frequency stability or lockout setup time tests; perform supplementary measurements on specific power supply branches or bias nodes; perform self-testing, calibration confirmation or alternative link cross-retesting on the test system; perform verification tests under different environmental conditions; perform fixed-point verification tests on suspected faulty nodes.

[0095] Therefore, the proactive testing policy network does not simply output whether to perform supplementary testing, but rather makes targeted supplementary testing decisions among multiple executable candidate actions based on the current diagnostic context.

[0096] In some embodiments, the active testing policy network operates as follows: Step 1: Use the comprehensive health modeling results obtained by integrating the outputs of various experts as the basis for current fault prediction; Step 2: Trigger a judgment to determine whether further testing is needed for the current diagnosis; Step 3: Upon triggering, construct current diagnostic decision state information that includes fault status, diagnostic uncertainty, test execution status, and candidate retest value; Step 4: Evaluate the value and score the strategies for the candidate retest actions; Step 5: Under the condition of satisfying the executable constraints, output the optimal supplementary test action; Step 6: Insert the supplementary test action into the current executable action sequence and execute it; Step 7: Update the test snapshot sequence, equipment health status diagram, and fault prediction results using the supplementary test snapshots; Step 8: Form a closed-loop proactive diagnosis and supplementary testing decision-making process.

[0097] Specifically, the execution process for each step is as follows: Step 1: The active testing strategy network uses the comprehensive health modeling result obtained after integrating the outputs of various experts in the previous stage as the basis for upstream diagnosis. This result can be expressed as: ; in: This represents the fault prediction result at the t-th test time. This indicates the comprehensive health modeling results obtained by gating and weighting the outputs of various experts.

[0098] In this embodiment, the fault prediction result is a fault category probability distribution, which is expressed as follows: ; Where: K represents the total number of candidate fault categories; This represents the predicted probability that the device belongs to the Kth type of fault at test time t; satisfying... .

[0099] Step 2: Before invoking the active testing strategy network, first determine whether the current fault prediction result meets the preset trigger condition. In this embodiment, the preset trigger condition is that the currently executed test action has failed to effectively reduce diagnostic uncertainty, which is expressed as: ; in, This represents the diagnostic entropy at the t-th test time. The larger the value, the more uncertain the current diagnosis. Indicates the probability of the fault category; The probability interval between the two types of head faults can also be defined as: ; in, This represents the index of the fault category with the highest predicted probability. This represents the index of the fault category with the second highest predicted probability. The smaller the value, the more difficult it is to distinguish in the current diagnosis.

[0100] If we consider the disagreements among experts, we can also define the degree of expert disagreement as: ; in, The number of experts is preferably 6 in this embodiment; This represents the fusion weight of the m-th expert at time t; This represents the output of the m-th expert; This represents the overall result after fusion; The larger the value, the more pronounced the disagreement among experts.

[0101] Step 3: When the triggering condition is met, construct the input state of the active testing strategy network. The current diagnostic decision state information can be represented as: ; in: This represents the current diagnostic decision state vector at the t-th test time. Indicates fault status information; Indicates information about uncertain diagnostic status; Indicates test execution status information; This indicates the value information of the candidate supplementary test; Fault status information is used to characterize the current fault prediction result itself, including: Fault Category Probability Distribution Risk vector of faulty nodes Fault edge risk vector Current health score Current risk level ; Therefore, it can be written as: .

[0102] Information on diagnostic uncertainty is used to characterize the stability and reliability of the current diagnostic conclusion, including: Diagnostic entropy probability interval Expert Disagreement Test credibility Fluctuation of historical diagnostic results ; Therefore, it can be written as: .

[0103] Test execution status information describes the current test resources and execution constraints, including: Action mask already executed Remaining test budget Remaining test time Current status of available instrument resources Current switchable path status Current test sequence position ; Therefore, it can be written as: .

[0104] Candidate supplementary testing value information is used to characterize the potential benefit of each candidate supplementary testing action to the current diagnosis; If the current set of candidate supplementary test actions is: ; Where N represents the number of candidate actions for retesting at the current moment; This represents the i-th candidate supplementary test action.

[0105] Each candidate retest action can then correspond to an action feature vector: ; in, Indicates candidate actions Basic attributes; Indicates candidate actions Overall value score, basic attributes It may include action type, target node or target path, target frequency band, required instrument resources, execution time, execution cost, action risk, and whether it depends on switching or calibration prerequisites.

[0106] Step 4: Before making a decision on candidate actions, the active testing policy network first calculates the value score for each candidate action. The value score can be expressed as: ; in, Let represent the comprehensive value score of the i-th candidate supplementary test action at time t; This indicates the expected decrease in diagnostic uncertainty after performing this action; This indicates the expected increase in the head candidate fault differentiation interval after performing this action; This indicates the expected reduction in expert disagreement after the action is performed. This indicates the coverage gain of the action on the currently suspected faulty node, suspected faulty edge, or critical mechanism; This indicates the cost of performing the action; Indicates the time taken to perform the action; This indicates the risk of performing the action or the cost of perturbing the current test sequence; These are the weighting coefficients for each item.

[0107] ; Where o represents the possible supplementary test result after performing the candidate supplementary test action; Indicates the current state Next action Predicted distribution of post-observation results; This indicates the uncertainty after taking a supplementary snapshot and re-diagnosing; The larger this value, the more likely the supplementary test is to reduce diagnostic ambiguity.

[0108] ; in, This represents the probability interval between the current best and second-best failures; This represents the probability interval obtained after the retest; The larger this value, the more helpful the supplementary test is in distinguishing the closest fault candidate.

[0109] ; in, Indicates the current degree of disagreement among experts; This indicates the degree of disagreement obtained from the expert reassessment after the supplementary testing; The larger this amount, the more helpful the action is in mitigating conflicts between different expert conclusions.

[0110] ; in, Indicates candidate actions The set of nodes involved; Indicates candidate actions The set of edges involved; This indicates the suspected fault risk of node n at the current moment; This indicates the potential fault risk of edge e at the current moment; This represents the weighting coefficient for node risk and edge risk; The larger the quantity, the more the candidate supplementary testing action can cover the current high-risk areas.

[0111] Step 5: After obtaining the state information and candidate action information, the active testing policy network outputs the corresponding decision score or selection probability for each candidate supplementary test action.

[0112] First, the current diagnostic decision state vector Encode the data to obtain a global state representation: ; in, This represents the hidden vector of the current global diagnostic state. Represents the state-encoded weight matrix; Indicates the bias term; This represents a nonlinear mapping function.

[0113] For each candidate retest action For its action feature vector Encode: ; in, This represents the hidden action vector of the i-th candidate action; This represents the action encoding weight matrix; This represents the action coding bias term.

[0114] Subsequently, the global state representation and the candidate action representation are fused to obtain the action-related context representation: ; in, Indicates candidate actions Joint representation in the current diagnostic context; This represents element-wise multiplication; Represents the absolute value of each element; These represent the weights and biases of the fusion layer, respectively.

[0115] Calculate a policy score for each candidate action: ; in, Represents the original policy score of the i-th candidate retest action; Indicates the output layer parameters; This indicates the output bias term.

[0116] Considering that some candidate actions may not be executable at the current moment, such as due to unavailable resources, unmet preconditions, insufficient budget, occupied paths, or security constraints, an executable mask is applied to the action scoring: ; in, This indicates the action score after constraints have been added; Indicates whether the i-th candidate action is executable at the current time; Indicates that it is executable; This indicates that the action is not executable. This represents a sufficiently large positive integer used to strongly suppress non-executable actions.

[0117] Normalize the constrained action scores to obtain the selection probability of candidate retest actions: ; in, This represents the probability that the i-th candidate action is selected; This represents the temperature parameter, used to adjust the sharpness of the action selection distribution; N represents the number of current candidate actions for retesting. Output a retest action .

[0118] Step 6: After the active testing strategy network outputs the supplementary testing action, it needs to be inserted into the current executable action sequence. Let the current action sequence to be executed be: ; in, This indicates the queue of test actions that have not yet been completed. This represents the existing test actions in the queue; L represents the length of the current action sequence.

[0119] For the retest action It can select the optimal insertion position while satisfying priority constraints, resource constraints, and safety constraints. : ; in, Indicates the optimal insertion position; This indicates that the supplementary test action will be inserted into the first... The additional time cost incurred by each location; This indicates the additional switching, calibration, or execution costs incurred when inserting at this location; This represents the cost of perturbations to the stability or risk control of the current testing sequence; This indicates the corresponding cost weight.

[0120] Insertion also requires the following conditions to be met: ; in, Indicates whether the dependencies between the preceding and following objects are satisfied; Indicates the first Whether the resource is available when the action is performed at a given location.

[0121] After the insertion is complete, the update action sequence is as follows: ; Then, a supplementary test is performed, and a snapshot of the supplementary test is taken.

[0122] Step 7: Perform the supplementary test. Then, collect the corresponding supplementary test snapshots. This data is then incorporated into the test snapshot sequence to update subsequent failure prediction results. ; in, This represents the set of test snapshots before the retest; This represents the set of test snapshots updated after the retest; This indicates a newly collected snapshot of the supplementary test.

[0123] At the same time, update the device health status graph based on the supplementary snapshot: ; in: This represents the equipment's health status before the retest. This represents the equipment health status diagram after the supplementary testing. This indicates a graph state update operation.

[0124] By re-executing the aforementioned spatiotemporal joint health modeling and expert fusion process, new fault prediction results are obtained: ; If the new fault prediction result still meets the preset triggering conditions, the active test strategy network can be invoked again.

[0125] Step S5, which describes the process of performing fault location and diagnosis by combining supplementary test snapshots, includes: Based on diagnostic correlation data, correlation calculations are performed, and graph attention mechanism is used to establish the correlation between anomalies and nodes and edges, and diagnostic results are output. The diagnostic correlation data is a dataset used to characterize test anomalies, retest feedback, historical baselines, link context, and model inference results. The diagnostic results include fault categories, fault location probability distributions, confidence scores, and remaining life estimates of the target component.

[0126] Step S5, the incremental learning process based on the diagnostic results, includes: Perform model training and updates based on model training data; The model training and update include self-supervised pre-training, supervised fine-tuning, and incremental learning; the model training data includes normal sample data, simulated fault sample data, and labeled feedback data. Physical consistency constraints are introduced during the model training and update process.

[0127] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automated testing and fault prediction method for frequency converter equipment, characterized in that, include: Perform initialization and loopback self-check on the test resources; Collect correlated data to construct a test snapshot sequence and device health status graph; Three operations—multimodal coding, spatiotemporal joint health modeling, and hybrid expert model fusion—are performed on the test snapshot sequence and equipment health status diagram to obtain fault prediction results. Based on the fault prediction results, generate supplementary test actions and collect supplementary test snapshots; Fault location diagnosis is performed by combining supplementary test snapshots, and incremental learning is performed based on the diagnosis results; The process of generating supplementary testing actions based on fault prediction results includes: The results from all experts will be used as the fault prediction result. When the fault prediction result meets the preset triggering conditions, the active testing strategy network is invoked. The proactive testing strategy network receives the current diagnostic decision status information and outputs supplementary testing actions. It inserts the supplementary testing actions into the executable action sequence and executes them to collect supplementary test snapshots. The working process of the active testing strategy network is as follows: The system uses the comprehensive health modeling results obtained from integrating the outputs of various experts as the basis for current fault prediction; it triggers a judgment on whether further supplementary testing is needed for the current diagnosis; upon triggering, it constructs current diagnostic decision state information including fault status, diagnostic uncertainty, test execution status, and candidate supplementary testing value; it evaluates the value and scores the strategies of candidate supplementary testing actions; under the condition of satisfying executable constraints, it outputs the optimal supplementary testing action; it inserts the supplementary testing action into the current executable action sequence and executes it; it updates the test snapshot sequence, equipment health status diagram, and fault prediction results using supplementary testing snapshots; thus forming a closed-loop proactive diagnosis and supplementary testing decision-making process. Before invoking the active testing strategy network, it is determined whether the current fault prediction result meets the preset trigger condition. The preset trigger condition is that the currently executed test actions have failed to effectively reduce diagnostic uncertainty, which is expressed as follows: ; This represents the diagnostic entropy at the t-th test time. Indicates the probability of the fault category; The probability interval between the two types of head failures is: ; This represents the index of the fault category with the highest predicted probability. This represents the index of the fault category with the second highest predicted probability. The degree of expert disagreement is defined as: ; Indicates the number of experts; This represents the fusion weight of the m-th expert at time t; This represents the output of the m-th expert; This represents the overall result after fusion; Before making a decision on candidate actions, the active testing policy network calculates the value score for each candidate action. The value score is represented as follows: ； Let represent the comprehensive value score of the i-th candidate supplementary test action at time t; This indicates the expected decrease in diagnostic uncertainty after performing this action; This indicates the expected increase in the head candidate fault differentiation interval after performing this action; This indicates the expected reduction in expert disagreement after the action is performed. This indicates the coverage gain of the action on the currently suspected faulty node, suspected faulty edge, or critical mechanism; This indicates the cost of performing the action; Indicates the time taken to perform the action; This indicates the risk of performing the action or the cost of perturbing the current test sequence; These are the weighting coefficients for each item; ;o represents the possible supplementary test result observed after performing the candidate supplementary test action; Indicates the current state Next action Predicted distribution of post-observation results; This indicates the uncertainty after taking a supplementary snapshot and re-diagnosing; ; This represents the probability interval between the current best and second-best failures; This represents the probability interval obtained after the retest; ; Indicates the current degree of disagreement among experts; This indicates the degree of disagreement obtained from the expert reassessment after the supplementary testing; ; Indicates candidate actions The set of nodes involved; Indicates candidate actions The set of edges involved; This indicates the suspected fault risk of node n at the current moment; This indicates the potential fault risk of edge e at the current moment; This represents the weighting coefficient for node risk and edge risk; After obtaining state information and candidate action information, the active testing policy network outputs a corresponding decision score or selection probability for each candidate retest action; and outputs the current diagnostic decision state vector. Encode the data to obtain a global state representation: ; This represents the hidden vector of the current global diagnostic state; Represents the state-encoded weight matrix; Indicates the bias term; Represents a nonlinear mapping function; For each candidate retest action For its action feature vector Encode: ; This represents the hidden action vector of the i-th candidate action; This represents the action encoding weight matrix; This represents the action coding bias term; By fusing the global state representation with the candidate action representation, we obtain the action-related context representation: ; Indicates candidate actions Joint representation in the current diagnostic context; This represents element-wise multiplication; Represents the absolute value of each element; These represent the weights and biases of the fusion layer, respectively. Calculate a policy score for each candidate action: ; Represents the original policy score of the i-th candidate retest action; Indicates the output layer parameters; Indicates the output bias term; Apply an executable mask to the action score: ; This indicates the action score after constraints have been added; Indicates whether the i-th candidate action is executable at the current time; Indicates that it is executable; This indicates that the action is not executable. This represents a sufficiently large positive integer used to strongly suppress non-executable actions; Normalize the constrained action scores to obtain the selection probability of candidate retest actions: ; This represents the probability that the i-th candidate action is selected; This represents the temperature parameter, used to adjust the sharpness of the action selection distribution; N represents the number of current candidate actions for retesting. Output a retest action ; After the proactive testing strategy network outputs a supplementary testing action, it inserts it into the currently executable action sequence. Let the current action sequence to be executed be: ; This indicates the queue of test actions that have not yet been completed. This represents the existing test actions in the queue; L represents the length of the current action sequence. For the retest action It can select the optimal insertion position while satisfying priority constraints, resource constraints, and safety constraints. : ; Indicates the optimal insertion position; This indicates that the supplementary test action will be inserted into the first... The additional time cost incurred by each location; This indicates the additional switching, calibration, or execution costs incurred when inserting at this location; This represents the cost of perturbations to the stability or risk control of the current testing sequence; Indicates the corresponding cost weight; Insertion also satisfies: ; Indicates whether the dependencies between the preceding and following objects are satisfied; Indicates the first The resource availability is checked when the action is performed at each location; after insertion, the update action sequence is as follows: .

2. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The process of initializing and performing a loopback self-test on the test resources includes: Detect the communication status of the test resources and load the corresponding instrument address configuration, test process template, routing table, device health status diagram, historical baseline database, and model threshold configuration; The control signal source outputs a standard signal, which forms a loop path through a switch matrix, and the loop data is collected by a spectrum analyzer and a power meter. The loopback data is constructed into a test snapshot and compared with the historical baseline database. Based on the comparison results, it is determined whether the loopback path meets the preset baseline conditions, and the loopback self-test results are output. The test resources include signal sources, spectrum analyzers, power meters, switch matrices, and the module under test.

3. The automated testing and fault prediction method for frequency converters according to claim 2, characterized in that, Perform the following operations before collecting related data: The test process template is parsed into a sequence of executable actions; The six-throw single-pole channel and the two-throw single-pole channel in the switch matrix are abstracted into a routing resource pool, and a mapping relationship between the logical routing table and the actual switch action is established. Execute path connection instructions, path disconnect instructions, and path query instructions according to the executable action sequence to establish the current routing path; The working mode and parameter configuration are sent to the module under test based on the executable action sequence.

4. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The process of constructing the test snapshot sequence includes: Related data is collected via a unified data bus and aligned to a unified timeline; A test snapshot is constructed based on identification information, path configuration information, measurement result information, and status monitoring information. Combine consecutive test snapshots into a test snapshot sequence.

5. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The process of constructing the device health status diagram includes: Treat the test system entity and the object under test as nodes; The relationships between entities, including radio frequency signal connection, control, power supply influence, clock synchronization, and thermal coupling, are treated as edges. The test system entity includes test instruments, routing switching units, transmission interconnection units, and environmental conditioning units, while the test object entity includes the test module and its internal functional sub-units.

6. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The process of performing multimodal encoding on the test snapshot sequence and device health status graph includes: The features within the test snapshot sequence and device health status map are divided into four categories: structured numerical features, spectrum and scan curve features, event flow features, and topological features. Different neural networks are used to process the four features respectively to obtain a unified state representation containing time information, frequency domain information, behavioral information, and structural information. Structured numerical features are encoded using a temporal convolutional network; The spectral and scanning curve features are encoded using a frequency domain transformer network; The event stream features are encoded using an event sequence encoding method; Topological features are encoded using a graph attention network.

7. The automated testing and fault prediction method for frequency converters according to claim 6, characterized in that, The process of performing spatiotemporal joint health modeling and hybrid expert model fusion on test snapshot sequences and device health status graphs includes: Based on a unified state representation, the state of each node and each edge is updated according to the test time to obtain the spatiotemporal joint health characteristics. The spatiotemporal joint health features are input into a hybrid expert model; The gating network assigns weights to each expert based on the current test context and merges the outputs of each expert.

8. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that: The current diagnostic decision status information includes at least one of the following: fault status information, diagnostic uncertainty status information, test execution status information, and candidate supplementary test value information.

9. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The process of performing fault location and diagnosis by combining supplementary test snapshots includes: Based on diagnostic correlation data, correlation calculations are performed, and graph attention mechanism is used to establish the correlation between anomalies and nodes and edges, and diagnostic results are output. The diagnostic correlation data is a dataset used to characterize test anomalies, retest feedback, historical baselines, link context, and model inference results. The diagnostic results include fault categories, fault location probability distributions, confidence scores, and remaining life estimates of the target component.

10. The automated testing and fault prediction method for frequency converters according to claim 1, characterized in that, The incremental learning process based on diagnostic results includes: Perform model training and updates based on model training data; The model training and update include self-supervised pre-training, supervised fine-tuning, and incremental learning; the model training data includes normal sample data, simulated fault sample data, and labeled feedback data. Physical consistency constraints are introduced during the model training and update process.

Images