An equipment intelligent evaluation system based on operation and state parameter fusion

Abstract

The application relates to an equipment intelligent evaluation system based on operation and state parameter fusion, and particularly relates to the field of industrial equipment predictive maintenance. The scheme breaks through the limitation of isolated analysis of equipment in traditional evaluation, constructs a real-time dependence network among equipment dynamically, and fuses individual states in the whole operation context for perception; the scheme can accurately quantify the chain reaction and potential influence of single-point equipment performance degradation in the whole cooperative network, realizes context-aware evaluation of equipment health, can also simulate failure propagation path in advance to reveal hidden cascading risks, and finally generates key ranking reflecting the real risk level of equipment through fusion of average loss and result uncertainty, so as to provide direct and reliable decision basis for predictive maintenance and accurate resource allocation from the global perspective, and effectively improve the overall operation resilience and safety guarantee level of complex equipment clusters.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance of industrial equipment, and more specifically, to an intelligent equipment evaluation system based on the fusion of operating and status parameters. Background Technology

[0002] In modern industrial and infrastructure systems, typical scenarios such as intelligent manufacturing production lines, smart building clusters, and power grid substations are characterized by the deep coupling of numerous equipment units through precise physical connections, continuous energy flow, and high-speed information flow, forming a complex system that operates collaboratively. In such scenarios, the functional output of a single device is not only a reflection of its own performance but also a prerequisite for the normal operation of downstream equipment or a key constraint on the overall system efficiency. For example, on an automated automobile assembly line, a delay in the cycle time of a tightening robot can directly lead to waiting and idling at subsequent painting and inspection stations, ultimately affecting the output efficiency and product quality of the entire production line. These dependencies between devices not only exist in static hardware connections but also dynamically evolve with changes in software logic such as production task scheduling and load allocation strategies, forming a complex network that is multi-layered, interwoven, and changes in real time.

[0003] However, existing equipment status assessment technologies primarily focus on the operating parameters and health indicators of individual devices, lacking the ability to model and analyze the complex dependencies between devices and their dynamic impacts at the system level. Currently widely used system reliability assessment methods, such as reliability block diagrams or fault tree analysis based on the assumption of device independence, typically treat devices as isolated probabilistic units with static and deterministic series and parallel logical relationships. These methods struggle to characterize the nonlinear transmission and amplification effects of equipment status degradation under actual physical coupling and dynamic control logic, and are even less capable of quantifying the cascading impact risks of a local performance degradation on the overall system output, stability, or energy efficiency. The fundamental technical problem lies in the fact that traditional assessment models fail to effectively integrate two key dimensions of information: first, real-time, multi-source intrinsic status data of individual devices; and second, dynamic correlation data between devices that change with business logic and physical constraints. This fragmentation leads to maintenance decisions often based on local health scores, potentially severely underestimating the potential destructive power of critical bottleneck devices or misdirecting maintenance resources to links with minimal impact on the overall system resilience. This can trigger cascading failures or even systemic paralysis when complex systems face disturbances. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an intelligent equipment evaluation system based on the fusion of operating and state parameters. The system utilizes a graph construction module, a graph neural network evaluation module, a cascaded failure simulation module, and a criticality ranking module to solve the problems mentioned in the background.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes: a graph construction module, a graph neural network evaluation module, a cascaded failure simulation module, and a key ranking module connected in sequence, wherein; Graph construction module: Receives the operating parameters, status parameters and association data describing the topological and logical relationships between devices. Based on the association data, it abstracts each device into nodes and the association relationship into edges to construct a dependency graph. Based on the real-time interaction data in the operating parameters, it dynamically calculates and updates the weight of each edge in the dependency graph, and generates and outputs a dynamic weighted multigraph. The graph neural network evaluation module obtains the temporal feature vector and dynamic weighted multigraph extracted from each device based on its operating and state parameters. The temporal feature vector is used as the initial feature of the corresponding node, and the dynamic weight of each edge in the dynamic weighted multigraph is used as the edge feature. The attention message passing mechanism that integrates edge weights is used to iteratively aggregate and update the node features, and outputs the network awareness health of each device. Cascaded Failure Simulation Module: Acquires a dynamic weighted multigraph and the network-aware health of each device, determines the probability of each device as the initial failure point based on the network-aware health, sets the failure propagation rules on the edges based on the dynamic weights of each edge in the dynamic weighted multigraph, performs multiple Monte Carlo simulations in combination with the probability and propagation rules, records and outputs the overall performance loss value caused by each simulation. Criticality ranking module: Based on the overall performance loss value, for each device, the probability distribution of the overall performance loss value caused by the device as the initial failure point is statistically analyzed. The influence entropy of the device is calculated based on the probability distribution, and the average performance loss caused by the device is calculated based on the probability distribution. The comprehensive criticality index of the device is calculated by combining the average performance loss and the influence entropy. All devices are ranked according to the comprehensive criticality index.

[0006] In a preferred embodiment, the graph construction module dynamically calculates and updates the weights of each edge in the dependency graph based on real-time interaction data in the running parameters. The specific process is as follows: For each edge in the dependency graph that represents a relationship, real-time interaction data between the corresponding two devices is received. The deviation between the current value of real-time interactive data and the long-term benchmark value obtained based on historical data is calculated. The long-term benchmark value includes the historical mean and the historical standard deviation. The deviation is calculated by subtracting the historical mean from the current value and then dividing by the sum of the historical standard deviation and a small constant. At the same time, the instantaneous rate of change of the current value of the real-time interactive data relative to the value of the previous moment is calculated, and the instantaneous rate of change is compressed and mapped using the hyperbolic tangent function, and then multiplied by an adjustable gain coefficient to obtain an adjustment amount that reflects the instantaneous change trend. Based on the sum of the aforementioned deviation and trend adjustment, the dynamic weight of this edge is calculated, and the weight attribute of the corresponding edge in the dependency graph is updated using this weight value, thereby generating and outputting a dynamically weighted multigraph.

[0007] In a preferred embodiment, the specific process of abstracting each device as a node and the association relationship as an edge based on the associated data to construct a dependency graph is as follows: From the received associated data, various relationships between devices are parsed out, including static topology relationships defined by physical connections and logical dependencies defined by process sequence. For any two devices that have a relationship, they are mapped to two nodes in the dependency graph; For each relationship between the two nodes, an independent directed edge is created in the graph to correspond to it, and different edge type identifiers are used to distinguish them, thereby constructing a multigraph structure that allows multiple edges of different types between two nodes. This structure is the topological skeleton of the dynamic weighted multigraph. The dynamic calculation and updating of weights are performed independently and in parallel for each edge with an edge type identifier in this dynamically weighted multigraph.

[0008] In a preferred embodiment, the process of iteratively aggregating and updating node features using an attention message passing mechanism that incorporates edge weights in the graph neural network evaluation module is as follows: First, for each directed edge pointing to the target node in the dynamic weighted multigraph, the attention coefficient of edge weight enhancement is calculated. The calculation process concatenates the feature vector held by the target node after the previous iteration, the feature vector held by the neighboring nodes connected to the target node through this edge after the previous iteration, and the dynamic weight vector carried by this edge in the dynamic weighted multigraph to form a comprehensive feature vector. Next, the comprehensive feature vector is multiplied by a pre-defined learnable attention weight vector, and then processed by a linear rectified activation function with a negative slope to obtain an original attention score. The original attention score reflects both the importance of the neighboring node state and the real-time strength of the connection edge. Secondly, normalized attention and neighbor information aggregation are performed. For the same target node, the original attention scores corresponding to all its incoming edges are input into the exponential normalization function for processing, so that all scores are converted into normalized attention weights that sum to one. Then, the feature vector held by each neighbor node after the previous iteration is linearly transformed through a learnable shared weight matrix to obtain the transformed feature vector of the neighbor node; the transformed feature vector of each neighbor node is multiplied by its corresponding normalized attention weight to obtain a weighted feature vector; all such weighted feature vectors are summed and then passed through a non-linear activation function to obtain a new feature vector of the target node after aggregating neighborhood information for the next iteration. Finally, the message passing steps are repeated multiple times to complete the iterative update of node features.

[0009] In a preferred embodiment, the specific process of outputting the network-aware health status of each device is as follows: After multiple rounds of iterative message passing, the feature vector representation of each device node in the final round is obtained; The final feature vector representation of each node is input into a final fully connected computation layer for linear transformation. The fully connected computation layer contains learnable weight vectors and bias terms. By applying the S-shaped growth function to the result of the linear transformation, the output value is mapped to a numerical range of zero to one, thereby obtaining the network-aware health of the corresponding device. Network health awareness integrates the device's own operating status characteristics, the device's location information in the network topology, and the strength of real-time dynamic dependencies between devices.

[0010] In a preferred embodiment, the cascading failure simulation module determines the probability of each device being the initial failure point based on the network's perceived health status. Specifically, this is calculated using a nonlinear probability mapping model, and the calculation process is as follows: For each device, the device's network awareness health is input into an exponential conversion function. The input to the conversion function is the product of the negative number of the network awareness health and a preset positive sensitivity coefficient. Then, the natural constant is calculated and the product result is used as the power of the exponent to obtain an intermediate value. Summing up the intermediate values ​​of all devices and adding the sum to a preset minimal positive smoothing constant to obtain the denominator value; Finally, dividing the median value of a single device by the denominator gives the probability that the device will be selected as the initial failure point in a single Monte Carlo simulation.

[0011] In a preferred embodiment, the process of setting the propagation rules of failure on edges based on the dynamic weights of each edge in the dynamically weighted multigraph is as follows: Define a conditional propagation probability model to calculate the probability that when a device fails, the failure effect will propagate along the dependency edges to the device's neighboring devices. For all directed edges in a dynamically weighted multigraph from a failed device to any of its neighboring devices, first find the edge with the largest dynamic weight value and use the largest dynamic weight value as one of the key inputs. At the same time, the network awareness health of the neighboring devices is obtained and converted into a vulnerability index. The vulnerability index is calculated by subtracting the network awareness health of the neighboring devices. The maximum dynamic weight value, the vulnerability index, and a preset basic constant are linearly combined, and the coefficients of the linear combination are three preset learnable logistic regression coefficients. The result of the linear combination is input into a logic function, which maps the result to a range of zero to one. The output value is the conditional probability that a failure propagates from the current device to a neighboring device along the edge.

[0012] In a preferred embodiment, the combination of probability and propagation rules is used to perform multiple Monte Carlo simulations, and the overall performance loss caused by each simulation is recorded and output. The specific process is as follows: First, the steps for performing a single Monte Carlo simulation are as follows: based on the probability distribution of the initial failure point, randomly select a device as the initial failure device for this simulation, and set the operating performance of the initial failure device to a decay level randomly generated within a preset range; Then, the cascading propagation phase begins. Starting from the set of currently failed devices, the failure is determined by random sampling based on the conditional propagation probability set on each outgoing edge from the device in the dynamic weighted multigraph. Performance decay is also set for the devices that successfully propagate. The iterative process continues until no new devices fail, resulting in a set containing the performance states of all devices after the simulation ends. Secondly, the overall performance loss value of a single simulation is calculated. The specific process is as follows: based on a preset performance calculation model, the overall performance index under the baseline state before the simulation is calculated, and the actual overall performance index calculated based on the set of equipment performance states after the simulation is completed. The overall performance loss is calculated by subtracting the ratio of the actual overall performance index to the benchmark overall performance index. Finally, the Monte Carlo simulation is executed independently and repeatedly a preset number of times. Each simulation records the identifier of the initially failed device and the calculated overall performance loss value, forming a record set containing the results of multiple simulations, which is then output.

[0013] In a preferred embodiment, the key ranking module calculates the influence entropy of the device based on the probability distribution, and the specific calculation process is as follows: First, for each device, extract all the overall performance loss values ​​corresponding to the device as the initial failure device identifier from the record set to form a performance loss sample set for the device. Secondly, based on the performance loss sample set, the probability distribution of the overall performance loss value caused by the equipment as the initial failure point is estimated by statistical methods. Next, the probability distribution is discretized, dividing the domain of the overall efficiency loss value from zero to one into multiple continuous, non-overlapping numerical intervals; based on the probability distribution, the probability value of the overall efficiency loss value falling within each numerical interval is calculated. Then, the sum of squares of the probability values ​​corresponding to all numerical intervals is calculated to obtain a purity index that reflects the degree of concentration of the probability distribution; Finally, the logarithm of the purity index with base 2 is calculated, and the negative value of the logarithm is taken. The result is the influence entropy of the equipment.

[0014] In a preferred embodiment, the specific process for calculating the comprehensive criticality index of the equipment by combining average performance loss and influence entropy is as follows: First, the average performance loss caused by the device is calculated based on the probability distribution. The average performance loss is the expected value of the overall performance loss value corresponding to the probability distribution. Next, the average performance loss of all devices is normalized. The normalization process is as follows: calculate the difference between the maximum and minimum average performance loss of all devices, then subtract the minimum value from the average performance loss of each device, and then divide by the difference, so that the maximum average performance loss is normalized to the value of one and the minimum is normalized to the value of zero, thus obtaining the normalized average performance loss of each device. At the same time, the influence entropy of all devices is normalized in the same way to obtain the normalized influence entropy of each device; Then, select a preset adjustable risk preference coefficient between zero and one. The risk preference coefficient is used to weigh the relative importance of average efficiency loss and impact entropy in the final evaluation. Next, using the normalized average efficiency loss of the equipment as the base and the risk preference coefficient as the exponent, we calculate its power value; using the normalized influence entropy of the equipment as the base and the one minus the risk preference coefficient as the exponent, we calculate its power value. Finally, the two calculated power values ​​are multiplied together, and the product is the overall criticality index of the equipment. After calculating the overall criticality index of all equipment, all equipment are sorted in descending order of the overall criticality index value.

[0015] The beneficial effects of this invention are as follows: This scheme breaks through the limitations of isolated analysis of equipment in traditional assessments. By dynamically constructing a real-time dependency network between equipment, it integrates the individual states into the overall operational context for fusion perception. It can accurately quantify the chain reactions and potential impacts that may be triggered by the performance degradation of a single point of equipment in the entire collaborative network. It not only achieves context-aware assessment of equipment health but also proactively simulates failure propagation paths to reveal hidden cascading risks. Finally, by integrating average loss and outcome uncertainty, it generates a critical ranking that reflects the true risk level of the equipment, thereby providing a direct and reliable decision-making basis for predictive maintenance and precise resource allocation from a global perspective, effectively improving the overall operational resilience and security level of complex equipment clusters. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] Example 1 This embodiment provides, for example Figure 1-2 The system illustrates an intelligent equipment evaluation system based on the fusion of operational and status parameters, specifically comprising: a graph construction module, a graph neural network evaluation module, a cascaded failure simulation module, and a criticality ranking module connected in sequence; wherein; Graph construction module: Receives the operating parameters, status parameters and association data describing the topological and logical relationships between devices. Based on the association data, it abstracts each device into nodes and the association relationship into edges to construct a dependency graph. Based on the real-time interaction data in the operating parameters, it dynamically calculates and updates the weight of each edge in the dependency graph, and generates and outputs a dynamic weighted multigraph. The graph neural network evaluation module obtains the temporal feature vector and dynamic weighted multigraph extracted from each device based on its operating and state parameters. The temporal feature vector is used as the initial feature of the corresponding node, and the dynamic weight of each edge in the dynamic weighted multigraph is used as the edge feature. The attention message passing mechanism that integrates edge weights is used to iteratively aggregate and update the node features, and outputs the network awareness health of each device. Cascaded Failure Simulation Module: Acquires a dynamic weighted multigraph and the network-aware health of each device, determines the probability of each device as the initial failure point based on the network-aware health, sets the failure propagation rules on the edges based on the dynamic weights of each edge in the dynamic weighted multigraph, performs multiple Monte Carlo simulations in combination with the probability and propagation rules, records and outputs the overall performance loss value caused by each simulation. Criticality ranking module: Based on the overall performance loss value, for each device, the probability distribution of the overall performance loss value caused by the device as the initial failure point is statistically analyzed. The influence entropy of the device is calculated based on the probability distribution, and the average performance loss caused by the device is calculated based on the probability distribution. The comprehensive criticality index of the device is calculated by combining the average performance loss and the influence entropy. All devices are ranked according to the comprehensive criticality index.

[0021] In this embodiment, it is specifically necessary to explain that in the graph construction module, the weights of each edge in the dependency graph are dynamically calculated and updated based on the real-time interaction data in the running parameters. The specific process is as follows: For each edge in the dependency graph that represents a relationship, real-time interaction data between the corresponding two devices is received. The deviation of the current value of real-time interactive data from a long-term benchmark value obtained based on historical data is calculated. The long-term benchmark value includes the historical mean and historical standard deviation. The deviation is calculated by subtracting the historical mean from the current value and then dividing by the sum of the historical standard deviation and a small constant. The historical mean and historical standard deviation are calculated based on the real-time interactive data sequence of all sampling times between the corresponding device pairs within a preset historical time window. The small constant is an extremely small positive number set to prevent the denominator from being zero, with a typical value of 0.0001. The process of calculating the deviation is essentially to standardize the current real-time interactive data value based on historical statistics to obtain a standardized score that represents the degree to which the current value deviates from the historical normal fluctuation range. Simultaneously, the instantaneous rate of change of the current value of the real-time interactive data relative to the value of the previous moment is calculated, and the instantaneous rate of change is compressed and mapped using the hyperbolic tangent function. Then, it is multiplied by an adjustable gain coefficient to obtain an adjustment amount that reflects the instantaneous change trend. The instantaneous rate of change is obtained by subtracting the value of the previous data acquisition cycle from the current value. The hyperbolic tangent function maps the instantaneous rate of change of any size to the interval between -1 and +1, achieving nonlinear compression and preventing drastic oscillations in weights due to data mutations. The adjustable gain coefficient is used to control the intensity of the influence of the instantaneous change trend on the final dynamic weight, and its typical value range is between 0.1 and 1. Based on the sum of the aforementioned deviation degree and trend adjustment amount, the dynamic weight of this edge is calculated, and the weight attribute of the corresponding edge in the dependency graph is updated using this weight value, thereby generating and outputting a dynamic weighted multigraph. The specific process of calculating the dynamic weight is as follows: the standardized score representing the historical deviation degree is directly added to the adjustment amount reflecting the instantaneous trend change to obtain a comprehensive, dimensionless dynamic weight value. The sign and magnitude of this dynamic weight value comprehensively reflect the inter-device correlation represented by the edge, whether its strength is higher or lower than the historical normal, and whether its trend is strengthening or weakening. The update process is to replace the original weight value stored in the dynamic weighted multigraph data structure with the newly calculated dynamic weight value. The specific process of abstracting each device as a node and the relationships as edges based on the associated data to construct a dependency graph is as follows: From the received associated data, various relationships between devices are parsed out. These relationships include static topology relationships defined by physical connections and logical dependencies defined by control logic or process sequence. Static topology relationships refer to the actual, fixed physical connections between devices, such as cable connections, pipe connections, or mechanical coupling. Logical dependencies refer to the functional dependency sequence defined by software control programs, production processes, or business rules, such as the order of sending and receiving control signals and the sequential connection of production steps. For any two devices that have at least one relationship, they are mapped to two nodes in the dependency graph; For each type of relationship between two nodes, a corresponding independent directed edge is created in the graph, distinguished by different edge type identifiers. This constructs a multigraph structure that allows multiple edges of different types between two nodes; this structure is the topological skeleton of the dynamically weighted multigraph. The direction of the created directed edge represents the direction of dependency or influence, such as from a signal transmitting device to a receiving device, or from an upstream process device to a downstream process device. The edge type identifier is a code used to uniquely distinguish different types of relationships; for example, a number one identifies an electrical connection relationship, and a number two identifies a control signal dependency relationship. The multigraph structure is the foundation of the dynamically weighted multigraph. It is stored in memory in the form of an adjacency list or adjacency matrix, recording information about all nodes, edges, and edge types. Dynamic weights are attached as attribute values ​​to each edge. Dynamic weight calculation and updating are performed independently and in parallel for each edge with an edge type identifier in this dynamic weighted multigraph. Independent parallel execution means that the system starts an independent weight calculation thread or task for each edge in the dynamic weighted multigraph. Each task obtains the corresponding real-time interactive data stream according to its edge type identifier and independently completes the entire process from calculating the degree of deviation, calculating the trend adjustment amount to summing to obtain the dynamic weight. Finally, the dynamic weight values ​​calculated for all edges are synchronously updated to the dynamic weighted multigraph data structure. This method greatly improves the efficiency of weight updating for massive relationships in large-scale systems.

[0022] In this embodiment, it is specifically necessary to explain the process of iteratively aggregating and updating node features using the attention message passing mechanism that incorporates edge weights in the graph neural network evaluation module: First, for each directed edge pointing to the target node in the dynamic weighted multigraph, attention coefficients for edge weight enhancement are calculated. This calculation process concatenates the feature vector held by the target node after the previous iteration, the feature vectors held by the neighboring nodes connected to the target node through this edge after the previous iteration, and the dynamic weight vector carried by this edge in the dynamic weighted multigraph, forming a comprehensive feature vector. Concatenation refers to connecting the three vectors sequentially in the dimensional direction to generate a longer-dimensional vector. The dynamic weight vector comes directly from the weight attribute values ​​stored for each edge in the dynamic weighted multigraph data structure, which are updated by the graph construction module before each graph neural network evaluation. In the first iteration, the "feature vectors held by the target node and its neighboring nodes after the previous iteration" are the temporal feature vectors extracted from the fusion of operating and state parameters of each device. Next, the comprehensive feature vector is multiplied by a pre-defined learnable attention weight vector, and then processed by a linear rectified activation function with a negative slope to obtain an original attention score. The original attention score reflects both the importance of the neighboring node states and the real-time strength of the connecting edges. The dot product operation is the calculation process of multiplying the corresponding elements of the comprehensive feature vector and the attention weight vector and then summing them. The linear rectified activation function with a negative slope outputs the original value when the input value is greater than zero, and outputs the input value multiplied by a small negative slope when the input value is less than or equal to zero. This negative slope is usually preset to a fixed decimal between 0.01 and 0.2. This function allows the calculated original attention score to be negative, thereby allowing the model to learn inhibitory attention relationships. Secondly, normalized attention and neighbor information aggregation are performed. For the same target node, the original attention scores corresponding to all its incoming edges are input into the exponential normalization function for processing, so that all scores are converted into normalized attention weights that sum to one. The exponential normalization function, namely the Softmax function, is calculated as follows: for each original attention score, the natural constant e is raised to the power of the score to obtain the exponential value; then the exponential value is divided by the sum of the exponential values ​​corresponding to all original attention scores, thus obtaining normalized attention weights that are between zero and one and have a sum of one. Then, the feature vector held by each neighbor node after the previous iteration is linearly transformed through a learnable shared weight matrix to obtain the transformed feature vector of that neighbor node. The learnable shared weight matrix is ​​a parameter matrix optimized by the gradient descent algorithm during model training. The linear transformation refers to matrix multiplication of the neighbor node's feature vector with this weight matrix. The transformed feature vector of each neighbor node is multiplied by its corresponding normalized attention weight to obtain a weighted feature vector. All such weighted feature vectors are summed and then passed through a nonlinear activation function to obtain a new feature vector of the target node after aggregating neighborhood information, which is used for the next iteration. The summation is an element-wise addition of vectors. The nonlinear activation function can be a linear rectified unit function, an exponential linear unit function, or a hyperbolic tangent function. Its role is to introduce nonlinear transformation and enhance the expressive power of the model. Finally, the message passing steps are repeated multiple times to complete the iterative update of node features. The number of rounds repeated is a preset hyperparameter, usually between two and four layers. The computation of each layer shares the same logic but uses an independent set of learnable parameters, namely an independent attention weight vector and a shared weight matrix. Through multi-layer iteration, each node can aggregate indirect information from multi-hop neighbors. The specific process for outputting the network-aware health status of each device is as follows: After multiple rounds of iterative message passing, the feature vector representation of each device node in the final round is obtained; The final feature vector representation of each node is input into a final fully connected computation layer for linear transformation. The fully connected computation layer contains learnable weight vectors and bias terms. The linear transformation process of the fully connected computation layer is as follows: the final feature vector representation is multiplied by the learnable weight vector, and then the learnable bias term scalar is added to obtain a scalar value. The sigmoid function is applied to the result of the linear transformation to map the output value to a numerical range of zero to one, thereby obtaining the network-aware health of the corresponding device. The sigmoid function is calculated as follows: the scalar value obtained by the linear transformation is used as input, and the negative power of one divided by one plus the natural constant e is calculated. The output value of the function is strictly limited to the open interval between zero and one. Network-aware health is a scalar value between zero and one. The closer the value is to one, the better the overall health status of the device in the network context. The closer the value is to zero, the worse the health status. Network health awareness integrates the device's own operating status characteristics, the device's location information in the network topology, and the strength of real-time dynamic dependencies between devices.

[0023] In this embodiment, it is specifically necessary to explain that in the cascading failure simulation module, the probability of each device being the initial failure point is determined based on the network's perceived health status. This is specifically calculated using a nonlinear probability mapping model, and the calculation process is as follows: For each device, the device's network-aware health score is input into an exponential transformation function. The input to the transformation function is the product of the negative value of the network-aware health score and a preset positive sensitivity coefficient. Then, the natural constant is calculated, with the product as the exponent, to obtain an intermediate value. The preset positive sensitivity coefficient is a real number greater than zero, used to adjust the sensitivity of the probability distribution to differences in health scores. Its typical value ranges from five to twenty, with a larger value indicating a stronger ability of the model to focus on unhealthy devices. The natural constant is the mathematical constant e, with a value of approximately 2.71828. The transformation function is an exponential function, which amplifies the input non-linearly, significantly amplifying the intermediate value corresponding to devices with lower network-aware health scores and significantly suppressing the intermediate value corresponding to devices with higher health scores. The intermediate values ​​of all devices are summed, and the sum is added to a preset minimal positive smoothing constant to obtain the denominator. The preset minimal positive smoothing constant is a positive decimal much smaller than one, typically 0.00000001. Its main function is to prevent calculation errors that would result in a zero denominator in extreme cases (such as when the sum of the intermediate values ​​of all devices is zero) and to ensure a slight smoothness of the probability distribution. Finally, the median value of a single device is divided by the denominator value to obtain the probability that the device is selected as the initial failure point in a single Monte Carlo simulation. The division calculation normalizes the probability, ensuring that the sum of the failure probabilities of all devices is theoretically "one". This calculation process is essentially a variant of the Softmax probability distribution, which takes the "unhealthiness" of the device (the negative value of the network-aware health) as input and transforms it into a probability distribution in which all devices compete to become the initial failure point through an exponential function and normalization. The probability value is between zero and one, and the sum of the probabilities of all devices is one. The lower the network awareness health value, the higher the calculated initial failure probability. The sensitivity coefficient controls the sensitivity of the network awareness health change to the probability. Based on the dynamic weights of each edge in a dynamically weighted multigraph, the rules for the propagation of failures on the edges are set. The specific process is as follows: Define a conditional propagation probability model to calculate the probability that when a device fails, the failure effect will propagate along the dependency edges to the neighboring devices of that device. The conditional propagation probability model is essentially a probability function with the connection strength between devices and the vulnerability of the target device as independent variables. For all directed edges from a failed device to any of its neighboring devices in a dynamically weighted multigraph, the edge with the largest dynamic weight value is first identified, and the largest dynamic weight value is used as one of the key inputs. The operation of "finding the largest dynamic weight value" is based on the consideration that there may be multiple types of dependencies between devices. It selects the strength of the most important and strongest dependency channel at the current moment as a representative indicator of the propagation risk. Simultaneously, the network awareness health of neighboring devices is obtained and transformed into a vulnerability index. The vulnerability index is calculated by subtracting the network awareness health of neighboring devices from one. The vulnerability index is a value between zero and one. When the neighboring device is completely healthy (network awareness health is one), the vulnerability is zero; when the neighboring device is extremely unhealthy (network awareness health is close to zero), the vulnerability is close to one. This index quantifies the ability of neighboring devices to resist failure impacts. The maximum dynamic weight value, vulnerability index, and a preset basic constant are linearly combined. The coefficients of the linear combination are three preset learnable or configurable logistic regression coefficients. The preset basic constant represents the inherent tendency of failure propagation and is independent of specific devices and connections. The three logistic regression coefficients need to be learned through training on historical failure data or set by domain experts based on experience. The linear combination is an operation of multiplying the coefficients by their corresponding variables and then adding them together. The result of a linear combination is input into a logic function, which maps the result to the interval between zero and one. The output value is the conditional probability that a failure propagates from the current device to neighboring devices along the edge. The logic function, namely the Sigmoid function, is calculated by taking the linear combination result as input and calculating the negative power of one divided by one plus the natural constant e. This function smoothly maps any real number input to the interval between zero and one, making it very suitable as a probability output. The conditional probability value depends on the dynamic strength of the connection between devices and the health status of the neighboring devices themselves. The greater the dynamic weight or the less healthy the neighboring devices are, the higher the conditional probability value. By combining probability and propagation rules, multiple Monte Carlo simulations are performed, and the overall performance loss caused by each simulation is recorded and output. The specific process is as follows: First, the steps for performing a single Monte Carlo simulation are as follows: Based on the probability distribution of the initial failure point, a device is randomly selected as the initial failure device for this simulation, and the operating performance of the initial failure device is set to a decay level randomly generated within a preset range; "random selection" is achieved through a roulette wheel algorithm: a uniformly random number between zero and one is generated, and based on the cumulative distribution of the initial failure probability of each device, the device whose probability interval the random number falls into is selected; the "preset range" is usually 0.3 to 0.7, indicating that the initial failure will cause the device performance to decrease by 30% to 70%, and the specific value is uniformly and randomly selected within this range to simulate different degrees of initial failure; Then, the cascading propagation phase begins. Starting from the set of currently failed or degraded devices, based on the conditional propagation probability set on each outgoing edge from the device in the dynamic weighted multigraph, a random sampling method is used to determine whether the failure propagates to the corresponding neighboring device. For devices that successfully propagate, a similar performance degradation is applied. This iterative process continues until no new devices fail or the preset propagation round limit is reached, resulting in a set containing the performance states of all devices after the simulation ends. "Random sampling" means that for each outgoing edge, a uniformly random number between zero and one is generated. If this random number is less than the conditional propagation probability set on the edge, the propagation is considered successful. The "preset propagation round limit" is a safety threshold to prevent infinite loops in the simulation, typically set to ten to twenty rounds. After successful propagation, the performance degradation level of neighboring devices can be set to the same as the source device, or a certain percentage of random degradation can be applied to it. Secondly, the overall performance loss value of a single simulation is calculated. The specific process is as follows: Based on a preset performance calculation model, the overall performance index under the baseline state before the simulation is calculated, and the actual overall performance index calculated based on the set of equipment performance states after the simulation is completed. The preset performance calculation model is a function used to calculate a scalar index characterizing the overall output, efficiency, or stability of the system based on the real-time performance state of all equipment in the system. For example, on a production line, this model may be a weighted product of the performance of each process equipment, or the performance value of the slowest equipment (bottleneck). The overall performance index under the baseline state refers to the value calculated by this model when all equipment is running healthily (performance is one) before the simulation. The overall performance loss value is calculated by subtracting the ratio of the actual overall performance index to the baseline overall performance index from one, resulting in a scalar value between zero and one, representing the proportion of performance loss suffered due to the cascading failure event in this simulation; this calculation reflects the relative degree of system performance degradation; a ratio of one indicates no performance loss, and zero indicates complete paralysis; the overall performance loss value intuitively quantifies the impact intensity of a single cascading failure event on system function; Finally, the Monte Carlo simulation is independently repeated a preset number of times. Each simulation records the identifier of the initial failed device and the calculated overall performance loss value, forming a record set containing the results of multiple simulations, which is then output. The preset number of simulations is usually several thousand to tens of thousands to ensure the stability of the statistical results. The record set is a data list, in which each record is associated with an initial failed device identifier and a corresponding overall performance loss value. This set is the original data basis for subsequent key ranking analysis.

[0024] In this embodiment, it is specifically necessary to explain the key ranking module, which calculates the influence entropy of the device based on the probability distribution. The specific calculation process is as follows: First, for each device, extract all the overall performance loss values ​​corresponding to the device as the initial failure device identifier from the record set to form the device's performance loss sample set; the record set is the output from the cascaded failure simulation module, and each record is clearly associated with an initial failure device identifier and an overall performance loss value; the performance loss sample set is a list of values ​​that contains all the system performance loss results corresponding to the simulations with that device as the initial failure point; Secondly, based on the performance loss sample set, the probability distribution of the overall performance loss value caused by the equipment as the initial failure point is estimated by statistical methods. The statistical method can be a parametric estimation method or a non-parametric kernel density estimation method. A preferred implementation is to use Gaussian kernel density estimation. The calculation process is as follows: take each sample point in the performance loss sample set as the center, place a Gaussian-shaped probability density function, and then superimpose all such Gaussian functions and take the average to obtain a smooth and continuous probability density function curve of the overall performance loss value. This curve is the estimate of the probability distribution. Next, the probability distribution is discretized, dividing the domain of the overall performance loss value from zero to one into multiple continuous, non-overlapping numerical intervals. Discretization can employ equal-width binning or equal-depth binning; a more adaptable method is equal-probability binning, calculated as follows: based on the estimated probability distribution, a series of quantiles are found such that the probability of the overall performance loss value falling between any two adjacent quantiles is equal, thus dividing the domain of zero to one into multiple intervals with equal probability mass; the number of intervals is a preset value, typically ten; based on the probability distribution, the probability value of the overall performance loss value falling within each numerical interval is calculated; the probability value is obtained by integrating the probability density function over the corresponding numerical interval, approximately reflecting the likelihood of the failure loss result falling into that interval. Then, the sum of squares of the probability values ​​corresponding to all numerical intervals is calculated to obtain a purity index that reflects the degree of concentration of the probability distribution. The sum of squares is calculated by multiplying the probability value of each interval by itself and then adding the results of all intervals. The closer the value of this purity index is to one, the more concentrated the probability distribution is in a few intervals, and the more certain the result is. The closer the value is to the reciprocal of the number of intervals, the more uniform the distribution is, and the less certain the result is. Finally, the logarithm of the purity index to base 2 is calculated, and the negative value of the logarithm is taken. The result is the influence entropy of the device. The operation of calculating the logarithm is to find the exponent that makes the logarithm of 2 equal to the purity index. The operation of taking the negative value makes the influence entropy zero when the purity index is one (completely certain), and the influence entropy maximum when the purity index is minimum (most uncertain). In information theory, this calculation corresponds to the calculation of Rényi entropy (order 2), which quantifies the average degree of "unexpectedness" or information contained when a loss result is randomly drawn from the probability distribution. Impact entropy is a non-negative real number. The larger the impact entropy value, the more dispersed and unpredictable the overall performance loss caused by the device as the initial failure point, that is, the higher the uncertainty of the failure consequences. High impact entropy means that the failure of the device may lead to any level of loss from minor to severe, and the consequences are difficult to predict. This unpredictability itself is a high-level risk dimension that needs to be paid attention to. The comprehensive criticality index of this equipment is calculated by combining average performance loss and impact entropy. The specific process is as follows: First, the average performance loss caused by the device is calculated based on the probability distribution. The average performance loss is the expected value of the overall performance loss corresponding to the probability distribution. The expected value is calculated by multiplying the representative value of each numerical interval (such as the midpoint of the interval) by the probability value of that interval, and then summing the products of all intervals. It represents the average severity caused by the failure of the device. Next, the average performance loss of all equipment is normalized. The normalization process is as follows: calculate the difference between the maximum and minimum average performance loss of all equipment, then subtract the minimum value from the average performance loss of each equipment, and divide by the difference, so that the maximum average performance loss is normalized to a value of one, and the minimum is normalized to a value of zero, thus obtaining the normalized average performance loss of each equipment. This normalization method is called range normalization. In extreme cases (when the average loss of all equipment is equal), the difference may be zero. In this case, the normalized average performance loss of all equipment can be defined as 0.5, or other smoothing methods can be used. Normalization makes the average losses of different equipment comparable on a uniform scale from zero to one. At the same time, the influence entropy of all devices is normalized in the same way to obtain the normalized influence entropy of each device; the normalized influence entropy is also between zero and one, reflecting the relative level of uncertainty of that device among all devices; Then, select a preset adjustable risk preference coefficient between zero and one. The risk preference coefficient is used to weigh the relative importance of average performance loss and impact entropy in the final evaluation. A typical value of the risk preference coefficient is 0.7, which means that the decision-maker values ​​the average loss more. If it is set to 0.5, it means that the two are equally important. If it is set to 0.3, it means that the uncertainty of the consequences is more important. This coefficient allows the system to be flexibly configured according to different operation and maintenance strategies (such as "avoiding black swan events" or "controlling average downtime"). Next, using the normalized average performance loss of the equipment as the base and the risk preference coefficient as the exponent, its power value is calculated; using the normalized influence entropy of the equipment as the base and the risk preference coefficient minus one as the exponent, its power value is calculated; the operation of calculating the power value is to raise the base to the power of the exponent; when the exponent is less than one, this operation has a "compression" effect on the base, and the smaller the value, the stronger the compression; this means that in the comprehensive evaluation, the poorer the dimension (smaller value), the greater the negative impact. Finally, the two calculated power values ​​are multiplied together, and the product is the overall criticality index of the equipment. After calculating the overall criticality index of all equipment, all equipment is sorted in descending order of the overall criticality index value. The multiplication operation constitutes the weighted geometric mean. Compared with the weighted arithmetic mean, the weighted geometric mean requires both normalized indicators to be high in order to obtain a high overall criticality index. If one indicator is very low, it will significantly reduce the product, which can more effectively identify those "high-risk management priority" equipment that have both high average loss and high uncertainty, while filtering out equipment with only one high indicator. The sorting result directly guides the priority allocation order of maintenance resources.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A device intelligent evaluation system based on the fusion of operating and status parameters, characterized in that, Specifically, it includes: The graph construction module, graph neural network evaluation module, cascaded failure simulation module, and key ranking module are connected in sequence, among which; Graph construction module: Receives the operating parameters, status parameters and association data describing the topological and logical relationships between devices. Based on the association data, it abstracts each device into nodes and the association relationship into edges to construct a dependency graph. Based on the real-time interaction data in the operating parameters, it dynamically calculates and updates the weight of each edge in the dependency graph, and generates and outputs a dynamic weighted multigraph. The graph neural network evaluation module obtains the temporal feature vector and dynamic weighted multigraph extracted from each device based on its operating and state parameters. The temporal feature vector is used as the initial feature of the corresponding node, and the dynamic weight of each edge in the dynamic weighted multigraph is used as the edge feature. The attention message passing mechanism that integrates edge weights is used to iteratively aggregate and update the node features, and outputs the network awareness health of each device. Cascaded Failure Simulation Module: Acquires a dynamic weighted multigraph and the network-aware health of each device, determines the probability of each device as the initial failure point based on the network-aware health, sets the failure propagation rules on the edges based on the dynamic weights of each edge in the dynamic weighted multigraph, performs multiple Monte Carlo simulations in combination with the probability and propagation rules, records and outputs the overall performance loss value caused by each simulation. Criticality ranking module: Based on the overall performance loss value, for each device, the probability distribution of the overall performance loss value caused by the device as the initial failure point is statistically analyzed. The influence entropy of the device is calculated based on the probability distribution, and the average performance loss caused by the device is calculated based on the probability distribution. The comprehensive criticality index of the device is calculated by combining the average performance loss and the influence entropy. All devices are ranked according to the comprehensive criticality index.

2. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 1, characterized in that: In the graph construction module, the weights of each edge in the dependency graph are dynamically calculated and updated based on real-time interaction data in the running parameters. The specific process is as follows: For each edge in the dependency graph that represents a relationship, real-time interaction data between the corresponding two devices is received. The deviation between the current value of real-time interactive data and the long-term benchmark value obtained based on historical data is calculated. The long-term benchmark value includes the historical mean and the historical standard deviation. The deviation is calculated by subtracting the historical mean from the current value and then dividing by the sum of the historical standard deviation and a small constant. At the same time, the instantaneous rate of change of the current value of the real-time interactive data relative to the value of the previous moment is calculated, and the instantaneous rate of change is compressed and mapped using the hyperbolic tangent function, and then multiplied by an adjustable gain coefficient to obtain an adjustment amount that reflects the instantaneous change trend. Based on the sum of the aforementioned deviation and trend adjustment, the dynamic weight of this edge is calculated, and the weight attribute of the corresponding edge in the dependency graph is updated using this weight value, thereby generating and outputting a dynamically weighted multigraph.

3. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 2, characterized in that: The specific process of abstracting each device as a node and the association relationship as an edge based on the associated data to construct a dependency graph is as follows: From the received associated data, various relationships between devices are parsed out, including static topology relationships defined by physical connections and logical dependencies defined by process sequence. For any two devices that have a relationship, they are mapped to two nodes in the dependency graph; For each relationship between the two nodes, an independent directed edge is created in the graph to correspond to it, and different edge type identifiers are used to distinguish them, thereby constructing a multigraph structure that allows multiple edges of different types between two nodes. This structure is the topological skeleton of the dynamic weighted multigraph. The dynamic calculation and updating of weights are performed independently and in parallel for each edge with an edge type identifier in this dynamically weighted multigraph.

4. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 3, characterized in that: In the graph neural network evaluation module, the process of iteratively aggregating and updating node features using an attention message passing mechanism that incorporates edge weights is as follows: First, for each directed edge pointing to the target node in the dynamic weighted multigraph, the attention coefficient of edge weight enhancement is calculated. The calculation process concatenates the feature vector held by the target node after the previous iteration, the feature vector held by the neighboring nodes connected to the target node through this edge after the previous iteration, and the dynamic weight vector carried by this edge in the dynamic weighted multigraph to form a comprehensive feature vector. Next, the comprehensive feature vector is multiplied by a pre-defined learnable attention weight vector, and then processed by a linear rectified activation function with a negative slope to obtain an original attention score. The original attention score reflects both the importance of the neighboring node state and the real-time strength of the connection edge. Secondly, normalized attention and neighbor information aggregation are performed. For the same target node, the original attention scores corresponding to all its incoming edges are input into the exponential normalization function for processing, so that all scores are converted into normalized attention weights that sum to one. Then, the feature vector held by each neighbor node after the previous iteration is linearly transformed through a learnable shared weight matrix to obtain the transformed feature vector of the neighbor node; the transformed feature vector of each neighbor node is multiplied by its corresponding normalized attention weight to obtain a weighted feature vector; all such weighted feature vectors are summed and then passed through a non-linear activation function to obtain a new feature vector of the target node after aggregating neighborhood information for the next iteration. Finally, the message passing steps are repeated multiple times to complete the iterative update of node features.

5. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 4, characterized in that: The specific process of outputting the network-aware health status of each device is as follows: After multiple rounds of iterative message passing, the feature vector representation of each device node in the final round is obtained; The final feature vector representation of each node is input into a final fully connected computation layer for linear transformation. The fully connected computation layer contains learnable weight vectors and bias terms. By applying the S-shaped growth function to the result of the linear transformation, the output value is mapped to a numerical range of zero to one, thereby obtaining the network-aware health of the corresponding device. Network health awareness integrates the device's own operating status characteristics, the device's location information in the network topology, and the strength of real-time dynamic dependencies between devices.

6. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 5, characterized in that: In the cascading failure simulation module, the probability of each device being the initial failure point is determined based on the network's perceived health status. This is specifically calculated using a nonlinear probability mapping model, and the calculation process is as follows: For each device, the device's network awareness health is input into an exponential conversion function. The input to the conversion function is the product of the negative number of the network awareness health and a preset positive sensitivity coefficient. Then, the natural constant is calculated and the product result is used as the power of the exponent to obtain an intermediate value. Summing up the intermediate values ​​of all devices and adding the sum to a preset minimal positive smoothing constant to obtain the denominator value; Finally, dividing the median value of a single device by the denominator gives the probability that the device will be selected as the initial failure point in a single Monte Carlo simulation.

7. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 6, characterized in that: The specific process of setting the failure propagation rules on edges based on the dynamic weights of each edge in a dynamically weighted multigraph is as follows: Define a conditional propagation probability model to calculate the probability that when a device fails, the failure effect will propagate along the dependency edges to the device's neighboring devices. For all directed edges in a dynamically weighted multigraph from a failed device to any of its neighboring devices, first find the edge with the largest dynamic weight value and use the largest dynamic weight value as one of the key inputs. At the same time, the network awareness health of the neighboring devices is obtained and converted into a vulnerability index. The vulnerability index is calculated by subtracting the network awareness health of the neighboring devices. The maximum dynamic weight value, the vulnerability index, and a preset basic constant are linearly combined, and the coefficients of the linear combination are three preset learnable logistic regression coefficients. The result of the linear combination is input into a logic function, which maps the result to a range of zero to one. The output value is the conditional probability that a failure propagates from the current device to a neighboring device along the edge.

8. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 7, characterized in that: The process of performing multiple Monte Carlo simulations by combining probability and propagation rules, recording and outputting the overall performance loss caused by each simulation, is as follows: First, the steps for performing a single Monte Carlo simulation are as follows: based on the probability distribution of the initial failure point, randomly select a device as the initial failure device for this simulation, and set the operating performance of the initial failure device to a decay level randomly generated within a preset range; Then, the cascading propagation phase begins. Starting from the set of currently failed devices, the failure is determined by random sampling based on the conditional propagation probability set on each outgoing edge from the device in the dynamic weighted multigraph. Performance decay is also set for the devices that successfully propagate. The iterative process continues until no new devices fail, resulting in a set containing the performance states of all devices after the simulation ends. Secondly, the overall performance loss value of a single simulation is calculated. The specific process is as follows: based on a preset performance calculation model, the overall performance index under the baseline state before the simulation is calculated, and the actual overall performance index calculated based on the set of equipment performance states after the simulation is completed. The overall performance loss is calculated by subtracting the ratio of the actual overall performance index to the benchmark overall performance index. Finally, the Monte Carlo simulation is executed independently and repeatedly a preset number of times. Each simulation records the identifier of the initially failed device and the calculated overall performance loss value, forming a record set containing the results of multiple simulations, which is then output.

9. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 8, characterized in that: In the key ranking module, the influence entropy of the device is calculated based on the probability distribution. The specific calculation process is as follows: First, for each device, extract all the overall performance loss values ​​corresponding to the device as the initial failure device identifier from the record set to form a performance loss sample set for the device. Secondly, based on the performance loss sample set, the probability distribution of the overall performance loss value caused by the equipment as the initial failure point is estimated by statistical methods. Next, the probability distribution is discretized, dividing the domain of the overall efficiency loss value from zero to one into multiple continuous, non-overlapping numerical intervals. Based on the probability distribution, calculate the probability that the overall efficiency loss value falls within each numerical interval; Then, the sum of squares of the probability values ​​corresponding to all numerical intervals is calculated to obtain a purity index that reflects the degree of concentration of the probability distribution; Finally, the logarithm of the purity index with base 2 is calculated, and the negative value of the logarithm is taken. The result is the influence entropy of the equipment.

10. The intelligent equipment evaluation system based on the fusion of operating and status parameters according to claim 9, characterized in that: The specific process for calculating the comprehensive criticality index of the equipment by combining average performance loss and impact entropy is as follows: First, the average performance loss caused by the device is calculated based on the probability distribution. The average performance loss is the expected value of the overall performance loss value corresponding to the probability distribution. Next, the average performance loss of all devices is normalized. The normalization process is as follows: calculate the difference between the maximum and minimum average performance loss of all devices, then subtract the minimum value from the average performance loss of each device, and then divide by the difference, so that the maximum average performance loss is normalized to the value of one and the minimum is normalized to the value of zero, thus obtaining the normalized average performance loss of each device. At the same time, the influence entropy of all devices is normalized in the same way to obtain the normalized influence entropy of each device; Then, select a preset adjustable risk preference coefficient between zero and one. The risk preference coefficient is used to weigh the relative importance of average efficiency loss and impact entropy in the final evaluation. Next, using the normalized average efficiency loss of the equipment as the base and the risk preference coefficient as the exponent, its power value is calculated. Using the normalized impact entropy of the equipment as the base, and the risk preference coefficient minus one as the exponent, calculate its power value. Finally, the two calculated power values ​​are multiplied together, and the product is the overall criticality index of the equipment. After calculating the overall criticality index of all equipment, all equipment are sorted in descending order of the overall criticality index value.

Images