Ship electronic equipment intelligent diagnosis method and system based on multi-source data fusion

Abstract

The application discloses a kind of ship electronic equipment intelligent diagnosis method and system based on multi-source data fusion, belong to electronic equipment test field.The method is by collecting radio frequency response, BIT state in machine and historical maintenance record, splice generation high-dimensional comprehensive feature vector, utilize Mahalanobis distance and kernel density estimation carry out health baseline evaluation;When equipment is abnormal, expert system and random forest model trained by SMOTE oversampling are used in parallel to make decision fusion, accurately locate fault unit;While introducing reinforcement learning intelligent scheduling agent, based on multidimensional resource state when network node fails, comprehensive reward optimization is carried out, test task is encapsulated as service package and issued to optimal replacement node, drive replacement node dynamically reconfigure radio matrix and FPGA logic to continue test.The application realizes the predictive accurate maintenance of equipment, and greatly improves the anti-damage network resilience of support system.

Description

Technical Field

[0001] This invention belongs to the field of electronic equipment testing and integrated support technology, specifically relating to an intelligent diagnostic method and system for shipboard electronic equipment based on multi-source data fusion. Background Technology

[0002] With the increasing complexity of modern electronic equipment systems (such as radar, electronic countermeasures, communications, and data links), their operational readiness directly impacts overall system effectiveness. Currently, field testing and maintenance of such equipment (docks, decks, etc.) primarily rely on manual troubleshooting using discrete instruments or Automatic Test System (ATS).

[0003] The existing technical solutions mainly have the following technical problems:

[0004] First, in terms of fault diagnosis, existing test diagnostics mainly rely on scalar status indicators such as "pass / fail" provided by the device's built-in test (BIT), or solely on single radio frequency parameters measured by external instruments. This diagnostic data source is relatively singular, lacking in-depth integrated analysis of external electromagnetic response, internal electrical status, and historical degradation patterns. This makes it difficult to accurately locate slow degradation of equipment performance, hidden intermittent faults, and complex faults, keeping maintenance efforts at the "retroactive repair" level and hindering the upgrade to "predictive maintenance."

[0005] Second, regarding system coordination and resilience, traditional test equipment is often designed as an independent "information silo," with its internal hardware topology (such as RF paths and processing modules) fixed from the initial design stage. When test nodes experience hardware anomalies, communication failures, or combat damage in harsh environments, the test mission will be forced to completely cease. Existing systems lack resource awareness from a global network perspective and cannot dynamically migrate and re-instantiate unfinished test missions onto other intact nodes, resulting in a lack of flexibility and resilience in the overall support system. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to address the problems of low fault diagnosis accuracy, lack of predictive ability, and easy interruption of support missions due to single-point failures of test nodes in the existing technology. This invention provides an intelligent diagnostic method and system for ship electronic equipment based on multi-source data fusion.

[0007] Technical Solution: To achieve the above objectives, this invention provides an intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion, comprising the following steps:

[0008] Step S1, Multi-source data acquisition and feature stitching: Acquire external RF response data generated during the test of the device under test, internal test BIT data of the device under test acquired through the data bus, and historical maintenance records of the device under test. Align the above heterogeneous data according to absolute timestamps and stitch them together to generate a comprehensive feature vector.

[0009] Step S2, Health status assessment based on multivariate model and hardware interruption: Calculate the degree of deviation between the high-dimensional comprehensive feature vector and the pre-established historical health baseline data of normal equipment. When the degree of deviation exceeds the preset health threshold, determine that the tested equipment is in an abnormal state, trigger the underlying hardware abnormal interruption to suspend the current test process, and enter the fault diagnosis and networked scheduling process.

[0010] Step S3, precise fault diagnosis by integrating mechanism and data: The high-dimensional comprehensive feature vector that determines the anomaly is input in parallel into the rule-based expert system and the pre-trained machine learning classification model. The LRU code of the faulty replaceable unit and its confidence level are output through the decision fusion mechanism to generate predictive maintenance suggestions.

[0011] Step S4, Reinforcement Learning-Based Network Node Hardware Reconstruction and Elastic Task Scheduling: When the underlying hardware abnormality or communication failure of the currently executing test node is detected, the intelligent scheduling agent running on the main control module obtains the state space of the global network nodes based on the reinforcement learning algorithm, outputs the optimal replacement target node, and encapsulates the unfinished test tasks into standardized service packages and sends them to the optimal replacement target node to drive the optimal replacement target node to allocate underlying physical resources for hardware reconstruction and continue to execute the test tasks.

[0012] This invention also provides an intelligent diagnostic system for electronic devices based on multi-source data fusion, used to implement the above method, comprising:

[0013] The multi-source data acquisition and feature stitching module is used to acquire external RF response data, in-machine test BIT data and historical maintenance records, and generate high-dimensional comprehensive feature vectors;

[0014] The health status assessment module is used to calculate the degree of deviation between the high-dimensional comprehensive feature vector and the health baseline data. When the deviation exceeds the limit, an anomaly is determined, an anomaly interruption is triggered, and the scheduling process is initiated.

[0015] The precise fault diagnosis module is used to input high-dimensional comprehensive feature vectors in parallel into expert systems and machine learning classification models, and output LRU codes and their confidence scores through decision fusion;

[0016] The networked elastic scheduling module is used to intelligently schedule agents to obtain the state space based on reinforcement learning and output the optimal replacement target node when a node fails. It then sends out service packages to drive the target node to allocate physical resources for hardware reconstruction and continue to execute tests.

[0017] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the above-described method.

[0018] Beneficial effects: Compared with the prior art, the advantages of the present invention are as follows:

[0019] This technology, which combines external RF response data, internal BIT data, and historical records to generate a high-dimensional comprehensive feature vector, breaks down the traditional "information silos" between internal and external equipment data. It aligns macroscopic parameters characterizing electromagnetic properties with microscopic parameters characterizing electrical status on a unified time axis. This enables multi-dimensional feature cross-validation, effectively overcoming the insensitivity of single BIT data to slow degradation and complex faults, allowing the system to gain insights into the deep health status of the equipment.

[0020] By calculating Mahalanobis distance and dynamically setting health thresholds in conjunction with kernel density estimation, the dimensional differences of various heterogeneous data dimensions are effectively eliminated while taking into account the correlation between features. Combined with dynamic threshold setting based on kernel density estimation, the system can adaptively adjust health judgment criteria in harsh environments such as wide temperature variations and strong electromagnetic interference on ships, thereby maintaining high sensitivity and low false alarm rate for early performance degradation warnings.

[0021] The hybrid diagnostic model and decision fusion technology of "rule-based expert system + random forest" is characterized by its complementary combination of deterministic domain physical mechanism knowledge (manufacturer's "IF-THEN" rules) and pure data-driven algorithm (random forest) for handling high-dimensional nonlinear mappings. Simultaneously, the SMOTE algorithm is used to overcome the severe data imbalance problem in the military industry. This enables second-level interpretable localization of known faults and high-confidence generalized diagnosis of unknown edge faults, accurately locating faults down to the LRU (Low-Replaceable Unit) level.

[0022] This technology utilizes reinforcement learning to enable intelligent agents to acquire global state, calculate a comprehensive reward function, and issue service packages to drive hardware reconfiguration of successor nodes. Its underlying principle is that the system transcends the limitations of a single-machine perspective, decoupling the test task from specific hardware using a "service-oriented" approach. Simultaneously, the reinforcement learning model comprehensively considers network channel capacity, amplifier lifespan, and latency, making globally optimal load balancing decisions. Upon issuing commands, the underlying FPGA dynamically loads and the RF matrix switches in milliseconds in tandem. This ensures that when a test node unexpectedly fails or is damaged, other nodes can seamlessly take over and instantly transform into the required form, guaranteeing uninterrupted task completion. Attached Figure Description

[0023] Figure 1 This is a structural block diagram of the underlying hardware architecture upon which the intelligent diagnosis and protection method for electronic devices based on multi-source data fusion of this invention relies.

[0024] Figure 2 This is a flowchart of an intelligent diagnostic and protection method for electronic devices based on multi-source data fusion, provided by an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the embodiments described.

[0026] Example 1: This example provides an intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion, deployed in harsh outdoor environments such as ship decks and docks. The underlying hardware architecture of this method adopts a "control-RF physical separation" design, meaning the hardened display and control terminal (control domain) is deployed in a secure area, while the RF front-end host is placed in a high-radiation testing area. The two communicate wirelessly / wiredly through a national-level cryptographic hardware security module (SM2 two-way authentication and SM3 / SM4 real-time encryption). The intelligent control domain is the system's human-computer interaction and mission command center, mainly including the display and control terminal, an external control computer, or an external networked node, providing a graphical interface for mission planning, parameter setting, status monitoring, and data display.

[0027] All hardware modules in the RF front-end domain are housed in a high-density ruggedized chassis and interconnected via a built-in high-speed PXIe backplane. For example... Figure 1As shown, the system features an innovative SWaP-C core architecture within the chassis. Its built-in high-speed backplane carries the PCI Express data stream and PXI timing bus, ensuring the real-time performance and integrity of high-capacity data transmission. Simultaneously, the constructed star-shaped clock network distributes the system master clock directly to each module via dedicated clock lines with equal-length paths, achieving extremely low phase deviation and jitter. This provides a high-precision reference for core timing synchronization. Combined with direct-connect control I / O and trigger buses, it ensures efficient and stable control signal transmission and deterministic, low-latency coordination of multi-module operations. Ultimately, this makes the system a highly reliable, high-performance, high-synchronization-accuracy, and easily expandable and maintainable mobile signal generation and processing platform. The specific design of its internal modules is as follows:

[0028] The main control module, serving as the system's control and data processing hub, operates throughout the entire testing process. It receives operation commands from authorized display and control terminals / external control computers / external networked nodes via a wireless encrypted link or a wired transmission link. After parsing, it sends control parameters to the multi-functional complex scene signal generation module via a high-speed dedicated backplane and automatically loads them according to the test items. The multi-functional complex scene signal generation module generates the required intermediate frequency signal based on the control parameters and directly controls various frequency conversion modules, shared broadband frequency integration modules, power amplifier modules, and programmable RF switching matrices via the high-speed customized dedicated backplane to generate the required output signal and output it through a designated path. Simultaneously, it aggregates measurement data and communication data uploaded by the spectrum analyzer, the multi-functional complex scene signal generation module, and the serial communication module, along with status information reported by each module, for fusion processing and storage. During task execution, this module continuously monitors the system's health. Upon detecting power anomalies, module failures, or communication interruptions, it immediately triggers preset emergency procedures, such as sending protective commands to the matrix switch and power module, ensuring the safety of the system and field equipment.

[0029] The reconfigurable signal processing module is the core of the system's software-defined waveform capability. It adopts a dual-threshold detection mechanism and consists of two large-scale programmable logic devices (FPGA1 and FPGA2) and a multi-channel high-speed AD / DA converter. FPGA1 / FPGA2 acts as a reconfigurable signal processing engine. According to the task code issued by the computer, it dynamically loads the corresponding algorithm kernel and directly generates baseband or intermediate frequency digital signal streams at the hardware level, including complex radar signal generation, data link signals (Link16, Link4A, PL16, etc.), radar target signal simulation, SAR echo signal simulation, navigation signal simulation, guidance signal simulation, etc. These digital streams are converted into intermediate frequency analog signals (IF) by four high-speed DAs (DA1-DA4) and sent to the subsequent frequency conversion modules 1, 2, and 3, respectively. At the same time, the high-speed AD channel receives the intermediate frequency signal after down-conversion from frequency conversion module 3 and sends it to FPGA1 for closed-loop analysis.

[0030] The shared broadband frequency synthesis module serves as the frequency reference for the entire RF link. It receives the 100MHz clock synchronization signal from the main control module and, through internal multi-loop phase-locked loop (PLL) technology, generates low-phase-noise local oscillator signals (LO1, LO2, LO3, LO4) covering the entire frequency band required by the system. These signals are then precisely distributed to frequency conversion modules 1, 2, and 3 via an equal-length RF blind-plug backplane. Furthermore, it generates a clean synchronization clock signal, which is output to the programmable logic device in the reconfigurable signal processing module. This ensures that the timing of the digital signal generation is strictly aligned with the local oscillator phase of the RF frequency conversion, enabling high-fidelity simulation of complex electromagnetic scenarios.

[0031] The frequency converter module and power amplifier module (RF link layer) consist of frequency converter module 1, frequency converter module 2, frequency converter module 3 and power amplifier module.

[0032] Frequency converter module 1 receives the intermediate frequency (IF) signals (0.1GHz~1.1GHz) output from DA1 and DA2, and performs independent two-way frequency conversion operations in conjunction with the LO signal allocated by the frequency synthesizer. The up-converted RF signals are then sent to power amplifier module 1 (covering 7-13GHz) and power amplifier module 2 (covering 2-18GHz) respectively through power divider 1 and power divider 2. Frequency converter module 2 receives the IF signal (0.1GHz~1.1GHz) from DA4, up-converts it, and sends it to power amplifier module 4 (covering 2-18GHz) for linear amplification. Frequency converter modules 1 and 2 mainly complete the frequency shifting and preliminary conditioning of the signals, providing RF signals that meet the input requirements for the power amplifier links.

[0033] Frequency conversion module 3 (bidirectional): In the transmitting direction, it receives the DA3 signal, up-converts it, and sends it to the power amplifier module 3 (covering 0.1-18GHz); in the receiving direction, it receives externally input high-frequency signals such as "radar excitation signals", down-converts them into intermediate frequency signals, and sends them back to the AD channel of the reconfigurable signal processing module.

[0034] The power amplifier modules are responsible for linearly amplifying the frequency-converted signal to achieve the radiated power level required for field testing. Each power amplifier module is optimized for its specific operating frequency band (Power Amplifier Module 1: 7-13GHz; Power Amplifier Module 2: 2-18GHz; Power Amplifier Module 3: 2-18GHz; Power Amplifier Module 4: 0.1-18GHz) to ensure sufficient output power, good linearity, and gain flatness across a wide frequency range. The modules integrate comprehensive power monitoring, temperature compensation, and protection circuits, enabling stable operation under the harsh environmental temperature changes of the ship's deck. They also automatically implement protective measures in case of overload or deterioration of the standing wave ratio (VSWR) to ensure equipment safety.

[0035] The programmable RF switching matrix module, serving as the intelligent routing hub of this system, employs a broadband solid-state MEMS microwave switch array (isolation >70dB, response <5ms). Its core function is to replace manual cable plugging and unplugging, enabling software-defined test topology. The matrix's inputs aggregate various high-power output signals from power amplifier modules 1 to 4, external radar excitation inputs, and signals under test. Based on the direct drive commands from the main control module, the matrix routes these signals millisecond-level to the physical output ports on the front panel (such as the 7-13G output of scenario 1, the 2-18G output of scenario 2, etc.) or back to the internal spectrum analyzer module via the closing of internal cross-point switches.

[0036] As a built-in measurement unit of the system, the spectrum analyzer module provides standard frequency domain signal analysis capabilities. It receives the signal under test routed by the matrix switch and can measure multiple indicators such as spectrum, power, phase noise, and spurious signals in real time. The results are then uploaded to the main control module through a dedicated backplane. This allows operators to perform self-testing of the generated signal quality or preliminary evaluation of the response of the device under test within the system without connecting an external oscilloscope or spectrum analyzer, thus achieving integrated "test-analysis".

[0037] The IFF (Identification Friend or Foe) module is a dedicated signal processing unit that fully implements the simulation and calculation functions of the IFF interrogation and response protocol. It can generate standard-compliant interrogation signal waveforms according to instructions, amplify them, and transmit them. Simultaneously, its receiving channel can process and decode signals from the transponder in real time, completing identification and information extraction, and reporting the results to the main control system. The integration of this module enables the platform to have independent IFF equipment testing and verification capabilities.

[0038] To overcome the drawbacks of low equipment availability and high troubleshooting costs caused by the traditional "post-incident maintenance" model in field support, and to address the increasingly prominent challenges of latent faults and performance degradation in modern complex equipment, this system integrates an intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion. It aims to fully utilize the massive amounts of multi-dimensional data generated during automated testing, combined with equipment mechanism knowledge and data-driven algorithms, to achieve real-time monitoring of equipment health status, early warning of faults, precise fault location, and prediction of remaining service life. This will drive a fundamental shift in support models towards "pre-emptive prediction" and "condition-based maintenance," significantly improving combat readiness and optimizing support resource allocation. Figure 2 As shown, the method specifically includes the following steps:

[0039] Step S1: Multi-source data acquisition and feature stitching based on interfaces.

[0040] During the test execution phase, the system ensures that it synchronously collects two types of data using the PXIe high-precision clock synchronization network (sub-microsecond time synchronization reference):

[0041] (1) External RF response data: obtained directly through the system's built-in spectrum analyzer module and power monitoring link. This includes: the radiated power of the radar transmitted signal; the single-sideband phase noise (in dBc / Hz) at frequency offsets of 1kHz, 10kHz, and 100kHz measured using the phase detection method; and the power efficiency calculated by monitoring the power amplifier output power and DC power consumption, and extracting characteristic curves such as peak value, valley value, and inflection point within its operating frequency band.

[0042] (2) Internal Test (BIT) Data: The internal status of the device under test is obtained through the 1553B bus or Gigabit Ethernet interface. This includes: analog quantities such as the operating voltage and current of the internal TR components of the radar; digital quantities such as channel status flags and alarm codes; and performance indicators such as the target detection probability (obtained through multiple independent measurements under a preset signal-to-noise ratio) and the signal-to-noise ratio of the receiving channel, which are statistically obtained by the radar signal processor based on the constant false alarm rate (CFAR) detection principle.

[0043] Simultaneously, the serial number of the device under test (used to distinguish individual differences) is collected, and the internal test (BIT) status words reported by each replaceable unit (LRU) within the equipment are obtained via the 1553B bus or Ethernet interface. These include, but are not limited to, the operating status parameters (such as voltage, current, channel status flags, alarm codes, etc.) of the signal processing unit, receiving unit, antenna unit, and power supply unit. All data is accompanied by a globally unified timestamp issued by the PTP protocol to ensure an absolute correspondence between external performance characterization and internal physical state on the timeline. The above data with absolute timestamps are aligned and concatenated to generate a high-dimensional comprehensive feature vector. .in The feature dimension typically ranges from 20 to 50. These correspond to the feature components of various types of collected data.

[0044] Step S2: Health status assessment based on multivariate model and hardware interrupt.

[0045] The model computation layer is the intelligent core of this technology, integrating three types of collaborative mathematical models, corresponding to health assessment, trend prediction, and fault diagnosis, respectively.

[0046] The health baseline model is established based on multivariate statistical process control theory. By analyzing a large amount of characteristic data from similar health equipment, its mean vector is calculated. Covariance Matrix Construct a multivariate Gaussian reference distribution to represent "health status". .

[0047] During online testing, the feature vector of the device under test is calculated. Mahalanobis distance from the healthy baseline:

[0048]

[0049] Mahalanobis distance, through the inverse of the covariance matrix, effectively eliminates the influence of different dimensions of various indicators, while also taking into account the correlation between features.

[0050] The system uses kernel density estimation (KDE) to dynamically determine the health threshold. Specifically, at a preset significance level (e.g., ... Under the condition that the integral of the probability density function is satisfied, critical distance This is the health threshold. When the Mahalanobis distance exceeds the health threshold, an abnormal interruption is triggered in the underlying hardware, and the feature vector is input into the pre-trained random forest classification model. The feature vector x is constructed as described above, containing multi-dimensional information such as radio frequency response features, BIT status word, ambient temperature, and equipment serial number.

[0051] In addition to using Mahalanobis distance for instantaneous anomaly detection, the system also includes predictive models for the degradation trends of key performance indicators. Specifically, for degradation data with trends or periods, the system uses exponential smoothing for short-term predictions; for more complex non-stationary degradation processes, the system extracts historical time-series sequences of key performance indicators (such as transmit power and noise figure) and uses an autoregressive integral moving average (ARIMA) model for medium- and long-term trend extrapolation. By performing d-order differencing on the original sequences to make them stationary, a predictive model is fitted, thereby predicting the future evolution trajectory of the indicators, estimating the remaining useful life of the equipment or the time of performance exceeding limits, and achieving true pre-emptive prevention.

[0052] To ensure accuracy during long-term field operation, the system implements an adaptive update mechanism: using stable test data that has not triggered anomalies, combined with equipment service life labels, it periodically refits a Gaussian distribution and dynamically corrects the data. and .

[0053] When a certain calculation When this occurs, it indicates that the equipment has experienced early degradation or potential failure. The system immediately triggers a hardware abnormality interrupt signal on the underlying dedicated backplane, suspends the regular test sequence, and triggers step S3.

[0054] Step S3: Accurate fault diagnosis by integrating mechanisms and data.

[0055] After obtaining the abnormal feature vectors, they are input in parallel into the dual diagnostic core of the decision fusion mechanism:

[0056] Engine A (Rule-Based Expert System): Encodes the fault tree and code table from the manufacturer's equipment maintenance manual into "IF-THEN" production rules. For example, "IF Sensitivity decrease >3dB AND BIT Receiver Channel Voltage Normal AND Radiated Power Normal THEN Suspected Low-Noise Amplifier Degradation". Outputs rule matching results and confidence levels.

[0057] Engine B (Machine Learning Classification Model): Employs a random forest classifier. For an input feature vector x, the random forest performs a diagnosis by aggregating the voting results of multiple decision trees. ,in This indicates the mode selection operation. Due to the inherent "extreme imbalance" of the samples (a large number of normal samples and a very small number of samples of a certain type of fault), the SMOTE (Synthetic Minority Oversampling Technique) method is used to interpolate and augment the minority class fault features before training. Subsequently, a random forest containing 100-500 decision trees is trained, with the maximum number of features per tree selected as the maximum. Maximum tree depth Hyperparameters are optimized through grid search and 5-fold cross-validation, enabling the model to fit a nonlinear mapping from high-dimensional degenerate features to specific LRU failures.

[0058] During the online diagnostic phase, after an abnormal interruption is triggered, the current feature vector is fed in parallel into K independent decision trees (K = nestimators) of a random forest for inference. Each decision tree outputs a predicted LRU fault code. The system uses a majority voting mechanism to determine the final fault unit code and outputs the vote percentage as the confidence level of the diagnostic conclusion (e.g., "Variable frequency module 1 fault, confidence level 95%"). The diagnostic results are synchronously stored in the historical database for subsequent maintenance analysis and model iteration optimization.

[0059] Step S4: Hardware reconstruction of network nodes and elastic scheduling of tasks based on reinforcement learning.

[0060] This system supports multi-node networking to form an internal support network. This step is triggered when a test node experiences a serious anomaly while performing a task (e.g., power failure, communication link interruption, or damage).

[0061] The intelligent scheduling agent running within the system's main control unit acquires the real-time status of global network nodes. The state space is constructed as follows: dimensional global state vector ( (Number of nodes). For nodes. Its six components are: the number of remaining available radio frequency channels. Amplifier remaining lifespan percentage Current battery level Task queue length, highest priority, online status flag.

[0062] Define the action space: When execution node j fails, the Agent selects a specific node as the migration target from the set of remaining available nodes (i.e., all online nodes except j) for the interrupted guarantee task; if there is no replacement node, define a special action a=0 (no node to migrate), and the action space size is N-1.

[0063] This invention employs the Proximal Policy Optimization (PPO) algorithm to construct a neural network model. For scenarios with a small number of nodes (N≤5), a Deep Q-Network (DQN) is used instead of PPO to reduce model complexity. PPO uses an Actor-Critic architecture: the input layer receives the normalized global state vector; 2-3 shared hidden layers are set (typically 256-dimensional → 128-dimensional), with the ReLU activation function; the Actor network (policy network) outputs the probability distribution of each action. The Critic network (value network) outputs the value V(s) of the current state. Model training parameter settings: network weights are initialized using Xavier, the optimizer is Adam; the Actor network learning rate is 3e-4, and the Critic network learning rate is 1e-3; discount factor γ=0.99, GAE parameter λ=0.95, and PPO pruning threshold ε=0.2. If DQN is used, training is completed in conjunction with an experience replay mechanism and the target network.

[0064] Before formal deployment, the model needs to be pre-trained in a high-fidelity digital twin simulation environment. Three types of models—resource consumption, fault injection, and communication latency—are built based on historical node operation data, and task flows (including task type, priority, and execution duration) that conform to real-world scenarios are randomly generated. During training, complex situations such as node failure and communication link interruption are randomly simulated, allowing the agent to learn through repeated trial and error in a virtual environment.

[0065] The experience trajectory generated by the interaction between the agent and the simulation environment ( The data are stored in an experience pool with a capacity of 2048. The advantage estimate is calculated using the GAE method. Data in batches of 64 are drawn from the experience pool to update the network. Each batch of data is repeated 10 times until the model converges (judgment criteria: the average cumulative reward no longer increases after 100 consecutive training rounds, or the success rate of the validation set scheduling is ≥95%).

[0066] To enable the Agent to make optimal scheduling decisions, a comprehensive reward function is designed. The core objective is to maximize the overall network guarantee efficiency. The total reward r consists of four components: task success reward... Penalties for mission failure Recovery time penalty Resource Balance Rewards The calculation formula is:

[0067]

[0068] Among them, the reward for successful task If the task successfully migrates to the target node and completes execution, a positive reward of +1.0 is given; a penalty is applied if the task fails. If the migration process fails due to communication interruption, node inability to reconstruct, or other reasons, a negative penalty of -1.0 is applied. The pre-trained, converged reinforcement learning model has its weights fixed and deployed to each support node, enabling it to output the optimal replacement node decision in milliseconds under complex constraints. Recovery time penalty. , The penalty coefficient is set to 0.001 in this embodiment through simulation debugging and verification, in order to coordinate the penalty of recovery time (seconds) with other reward items (usually ±1) on a magnitude scale; The time (in seconds) from node failure to task recovery. Resource balancing reward interval. ,in This is the weighting coefficient, which is set to 0.1 in this embodiment. The number of remaining available radio frequency channels for the target node. This represents the maximum number of radio frequency channels for a node. This represents the percentage of remaining lifespan of the target node's power amplifier. The target node's current battery level.

[0069] When running online, the Agent performs scheduling and task migration according to the following process:

[0070] The first step is to monitor network node failure events in real time (through heartbeat detection or fault reporting interface).

[0071] The second step is that after a node fails, the agent queries the latest resource status of all online nodes through the internal security network to construct a global state vector.

[0072] The third step is to input the global state vector into the Actor network and output the optimal transfer action a* with the highest probability.

[0073] The fourth step is to package the incomplete task information (including task ID, test script, algorithm IP core, etc.) of the failed nodes into a task service package;

[0074] The fifth step is to distribute the task service package to the target successor node via a secure communication link;

[0075] Step 6: After receiving the data, the target node automatically completes hardware link reconstruction and resumes task execution within seconds.

[0076] The entire decision-making process has a delay of less than 100ms, and the overall task recovery time is less than 10s, achieving seamless migration of the support task.

[0077] After the scheduling decision is generated, the original system master controller packages the incomplete "test scripts, required signal processing algorithm IP cores, and data analysis plugins" of the failed node into a standardized "service package" and sends it to the selected replacement node through the gateway with encryption.

[0078] Upon receiving the service packet, the replacement node instantly triggers a reconfiguration of the underlying hardware: 1) The main control processor drives a fully programmable RF switching matrix (composed of MEMS solid-state switches) to establish a physical path for the specified frequency band within milliseconds; 2) The algorithm IP cores in the service packet are dynamically loaded into the reconfigurable partition of the local high-speed FPGA; 3) A shared broadband frequency synthesis module is simultaneously configured to output the corresponding local oscillator frequency. This hardware and software collaboration enables the replacement node to "transform" into a dedicated test environment for the interrupted test task within seconds, seamlessly restoring the support operation.

[0079] Example 2: Based on the above method, this example provides an intelligent diagnostic and support system for shipboard electronic equipment based on multi-source data fusion. The system includes: a multi-source data acquisition and feature stitching module, used to acquire external RF response data, in-system test BIT data, and historical maintenance records, aligning and stitching them according to absolute timestamps to generate a high-dimensional comprehensive feature vector; a health status assessment module, used to calculate the Mahalanobis distance between the high-dimensional comprehensive feature vector and the health baseline data, determining an anomaly and triggering a low-level hardware anomaly interruption when the distance exceeds a threshold; a precise fault diagnosis module, used to input the high-dimensional comprehensive feature vector in parallel into an expert system and a machine learning classification model, outputting the LRU code of the faulty replaceable unit and its confidence level; and a networked elastic scheduling module, used to calculate the optimal replacement target node based on reinforcement learning algorithms and the global node state space through an intelligent scheduling agent when a node experiences an abnormal interruption, and to issue service packages to drive the optimal replacement target node to perform hardware reconstruction and continue executing test tasks.

[0080] Furthermore, the present invention also provides a computer device, which can be a ruggedized industrial control computer or a portable display and control terminal. The device internally includes a memory, a processor, and a communication bus. The memory stores an executable computer program, and when the processor executes the program, it can implement all the method steps described in Embodiment 1.

[0081] Example 3: To further illustrate the engineering feasibility of the intelligent diagnostic mechanism based on multi-source data fusion in this invention, this example provides a specific application scenario for shipborne radar and electronic countermeasures equipment.

[0082] In this scenario, the system uses multi-source heterogeneous data fusion to perform precise unit-level positioning of the replaceable unit (LRU) within the device. The specific implementation process is as follows:

[0083] The first step is the extraction and fusion of multi-source data.

[0084] The system synchronously acquires the following three types of heterogeneous data to construct diagnostic feature vectors:

[0085] (1) External performance test data: For radar equipment, this includes the measured system sensitivity, the error rate of the XX-16 missile data link communication, the external radiation power, the target acquisition time and success rate; for electronic countermeasures equipment, this includes the single / multi-target alarm response time and accuracy, the measured value of the jamming power, and the beam pointing error.

[0086] (2) In-system test (BIT) data: BIT status words, key node voltage / temperature analog quantities and subsystem self-test results read in real time through the 1553B data bus.

[0087] (3) Manufacturer's fault knowledge base: Convert the fault code table and fault tree analysis results provided by the original equipment manufacturer into structured data.

[0088] The second step is to define typical failure modes and generate rules.

[0089] The system is pre-set and can identify a variety of typical low-level fault types.

[0090] Preset typical fault types (F) that can be injected / simulated:

[0091] Radar category (F1-F4): F1: Reduced gain of the low-noise amplifier in the receiving channel (affecting system sensitivity); F2: Reduced efficiency of the power amplifier in the transmitting channel (affecting the external radiated power); F3: Abnormal data link terminal transceiver module (affecting the XX-16 data link function); F4: Resource bottleneck of the signal processing board (affecting multi-target interception capability).

[0092] Electronic countermeasures (F5-F7): F5: Decreased sensitivity of the detection channel (affects alarm distance); F6: Insufficient output power of the jamming transmitter (affects jamming power); F7: Beam control driver deviation (affects beam pointing accuracy).

[0093] To address the aforementioned faults, the knowledge base in the rule-based expert system is transformed into "IF-THEN" production rules. For example, a specific rule configuration could be: "IF External measurement of system sensitivity decreases by > 3dB AND 1553B bus reports 'Receive channel A voltage is normal' AND External measurement of radiated power is normal THEN Diagnostic conclusion: Suspected low-noise amplifier (LNA) performance degradation."

[0094] The third step is to integrate diagnostic implementation with indicator evaluation.

[0095] When the device triggers a health threshold alarm, the system automatically combines the currently extracted comprehensive feature vector and BIT status word, and sends them in parallel to the expert system rule engine and random forest model. If the expert system rule matches successfully, it assigns high confidence; if the rule base does not cover the complex feature, it outputs the predicted fault category (such as F1 or F2) and probability based on the nonlinear mapping of the random forest.

[0096] The diagnostic center performs a weighted fusion of the results from both sources, ultimately outputting a high-confidence single fault unit report. The system in this embodiment possesses closed-loop evaluation capabilities. After completing the diagnosis, it automatically calculates the overall diagnostic accuracy (number of correct diagnoses / total number of tests), the recall rate for various fault types, and generates a confusion matrix. This allows for a detailed analysis of the intelligent diagnostic model's ability to distinguish between different fault types (such as F1 and F5), thus providing data support for the continuous iterative optimization of the diagnostic algorithm.

[0097] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for intelligent diagnosis of shipboard electronic equipment based on multi-source data fusion, characterized in that, include: Step S1, Multi-source data acquisition and feature stitching: Acquire external RF response data generated during the test of the device under test, internal test BIT data of the device under test acquired through the data bus, and historical maintenance records of the device under test. Align the above heterogeneous data according to absolute timestamps and stitch them together to generate a comprehensive feature vector. Step S2, Health status assessment based on multivariate model and hardware interruption: Calculate the degree of deviation between the high-dimensional comprehensive feature vector and the pre-established historical health baseline data of normal equipment. When the degree of deviation exceeds the preset health threshold, determine that the tested equipment is in an abnormal state, trigger the underlying hardware abnormal interruption to suspend the current test process, and enter the fault diagnosis and networked scheduling process. Step S3, precise fault diagnosis by integrating mechanism and data: The high-dimensional comprehensive feature vector that determines the anomaly is input in parallel into the rule-based expert system and the pre-trained machine learning classification model. The LRU code of the faulty replaceable unit and its confidence level are output through the decision fusion mechanism to generate predictive maintenance suggestions. Step S4, Reinforcement Learning-Based Network Node Hardware Reconstruction and Elastic Task Scheduling: When the underlying hardware abnormality or communication failure of the currently executing test node is detected, the intelligent scheduling agent running on the main control module obtains the state space of the global network nodes based on the reinforcement learning algorithm, outputs the optimal replacement target node, and encapsulates the unfinished test tasks into standardized service packages and sends them to the optimal replacement target node to drive the optimal replacement target node to allocate underlying physical resources for hardware reconstruction and continue to execute the test tasks.

2. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 1, characterized in that, In step S1: The external radio frequency response data is acquired in real time through the built-in measurement module and power monitoring link. The external radio frequency response data includes the radiated power of the radar transmitted signal, the single-sideband phase noise at a specified frequency offset, and the peak-valley characteristics within the operating frequency band. The in-machine test BIT data includes the analog values ​​of the operating voltage and current of the replaceable unit LRU, channel status flags, alarm codes, and the target detection probability statistically obtained based on the constant false alarm rate detection principle. The high-dimensional comprehensive feature vector also includes the serial number of the device under test and the current ambient temperature information.

3. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 1, characterized in that, In step S2: The historical health baseline data of normal equipment includes a baseline mean vector. Covariance Matrix ; The deviation between the high-dimensional integrated feature vector and the pre-established historical health baseline data of normal devices is calculated using Mahalanobis distance, the formula for which is: ,in, This is the currently acquired high-dimensional comprehensive feature vector; The preset health threshold is set based on the kernel density estimation method and is a distance critical value that satisfies the cumulative probability integral condition at a preset significance level.

4. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 3, characterized in that, Step S2 further includes a dynamic adaptive update mechanism for the health baseline: the system continuously collects normal test data that does not trigger the underlying hardware abnormal interruption, and periodically supplements it to the historical database by combining the current ambient temperature and service stage information of the device; the baseline mean vector is dynamically updated by refitting the multivariate Gaussian distribution model. and the covariance matrix In order to adaptively adjust the preset health threshold.

5. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 1, characterized in that, In step S3: The rule-based expert system transforms the manufacturer's fault code table into production rules for feature matching and reasoning. The machine learning classification model is a random forest classifier trained using the SMOTE oversampling method to process the imbalanced historical fault dataset. The random forest classifier is configured to contain multiple independent decision trees, and its hyperparameters are set according to the following rules: the number of decision trees is set between 100 and 500, and the maximum number of features selected when training a single decision tree is [not specified]. ,in The feature dimension of the high-dimensional comprehensive feature vector; The decision fusion mechanism extracts the voting results and probabilities of fault codes output by multiple decision trees in the random forest classifier, and performs weighted fusion with the rule reasoning results generated by the expert system based on the manufacturer's fault code table. When the two results conflict, the final diagnostic conclusion is output according to the preset arbitration rules.

6. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 1, characterized in that, In step S4: The state space of the global network nodes obtained by the intelligent scheduling agent is constructed as follows: global state vector ,in The total number of network nodes; for any network node, its corresponding 6 components are composed of the following physical quantities in a fixed order: the number of remaining available RF channels, the percentage of remaining power amplifier lifespan, the current battery level, the queue length of pending tasks, the highest task priority, and an online state binary variable characterizing whether the node has failed.

7. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 6, characterized in that, In step S4: The intelligent scheduling agent employs either the Proximal Policy Optimization (PPO) algorithm or the Deep Q-Network (DQN) algorithm, which utilizes a built-in comprehensive reward function. Output the optimal replacement target node; The comprehensive reward function The calculation formula is: ; in, The positive reward value for successfully migrating the task to the target node and completing the execution; The negative penalty value for task interruption due to migration failure; The formula for the recovery time penalty value is as follows: , The penalty coefficient is... The physical time from node failure to task resumption; The resource equilibrium reward value is calculated using the following formula: ;in, These are the weighting coefficients. The number of remaining available radio frequency channels for the target node. This represents the maximum number of radio frequency channels for a node. This represents the percentage of remaining lifespan of the target node's power amplifier. The target node's current battery level.

8. The intelligent diagnostic method for shipboard electronic equipment based on multi-source data fusion according to claim 1, characterized in that, In step S4: The process of packaging incomplete test tasks into standardized service packages refers to packaging the signal processing algorithm IP cores, test sequence scripts, and data analysis plugins required for the test task. The process of driving the optimal replacement target node to allocate underlying physical resources for hardware reconstruction includes: after receiving the service packet, the optimal replacement target node dynamically switches the on / off state of the fully programmable radio frequency switching matrix through the internal control bus to establish a physical test path, and dynamically loads the signal processing algorithm IP core into the reconfigurable partition of the local field-programmable gate array FPGA to reconstruct its own hardware topology into a dedicated test environment required to execute the test task.

9. A shipboard electronic equipment intelligent diagnostic system based on multi-source data fusion, characterized in that, include: The multi-source data acquisition and feature stitching module is used to acquire external radio frequency response data generated when the device under test is tested, the internal test BIT data of the device under test acquired through the data bus, and the historical maintenance records of the device under test. The module aligns the above heterogeneous data according to the absolute timestamp and stitches them together to generate a high-dimensional comprehensive feature vector. The health status assessment module is used to calculate the degree of deviation between the high-dimensional comprehensive feature vector and the pre-established historical health baseline data of normal equipment. When the degree of deviation exceeds the preset health threshold, the tested equipment is determined to be in an abnormal state, triggering an underlying hardware abnormal interruption to suspend the current test process and enter the fault diagnosis and networked scheduling process. The precise fault diagnosis module is used to input the high-dimensional comprehensive feature vector that determines the anomaly into a rule-based expert system and a pre-trained machine learning classification model in parallel, and output the LRU code of the faulty replaceable unit and its confidence level through a decision fusion mechanism to generate predictive maintenance suggestions. The networked elastic scheduling module is used to detect when the underlying hardware abnormality or communication failure of the current test support node is detected. It obtains the state space of the global network nodes through the intelligent scheduling agent running on the main control module based on the reinforcement learning algorithm, outputs the optimal replacement target node, and encapsulates the unfinished test tasks into standardized service packages and sends them to the optimal replacement target node. This drives the optimal replacement target node to allocate underlying physical resources for hardware reconstruction and continue to execute the test tasks.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

Images