Equipment quality and supplier intelligent evaluation method based on space-time coupling Bayesian model and two-ticket graph database

Abstract

The invention provides an equipment quality and supplier intelligent evaluation method based on a space-time coupling Bayesian model and a two-ticket graph database, and belongs to the technical field of intelligent databases. The two-ticket graph database is firstly constructed, and a station environment, equipment full life cycle maintenance and supplier data are integrated; obtaining a supplier initial score by using an entropy weight method and a Sigmoid function; constructing a space-time coupling Bayesian model by means of ST-LSTM and VAE, and clustering equipment health states; introducing an environment coefficient and a KL divergence gradient to correct a score, and separating environment responsibilities from supplier responsibilities; calculating an equipment target score by combining an aging model and cosine similarity; and aggregating scores through a graph attention network, considering quality fluctuation, and generating a supplier global score. According to the invention, dynamic, precise and traceable evaluation is realized, a basis is provided for power grid supply chain management and the like, and the method is suitable for supplier evaluation in electric power and other industrial fields.

Description

Technical Field

[0001] This invention relates to the field of intelligent database technology, specifically to an intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-vote graph database. Background Technology

[0002] In the field of power equipment supply chain management, accurate evaluation of equipment quality and supplier performance is a core element in ensuring the safe and stable operation of the power grid and optimizing procurement decisions. With the rapid development of new power systems, equipment types are becoming increasingly diverse (such as new photovoltaic inverters and energy storage devices), and operating environments are becoming more complex and variable (such as coastal salt spray corrosion and extreme weather in wind farms), placing higher demands on the dynamism, comprehensiveness, and accuracy of supplier evaluation. Currently, supplier evaluation techniques based on Bayesian models have emerged in the industry. Typical examples include the supplier performance evaluation method CN114240016A based on a weighted Naive Bayes model for the operation and maintenance process, and the supplier performance score determination method CN111292000A based on a Bayesian averaging algorithm. These methods have made some progress in quantitative evaluation and reducing subjective intervention, but they still reveal many deep-seated defects in practical applications. The core problems lie in the limitations of the evaluation data support system and the insufficient adaptability of the evaluation model.

[0003] The existing supplier evaluation methods lack a systematic design for underlying data support and have failed to build a database architecture that is compatible with the entire lifecycle management of equipment and multi-dimensional correlation analysis, resulting in a weak evaluation foundation. 1. Limited and Static Data Dimensions: Existing technologies rely on data sources that only cover basic indicators such as equipment commissioning time, defect count, and unplanned downtime events. They fail to incorporate key data reflecting the equipment's entire lifecycle status and supplier service quality, such as dynamic aging rate, environmental sensitivity coefficient, and spare parts supply timeliness. For example, a supplier's equipment may age rapidly in high-humidity environments but maintain a low defect count through frequent maintenance. Existing data systems cannot capture such hidden quality risks, leading to evaluation results that deviate from reality.

[0004] 2. Fragmented Data Relationships: A multi-dimensional data model linking "supplier-equipment-site-environment" has not been established. Data such as equipment maintenance records, changes in environmental parameters, and supplier service behaviors are stored in a scattered manner, making it impossible to quantify the coupling effects between various elements. For example, the causal relationship between maintenance delays caused by spare parts supply delays and the number of equipment failures is fragmented into independent events in the existing data system, making it difficult to trace the supplier's fundamental responsibility.

[0005] 3. Insufficient dynamic data update and adaptation capabilities: In the face of the rapid technological iteration of power equipment (such as the significant differences between the failure modes of new energy storage equipment and traditional equipment), the existing database lacks a data adaptation mechanism for new equipment types and new failure modes. The sparsity of historical data is prominent, resulting in the evaluation model built based on traditional data having a serious lack of risk identification capability for new equipment.

[0006] Given the lack of sufficient data support, existing evaluation algorithms based on Bayesian models further amplify evaluation bias, mainly in the following aspects: 1. Incomplete Evaluation Dimension Coverage and Contradictory Independence Assumptions: The weighted Naive Bayes model used in CN114240016A is limited by the assumption of indicator independence. However, in real-world scenarios, equipment status is strongly coupled with supplier services (such as maintenance response speed and spare parts quality) and environmental factors (such as salt spray corrosion and humidity). This model fails to cover key dimensions such as equipment aging rate and environmental sensitivity coefficients, and incorrectly treats coupled indicators as independent variables, leading to inaccurate quantification of supplier responsibility. For example, if a supplier's delay in handling faults due to logistical issues triggers a chain of defects, the model only calculates the number of faults and deducts points, underestimating the supplier's overall responsibility.

[0007] 2. Lack of Environmental Adaptability and Dynamic Correction Mechanisms: Existing models cannot dynamically adapt to the spatiotemporal variations in the site environment. CN114240016A lacks a dynamic adjustment mechanism for environmental parameters. When salt spray corrosion intensifies at coastal sites (enhancing the environmental coefficient β), it still uses a fixed evaluation benchmark, leading to suppliers being wrongly penalized due to uncontrollable environmental factors. Although CN111292000A uses Bayesian averaging for partial correction, it does not explicitly distinguish the boundary between environmental disturbances and the supplier's own responsibility. For example, a surge in failure rates caused by extreme weather at the wind farm is directly attributed to the supplier, resulting in unfair evaluation results.

[0008] 3. Insufficient Quantification of Systemic Risks and Correlation Effects: Existing methods simplify supplier evaluation to the sum of scores for individual devices, failing to consider the risk of fault propagation between devices and the cascading effects of batch quality defects. Due to a lack of correlation data models and graph network analysis capabilities, it is impossible to identify interconnected failures in multiple devices caused by design defects in a particular batch of equipment. Simply accumulating deductions for individual devices is insufficient to reflect the supplier's systemic quality problems. Furthermore, the lack of differentiated quantification based on equipment importance (such as the weighting of main transformers and low-voltage switchgear) leads to an underestimation of the impact of failures in highly important equipment.

[0009] 4. Poor interpretability and traceability of evaluation results: The scoring process of existing models is "black box" in nature, lacking a traceability link from the scoring results to the original data. When deductions occur, it is impossible to clearly determine whether the deductions are due to equipment quality issues, sudden environmental changes, improper maintenance operations, or supplier service negligence. This is not conducive to targeted improvements by suppliers, nor can it support subsequent audits and refined decision-making.

[0010] 5. Conflict between historical data dependence and new equipment adaptation: CN111292000A relies on historical samples to construct a prior distribution, but the technical path and failure mode of new power equipment are significantly different from those of traditional equipment. The sparsity of historical data leads to prior estimation bias, making it impossible to accurately assess the quality risk of new equipment and limiting the applicability of the evaluation method in the context of technological iteration.

[0011] In summary, existing supplier evaluation technologies suffer from significant deficiencies in both data support systems and core algorithm models, resulting in insufficient accuracy, adaptability, and guidance in the evaluation results. This makes it difficult to meet the demands of modern power systems for refined supply chain management. Therefore, there is an urgent need to construct an intelligent evaluation system with multi-dimensional correlated data support, dynamic environment adaptability, systemic risk quantification capabilities, and full-process traceability. Building a database kernel adapted to this system and optimizing the evaluation model are key to solving the aforementioned technical problems. Summary of the Invention

[0012] The technical problem to be solved by this invention is to provide a method for intelligent evaluation of equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database, and to construct an accurate intelligent supplier evaluation model.

[0013] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for intelligent evaluation of equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database includes: Step 1: Calculate the initial score of the supplier. Based on the historical data in the two-ticket graph database, the initial score of the supplier is quantified by extracting key parameters such as specific site, operated component, component code and operation type. According to the equipment failure frequency recorded at the specific site, such as the proportion of "fault handling" operation times in "mechanical work ticket" and the differentiated weights of maintenance intensity such as "three-month scheduled inspection" and "replacement" operation types, the weight value of each type of maintenance operation is automatically calculated by the entropy weight method. Step 2: Establish a spatiotemporal coupled Bayesian model; based on the time series data generated in the two-ticket graph database, such as the maintenance operation intervals listed in the "work ticket permission time" and the maintenance records of equipment component codes, construct a spatiotemporal encoder ST-LSTM to model the dynamic state changes within the equipment life cycle; Step 3: Dynamic Score Correction; To address the indirect impact of site environment differences on equipment status, such as latent aging caused by salt spray corrosion, the supplier score is dynamically adjusted through an environmental compensation function; Based on the site environment coefficient, a hyperbolic tangent function is introduced to constrain the compensation amplitude, and the KL divergence gradient is combined to measure the disturbance of environmental changes on the distribution of latent variables; For example, in a site where a supplier encountered extreme rainfall, maintenance records showed a surge in "fault handling" operations in a short period of time. The model automatically increases the compensation factor and incorporates it into the score correction to balance the risk of misjudging supplier responsibility due to external environmental factors; Step 4: Calculate the target score for the equipment; Based on the dynamically adjusted supplier scores, calculate the quality score for individual equipment by integrating the degree of equipment aging and the similarity of operating modes. Step 5: Utilize Graph Attention Network (GAT) to score suppliers. Based on a supplier-equipment-site association network constructed from a two-vote graph database, nodes represent supplier and equipment codes, and edges represent operation types and time-series associations. The system integrates individual equipment scores with the systemic risk propagation effect to generate a global supplier score. The GAT dynamically allocates the influence weights between equipment and quantifies supplier risk volatility by combining the standard deviation of equipment scores. For example, when a "legacy issue handling" operation is associated with transformers of the same model from multiple sites, the scoring system will amplify the supplier's overall risk level and output a downgrade recommendation, thereby achieving a predictive assessment of potential supply chain risks.

[0014] In Step 1 above, the equipment failure frequency calculation uses equipment from a single supplier as the statistical unit. It incorporates normalization processes based on different site environmental factors, such as the coefficient for accelerated aging due to humidity. Finally, the initial score is constrained to a standardized range of (0, 100) using a Sigmoid function to avoid scoring distortion in extreme failure scenarios. For example, if the failure handling frequency of equipment supplied by a certain wind turbine supplier to a power plant exceeds a threshold, the initial score will significantly decrease and trigger an early warning.

[0015] The spatiotemporally coupled Bayesian model in Step 2 above adjusts the importance of input features through spatial gating mechanisms, such as assigning higher environmental weights to humidity-sensitive power plant data, and combines this with a Long Short-Term Memory (LSTM) network to capture the temporal correlation of equipment maintenance cycles. The output latent features are input into a Variational Autoencoder (VAE), which generates latent variable distributions through an inference network, such as the normal state clustering of the "7L-082F" wind turbine unit and the abnormal state clustering of the frequently replaced "5L-051F" component. This process achieves equipment quality pattern recognition in an unsupervised state, for example, detecting similar abnormal vibration characteristics in a batch of gearboxes in a high-temperature environment.

[0016] The construction of the two-vote graph database based on two votes in Step 1 above includes: The two types of data are derived from the operation and maintenance records of each site, covering information such as the specific site, the type of the two types of data, the components being operated (e.g., wind turbines, oil tanks, gearboxes, generators, frequency converters), component codes, operation types (e.g., fault handling, replacement, scheduled maintenance), and permitted time. Using graph database technology, the data is modeled as a network structure of nodes and edges. 1) Nodes: These include sites, equipment components, suppliers, operation types, etc. For example, a site node contains the environmental factor β attribute, and an equipment component node records attributes such as usage duration and design life.

[0017] 2) Edge: Represents the affiliation between equipment and the station, the supply relationship between equipment and the supplier, and operation records, including timestamps, operation types, and component codes; for example, an edge can describe "Wind turbine F22 performed a fault handling operation at Baishatan Power Plant on 2022-07-19".

[0018] 3) Weighting: Different weights are assigned to operation types according to maintenance intensity, and the weights are dynamically adjusted using the entropy weighting method to ensure that high-frequency, high-weight operations such as fault handling significantly affect the score. Graph databases support efficient querying and statistics, such as counting the number of faults handled by a supplier's equipment or the average maintenance interval of a site component, providing a data foundation for subsequent scoring.

[0019] In Step 4 above, the basic aging rate is quantified by using the ratio of years to design life. For example, for a fan that has been used for 3 years, α=3 / 5=0.6. An exponential decay factor is generated by combining a preset aging sensitivity coefficient such as λ=1.5 to ensure that the score of old equipment decreases. At the same time, the cosine similarity between the operating characteristics corresponding to the equipment code and its cluster center is compared to identify abnormal equipment that deviates from the standard pattern. Finally, the equipment status is comprehensively evaluated by three weighted scores.

[0020] The formula for calculating the initial supplier score in Step 1 above is: ; in: Failure rate This directly reflects the reliability of the equipment; Maintenance intensity ; The weight is the operation type weight, representing the maintenance strength. The frequency of operation types, such as fault handling and replacement, represents the number of maintenance operations; the weight of each operation type is automatically learned using the entropy weight method. and in combination with the number of operations Weighted calculation; for example, high-frequency fault handling has a greater impact on m than scheduled maintenance.

[0021] Normalization constant Control the scoring range; Sigmoid function Constrain the score within the (0, 100) range to eliminate the influence of extreme values; , For sensitivity and maintenance factor.

[0022] The specific process of establishing the spatiotemporally coupled Bayesian model in Step 2 above includes: Step 2.1: Extract the spatiotemporal features of the device using a spatiotemporal encoder; Step 2.2: Variational Autoencoder (VAE) implements unsupervised clustering.

[0023] Step 2.1 above specifically includes: Step 2.1.1: Input data is a time series. The license interval Indicates maintenance cycle; operation type weight. A one-hot vector representing the maintenance type; site environmental coefficient. This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value; Step 2.1.2, the weighting of the impact of the spatial gating dynamic adjustment environment on the operation, the formula is: ; in, It is the sigmoid function. This is the current input. It is the hidden state of the previous moment; the parameters of the spatial gating formula. , , These parameters are obtained through model training; specifically, they are automatically learned during the training process using the backpropagation algorithm and optimizer. The weight matrix input to the gate determines the degree of influence of the current input data, such as time interval, operation type weight, and environment coefficient, on the gate value. The weight matrix from the hidden state to the gate is used to link the hidden state of the previous time step with the current gate calculation, thereby capturing temporal dependencies. The bias term for gating is used to adjust the gating signal; Training samples are constructed using historical data. Each sample is a time series containing historical operation records of the device, such as time intervals, operation type weights, and environmental coefficients. The loss function adopted is the loss function of the Variational Autoencoder (VAE), including reconstruction loss and KL divergence loss. The reconstruction loss measures the encoder's ability to reconstruct the input data; the KL divergence loss measures the difference between the latent variable distribution and the prior distribution. The loss function is calculated using the backpropagation algorithm. , , The gradients are calculated, and the optimizer is used to update these parameters; The learned parameters can capture the impact of different operation types and environmental factors on equipment status. If fault handling operations (with higher weight) are performed in a humid environment (…),… It is more likely to occur at higher (higher) levels. The corresponding operation type and environment coefficient will have a larger weight. It captures time-dependent relationships, such as two consecutive fault handlings possibly indicating a deterioration in the equipment's condition.

[0024] Step 2.1.3, LSTM timing modeling: Adjusted input Latent features generated by LSTM network Characterizes the spatiotemporal operating status of the equipment: .

[0025] Step 2.2 above specifically includes: Step 2.2.1, Inference Network, i.e., Encoder Will Mapping to latent variables ,in: ; ; ; The inference network is used to extract the feature vectors output by the ST-LSTM. Mapping to latent variables The distribution parameters, mean and variance ; To calculate the weight matrix for the latent variable mean, To calculate the bias of the latent variable mean, To calculate the weight matrix for the variance of the latent variables, This is used to calculate the bias of the latent variable variance.

[0026] The encoder will identify the characteristics of the device maintenance sequence. This is transformed into a probability distribution such as a Gaussian distribution, which represents the health status of the device; for example, for a gearbox that frequently fails, the encoder may map it to a specific region in the latent space, such as a high-risk category.

[0027] Step 2.2.2: Generate the network, i.e., the decoder: ; Reconstruction features: ; The function of the decoder is to extract hidden variables If the reconstruction error of the feature vector of the reconstructed device is small, it indicates that the latent variables... It can accurately represent the health status of the equipment; The optimization objective is to minimize the reconstruction loss and the KL divergence loss. ; This process automatically identifies device clusters, such as stable and high-risk.

[0028] In Step 3 above, when performing dynamic score correction, it is based on the environmental coefficient. The model sensitivity adjustment score satisfies: ; Among them, the index term quantifies the impact of environmental severity, and the site environmental coefficient This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value. To compensate for the intensity, To compensate for the growth rate; the gradient of KL divergence with respect to the environmental coefficient. The sensitivity of the model to the environment is measured by the environment sensitivity gradient calculated during the VAE training process in Step 2.

[0029] In Step 4 above, when calculating the target score for the equipment, the formula for calculating the equipment score by combining supplier quality and its own condition is as follows: ; Among them, aging rate This reflects the consumption of equipment lifespan, including equipment usage time. Indicates the number of years the equipment has been used, and its design life. Indicates the rated lifespan. For aging sensitivity coefficient; pattern matching term Characteristics of computing devices With cluster center Cosine similarity; if the similarity is below a threshold, an anomaly warning is triggered; spatiotemporal features The device feature vector derived from the ST-LSTM output in Step 2 represents the device operating status features; cluster centers The device cluster centers generated by VAE in Step 2 identify the baseline status of devices of the same type.

[0030] In Step 5 above, when using the Graph Attention Network (GAT) to score suppliers, device scores are aggregated and risk propagation is assessed: ; Among them, attention mechanism Quantify the risk of fault propagation between equipment; standard deviation term To measure the variability in equipment quality.

[0031] The present invention discloses a method for intelligent evaluation of equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database. The technical improvements of this invention are: 1. Dynamic spatiotemporally coupled modeling based on a two-ticket graph database. This system innovatively constructs a two-ticket graph database spanning different work sites and equipment types. By parsing work tickets and data such as work site information, operation type, operated component codes, and time series, it maps equipment maintenance records into a heterogeneous graph structure containing spatiotemporal attributes. Nodes in the graph encompass work site environmental factors, equipment entities, and operation types, while edge weights are dynamically calculated based on operation frequency, maintenance time, and fault severity. Combined with a spatiotemporally gated LSTM model, the system automatically extracts the operating mode characteristics of equipment under specific environments. For example, in salt spray corrosion work sites, the failure interval of fan gearboxes is 30% shorter than in dry environments. This spatiotemporal coupling effect is used to generate a mixture of Gaussian prior distributions of equipment health status through latent variable clustering (VAE) technology, providing a data-driven dynamic baseline for subsequent scoring.

[0032] 2. Maintenance Intensity Quantization of Entropy Weight Fusion for Multi-Source Operation Types To address the diverse operational types involved in the two-ticket system, this system proposes a dynamic weight allocation mechanism based on the entropy weight method. By statistically analyzing the frequency distribution of various operations at different sites, the information entropy value is calculated and a weight coefficient is generated: high-frequency, low-entropy operations have significantly higher weights than low-frequency, high-entropy operations. For example, the entropy weight for fault handling is 0.46, while that for scheduled maintenance is only 0.12. This ultimately generates a maintenance intensity index, effectively distinguishing between quality defects in supplier equipment and normal operational needs, avoiding scoring biases caused by the simple accumulation of maintenance frequencies in traditional methods.

[0033] 3. Nonlinear scoring correction mechanism driven by site environmental factors To address the impact of environmental variations on equipment aging, this system introduces a site environmental factor β and designs an exponential compensation function to dynamically adjust supplier scores. By analyzing the correlation between historical fault data and environmental parameters, the system automatically calculates the environmental gradient sensitivity factor and uses a hyperbolic tangent function to constrain the correction amplitude. This both counteracts environmental interference and avoids inflated scores due to overcompensation. The environmental factor β amplifies or weakens the score through an exponential function, while the KL divergence gradient constrains the correction amplitude, balancing environmental interference and supplier responsibility.

[0034] 4. Decoupled Equipment Aging Model and Risk Transmission Network This system innovatively decomposes the equipment aging rate into a time baseline term (α=t). use / t life The evaluation includes an environmental impact component, where the time baseline is quantified by a Weibull decay model to measure natural aging, while the environmental impact is handled by an independent correction module. Simultaneously, a supplier-equipment bipartite graph is constructed based on component codes and operation records from both sets of votes. A graph attention network (GAT) is used to quantify the risk of fault propagation between equipment. For example, if a gearbox failure leads to generator overload, high-weighted edges are used to increase the risk score of related equipment, thus accurately reflecting the cascading effects of batch quality defects in the supplier evaluation.

[0035] 5. A data collaboration mechanism for end-to-end closed-loop optimization From data extraction and graph database construction to score generation, this system achieves closed-loop data optimization through bidirectional feedback. For example, when a batch of equipment has high-frequency fault handling records in multiple sites, the system automatically adjusts its latent variable cluster centers, triggering an increase in VAE reconstruction loss, and updates the spatiotemporal gating weights through backpropagation, making the model more sensitive to abnormal signals of this type of component. At the same time, the supplier scoring results drive the priority allocation of the two work orders, forming a complete link from data collection and analysis to decision execution, which can effectively improve the efficiency of operation and maintenance response compared to traditional methods.

[0036] The present invention has the following beneficial effects: By constructing a multi-source data fusion system based on a two-ticket graph database, the system achieves dynamic, precise, and intelligent evaluation of equipment quality and suppliers. Specifically, the system uses historical work ticket data from the site as its core. By analyzing ticket types such as "mechanical work tickets," "electrical work tickets," and "wind turbine generator work tickets," it extracts maintenance records of the "operated components" and their corresponding "operation types." Combined with the "work ticket permit time" sequence, it constructs a full lifecycle behavior profile of the equipment, thereby quantifying the frequency of equipment failures, maintenance intensity, and aging degree. Through a spatiotemporally coupled Bayesian model, the system introduces site environmental coefficients to dynamically correct the evaluation results. It uses VAE unsupervised clustering technology to automatically learn maintenance weights from the distribution of operation types and combines graph attention network (GAT) to analyze the risk of fault propagation between equipment, ultimately achieving environmental adaptability and risk warning for supplier scoring. This technology, through an automated, data-driven evaluation system, upgrades traditional quarterly manual assessments to minute-level real-time calculations, effectively improving the accuracy of equipment anomaly identification and the timeliness of supplier risk prediction. It provides a new generation of technological paradigms for enterprises to optimize supply chain management and reduce unplanned downtime losses. Attached Figure Description

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the process of establishing a spatiotemporally coupled Bayesian model according to the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0039] Example 1: 1. Construction of the Two-Vote Graph Database The two types of data are derived from the operation and maintenance records of each site, covering information such as the specific site, the type of the two types of data, the components being operated (e.g., wind turbines, oil tanks, gearboxes, generators, frequency converters, etc.), component codes, operation types (e.g., fault handling, replacement, scheduled maintenance, etc.), and permitted time. Using graph database technology, the data is modeled as a network structure of nodes and edges. (1) Nodes: These include sites, equipment components, suppliers, operation types, etc. For example, the site node contains the environmental factor β attribute, and the equipment component node records attributes such as usage duration and design life.

[0040] (2) Edge: Represents the affiliation between equipment and the station, the supply relationship between equipment and the supplier, and the operation record (including timestamp, operation type, component code). For example, an edge can describe "Wind turbine F22 performed a fault handling operation at Baishatan Power Plant on 2022-07-19".

[0041] (3) Weighting: Different weights are assigned to operation types according to maintenance intensity, and the weights are dynamically adjusted by the entropy weight method to ensure that high-frequency, high-weight operations (such as fault handling) significantly affect the score.

[0042] This database supports efficient querying and statistics, such as counting the number of faults handled by a supplier's equipment or the average maintenance interval of a site component, providing a data foundation for subsequent scoring.

[0043] 2. Calculation of Initial Supplier Score The supplier's initial score reflects the basic quality of their equipment, and is calculated using the following formula: ; (1) Failure rate This directly reflects the reliability of the equipment.

[0044] (2) Maintenance intensity , The weight is the operation type weight, representing the maintenance strength. The frequency of operation type (e.g., troubleshooting, replacement, etc.) represents the number of maintenance operations. The weights of each operation type are automatically learned using the entropy weight method. and in combination with the number of operations Weighted calculation. For example, high-frequency fault handling has a greater impact on m than scheduled maintenance.

[0045] (3) Normalization constant Control the scoring range.

[0046] (4) Sigmoid function Constrain the score within the (0, 100) range to eliminate the influence of extreme values.

[0047] (5) , For sensitivity and maintenance factor.

[0048] 3. Spatiotemporally Coupled Bayesian Model 3.1 Spatiotemporal features extracted by the spatiotemporal encoder (1) Input data is a time series The license interval Indicates the maintenance cycle. Operation type weight. This is a one-hot vector representing the maintenance type. Site environmental coefficient. This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value.

[0049] (2) Spatial gating The weight of the impact of the environment on operation is dynamically adjusted.

[0050] It is the sigmoid function. This is the current input. It is the hidden state from the previous moment.

[0051] Parameters of the space gating formula , , These parameters are obtained through model training. Specifically, they are automatically learned during the training process using the backpropagation algorithm and optimizer. The weight matrix is ​​the input to the gating, which determines the degree of influence of the current input data (such as time interval, operation type weight, and environment coefficient) on the gating value. The weight matrix from the hidden state to the gate is used to link the hidden state of the previous time step with the current gate calculation, thereby capturing temporal dependencies. The bias term for gating is used to adjust the gating signal.

[0052] Training samples are constructed using historical data, with each sample being a time series containing historical operation records of the device (time intervals, operation type weights, environmental coefficients, etc.). The loss function adopted is the loss function of a variational autoencoder (VAE), including reconstruction loss and KL divergence loss. Reconstruction loss measures the encoder's ability to reconstruct the input data. KL divergence loss measures the difference between the latent variable distribution and the prior distribution. The loss function is calculated using the backpropagation algorithm. , , The gradient is calculated, and the optimizer is used to update these parameters.

[0053] The learned parameters can capture the impact of different operation types and environmental factors on equipment status. If fault handling operations (with higher weight) are performed in a humid environment (…),… It is more likely to occur at higher (higher) levels. The corresponding operation type and environment coefficient will have a larger weight. It captures time-dependent relationships, such as two consecutive fault handlings possibly indicating a deterioration in the equipment's condition.

[0054] (3) LSTM timing modeling: adjusted input Latent features generated by LSTM network This characterizes the spatiotemporal operating status of the equipment.

[0055] ; 3.2 Variational Autoencoder (VAE) for Unsupervised Clustering (1) Inference Network (Encoder) Will Mapping to latent variables ,in ; ; ; The inference network is used to extract the feature vectors output by the ST-LSTM. Mapping to latent variables Distribution parameters (mean) and variance ). To calculate the weight matrix for the latent variable mean, To calculate the bias of the latent variable mean, To calculate the weight matrix for the variance of the latent variables, This is used to calculate the bias of the latent variable variance.

[0056] The encoder will identify the characteristics of the device maintenance sequence. This is transformed into a probability distribution (Gaussian distribution) that represents the health status of the device. For example, for a gearbox that frequently fails, the encoder might map it to a specific region in the latent space (high-risk category).

[0057] (2) Generating network (decoder) ; Reconstruction features: ; The function of the decoder is to extract hidden variables If the reconstruction error of the feature vector of the reconstructed device is small, it indicates that the latent variables... It can accurately indicate the health status of the equipment.

[0058] The optimization objective is to minimize the reconstruction loss and the KL divergence loss. ; This process automatically identifies device clusters (such as stable and high-risk types).

[0059] The loss function can simultaneously train all parameters of the spatially gated ST-LSTM, the inference network (encoder), and the generative network (decoder). The gradient-passive design enables end-to-end joint training, and a single loss function can uniformly optimize all parameters, avoiding information loss during staged training. The transmission chain is as follows: ->Decoder->Encoder->ST-LSTM->Spatial Gating, which can guarantee the underlying parameters (such as...) Perceive high-level objectives.

[0060] By simultaneously training the parameters of the spatial gating, inference, and generative networks using a unified loss function (VAE loss), the model is able to extract spatiotemporal features from the maintenance sequence and map them to the latent space for device health status clustering.

[0061] The spatiotemporal encoder considers the event sequence and spatial (environmental) impacts of maintenance events. Through a gating mechanism, the environmental coefficient... This affects the input, causing maintenance events in harsh environments to be assigned different weights. Latent variables generated through VAE clustering... It will be converted into a device feature vector. (Decoder output), this Simultaneously used for cosine similarity calculation in S104 and graph network node features in S105, this design ensures a smooth transition from single-device assessment to system-level risk assessment.

[0062] 4. Dynamic score correction Based on environmental coefficient Adjusted score for model sensitivity: ; (1) The index term quantifies the impact of environmental severity, including the site environmental coefficient. This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value. To compensate for the intensity, To compensate for the growth rate.

[0063] (2) Gradient of KL divergence with respect to the environmental coefficient The sensitivity of the model to the environment is measured by the environment sensitivity gradient calculated during the VAE training process of S102.

[0064] This mechanism avoids scoring distortion caused by environmental interference. For example, equipment failure rates are naturally higher in humid environments, but supplier responsibility should not be overemphasized. Through environmental gradient feedback and index compensation mechanisms, it achieves precise decoupling between equipment operating environment and supplier responsibility, providing an environmentally adaptable correction paradigm for supplier evaluation.

[0065] 5. Equipment Target Score The equipment score is a combination of supplier quality and its own condition. ; (1) Aging rate This reflects the consumption of equipment lifespan, including equipment usage time. Indicates the number of years the equipment has been used, and its design life. Indicates the rated lifespan. This represents the aging sensitivity coefficient.

[0066] (2) Pattern matching items Characteristics of computing devices With cluster center The cosine similarity is used. If the similarity is below a threshold, an anomaly warning is triggered. Spatiotemporal features. The device feature vector, derived from the ST-LSTM output of S102, represents the device's operating status characteristics. Cluster centers. The device cluster centers generated by the VAE from S102 identify the baseline state of similar devices. For example, if a gearbox frequently fails, its features deviate from the healthy cluster centers, resulting in a similarity of only 0.75 and reduced device gain.

[0067] 6. Network Provider Rating Aggregate equipment scores and assess risk propagation: ; (1) Attention mechanism Quantify the risk of fault propagation between devices. For example, if a low-scoring device increases its attention weight to its physically related devices, it will increase the risk contribution.

[0068] (2) Standard deviation To measure the variability in equipment quality.

[0069] Example 2: Suppose a supplier provides 10 gearboxes, 10 generators, and 10 frequency converters, for a total of 30 devices, with an environmental factor of 1.2.

[0070] The types and quantities of the equipment are shown in the table below:

[0071] Its operation type weight is:

[0072] Its operation log includes:

[0073] Failure rate ; Average maintenance intensity per unit ; Normalization constant: ; ; intermediate variables ; Initial rating 71.53.

[0074] The second step requires time-series data, and in actual scoring, each device needs to be calculated independently. For simplicity, we'll use the gearbox GB-001 as an example:

[0075] Taking the gearbox GB-001 as an example, the weighting parameters are:

[0076] Taking the GB-001 gearbox as an example: Initial state: ; Time step 1 (fault handling): enter: ; Gating calculation:

[0077] ; Adjust input: ; LSTM output: ; Time step 2 (replace): enter: ; Gating calculation:

[0078] ; Adjust input: ; LSTM output: ; Time step 3 (routine inspection): enter: ; Gating calculation:

[0079] ; Adjust input: ; LSTM output: .

[0080] Taking the GB-001 gearbox as an example, and :

[0081] Latent variable distribution calculation: ; ; .

[0082]

[0083] have: ; After dynamic correction of the fractions, we have:

[0084] Equipment aging rate:

[0085] Spatiotemporal features (from the second-step model):

[0086] Device attenuation value, similarity, and target score:

[0087] have:

[0088] In summary:

[0089] Total attention: ; Risk coefficient calculation: ; Average score: ; Standard deviation: ; Final rating: .

Claims

1. A method for intelligent evaluation of equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database, characterized in that, include: Step 1: Initial Supplier Score Calculation; Based on historical data in the two-ticket graph database, the initial supplier score is quantified by extracting key parameters such as specific site, operated component, component code, and operation type. According to the equipment failure frequency and maintenance intensity recorded at the specific site, the weight value of various maintenance operations is automatically calculated using the entropy weight method. Step 2: Establish a spatiotemporal coupled Bayesian model; based on the time series data generated in the two-ticket graph database, construct a spatiotemporal encoder ST-LSTM to model the dynamic state changes during the device's life cycle; Step 3: Dynamic score correction; To address the indirect impact of site environment differences on equipment status, supplier scores are dynamically adjusted through an environmental compensation function; Based on the site environment coefficient, a hyperbolic tangent function is introduced to constrain the compensation amplitude, and the KL divergence gradient is combined to measure the disturbance of environmental changes on the distribution of latent variables. Step 4: Calculate the target score for the equipment; Based on the dynamically adjusted supplier scores, calculate the quality score for individual equipment by integrating the degree of equipment aging and the similarity of operating modes. Step 5: Use Graph Attention Network (GAT) to score suppliers; Based on the supplier-equipment-site association network constructed in the two-vote graph database, integrate individual equipment scores with the systemic risk propagation effect to generate a global supplier score; Dynamically allocate the influence weights between equipment through the Graph Attention Network (GAT), and combine the standard deviation of equipment scores to quantify the supplier risk volatility.

2. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, is characterized in that... In Step 1, the equipment failure frequency calculation uses equipment provided by a single supplier as the statistical unit. Combined with the normalization processing of environmental coefficients of different sites, the initial score is finally constrained to the standardized range of (0,100) by Sigmoid function mapping to avoid scoring distortion under extreme failure scenarios.

3. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, is characterized in that... In Step 2, the spatiotemporally coupled Bayesian model adjusts the importance of input features through a spatial gating mechanism and combines a Long Short-Term Memory (LSTM) network to capture the temporal correlation of equipment maintenance cycles; the output latent features are input into a variational autoencoder (VAE), and the latent variable distribution is generated through an inference network.

4. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, is characterized in that... The construction of the two-vote graph database based on two votes in Step 1 includes: The two types of data are derived from the operation and maintenance records of each site, covering the specific site, the type of the two types of data, the component being operated on, the component code, the operation type, and the license time information; using graph database technology, the data is modeled as a network structure of nodes and edges: 1) Nodes: including site, equipment components, suppliers, and operation type; 2) Edge: Indicates the subordinate relationship between equipment and the site, the supply relationship between equipment and the supplier, and operation records; 3) Weighting: Different weights are assigned to operation types based on maintenance intensity, and the weights are dynamically adjusted using the entropy weighting method to ensure that high-frequency, high-weight operations have an impact on the score; Graph databases support efficient querying and statistics, providing a data foundation for subsequent scoring.

5. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, characterized in that, In Step 4, the basic aging rate is quantified by using the ratio of years to design life, and an exponential decay factor is generated by combining a preset aging sensitivity coefficient to ensure that the score of old equipment decreases. At the same time, the cosine similarity between the operating characteristics corresponding to the equipment code and its cluster center is compared to identify abnormal equipment that deviates from the standard pattern. Finally, the equipment status is comprehensively evaluated through three weighted scores.

6. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 3, is characterized in that... The formula for calculating the initial supplier score in Step 1 is as follows: ； in: Failure rate This reflects the reliability of the equipment; Maintenance intensity ; The weight is the operation type weight, representing the maintenance strength. Operation type frequency represents the number of maintenance operations; the weights of each operation type are automatically learned using the entropy weight method. and in combination with the number of operations Weighted calculation; Normalization constant Control the scoring range; Sigmoid function Constrain the score within the (0, 100) range to eliminate the influence of extreme values; , For sensitivity and maintenance factor.

7. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, is characterized in that... The specific process of establishing the spatiotemporally coupled Bayesian model in Step 2 includes: Step 2.1: Extract the spatiotemporal features of the device using a spatiotemporal encoder; Step 2.2: Variational Autoencoder (VAE) implements unsupervised clustering.

8. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 7, is characterized in that... Step 2.1 specifically includes: Step 2.1.1: Input data is a time series. The license interval Indicates maintenance cycle; operation type weight. A one-hot vector representing the maintenance type; site environmental coefficient. This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value; Step 2.1.2, the weighting of the impact of the spatial gating dynamic adjustment environment on the operation, the formula is: ； in, It is the sigmoid function. This is the current input. It is the hidden state of the previous moment; the parameters of the spatial gating formula. , , These parameters are obtained through model training; specifically, they are automatically learned during the training process using the backpropagation algorithm and optimizer. The weight matrix is ​​the input to the gate, which determines the degree of influence of the current input data on the gate value; The weight matrix from the hidden state to the gate is used to link the hidden state of the previous time step with the current gate calculation, thereby capturing temporal dependencies. The bias term for gating is used to adjust the gating signal; Training samples are constructed using historical data, with each sample being a time series containing historical operation records of the device. The loss function adopted is the loss function of a variational autoencoder (VAE), including reconstruction loss and KL divergence loss. The reconstruction loss measures the encoder's ability to reconstruct the input data; the KL divergence loss measures the difference between the latent variable distribution and the prior distribution. The loss function is calculated using the backpropagation algorithm. , , The gradients are calculated, and the optimizer is used to update these parameters; Step 2.1.3, LSTM timing modeling: Adjusted input Latent features generated by LSTM network Characterizes the spatiotemporal operating status of the equipment: 。 9. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 7, is characterized in that... Step 2.2 specifically includes: Step 2.2.1, Inference Network, i.e., Encoder Will Mapping to latent variables ,in: ； ； ； The inference network is used to extract the feature vectors output by the ST-LSTM. Mapping to latent variables The distribution parameters; To calculate the weight matrix for the latent variable mean, To calculate the bias of the latent variable mean, To calculate the weight matrix for the variance of the latent variables, To calculate the bias of the latent variable variance; The encoder will identify the characteristics of the device maintenance sequence. This is converted into a probability distribution, which represents the health status of the device. Step 2.2.2: Generate the network, i.e., the decoder: ； Reconstruction features: ； The function of the decoder is to extract hidden variables The feature vector of the reconstructed device; The optimization objective is to minimize the reconstruction loss and the KL divergence loss. ； This process automatically identifies device clusters.

10. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 3, is characterized in that... In Step 3, when performing dynamic score correction, it is based on the environmental coefficient. The model sensitivity adjustment score satisfies: ； Among them, the index term quantifies the impact of environmental severity, and the site environmental coefficient This indicates the severity of the environment; the more severe the environment, the higher the severity. The higher the value; To compensate for the intensity, To compensate for the growth rate; the gradient of KL divergence with respect to the environmental coefficient. The sensitivity of the model to the environment is measured by the environment sensitivity gradient calculated during the VAE training process in Step 2.

11. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 3, is characterized in that... In Step 4, when calculating the target score for the equipment, the formula for calculating the equipment score, which combines supplier quality and its own condition, is as follows: ； Among them, aging rate This reflects the consumption of equipment lifespan, including equipment usage time. Indicates the number of years the equipment has been used, and its design life. Indicates the rated lifespan. For aging sensitivity coefficient; pattern matching term Characteristics of computing devices With cluster center Cosine similarity; if the similarity is below a threshold, an anomaly warning is triggered; spatiotemporal features The device feature vector derived from the ST-LSTM output in Step 2 represents the device operating status features; cluster centers The device cluster centers generated by VAE in Step 2 identify the baseline status of devices of the same type.

12. The intelligent evaluation method for equipment quality and suppliers based on a spatiotemporally coupled Bayesian model and a two-ticket graph database as described in claim 1, characterized in that, In Step 5, when using the Graph Attention Network (GAT) to score suppliers, device scores are aggregated and risk propagation is assessed. ； Among them, attention mechanism Quantify the risk of fault propagation between equipment; Standard deviation term To measure the variability in equipment quality.

Images