Power station equipment health management system based on multi-modal perception and hybrid reasoning

Abstract

The invention is suitable for the field of power station equipment management, and provides a power station equipment health management system based on multi-modal perception and hybrid reasoning, and the system comprises a data perception layer which is used for obtaining and preprocessing multi-modal perception data of power station target equipment, and obtaining multi-modal feature data; the hybrid reasoning layer is used for taking the multi-modal characteristic data as input of a pre-constructed hybrid reasoning model to obtain an equipment fault diagnosis result; the decision management layer is used for generating an equipment health management decision according to the diagnosis result and outputting the decision; the hybrid reasoning model comprises a rule reasoning sub-layer, a deep learning sub-layer and a case reasoning sub-layer, the deep learning sub-layer adopts a lightweight Transform model, and the lightweight Transform model comprises a cross-modal attention enhancement module and a fault feature intensified training module. In conclusion, the hybrid reasoning architecture is constructed, so that the fault diagnosis accuracy of the power station equipment is obviously improved, the false alarm rate is reduced, and the technical problems of few unconventional fault samples and difficulty in diagnosis of the power station are solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and relates to the field of power plant equipment management technology. Specifically, it provides a power plant equipment health management system that combines multimodal perception and hybrid reasoning. Background Technology

[0002] As a supporting technology in the energy production field, the power plant equipment health management system plays a crucial role in real-time monitoring of equipment operating status, predicting potential failure risks, and optimizing operation and maintenance strategies. Against the backdrop of the accelerated transformation of the energy structure towards cleaner and more intelligent systems, the power plant equipment health management system not only concerns the safe and stable operation of the power plant but also directly impacts energy production efficiency and economic benefits.

[0003] Currently, the main shortcomings of power plant equipment health management systems in related technologies are as follows: Limited perception dimension: In actual operation scenarios, traditional power plant equipment health management systems mostly deploy only conventional physical quantity monitoring equipment such as vibration sensors and temperature sensors. A single monitoring dimension cannot effectively capture key indicators, leading to a high rate of missed early fault warnings. Isolated inference mechanism: Rule-based inference relies on pre-set experience rules by engineers. However, as the operating years of power plant equipment increase, wear and aging of mechanical parts can cause changes in fault modes, making it difficult for the originally set rules to adapt to new fault scenarios. Moreover, in niche fault scenarios, due to the scarcity of historical data samples and insufficient model training, the accuracy of fault diagnosis is low. Poor coordination between perception and inference: The coordination between the perception module and the inference module in current power plant management systems is poor. Data collected by the sensing devices is directly input into the inference module without processing, resulting in a large amount of data noise interfering with the accuracy of the inference results. Furthermore, the perception system cannot dynamically adjust the collection parameters according to the diagnostic needs of the inference module, making the system unable to meet the requirements of real-time monitoring. Summary of the Invention

[0004] The purpose of this invention is to provide a power plant equipment health management system with multimodal perception and hybrid reasoning, so as to at least solve the technical problems of existing power plant equipment health management systems, such as single perception dimension, isolated reasoning mechanism and poor coordination between perception and reasoning.

[0005] To achieve the above objectives, the embodiments of this application provide the following technical solutions.

[0006] According to one embodiment of this application, a power plant equipment health management system based on multimodal perception and hybrid reasoning is provided, comprising: The data perception layer is used to acquire and preprocess multimodal perception data of the target equipment in the power station to obtain multimodal feature data. A hybrid inference layer is used to input multimodal feature data as input to a pre-built hybrid inference model to obtain equipment fault diagnosis results. The hybrid inference model includes a rule-based inference sublayer, a deep learning sublayer, and a case-based inference sublayer. The rule-based inference sublayer filters and retains abnormal operating condition data. The deep learning sublayer uses a lightweight Transformer model, inputting multimodal feature data and equipment auxiliary data, capturing cross-modal correlation features through the lightweight Transformer model, and outputting preliminary fault diagnosis results. The lightweight Transformer model includes a cross-modal attention enhancement module and a fault feature reinforcement training module. The cross-modal attention enhancement module sets physically constrained learnable modal weight factors for each multimodal feature, calculates the contribution of different modal features to fault diagnosis, and dynamically allocates attention weights. The fault feature reinforcement training module perturbs and generates multimodal features of unconventional faults to expand the training sample set. The case-based inference sublayer is used to establish a fault case library, optimizing the retrieval of similar historical cases through dynamically allocated weight factors, verifying and correcting the deep learning diagnosis results. The decision management layer is used to generate and output equipment health management decisions based on diagnostic results.

[0007] Furthermore, in the hybrid inference layer, the rule-based inference sublayer is used to filter and retain abnormal operating condition data, including the following steps: A dynamic rule base for device structure association is constructed, which includes basic rules, component association rules, and a modal sensitivity coefficient library; The multimodal feature data output by the data perception layer is subjected to rule matching verification, including: filtering preliminary abnormal data based on the real-time normal threshold and modal sensitivity coefficient of the basic rules, comparing the component association features corresponding to the preliminary abnormal data with the collaborative constraint rules of the component association rules, and retaining abnormal data with single features exceeding the threshold and abnormal collaborative association features. A rule and feature feedback verification mechanism is established with the deep learning sub-layer. Abnormal working condition candidate data is synchronized to the cross-modal attention enhancement module of the deep learning sub-layer. The modal weight distribution output by the cross-modal attention enhancement module is obtained. Based on the sensitivity coefficient of the corresponding fault type in the modal sensitivity coefficient library, the matching degree between the modal weight distribution and the sensitivity coefficient is calculated to determine the abnormal working condition data that needs to be retained.

[0008] Furthermore, the modal sensitivity coefficient library is constructed based on historical fault data of power plant equipment, equipment fault mechanism manuals, and physical correlation characteristics of components, including: Establish the correlation matrix between fault types and modal characteristics; calculate the initial sensitivity coefficient; wherein, based on the analytic hierarchy process (AHP), statistical weights are assigned to historical fault data, and fault tree analysis is introduced for physical mechanism correction to obtain the corrected sensitivity coefficient, expressed as: In the formula, represents the correction sensitivity coefficient for the j-th mode under the i-th type of fault; This represents the AHP statistical weight for the j-th mode under the i-th type of fault; The fault tree physical correction factor represents the fault tree physical correction factor for the j-th mode under the i-th type of fault. This represents the physical correlation coefficient between the component belonging to the j-th mode and the core component under the i-th type of fault; It receives decision data from the decision management team in real time, traces the source of misjudgments, and builds a modality sensitivity coefficient library using the corrected modality sensitivity coefficient.

[0009] Furthermore, the steps for capturing cross-modal correlation features in the deep learning sub-layers include: The cross-modal attention enhancement module receives multimodal feature data and equipment auxiliary data output from the data perception layer, and sets physical constraints and learnable modal weight factors for each modal feature. Among them, the component spatial correlation matrix is ​​generated based on the three-dimensional structural model of the equipment as a prior constraint of the weight factors. Combined with the sensitivity priority of the corresponding fault type in the modal sensitivity coefficient library, the optimization range of the weight factors is limited. The statistical correlation of multimodal features is calculated by using a self-attention mechanism, and the weight factors are dynamically updated by combining physical constraints to output cross-modal fusion features.

[0010] Furthermore, in the deep learning sub-layer, the steps for outputting preliminary fault diagnosis results include: In the dynamic inference and output of the lightweight Transformer, the network structure of the lightweight Transformer is dynamically adjusted based on the real-time load level and fault risk prediction value in the equipment auxiliary data. The output results include cross-modal correlation feature matrix, preliminary fault diagnosis results and weight distribution of each modality, and will be synchronized to the case inference sub-layer for similar case retrieval. It also receives matching degree data fed back from the case inference sub-layer for iterative optimization of the physical constraint coefficient of the cross-modal attention enhancement module.

[0011] Furthermore, in the fault feature enhancement training module, the fault feature enhancement training module is based on cross-modal fusion features and generates enhancement feature learning through virtual samples guided by a mechanism model, including: Embedded device fault physical model, defining the physical boundaries of characteristic changes; For unconventional faults with scarce samples, a fault degree-feature change curve is generated based on the mechanism model, and intermediate state virtual samples are generated by interpolation based on real samples. A physical consistency loss function is constructed to incorporate the deviation between virtual samples and the predicted values ​​of the mechanism model into the training loss, thereby forcing the model to learn fault characteristics that conform to physical laws.

[0012] Furthermore, the step of generating a fault degree-feature change curve based on the mechanism model and interpolating to generate intermediate state virtual samples based on real samples includes: From the cross-modal fusion features, multimodal coupling relationships strongly correlated with fault types are selected as generation anchors. The correlation strength of the coupling anchors is verified by the Pearson correlation coefficient, and the corresponding coupling relationship is selected as the generation benchmark based on the correlation coefficient. Multiple real samples of the same fault type are obtained, and a two-dimensional interpolation curve is constructed with the fault degree as the horizontal axis and the feature values ​​of each modal coupling anchor as the vertical axis. The interpolation calculation must meet the coupling constraint condition: when generating intermediate state virtual samples, the changes in the feature values ​​of each mode must synchronously follow the correlation law of the coupling anchors. Along the fault severity-feature change curve, intermediate virtual samples are generated between real samples according to gradient. Each virtual sample contains multimodal features, and the modal features of each virtual sample must meet the following requirements: the value of a single modal feature conforms to the physical model of equipment fault; and the multimodal features conform to a defined coupling anchor relationship. The resulting virtual sample set is mixed proportionally with the real sample set for reinforcement training of the lightweight Transformer model.

[0013] Furthermore, the fault feature enhancement training module also includes a sample verification and screening unit, which is used to implement the following steps: The training of the lightweight Transformer model is divided into an initialization phase, an iterative optimization phase, and a convergence and stabilization phase. The virtual sample requirements for each phase are determined by combining the modal sensitivity coefficient library. Specifically, for the initialization phase, only virtual samples at both ends of the fault severity-feature change curve are selected. In the iterative optimization phase, virtual samples in the middle segment of the fault severity are introduced. In the convergence and stabilization phase, edge virtual samples that closely resemble the real fault scenario are selected. The quality of virtual samples is verified in multiple dimensions, including secondary verification of physical consistency, modality-sensitive matching verification, and training adaptability verification; the mixed sample set is dynamically updated based on the verification results.

[0014] Furthermore, in the decision-making and management level, the analytic hierarchy process (AHP) is used to obtain the equipment health status, and combined with the health status and equipment operating load characteristics, a predictive model is used to predict the remaining service life of key components.

[0015] Furthermore, the prediction model includes: The working condition classification feature encoding unit is used to receive the equipment health status sequence and equipment operating load characteristics output by the decision management layer. It is divided into three-level working condition subspaces according to the load fluctuation characteristics. For stable low load working conditions, a single-layer CNN is used to extract the static health status features of the load steady segment, focusing on capturing the slow deterioration trend. For fluctuating load working conditions, a bidirectional LSTM is used to encode the temporal correlation between load and health status, and the coupling relationship between the load fluctuation peak and the sudden drop in health status is strengthened through the attention mechanism. For shock high load working conditions, the residual network submodule is activated and two skip connection layers are added to capture nonlinear deterioration patterns. The dynamic weighting unit is used to calculate the membership degree of the current working condition in real time and dynamically allocate the output weights of the three-level working condition subspace based on the membership degree. Among them, the weight coefficient of the high load period is corrected to 1.2 times; the low load weight of newly commissioned equipment is increased to 0.8; and the multi-working condition collaborative feature vector is obtained by weighted fusion to solve the prediction jump problem of the single working condition model when the load is switched. An embedded physical constraint layer is used to constrain the effectiveness of features through physical rules before inputting the collaborative feature vector into the Transformer-based time-series prediction model. These constraints include hard and soft constraints. Under hard constraints, the predicted remaining lifetime must not exceed the equipment's design lifetime and must not be less than the minimum remaining lifetime calculated based on Miner's fatigue rule. Under soft constraints, when health status indicators exceed 80% of the warning threshold, the prediction results are forced to undergo exponential adjustments, conforming to the physical acceleration characteristics of fault development. Among these, the minimum remaining lifetime of key components... Represented as: In the formula, the cumulative fatigue damage of the key components Represented as; Indicates the design life of key components; This represents the actual runtime of the component during the t-th time period; This represents the load factor in the t-th time period; T represents the total number of time periods for cumulative statistics. The results prediction layer is used to output the remaining service life of key components.

[0016] Compared with existing technologies, the beneficial effects of the power plant equipment health management system based on multimodal perception and hybrid reasoning in this application are as follows: In the hybrid reasoning architecture constructed by this invention, redundant data is coarsely screened through the rule-based reasoning sublayer, complex multimodal correlations are captured using the deep learning sublayer, and black-box results are verified and corrected based on the case-based reasoning sublayer, forming a collaborative closed loop of data screening, accurate diagnosis, and experience verification. This not only solves the problem of insufficient adaptation of rule-based reasoning to complex faults, but also makes up for the sample dependence defects of deep learning and case-based reasoning, significantly improving the accuracy of power plant equipment fault diagnosis and reducing the false alarm rate. Furthermore, this invention, through a lightweight Transformer model adapted to edge computing power, introduces a cross-modal attention enhancement module and a fault feature enhancement training module into the model. While ensuring low latency inference at the edge, it fully explores the collaborative diagnostic value of multimodal data, avoids the information limitations of single-modal features, and solves the technical problem of few unconventional fault samples and difficulty in diagnosis in power plants. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] In the attached diagram: Figure 1 This is a structural block diagram of the power plant equipment health management system based on multimodal perception and hybrid reasoning of the present invention; Figure 2 The structural block diagram of the hybrid inference sublayer provided by the present invention; Figure 3 This is a flowchart illustrating an implementation of the rule-based reasoning sublayer in an embodiment of the present invention. Figure 4 This is a flowchart illustrating an implementation of the deep learning sublayer in an embodiment of the present invention. Figure 5 This is a flowchart of an implementation of the fault feature enhancement training module in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] According to the embodiments of this application, an embodiment of a power plant equipment health management system with multimodal perception and hybrid reasoning is provided. It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0022] Please refer to Figure 1 According to one embodiment of this application, a power plant equipment health management system with multimodal perception and hybrid reasoning is provided. The power plant equipment health management system includes a data perception layer 101, a hybrid reasoning layer 102, and a decision management layer 103. The data perception layer 101 is used to acquire and preprocess the multimodal perception data of the target equipment in the power station to obtain multimodal feature data; The multimodal sensing data provided in this application embodiment includes: Vibration data of power plant equipment are collected using fiber Bragg grating-based vibration sensors; A microfluidic chip-type oil sensor is installed in the lubrication circuit to detect dielectric loss, metal abrasive concentration, and moisture content in real time. An ultra-high frequency partial discharge sensor is installed in the cabinet of the high-voltage equipment in the power station to simultaneously collect ultra-high frequency signals and pulse currents. MEMS acoustic fingerprint sensors are installed on the surfaces of equipment such as fans and pumps to capture abnormal acoustic fingerprint features; Furthermore, in this embodiment, the sampling frequency, data transmission priority, and sleep strategy of each sensor can be dynamically adjusted based on the fault diagnosis results output by the hybrid inference layer 102. For example, when the hybrid inference layer 102 determines that a bearing has a risk of early wear, it automatically increases the sampling rate of the bearing vibration sensor from the conventional 1kHz to 5kHz, while reducing the sampling frequency of non-critical equipment sensors, thereby reducing system energy consumption and data transmission pressure while ensuring fault monitoring accuracy.

[0023] Furthermore, in the preprocessing of multimodal sensing data, a spatiotemporal calibration algorithm based on device runtime sequence marking is used for calibration and alignment to address the differences in data acquisition time and spatial location between different sensors, thus solving the fusion error problem caused by data spatiotemporal asynchrony in traditional systems. The aligned data is processed using hardware filtering, adaptive threshold denoising, and cross-modal feature correlation enhancement strategies. In hardware filtering, high-frequency electromagnetic interference is suppressed through a hardware-level RC filter circuit. In adaptive threshold denoising, a health feature library is constructed based on multimodal data under normal device operation, and an adaptive threshold algorithm is used to remove abnormal noise data that deviates from the health baseline. In cross-modal feature correlation enhancement, the correlation between cross-modal data is used for feature enhancement. For example, the oil metal abrasive particle concentration data is correlated with the high-frequency impact characteristics of vibration signals to enhance the characteristic signals of early wear faults.

[0024] Furthermore, the hybrid inference layer 102 in this embodiment is used to take multimodal feature data as input to a pre-built hybrid inference model to obtain equipment fault diagnosis results; wherein, the multimodal feature data needs to be processed by the hybrid inference layer 102 to be transformed from structured data into useful information that can be used for fault diagnosis; at the same time, it provides data input to the decision management layer.

[0025] Specifically, such as Figure 2 As shown, in one embodiment of this application, the hybrid reasoning model includes a rule-based reasoning sublayer 201, a deep learning sublayer 202, and a case-based reasoning sublayer 203; In this embodiment, the rule reasoning sublayer 201 is used to filter and retain abnormal operating condition data; Please refer to Figure 2 In one implementation of this application, the step of filtering and retaining abnormal operating condition data in the rule-based inference sublayer within the hybrid inference layer includes: Step S301: Construct a dynamic rule base for device structure association, wherein the dynamic rule base includes basic rules, component association rules, and a modal sensitivity coefficient library; The basic rule is to target the multimodal characteristics of a single component of power plant equipment (such as bearings, rotors, windings, etc.), and further screen out the abnormal characteristics of the single component from the multimodal characteristic data. When setting differentiated dynamic thresholds for different modal characteristics of a single component, the threshold fluctuation tolerance range is smaller for the mode with higher sensitivity coefficient. The component association rule is a multimodal feature collaborative anomaly judgment rule constructed based on the physical transmission / coupling relationship of the three-dimensional structure of power plant equipment. It is used to eliminate false judgment data where a single feature of a single component exceeds the threshold but there is no collaborative anomaly across components, thus solving the false alarm problem caused by the traditional rule's isolated judgment of a single component and neglect of physical association. The modal sensitivity coefficient library is a weighted library of fault types and modal characteristics built based on historical fault data of power plant equipment, equipment fault mechanism manuals and component physical correlation characteristics. It provides a unified benchmark for the threshold strictness of basic rules and the collaborative verification weight of component correlation rules. At the same time, it ensures the long-term adaptability of the rule library through closed-loop iterative optimization, solving the problems of traditional rules having no objective weight basis and being unable to be dynamically updated. Specifically, in one implementation of this application, the modal sensitivity coefficient library is constructed based on historical fault data of power plant equipment, equipment fault mechanism manuals, and physical correlation characteristics of components. Specifically, the steps in constructing the modal sensitivity coefficient library include: Establish the correlation matrix between fault types and modal characteristics; calculate the initial sensitivity coefficient; wherein, based on the analytic hierarchy process (AHP), statistical weights are assigned to historical fault data, and fault tree analysis is introduced for physical mechanism correction to obtain the corrected sensitivity coefficient, expressed as: In the formula, represents the correction sensitivity coefficient for the j-th mode under the i-th type of fault; This represents the AHP statistical weight for the j-th mode under the i-th type of fault; The fault tree physical correction factor represents the fault tree physical correction factor for the j-th mode under the i-th type of fault. This represents the physical correlation coefficient between the component belonging to the j-th mode and the core component under the i-th type of fault; It receives decision data from the decision management team in real time, traces the source of misjudgments, and builds a modality sensitivity coefficient library using the corrected modality sensitivity coefficient.

[0026] Please continue to refer to Figure 2 In the hybrid inference layer, the rule-based inference sublayer, used to filter and retain abnormal operating condition data, also includes the following steps: Step S302: Perform rule matching verification on the multimodal feature data output by the data perception layer; In this embodiment, step S302 specifically includes: filtering preliminary abnormal data based on the real-time normal threshold and modal sensitivity coefficient of the basic rules, comparing the component association features corresponding to the preliminary abnormal data with the collaborative constraint rules of the component association rules, and retaining abnormal data where a single feature exceeds the threshold and the association features are collaboratively abnormal. In step S302 of this embodiment, from the multimodal feature data output by the data perception layer, single-modal noise false alarms and isolated anomalies are removed, and real anomaly data with component collaboration characteristics are retained, including: Preliminary abnormal data is screened based on basic rules and modal sensitivity coefficients. Specifically, the real-time normal thresholds of basic rules in the dynamic rule base and the priority of modal sensitivity coefficients are used to distinguish between true anomalies and modal noise, avoiding misjudgments caused by traditional fixed thresholds. First, it is determined whether each modal feature exceeds the real-time normal threshold, and then the modal sensitivity coefficient is used to determine whether to include it in the preliminary abnormal data. Preliminary abnormal data includes high-sensitivity modal features exceeding the real-time threshold and medium-sensitivity modal features exceeding the threshold, while high-sensitivity modal features are close to the threshold. Data that only exceeds the threshold for low-sensitivity modes is directly removed, thus obtaining preliminary abnormal data for a single component. For the obtained preliminary abnormal data, the collaborative constraint rules based on component association rules are compared to determine whether the preliminary abnormal data is accompanied by collaborative anomalies of related components, thus removing non-real fault data such as isolated fluctuations of a single component. Step S303: Establish a rule and feature feedback verification mechanism with the deep learning sub-layer, synchronize the abnormal working condition candidate data to the cross-modal attention enhancement module of the deep learning sub-layer, obtain the modal weight distribution output by the cross-modal attention enhancement module, and calculate the matching degree between the modal weight distribution and the sensitivity coefficient based on the sensitivity coefficient of the fault type in the modal sensitivity coefficient library to determine the abnormal working condition data that needs to be retained.

[0027] Step S303 of this embodiment combines cross-modal feature analysis with the deep learning sub-layer, uses data-driven modal weights to verify the rule-driven sensitivity coefficient, and further eliminates rule-misjudged data; abnormal operating condition candidate data includes the multimodal feature matrix of candidate data and the corresponding component fault type, and synchronizes the abnormal operating condition candidate data to the cross-modal attention enhancement module of the deep learning sub-layer. In this embodiment, the modal weight distribution is obtained by dynamically allocating attention weights through the cross-modal attention enhancement module. Specifically, the cross-modal attention enhancement module dynamically allocates attention weights based on the multimodal features of the candidate data and the device's three-dimensional structural correlation matrix, and outputs the modal weight distribution. In the step of calculating the matching degree, the matching degree is calculated by using the sensitivity coefficient of the corresponding fault type in the modal sensitivity coefficient library as the benchmark, and by weighted cosine similarity or weight ratio deviation. Finally, the abnormal operating condition data to be retained is determined based on the matching degree. If the matching degree is ≥80%, the candidate data is retained as the final abnormal operating condition data; if the matching degree is <80%, it is determined to be data misjudged by the rule and the data is removed.

[0028] Please continue to refer to Figure 2 In this embodiment, the deep learning sub-layer 202 of the hybrid inference model adopts a lightweight Transformer model. It takes multimodal feature data and device auxiliary data as input, captures cross-modal correlation features through the lightweight Transformer model, and outputs preliminary fault diagnosis results. The input data of the lightweight Transformer model is a structured embedding vector after co-processing of multimodal feature data and equipment auxiliary data. It is used to solve the problems of difficulty in fusing multi-source heterogeneous data and feature mismatch caused by operating condition fluctuations.

[0029] Specifically, in the embodiments of this application, the lightweight Transformer model includes a cross-modal attention enhancement module and a fault feature enhancement training module; the cross-modal attention enhancement module sets physical constraint-based learnable modal weight factors for each of the multimodal features, calculates the contribution of different modal features to fault diagnosis, and dynamically allocates attention weights; the fault feature enhancement training module perturbs and generates multimodal features of unconventional faults to expand the training sample set; In this embodiment, the learnable modal weight factor of physical constraints is determined in the cross-modal attention enhancement module based on the physical correlation matrix of the equipment three-dimensional structure, combined with the modal sensitivity coefficient calculated by the fusion of the analytic hierarchy process (AHP) and the fault tree analysis (FTA) to limit the optimization range, and then the statistical correlation of multimodal features is calculated through the self-attention mechanism. The initial weight of physical constraints and the statistical correlation weight are obtained by weighting and fusing them with the gate coefficient. This invention utilizes a lightweight Transformer model adapted to edge computing power, incorporating a cross-modal attention enhancement module and a fault feature reinforcement training module. While ensuring low latency inference at the edge, it fully leverages the collaborative diagnostic value of multimodal data, avoids the information limitations of single-modal features, and solves the technical problem of limited and difficult-to-diagnose unconventional fault samples in power plants.

[0030] Please continue to refer to Figure 2 In this embodiment, the case reasoning sublayer 303 of the hybrid reasoning model is used to establish a fault case library, optimize the retrieval of similar historical cases through dynamically allocated weight factors, and verify and correct the deep learning diagnostic results. The power plant equipment fault case library established in this embodiment includes information such as multimodal features of historical faults, handling solutions, and fault consequences. After the deep learning sublayer outputs the fault diagnosis results, the most similar historical cases are retrieved through the case similarity matching algorithm to verify and correct the deep learning diagnosis results. At the same time, it provides reference solutions for fault handling, which greatly improves the credibility of the diagnosis results.

[0031] Please continue to refer to Figure 1 The power plant equipment health management system provided in this embodiment also includes: The decision management layer 103 is used to generate and output equipment health management decisions based on the diagnostic results. Specifically, in this embodiment, a multi-dimensional health status assessment index system is constructed based on the diagnostic results of the hybrid inference layer. The assessment index system includes fault severity, fault development trend, and equipment importance weight. Then, the analytic hierarchy process (AHP) is used to calculate the overall health index of the equipment and display it in the form of a visual chart, so that maintenance personnel can intuitively grasp the health status of the equipment.

[0032] In summary, the hybrid reasoning architecture constructed in this invention uses a rule-based reasoning sublayer to coarsely screen redundant data, a deep learning sublayer to capture complex multimodal relationships, and a case-based reasoning sublayer to verify and correct black-box results, forming a collaborative closed loop of data screening, accurate diagnosis, and experience verification. This not only solves the problem of rule-based reasoning's inability to adapt to complex faults, but also makes up for the sample dependence defects of deep learning and case-based reasoning, significantly improving the accuracy of power plant equipment fault diagnosis and reducing the false alarm rate.

[0033] Please refer to Figure 4 In one implementation of this application, the step of capturing cross-modal correlation features in the deep learning sublayer 202 includes: Step S401: The cross-modal attention enhancement module receives multimodal feature data and device-aided data output from the data perception layer, and sets physical constraint learnable modal weight factors for each modal feature; In step S401 of this embodiment, the equipment auxiliary data includes real-time load level, running time, equipment model, etc. When configuring physical constraint learnable modal weight factors for each type of modal feature, the initial value and optimization range of the weight factors are determined based on the three-dimensional structural correlation matrix of the power plant equipment and the modal sensitivity coefficient library. The physical constraints are used as the prior constraints of the weight factors. Combined with the sensitivity priority of the corresponding fault type in the modal sensitivity coefficient library, the optimization range of the weight factors is limited. Step S402: Calculate the statistical correlation of multimodal features through a self-attention mechanism, dynamically update the weight factors in combination with physical constraints, and output cross-modal fusion features.

[0034] In this embodiment of the application, the step of outputting preliminary fault diagnosis results in the deep learning sub-layer includes: in the dynamic inference and result output of the lightweight Transformer, dynamically adjusting the network structure of the lightweight Transformer based on the real-time load level and fault risk prediction value in the equipment auxiliary data; Furthermore, in this embodiment, the output results include a cross-modal association feature matrix, preliminary fault diagnosis results, and weight distribution of each modality. These results will be synchronized to the case reasoning sub-layer for similar case retrieval, and the matching degree data fed back from the case reasoning sub-layer will be received to iteratively optimize the physical constraint coefficients of the cross-modal attention enhancement module.

[0035] Please refer to Figure 5 In this embodiment, the fault feature enhancement training module is based on cross-modal fusion features and generates enhancement feature learning through virtual samples guided by a mechanism model, including: Step S501: Embed the physical model of device faults and define the physical boundaries of feature changes; In step S501, the physical model of the embedded equipment fault is first determined. For example, for bearing wear faults, Miner's fatigue accumulation law is embedded; for transformer insulation aging faults, the oil gas dissolution balance equation is embedded; for turbine blade crack faults, the fracture mechanics crack propagation model is embedded. These models belong to the existing power plant equipment fault mechanism manual or industry standard. Further, based on the embedded physical model, the reasonable range of values ​​and variation law of each modal characteristic parameter are quantitatively defined to realize the definition of the physical boundary of characteristic change. Step S502: For unconventional faults with scarce samples, generate a fault degree-feature change curve based on the mechanism model, and interpolate to generate intermediate state virtual samples based on real samples. In step S502 of this embodiment, firstly, multimodal coupling anchor points strongly correlated with unconventional faults are screened from the cross-modal fusion features. Then, a two-dimensional curve is generated by fitting real sample data with fault severity as the horizontal axis and the modal feature values ​​of each coupling anchor point as the vertical axis. The curve type must conform to the laws of the physical model. Within the fault severity range corresponding to the real sample, interpolation nodes are evenly divided according to 5-10 gradients. Based on the constructed fault severity-feature change curve, the multimodal feature values ​​corresponding to each node are calculated. The calculated multimodal feature values ​​must simultaneously satisfy coupling constraints, including that the multimodal features of the generated intermediate samples must synchronously follow the coupling anchor point relationship. The finally generated virtual samples can be used for model training. Specifically, in this embodiment, the steps of generating a fault severity-feature change curve based on the mechanism model and interpolating to generate intermediate virtual samples based on real samples include: selecting multimodal coupling relationships strongly correlated with fault types from cross-modal fusion features as generation anchors, wherein the correlation strength of the coupling anchors is verified by the Pearson correlation coefficient, and the corresponding coupling relationship is selected as the generation benchmark based on the correlation coefficient; obtaining multiple real samples of the same fault type, constructing a two-dimensional interpolation curve with fault severity as the horizontal axis and the feature values ​​of each modal coupling anchor as the vertical axis; wherein the interpolation calculation must meet the coupling constraint condition: when generating intermediate virtual samples, the changes in each modal feature value must synchronously follow the correlation law of the coupling anchor; generating intermediate virtual samples in gradients between real samples along the fault severity-feature change curve, wherein each virtual sample contains multimodal features, and the modal features of each virtual sample must satisfy: the single modal feature value conforms to the physical model of equipment fault; the multimodal features conform to the determined coupling anchor relationship; and finally, the generated virtual sample set is mixed with the real sample set in proportion for reinforcement training of the lightweight Transformer model.

[0036] Step S503: Construct a physical consistency loss function, incorporate the deviation between virtual samples and the predicted values ​​of the mechanism model into the training loss, and force the model to learn fault characteristics that conform to physical laws; In step S503 of this embodiment, the physical consistency loss function is constructed. , is represented as: In the formula, N represents the number of virtual samples in the training batch; M represents the multimodal feature data of the virtual samples; This represents the actual feature value of the j-th modality in the i-th virtual sample; This represents the theoretical prediction value for the j-th mode of the i-th virtual sample based on the embedded device fault physical model; To ensure the accuracy of fault classification, this embodiment introduces a classification cross-entropy loss in addition to the physical consistency loss function; the total loss is expressed as: In the formula, This represents the classification cross-entropy loss, which measures the accuracy of the model in diagnosing fault types based on virtual samples, ensuring that the model can accurately identify fault categories. The physical consistency loss measures the deviation between the modal feature values ​​of virtual samples and the predicted values ​​of the embedded physical model. This deviation is calculated using the mean squared error; the larger the deviation, the greater the physical consistency loss. The larger; This represents the weighting coefficient, used to balance the importance of the two types of losses; During model training in this embodiment, the backpropagation algorithm will simultaneously minimize and If the model learns features that violate the laws of physics, then The error will increase significantly. At this point, the model parameters are adjusted until the deviation between the learned features and the predicted values ​​of the physical model is controlled within the allowable range, ensuring that the fault characteristics output by the model conform to the statistical laws of data and do not violate the objective engineering laws of power plant equipment.

[0037] In this embodiment, the fault feature enhancement training module further includes a sample verification and screening unit, which is used to implement the following steps: The training of the lightweight Transformer model is divided into an initialization phase, an iterative optimization phase, and a convergence and stabilization phase. The virtual sample requirements for each phase are determined by combining the modality sensitivity coefficient library. This avoids the problem that the lightweight Transformer model is difficult to converge in the early stage due to the use of complex samples, and that the model learning stagnates in the later stage due to the use of simple samples. In this embodiment, during the initialization phase, only virtual samples at both ends of the fault severity-feature change curve are selected. In the initial phase, the parameters are randomly distributed during model initialization, and the model quickly establishes a basic understanding of fault characteristics through extreme state samples. During the iterative optimization phase, virtual samples in the middle segment of the fault severity are introduced. During the iterative optimization phase, the model learns the characteristics of gradual changes in fault severity through intermediate state samples, improving its ability to distinguish the fault development stage. During the convergence and stabilization phase, edge virtual samples that closely resemble the real fault scenario are selected. When the model is close to convergence, the generalization ability is enhanced through edge samples. The quality of virtual samples is verified in multiple dimensions, including secondary verification of physical consistency, modality-sensitive matching verification, and training fit verification; the mixed sample set is dynamically updated based on the verification results. In this embodiment, the physical consistency secondary verification is used to ensure that the samples do not deviate from the mechanistic model. The verification object is the deviation between the multimodal feature values ​​of the virtual samples and the predicted values ​​of the embedded physical model of device faults. The modal sensitivity matching verification is used to ensure that the samples conform to the fault sensitivity features. The verification object is the matching degree between the multimodal weight distribution of the virtual samples and the modal sensitivity coefficient library. The training fit verification is used to ensure that the samples help the model converge. The verification object is the effectiveness of the virtual samples in model training, avoiding the introduction of samples that cause loss fluctuations or slow convergence.

[0038] In this embodiment, after multi-dimensional verification, the sample verification and screening unit dynamically adjusts the mixed sample set to ensure that the sample set always matches the model training requirements.

[0039] Furthermore, in the decision-making and management level, the analytic hierarchy process is used to obtain the equipment health status, and combined with the health status and equipment operating load characteristics, a predictive model is used to predict the remaining service life of key components. In the step of obtaining equipment health status using the analytic hierarchy process (AHP), AHP is used to transform the multi-dimensional fault diagnosis results output by the hybrid inference layer into health status indicators. The input of AHP is the diagnosis results from the hybrid inference layer, including core fault information, modal contribution information, and case matching information, to ensure a more comprehensive health status assessment.

[0040] Specifically, in the step of predicting the remaining service life of key components using a prediction model in this embodiment, the prediction model includes: The operating condition classification feature encoding unit receives the equipment health status sequence and equipment operating load characteristics output by the decision management layer. It divides the operating condition subspace into three levels based on load fluctuation characteristics: stable low load, fluctuating medium load, and high-load shock. Specifically: for the stable low load condition, a single-layer CNN is used to extract the static health status features during periods of stable load, focusing on capturing slow degradation trends; for the fluctuating medium load condition, a bidirectional LSTM is used to encode the temporal correlation between load and health status, strengthening the coupling relationship between load fluctuation peaks and sudden drops in health status through an attention mechanism; for the high-load shock condition, the residual network submodule is activated, adding two skip connection layers to capture nonlinear degradation patterns. The dynamic weighting unit is used to calculate the membership degree of the current working condition in real time and dynamically allocate the output weights of the three-level working condition subspace based on the membership degree. Among them, the weight coefficient of the high load period is corrected to 1.2 times; the low load weight of newly commissioned equipment is increased to 0.8; and the multi-working condition collaborative feature vector is obtained by weighted fusion to solve the prediction jump problem of the single working condition model when the load is switched. An embedded physical mechanism constraint layer is used to constrain the effectiveness of collaborative feature vectors through physical rules, including hard and soft constraints, before inputting them into the Transformer-based time series prediction model; where: Under hard constraints, the predicted remaining life must not exceed the equipment design life, and must not be less than the minimum remaining life of critical components calculated based on Miner's fatigue rule. In soft constraints, when the health status index exceeds 80% of the warning threshold, the prediction results are forced to be adjusted exponentially, which conforms to the physical acceleration characteristics of fault development. In this embodiment, the minimum remaining lifespan of the key component is... Represented as: In the formula, the cumulative fatigue damage of the key components Represented as; Indicates the design life of key components; This represents the actual runtime of the component during the t-th time period; This represents the load factor in the t-th time period; T represents the total number of time periods for cumulative statistics. The results prediction layer is used to output the remaining service life of key components.

[0041] In summary, this invention constructs a hybrid reasoning collaborative architecture that uses a rule-based reasoning sublayer to coarsely screen redundant data, a deep learning sublayer to capture complex multimodal correlations, and a case-based reasoning sublayer to verify and correct black-box results. This architecture not only solves the problem of rule-based reasoning's inability to adapt to complex faults but also compensates for the sample dependence of deep learning and case-based reasoning, significantly improving the accuracy of power plant equipment fault diagnosis and reducing the false alarm rate. Furthermore, by adapting a lightweight Transformer model to edge computing power, this invention fully explores the collaborative diagnostic value of multimodal data while ensuring low latency in edge reasoning and avoiding the information limitations of single-modal features, effectively solving the technical problem of few samples and difficulty in diagnosing unconventional faults in power plants.

[0042] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A power plant equipment health management system based on multimodal perception and hybrid reasoning, characterized in that, include: The data perception layer is used to acquire and preprocess multimodal perception data of the target equipment in the power station to obtain multimodal feature data. The hybrid inference layer takes multimodal feature data as input to a pre-built hybrid inference model to obtain equipment fault diagnosis results. The hybrid inference model includes a rule-based inference sublayer, a deep learning sublayer, and a case-based inference sublayer. The rule-based inference sublayer filters and retains abnormal operating condition data. The deep learning sublayer uses a lightweight Transformer model to capture cross-modal correlation features and output preliminary fault diagnosis results. The lightweight Transformer model includes a cross-modal attention enhancement module and a fault feature reinforcement training module. The cross-modal attention enhancement module sets physically constrained learnable modal weight factors for each multimodal feature, calculates the contribution of different modal features to fault diagnosis, and dynamically allocates attention weights. The fault feature reinforcement training module perturbs and generates multimodal features of unconventional faults to expand the training sample set. The case-based inference sublayer establishes a fault case library, optimizes the retrieval of similar historical cases through dynamically allocated weight factors, and verifies and corrects the deep learning diagnosis results. The decision management layer is used to generate and output equipment health management decisions based on diagnostic results.

2. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 1, characterized in that, In the hybrid inference layer, the rule-based inference sublayer is used to filter and retain abnormal operating condition data, including the following steps: A dynamic rule base for device structure association is constructed, which includes basic rules, component association rules, and a modal sensitivity coefficient library; Rule matching and verification are performed on the multimodal feature data output by the data perception layer to obtain abnormal data; Establish a rule and feature feedback verification mechanism with the deep learning sub-layer, synchronize abnormal data to the cross-modal attention enhancement module of the deep learning sub-layer, obtain the modal weight distribution output by the cross-modal attention enhancement module, and calculate the matching degree between the modal weight distribution and the sensitivity coefficient based on the sensitivity coefficient of the fault type in the modal sensitivity coefficient library to determine the abnormal working condition data that need to be retained.

3. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 2, characterized in that, The modal sensitivity coefficient library is constructed based on historical fault data of power plant equipment, equipment fault mechanism manual, and physical correlation characteristics of components. include: Establish the correlation matrix between fault types and modal characteristics; calculate the initial sensitivity coefficient; wherein, based on the analytic hierarchy process (AHP), statistical weights are assigned to historical fault data, and fault tree analysis is introduced for physical mechanism correction to obtain the corrected sensitivity coefficient, expressed as: In the formula, represents the correction sensitivity coefficient for the j-th mode under the i-th type of fault; This represents the AHP statistical weight for the j-th mode under the i-th type of fault; The fault tree physical correction factor represents the fault tree physical correction factor for the j-th mode under the i-th type of fault. This represents the physical correlation coefficient between the component belonging to the j-th mode and the core component under the i-th type of fault; It receives decision data from the decision management team in real time, traces the source of misjudgments, and builds a modality sensitivity coefficient library using the corrected modality sensitivity coefficient.

4. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 3, characterized in that, In deep learning sublayers, the steps for capturing cross-modal correlation features include: The cross-modal attention enhancement module receives multimodal feature data and device-aided data output from the data perception layer, and sets physical constraints and learnable modal weight factors for each modal feature. The statistical correlation of multimodal features is calculated by using a self-attention mechanism, and the weight factors are dynamically updated by combining physical constraints to output cross-modal fusion features.

5. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 4, characterized in that, In the deep learning sublayer, the steps for outputting preliminary fault diagnosis results include: In the dynamic inference and output of the lightweight Transformer, the network structure of the lightweight Transformer is dynamically adjusted based on the real-time load level and fault risk prediction value in the equipment auxiliary data. The output results include cross-modal correlation feature matrix, preliminary fault diagnosis results and weight distribution of each modality, and will be synchronized to the case inference sub-layer for similar case retrieval. It also receives matching degree data fed back from the case inference sub-layer for iterative optimization of the physical constraint coefficient of the cross-modal attention enhancement module.

6. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 5, characterized in that, In the fault feature enhancement training module, the fault feature enhancement training module is based on cross-modal fusion features and generates enhancement feature learning through virtual samples guided by a mechanism model, including: Embedded device fault physical model, defining the physical boundaries of characteristic changes; For unconventional faults with scarce samples, a fault degree-feature change curve is generated based on the mechanism model, and intermediate state virtual samples are generated by interpolation based on real samples. A physical consistency loss function is constructed to incorporate the deviation between virtual samples and the predicted values ​​of the mechanism model into the training loss, thereby forcing the model to learn fault characteristics that conform to physical laws.

7. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 1, characterized in that, The steps of generating a fault degree-feature change curve based on the mechanism model and interpolating to generate intermediate state virtual samples based on real samples include: Along the fault severity-feature change curve, intermediate virtual samples are generated between real samples according to gradient. Each virtual sample contains multimodal features, and the modal features of each virtual sample must meet the following requirements: the value of a single modal feature conforms to the physical model of equipment fault; and the multimodal features conform to a defined coupling anchor relationship. The resulting virtual sample set is mixed proportionally with the real sample set for reinforcement training of the lightweight Transformer model.

8. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 7, characterized in that, The fault feature enhancement training module also includes a sample verification and screening unit, which is used to implement the following steps: The training of the lightweight Transformer model is divided into an initialization phase, an iterative optimization phase, and a convergence and stabilization phase. The virtual sample requirements for each phase are determined by combining the modality sensitivity coefficient library. During the initialization phase, only virtual samples at both ends of the fault severity-feature change curve are selected; During the iterative optimization phase, virtual samples representing intermediate fault levels are introduced. During the convergence and stabilization phase, edge virtual samples that closely resemble real-world fault scenarios are selected.

9. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 8, characterized in that, In the decision-making and management level, the analytic hierarchy process (AHP) is used to obtain the health status of the equipment, and combined with the health status and the characteristics of the equipment's operating load, a predictive model is used to predict the remaining service life of key components.

10. The power plant equipment health management system based on multimodal perception and hybrid reasoning according to claim 9, characterized in that, The prediction model includes: The working condition classification feature coding unit is used to receive the equipment health status sequence and equipment operating load characteristics output by the decision management layer, and divides them into three-level working condition subspaces according to load fluctuation characteristics. The dynamic weighting unit is used to calculate the membership degree of the current working condition in real time and dynamically allocate the output weights of the three-level working condition subspace based on the membership degree. An embedded physical mechanism constraint layer is used to constrain the effectiveness of features through physical rules before inputting the collaborative feature vector into the Transformer-based time series prediction model, including hard constraints and soft constraints. In the hard constraints, the predicted remaining life value must not exceed the equipment design life and must not be less than the minimum remaining life of key components calculated based on Miner's fatigue rule. In the soft constraints, the prediction results are adjusted exponentially. Among them, the minimum remaining life of key components Represented as: In the formula, the cumulative fatigue damage of the key components Represented as; Indicates the design life of key components; This represents the actual running time of the component during the t-th time period; This represents the load factor in the t-th time period; T represents the total number of time periods for cumulative statistics. The results prediction layer is used to output the remaining service life of key components.

Images