A knowledge graph-based intelligent fault diagnosis system for multi-stage catalytic coupling in hazardous waste treatment.

The knowledge graph-based intelligent fault diagnosis system for multi-stage catalytic coupling hazardous waste treatment achieves deep fusion of multi-physics field data and cross-stage time delay propagation modeling, solving the problems of early identification and root cause localization of faults in multi-stage catalytic coupling equipment, and providing accurate fault diagnosis and safe life prediction.

CN123115693APending Publication Date: 2026-11-17DONGYANG NAHAI ENVIRONMENTAL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611522917.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are unable to effectively handle the correlation anomalies between multiple physical fields in multi-stage catalytic coupling hazardous waste treatment equipment, and cannot explicitly model cross-stage causal propagation, resulting in early failure to report faults and difficulty in accurately identifying the propagation path and root cause of faults. They cannot meet the requirements for accurate fault diagnosis and safe life prediction under complex operating conditions.

Method used

The knowledge graph-based multi-level catalytic coupling hazardous waste treatment fault intelligent diagnosis system collects data in real time through a multi-source physical field decoupling module, decouples and calculates comprehensive thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics, uses a graph representation activation module to perform feature fusion, a cross-level time delay inference module calculates the topological pointing relationship between nodes, and a root cause prediction and diagnosis module locates the root cause fault and calculates the remaining safe operating life.

Benefits of technology

It improves the robustness and accuracy of early fault identification, accurately depicts the cross-stage propagation process of faults in multi-stage catalytic equipment, reduces the risk of root cause misjudgment, and provides continuous and stable lifetime prediction, meeting the requirements of industrial safety interlocking and operation monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115693A_ABST
    Figure CN123115693A_ABST
Patent Text Reader

Abstract

This invention relates to the field of fault diagnosis technology for hazardous waste treatment equipment, and discloses a multi-level catalytic coupling hazardous waste treatment fault intelligent diagnosis system based on knowledge graphs. The system includes a multi-source physical field decoupling module, a graph characterization activation module, a cross-level time-delay inference module, and a root cause prediction and diagnosis module. The multi-source physical field decoupling module collects temperature, pressure difference, pressure, and flow rate data, calculates comprehensive thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics, and introduces a flow rate lower limit protection to avoid low-flow division-to-zero singularities. The graph characterization activation module fuses the two types of features into a probability value of the activation intensity of the characterization node. The cross-level time-delay inference module weights and aggregates the probabilities of the predecessor parent nodes along the directed edges of the directed acyclic knowledge graph, combining inter-level aerodynamic time delays, and outputs the fault probability. The root cause prediction and diagnosis module locates the root cause node based on the intrinsic activation intensity, calculates the remaining lifetime by combining probability, temperature, running time, and thermal penalty, and balances safety and stability through fuse-based direct connection and inertial smoothing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for hazardous waste treatment equipment, and particularly relates to a multi-level catalytic coupling hazardous waste treatment fault intelligent diagnosis system based on knowledge graphs. Background Technology

[0002] Multi-stage catalytic coupling is one of the core technologies in the field of hazardous waste treatment. It achieves the step-by-step transformation and deep purification of pollutants through a series of multi-stage catalytic beds. Due to the large fluctuations in hazardous waste composition, high operating temperature, strong corrosiveness of flue gas, and strong coupling between stages, the equipment faces risks of failure such as catalyst sintering, bed blockage, and flow deviation collapse during long-term operation. In existing technologies, diagnostic methods based on single-parameter thresholds, single physical field statistical characteristics, or shallow data models are difficult to effectively handle the correlation anomalies between multiple physical fields, and cannot explicitly model cross-stage causal propagation. This leads to easy missed detection in the early stages of failure and difficulty in accurately identifying the propagation path and root cause of the failure. Therefore, it cannot meet the urgent need for accurate fault diagnosis and safe life prediction of multi-stage catalytic coupling hazardous waste treatment equipment under complex operating conditions. Summary of the Invention

[0003] The purpose of this invention is to provide a knowledge graph-based intelligent diagnostic system for multi-level catalytic coupling hazardous waste treatment faults, aiming to solve at least one of the above-mentioned problems.

[0004] This invention is implemented as follows: a multi-level catalytic coupling hazardous waste treatment fault intelligent diagnosis system based on a knowledge graph, comprising: a multi-source physical field decoupling module, used to collect physical field data of each level of nodes in the multi-level catalytic equipment in real time, and decouple and calculate the comprehensive thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics of each level of node; a graph characterization activation module, used to fuse the comprehensive thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics, and calculate the activation intensity probability value of each characterization node in the directed acyclic knowledge graph; a cross-level time delay reasoning module, used to extract the topological pointing relationship between nodes in the knowledge graph, calculate the aerodynamic time delay between adjacent nodes, and perform weighted aggregation calculation of the time delay alignment probability of the predecessor parent node set along the directed edges of the graph, and output the occurrence probability of each target fault node; and a root cause prediction and diagnosis module, used to locate the root cause fault node according to the activation intensity probability value of each node, and calculate the remaining safe operating life based on the occurrence probability and thermodynamic state of the root cause node, combined with the equipment's operating time.

[0005] A further technical solution involves the multi-source physical field decoupling module where: the comprehensive thermodynamic anomaly feature includes quantification of the degree of spatial dispersion and bidirectional deviation of the mean temperature; the transient differential feature of aerodynamic flow resistance introduces time-domain sliding filtering for noise reduction and singularity protection; for any i-th stage catalytic node, the real-time acquired physical field data includes: the gas mass flow rate at time t, the absolute pressure difference between the inlet and outlet of the catalytic bed at that stage, the absolute static pressure of the catalytic bed at that stage, and the temperatures of N thermocouples located on the grid node at that stage; the comprehensive thermodynamic anomaly feature is a linear weighted combination of a temperature spatial dispersion penalty term and a mean temperature bidirectional deviation penalty term; wherein, the temperature spatial dispersion penalty term is used to quantify the degree of fluctuation and dispersion of each thermocouple temperature relative to the mean temperature of the current stage, and the mean temperature bidirectional deviation penalty term is used to quantify the absolute deviation of the mean temperature of the current stage from the preset design mean temperature of the catalytic bed, so as to achieve isomorphic measurement of the bidirectional deviation of overheating and low temperature.

[0006] A further technical solution is that the transient differential characteristic of the pneumatic flow resistance is characterized as the relative rate of change of the pneumatic flow resistance over time after time-domain sliding filtering and noise reduction; the pneumatic flow resistance is dynamically equivalent based on the ratio of the absolute pressure difference between the inlet and outlet of the catalytic bed to the actual gas volume flow rate, wherein the actual gas volume flow rate is jointly derived from the gas mass flow rate, absolute static pressure and space average temperature through the ideal gas state mechanism, and a small positive flow rate lower limit constraint is applied to the gas mass flow rate during the derivation process to eliminate the division-to-zero singularity calculation collapse caused by the flow rate approaching zero under low frequency or shutdown conditions of the equipment.

[0007] In a further technical solution, in the spectrum characterization activation module: the activation intensity probability value of the i-th level characterization node is calculated by a monotonic nonlinear mapping function with probability saturation characteristics, and the range of the nonlinear mapping function is strictly converged to the interval [0,1); the input of the independent variable of the mapping function is the weighted sum of the comprehensive thermodynamic anomaly characteristics after dimensionless processing and the transient differential characteristics of aerodynamic flow resistance after taking the absolute value.

[0008] A further technical solution involves taking the absolute value of the transient differential characteristics of aerodynamic flow resistance, aiming to construct a two-sided response mechanism to equally capture catalytic bed blockage faults characterized by increased flow resistance and catalytic bed collapse and flow deviation faults characterized by sudden decreases in flow resistance.

[0009] A further technical solution is provided in the cross-level time-delay inference module: for any i-th level target node and its predecessor parent node set in the directed acyclic knowledge graph, the aerodynamic time delay between the predecessor node and the target node is determined by the ratio of the effective physical volume between the two nodes to the actual gas volume flow rate of the predecessor node after smoothing by the moving average and constrained by the small positive flow rate lower limit, so as to reflect the pure delay time of gas transmission between levels; the occurrence probability of the i-th level target fault node is calculated by the probability constraint transfer function, which linearly weights and fuses the activation intensity probability value of the current level characterization node with the occurrence probability of the predecessor parent node set after alignment by aerodynamic time delay, and further performs nonlinear probability space compression; the coupling weight and directed causal transmission weight of the current level characterization node participating in the weighted fusion are strictly limited to non-negative real numbers, and satisfy the local normalization constraint condition of summing to one within the local topology structure they point to, so as to prevent probability divergence and saturation effect in the recursive inference process from the network topology level.

[0010] In a further technical solution, in the root cause prediction and diagnosis module: the system locks the node with the highest intrinsic activation intensity probability value as the root cause fault node, extracts its aggregation probability and spatial average temperature, and determines the upper limit of the remaining theoretical lifespan of the equipment based on the maximum theoretical design life and the accumulated operating time.

[0011] In a further technical solution, the root cause prediction and diagnosis module constructs an inverse analytical operator for a nonlinear mapping space, which inversely restores the aggregation probability with a limited value range to an unbounded linear degenerate expansion domain, and calculates the first-order time difference of the characteristic trajectory of the linear degenerate expansion domain to obtain the degenerate derivative of the mapping space, which characterizes the true evolution rate of the current fault.

[0012] A further technical solution is that the root cause prediction and diagnosis module calculates the remaining safe operating life through the following multi-branch safety logic: Condition 1: When the probability of aggregation is greater than or equal to the preset safe shutdown tolerance probability threshold, or the average temperature of the root cause node space is greater than or equal to the extreme damage average temperature, a physical circuit breaker mechanism is triggered, forcing the final remaining life to be zero; Condition 2: When the circuit breaker mechanism is not triggered, and the probability of aggregation is less than or equal to the preset health probability upper limit. Simultaneously, when the degradation derivative of the mapping space is less than or equal to a small positive threshold for the stability of the preset judgment derivative, the equipment is determined to be in a healthy state without deterioration trend under the constraint of a low-risk range, triggering the bottom-line mechanism, and directly outputting the upper limit of the remaining theoretical lifespan, which is discounted and corrected by the local thermal penalty factor, as the original lifespan sequence; Condition 3: If the above two conditions are not met, the remaining degradation time is calculated based on the difference between the safe shutdown tolerance probability threshold and the current aggregation probability in the linear degradation expansion domain, combined with the degradation derivative of the mapping space constrained by the conservative benchmark lower limit, and the original lifespan sequence is calculated by combining the upper limit of the remaining theoretical lifespan and the local thermal penalty factor; Finally, only the original lifespan sequence calculated under conditions 2 or 3 is subjected to a first-order inertial smoothing filter, and the final remaining safe operating lifespan is output; Among them, the local thermal penalty factor characterizes the lifespan reduction based on the exponential thermal stress accelerated degradation mechanism, and its reduction intensity increases nonlinearly with the approach of the root cause node spatial average temperature to the ultimate failure average temperature.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by integrating thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics protected against anomalies at low flow rates, can simultaneously capture overheating and low temperature anomalies, and maintain numerical stability under low flow or shutdown conditions, thereby improving the robustness and accuracy of early fault identification.

[0014] 2. This invention utilizes a directed acyclic knowledge graph to explicitly model inter-stage topological relationships and time-varying aerodynamic delays. It aligns the failure probabilities of precursor nodes according to actual transport time and then performs weighted aggregation, thereby more accurately depicting the cross-stage propagation process of failures in multi-stage catalytic devices and reducing the risk of root cause misjudgment.

[0015] 3. This invention introduces a multi-branch safety logic of physical fuse direct connection, health guarantee and inertial smoothing in the remaining life calculation, which not only ensures that zero life is output immediately when the safety threshold is reached, but also provides continuous and stable life prediction in non-fuse state. At the same time, it takes into account the running time and thermal stress accelerated degradation effect, which meets the dual requirements of industrial safety interlocking and operation monitoring. Attached Figure Description

[0016] Figure 1 This invention provides an overall architecture diagram of a knowledge graph-based intelligent diagnostic system for multi-level catalytic coupling hazardous waste treatment faults. Figure 2 The logical calculation flowchart of the cross-level time-delay inference module provided by this invention; Figure 3 The multi-branch logic flowchart of the remaining safe operating life of the root cause prediction and diagnosis module provided by the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] During the operation of multi-stage catalytic coupling hazardous waste treatment equipment, the equipment faces risks of failure such as catalyst sintering, bed blockage, and flow deviation collapse due to large fluctuations in hazardous waste composition, high operating temperatures, and strong corrosiveness of flue gas. Existing fault diagnosis methods have significant shortcomings. Alarm methods based on single-parameter thresholds are unable to reflect anomalies in the correlation between multiple physical fields, leading to early missed faults. Methods based on the statistical characteristics of a single physical field are not adaptable to low-temperature anomalies and low-flow conditions, and do not explicitly model cross-stage causal propagation. Methods based on fault trees or Bayesian networks are difficult to handle time-varying and time-delayed dynamic propagation processes. In addition, remaining lifetime prediction methods are prone to numerical distortion during the healthy phase, and the predicted output lacks effective coordination with safety interlock actions, affecting system safety and reliability.

[0019] For example, in the actual operation of a hazardous waste treatment plant, when a multi-stage catalytic converter processes hazardous waste containing high concentrations of organic matter, the first-stage catalytic bed experiences a sudden increase in heat load, leading to uneven temperature distribution and increased spatial dispersion. While physical field data is collected in real time, the existing system only makes judgments based on single-point temperature thresholds, failing to correlate with changes in aerodynamic flow resistance, and misinterpreting these as normal fluctuations. Subsequently, this anomaly propagates to the second stage via flue gas flow, causing bed blockage, abnormally increasing pressure differentials, and ultimately resulting in emissions exceeding standards. In this case, the inter-stage propagation process of the fault was not effectively modeled, and the root cause localization error prevented the targeted implementation of maintenance measures.

[0020] If the above problems are not addressed, inaccurate fault diagnosis and distorted lifespan prediction will lead to root cause bias, making targeted maintenance impossible and exacerbating equipment degradation. Inter-stage fault propagation can trigger a chain reaction, causing permanent catalyst damage or equipment structural failure, seriously threatening operational safety and environmental compliance. Therefore, there is an urgent need for an intelligent diagnostic solution that can integrate multiphysics data, accurately model cross-stage time-delay propagation, and achieve accurate root cause localization.

[0021] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0022] like Figure 1 As shown, an embodiment of the present invention provides a multi-level catalytic coupling hazardous waste treatment fault intelligent diagnosis system based on knowledge graphs, comprising:

[0023] A multi-source physical field decoupling module is used to acquire physical field data from each level of a multi-stage catalytic converter in real time, and to decouple the calculation of the comprehensive thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic resistance at each level. In one implementation, physical field data can be acquired using conventional sensors deployed at each level of the converter, such as temperature data obtained through thermocouple arrays, differential pressure data obtained through differential pressure sensors, and gas mass flow rate data obtained through flow meters. This raw data is then transmitted to a data processing unit. The decoupling calculation can use preset empirical formulas or simple linear models to perform preliminary processing on the acquired temperature, pressure, and flow rate data to generate indicators reflecting thermodynamic state and aerodynamic resistance changes, respectively. For example, the deviation of the average temperature of each catalytic bed from the design temperature can be calculated as a preliminary indicator of thermodynamic anomaly, and the rate of change of the pressure-to-flow ratio can be calculated as a preliminary indicator of aerodynamic resistance change.

[0024] Comprehensive thermodynamic anomaly characteristics are used to quantify the degree to which the thermodynamic state of each stage of a multi-stage catalytic converter deviates from normal operating conditions. This may comprehensively consider the spatial uniformity of the temperature field distribution and the deviation of the overall average temperature from the design value to fully reflect thermodynamic anomalies. Transient differential characteristics of aerodynamic flow resistance are used to describe the rate of change of aerodynamic flow resistance over time at each stage of a multi-stage catalytic converter. By analyzing the instantaneous changes in flow resistance, dynamic aerodynamic anomalies such as blockage, collapse, or flow deviation can be detected.

[0025] The graph representation activation module is used to fuse comprehensive thermodynamic anomaly features with transient differential features of aerodynamic resistance, calculating the activation intensity probability value of each representation node in the directed acyclic knowledge graph. Specifically, this module can employ a simple fusion strategy, such as performing an arithmetic average or logical OR operation on the comprehensive thermodynamic anomaly features and the transient differential features of aerodynamic resistance. For example, when any feature value exceeds a preset single threshold, the node is considered to have an anomaly, and its activation intensity probability value is set to a fixed high value. Alternatively, a simple linear function can be used to map the fused feature values ​​to a range of 0 to 1, serving as the activation intensity probability value of the representation node.

[0026] A directed acyclic knowledge graph (DAG) is a special type of knowledge graph structure where the connections between nodes have a clear direction and there are no loop paths. This structure is particularly suitable for describing systems with clear causal chains or unidirectional propagation processes, such as the cascading propagation of faults in multi-stage catalytic converters. The activation intensity probability value is a numerical value between 0 and 1, used to quantify the degree to which a given node in the knowledge graph is activated or the probability of the anomaly it represents occurring. A higher value indicates a greater probability of the anomaly or fault corresponding to that node occurring.

[0027] The cross-level time-delay inference module is used to extract the topological pointing relationships between nodes in the knowledge graph, calculate the aerodynamic time delay between adjacent nodes, and perform weighted aggregation calculation on the time delay alignment probabilities of the predecessor parent node set along the directed edges of the graph, outputting the occurrence probability of each target fault node. In one implementation, the topological pointing relationships between nodes in the knowledge graph can be pre-defined through manual configuration or based on equipment design drawings and stored in a database. The calculation of aerodynamic time delay can be estimated based on empirical values ​​or fixed time delays, for example, setting a constant time delay based on the physical distance between two nodes and the average gas flow rate. The weighted aggregation calculation can use a simple average or fixed weight to accumulate the occurrence probabilities of predecessor parent nodes, for example, simply summing the historical occurrence probabilities of all predecessor nodes and directly outputting the accumulated value as the occurrence probability of the target fault node.

[0028] Topological pointing relationships in a knowledge graph refer to the causal, sequential, or dependent relationships between nodes, where directed edges connect them. For example, an upstream catalytic stage failure can lead to an anomaly in a downstream catalytic stage; this relationship is represented by a topological pointing relationship. Gas lag refers to the time required for gas to flow from one node in a multi-stage catalytic converter to its next adjacent node. This lag can be dynamically changing due to variations in equipment structure and operating conditions. The predecessor parent node set in a directed acyclic knowledge graph refers to the set of all upstream nodes that directly point to a target node. The state of these predecessor nodes directly influences the current state of the target node. Lag alignment probability is used when performing cross-stage reasoning to align the historical states of predecessor parent nodes with the current state of the target node, considering the time delay of fault propagation. The lag alignment probability refers to the probability of a predecessor node influencing the current state of the target node after considering gas lag. Weighted aggregation computation is a method that combines multiple input values ​​according to preset weights. In this system, it is used to perform a weighted summation of the time-delay alignment probabilities of the predecessor parent node set to comprehensively evaluate the occurrence probability of the target node. The occurrence probability of the target faulty node is the likelihood of each node failing at the current moment, calculated by the cross-level time-delay inference module. This probability value reflects the comprehensive failure risk of the node after considering its own activation strength and the impact of upstream propagation.

[0029] The root cause prediction and diagnosis module is used to locate the root cause failure node based on the activation intensity probability value of each node, and calculate the remaining safe operating life based on the occurrence probability and thermodynamic state of the root cause node, combined with the equipment's operating time. Specifically, this module can traverse all nodes and select the node with the highest activation intensity probability value as the root cause failure node. The calculation of the remaining safe operating life can be based on a fixed maximum theoretical design life, minus the equipment's operating time to obtain a base life value. Subsequently, this base life value can be adjusted linearly according to the occurrence probability of the root cause node and the space average temperature. For example, when the occurrence probability increases or the temperature deviates from the design value, the remaining life is linearly reduced. The root cause failure node is the initial failure node in a multi-stage catalytic converter that leads to a series of chain reactions or downstream anomalies. Accurate identification of the root cause is crucial for taking effective maintenance measures. The remaining safe operating life refers to the estimated time that the equipment can continue to operate safely under the current operating conditions without serious failure or reaching the safety limit. This life value is dynamic and is affected by various factors such as equipment health status and operating conditions.

[0030] The following specific example will provide a more in-depth explanation of the above technical solution:

[0031] Suppose a multi-stage catalytic coupling hazardous waste treatment system contains three catalytic stages in series: the first stage, the second stage, and the third stage. During system operation, the catalyst in the second-stage catalytic bed may experience slight sintering, leading to a localized increase in temperature and flow resistance, but not yet reaching the threshold for a traditional single-parameter alarm.

[0032] At this point, the multi-source physics decoupling module comes into play. This module acquires real-time temperature, pressure, and flow data for the first, second, and third stage catalytic beds. For the second stage catalytic bed, it may detect slight differences between multiple thermocouple temperature readings and a slight increase in the instantaneous rate of change of its inlet and outlet pressure differential. The module then decouples and calculates the comprehensive thermodynamic anomaly characteristics and the transient differential characteristics of the aerodynamic flow resistance for the second stage. For example, the comprehensive thermodynamic anomaly characteristics may show local temperature deviations from the mean in the second stage catalytic bed, while the transient differential characteristics of the aerodynamic flow resistance indicate slight fluctuations in flow resistance.

[0033] Next, the graph characterization activation module receives these features. This module fuses the comprehensive thermodynamic anomaly features of the second stage with the transient differential features of aerodynamic resistance and calculates the activation intensity probability value of the characterization node representing the second-stage catalytic bed in the knowledge graph. This probability value may be moderately high, indicating the presence of potential anomalies in the second-stage catalytic bed.

[0034] Subsequently, the cross-stage time-delay inference module intervenes. This module utilizes a pre-constructed knowledge graph that clarifies the topological pointing relationships from the first stage to the second stage, and from the second stage to the third stage. It calculates the aerodynamic time delays for gas flow from the first stage to the second stage and from the second stage to the third stage. Assuming that the first-stage catalyst bed also experienced a slight anomaly earlier, the module considers the impact of the first-stage anomaly on the second stage and performs time alignment based on the aerodynamic time delay. By performing a weighted aggregation calculation of the time-delay alignment probability of the first stage as the precursor parent node, combined with the activation intensity of the second stage itself, the module finally outputs the occurrence probability of the second-stage catalyst bed. For example, if the first stage has a slight anomaly at T-5 minutes and the aerodynamic time delay is 5 minutes, then at time T, the anomaly of the first stage at T-5 minutes will be considered in the calculation of the occurrence probability of the second stage at time T.

[0035] Finally, the root cause prediction and diagnosis module identifies the node with the highest activation intensity probability as the root cause failure node based on the activation intensity probability values ​​of all nodes. In this example, the second-stage catalyst bed is identified as the root cause failure node. The module then extracts the polymerization probability and its spatial average temperature of the second-stage catalyst bed, and combines this with the equipment's operating time to calculate the remaining safe operating life. For example, if the polymerization probability of the second-stage catalyst bed continues to rise, and its spatial average temperature deviates from the design value, the calculated remaining safe operating life will be shortened accordingly. In this way, the system can provide a dynamic, multi-source information fusion-based life prediction, providing a basis for equipment maintenance and safety decisions.

[0036] Based on the above examples, the technical contributions of this solution to the prior art are reflected in the following aspects:

[0037] Firstly, regarding the decoupling of multi-source physical fields, this technical solution acquires physical field data from each node of a multi-stage catalytic converter in real time and decouples and calculates comprehensive thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic flow resistance, achieving deep fusion of multi-physical field data and refined extraction of anomaly features. Compared to existing methods based on single-parameter threshold alarms or single physical field statistical features, this solution can detect equipment anomalies more comprehensively and earlier, avoiding missed and false alarms caused by isolated information. For example, in the above example, even if the local temperature rise and flow resistance increase in the second-stage catalytic bed do not reach the traditional alarm threshold, this solution can still identify potential anomalies through the comprehensive features calculated by decoupling.

[0038] Secondly, regarding graph representation activation and cross-level time-delay inference, this technical solution introduces a directed acyclic knowledge graph to represent the topological pointing relationships between nodes and calculates the aerodynamic time delays between adjacent nodes. This is then used to perform weighted aggregation calculations on the time-delay alignment probabilities of the predecessor parent node set. This effectively solves the problems in existing technologies that treat each level of equipment as independent, fail to explicitly model cross-level causal propagation, and struggle to handle time-varying time-delay dynamic propagation processes. By considering inter-level fault propagation and time delays, this solution can more accurately identify the fault propagation path and impact range, avoiding the influence of downstream node aggregation of upstream anomalies on root cause localization. In the example, the system can distinguish whether the anomaly in the second-stage catalyst bed originates from itself or propagates from the first stage, thus achieving more accurate fault localization.

[0039] Furthermore, in terms of root cause prediction and diagnosis, this technical solution locates the root cause failure node based on the activation intensity probability value of each node, and calculates the remaining safe operating life based on the occurrence probability and thermodynamic state of the root cause node, combined with the equipment's operating time. This overcomes the limitations of existing life prediction methods, such as numerical distortion during the healthy phase and failure to simultaneously consider the equipment's operating time and the cumulative effect of thermal stress. The remaining safe operating life provided by this solution not only considers the probability of failure but also incorporates the equipment's thermodynamic state and operating history, making the life prediction results more reliable and practical. For example, in the example, the system not only identifies the second-stage catalytic bed as the root cause but also provides a specific remaining safe operating life based on its degree of abnormality and operating time, providing a quantitative basis for maintenance decisions.

[0040] In summary, this technical solution provides an intelligent diagnostic system that simultaneously addresses multi-physics field fusion, cross-level time-delay propagation, accurate root cause localization, and safe lifetime output through the systematic integration of multi-source physical field decoupling, knowledge graph representation activation, cross-level time-delay inference, and root cause prediction and diagnosis. This system demonstrates significant technological advancements in fault diagnosis and lifetime management of multi-stage catalytically coupled hazardous waste treatment equipment.

[0041] In a preferred embodiment of the present invention, the multi-source physical field decoupling module includes: the comprehensive thermodynamic anomaly feature comprising quantification of the degree of deviation between spatial dispersion and uniform temperature; the transient differential feature of aerodynamic flow resistance introducing time-domain sliding filtering for noise reduction and singularity protection; and for any i-th stage catalytic node, the real-time acquired physical field data includes: the gas mass flow rate at time t. The absolute pressure difference between the inlet and outlet of this stage catalyst bed The absolute static pressure of this stage of the catalytic bed and the temperatures of N thermocouples located on the grid nodes at this level. The comprehensive thermodynamic anomaly characteristics are a linear weighted combination of a temperature spatial dispersion penalty term and a mean temperature bidirectional deviation penalty term. The temperature spatial dispersion penalty term quantifies the degree of temperature fluctuation relative to the average temperature of the current stage, while the mean temperature bidirectional deviation penalty term quantifies the absolute deviation of the mean temperature of the current stage from the preset rated design mean temperature of the catalytic bed, thus achieving an isomorphic measurement of the bidirectional deviation between overheating and undertemperature. The transient differential characteristics of the pneumatic flow resistance are characterized as the relative rate of change of the pneumatic flow resistance over time after time-domain sliding filtering and denoising. The pneumatic flow resistance is dynamically equivalent based on the ratio of the absolute pressure difference between the inlet and outlet of the catalytic bed to the actual gas volumetric flow rate. The actual gas volumetric flow rate is jointly derived from the gas mass flow rate, absolute static pressure, and mean temperature of the current stage through an ideal gas state mechanism. During the derivation process, a small positive lower limit constraint is applied to the gas mass flow rate to eliminate the division-to-zero singularity calculation collapse caused by the flow rate approaching zero under low-frequency or shutdown conditions.

[0042] Preferably, the comprehensive thermodynamic anomaly characteristics of the i-th level node Transient differential characteristics of aerodynamic flow resistance The calculation formulas are as follows: In the formula, This represents the comprehensive thermodynamic anomaly characteristics of the i-th level node, and its unit is Kelvin square; This represents the temperatures of N thermocouples located on the grid nodes at this level; Represents the spatial average temperature of N grid nodes, in Kelvin; square term Used to quantify the degree of deviation of the temperature uniformity in both overheating and low temperature environments; This indicates the preset rated design uniform temperature, in Kelvin. and The feature fusion coefficient; The transient differential characteristic of the aerodynamic drag of the i-th stage node at time t is expressed in negative first power of seconds. This represents a moving average filter operator based on a preset number of sampling points or a time length w; This represents the first-order difference derivative operator with a time step of seconds. This indicates the absolute pressure difference between the inlet and outlet of the catalyst bed, expressed in Pascals. This indicates the absolute hydrostatic pressure of the catalyst bed, expressed in Pascals. To prevent singular flow lower limit protection function; This represents the mass flow rate of the gas at time t, expressed in kilograms per second. This is a preset lower limit value for a small positive flow rate, expressed in kilograms per second. This represents the gas constant, with units of joules per kilogram (Kelvin). This represents the reference current resistance calibration value, in units of kilograms multiplied by meters to the power of -4 and seconds to the power of -1.

[0043] The comprehensive thermodynamic anomaly feature analysis aims to fully capture anomalies in the temperature distribution within the catalytic bed. Spatial discretization quantifies the non-uniformity of the temperature field, such as localized overheating or undercooling; the bidirectional deviation of the mean temperature quantifies the extent to which the overall average temperature deviates from the design value, whether it is too high (overheating) or too low (insufficient cooling or inadequate reaction). This feature can quantify spatial discretization by calculating the variance or standard deviation of the temperature sensor array readings, and quantify mean temperature deviation by calculating the square of the difference between the mean temperature and the design temperature. Furthermore, signal processing methods based on wavelet analysis or Fourier transform can be used to identify spatial frequency anomalies in the temperature field, combined with the deviation of the weighted average value from the design value.

[0044] The transient differential characteristic of pneumatic flow resistance aims to capture the dynamic trend of resistance changes as fluid passes through a catalytic bed, which is crucial for identifying faults such as blockage, collapse, or flow deviation. Time-domain sliding filter denoising is introduced to eliminate the influence of sensor noise and transient disturbances on the differential calculation, ensuring the smoothness and reliability of the characteristic. Singularity protection is used to avoid numerical instability or infinity caused by zero or near-zero values ​​in the denominator of the flow resistance calculation formula at extremely low or zero flow rates. Time-domain sliding filter denoising can be achieved by preprocessing the raw flow resistance data using a moving average filter, exponential smoothing filter, or Kalman filter. Singularity protection can be achieved by setting a minimum flow threshold. When the actual traffic is lower than this threshold, use this threshold for calculation, or smooth the traffic data before calculation to ensure that it is always greater than zero.

[0045] The real-time acquired physical field data serves as the fundamental raw data for calculating the aforementioned comprehensive thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic flow resistance. Gas mass flow rate reflects the amount of material passing through the catalytic bed; the inlet and outlet absolute pressure difference directly reflects the flow resistance; absolute static pressure provides a reference for the fluid state; and the thermocouple temperature array provides information on the temperature distribution inside the catalytic bed. This data can be acquired in real-time by installing mass flow meters, differential pressure sensors, absolute pressure sensors, and multiple thermocouple sensor arrays, and simultaneously sampled through a data acquisition system. Furthermore, computational fluid dynamics (CFD) models can be used to supplement or correct some of the physical field data by combining a small amount of sensor data with model predictions, especially in locations where it is difficult to directly install sensors.

[0046] Comprehensive thermodynamic anomaly characteristics of the i-th level node Transient differential characteristics of aerodynamic flow resistance The calculation formulas are the specific mathematical implementations of the above features, which transform the raw physical field data into quantified anomaly indicators. The first term in the formula quantifies the spatial dispersion (variance) of temperature, and the second term quantifies the deviation of the average temperature from the design value. The formulas calculate the rate of change of current resistance over time after filtering and singularity protection. These formulas can be implemented directly in the diagnostic system's software modules using programming languages ​​(e.g., Python or C++) and calculated using real-time acquired data. Alternatively, they can be implemented in dedicated data processing units or hardware such as field-programmable gate arrays (FPGAs) to improve computational efficiency and real-time performance.

[0047] The key preset parameters and constants in the above formulas all have clear objective physical basis and conventional acquisition methods:

[0048] For the lower limit of small positive flow It is not an arbitrarily given constant, but is determined by extracting the instrument's 'zero dead zone' or 'small signal cutoff lower limit' based on the factory hardware parameters of the gas mass flow meters (such as Coriolis transmitters) actually deployed on site. In engineering, it is usually calibrated as a fraction of the rated design mass flow rate of the equipment at this level. This ensures that the lower limit of anti-singularity protection matches the actual physical boundary of the hardware.

[0049] For feature fusion coefficients and This is used to balance the weights of spatial dispersion and overall temperature deviation. During system initialization, steady-state operating data of the equipment under fault-free and healthy conditions is collected as a benchmark. An offline optimization method (GridSearch) is then employed to minimize the comprehensive thermodynamic anomalies under healthy conditions. The baseline noise amplitude is the optimization target, and the optimal coefficient pair after convergence is used as the preset value.

[0050] For the moving average filter operator The time length w is determined based on the aerodynamic characteristic time required for the flow field in the catalytic bed gas chamber to establish a steady state. It is usually obtained by consulting the flow field stabilization time constant in the equipment's fluid dynamics (CFD) simulation report.

[0051] For the reference current resistance calibration value This characteristic represents the inherent aerodynamic resistance of the catalytic bed under initial clean and healthy conditions. The specific method for obtaining this characteristic is as follows: During the initial installation and load-bearing trial operation (or clean water / cold state aerodynamic calibration) of the multi-stage catalytic equipment, the equipment is controlled to operate stably at the rated design mass flow rate. Gas mass flow rate, inlet and outlet absolute pressure difference, absolute static pressure, and average spatial temperature are synchronously recorded over a continuous reference time period (e.g., 1 hour) using field sensors. The steady-state data for this period is then substituted into the system flow resistance model to calculate the time average, which is then solidified as [the desired value]. The input parameters. This physical calibration process ensures the subsequent transient differential characteristics of the flow resistance. The calculation has an objective engineering benchmark anchor point, eliminating the calculation ambiguity caused by the difference in porosity of different bed layers.

[0052] The proposed solution first uses a multi-source physical field decoupling module to collect the gas mass flow rate at each node of the multi-stage catalytic device in real time. absolute pressure difference between inlet and outlet of the catalyst bed absolute static pressure of catalytic bed and the temperatures of N thermocouples on the grid nodes. The module then uses pre-defined calculation formulas to decouple and perform calculations on the raw physical field data. This is done in the calculation of the comprehensive thermodynamic anomaly characteristics. By quantifying spatial dispersion (e.g., temperature variance) and the degree of bidirectional deviation from the average temperature (e.g., the squared difference between the average temperature and the design temperature), the non-uniformity of the temperature distribution inside the catalyst bed and the overall temperature deviation from normal operating conditions can be comprehensively captured. This bidirectional quantification ensures that local overheating, insufficient local cooling, or overall excessively high or low temperatures can all be effectively identified. This is particularly useful in calculating the transient differential characteristics of aerodynamic flow resistance. First, a time-domain sliding filter is introduced to denoise the raw flow resistance data to eliminate the interference of sensor noise and transient fluctuations on the calculation of the flow resistance change rate, ensuring the stability and reliability of the characteristics. Simultaneously, an anti-singularity protection mechanism (e.g., a lower flow limit) is implemented. This avoids numerical instability caused by the denominator of the calculation formula approaching zero under low or no flow conditions, ensuring the robustness of the flow resistance differential characteristics. Finally, the module outputs a comprehensive thermodynamic anomaly characteristic that has been precisely calculated and processed. Transient differential characteristics of aerodynamic flow resistance These features serve as inputs to the subsequent map characterization activation module, providing high-quality, highly reliable anomaly signals for the entire fault diagnosis system.

[0053] The following is a specific example. In one implementation, the multi-source physics field decoupling module can be configured with a series of sensors. For example, gas mass flow rate. The absolute pressure difference between the inlet and outlet of the catalytic bed can be measured using a Coriolis mass flow meter. and absolute static pressure Temperatures from N thermocouples can be obtained using high-precision differential pressure transmitters and absolute pressure transmitters. These sensors can be arranged in K-type or T-type thermocouple arrays on different grid nodes of the catalyst bed. The analog signals acquired by these sensors are converted from analog to digital by a data acquisition card (e.g., based on the NIcRIO platform) and transmitted in real time to an industrial controller (e.g., an edge computing unit based on embedded Linux) at a preset sampling frequency. In this controller, a pre-programmed algorithm module processes the real-time data according to the above formula.

[0054] Assuming that at the current sampling time, the preset design uniform temperature of a certain stage of the catalytic bed is... The temperature was 480K, and the temperatures of the four grid nodes collected were 502K, 498K, 505K, and 495K, respectively. The spatial average temperature was calculated. Set the feature fusion coefficient to 500K. and Then, considering the comprehensive thermodynamic anomaly characteristics Calculated as Kelvin squared. Meanwhile, assuming the singularity flow lower limit protection function is effective (or the flow rate is normal), combined with the gas constant... Compared with real-time pressure and flow data, the aerodynamic resistance term after moving average filtering steadily increased from 2500 to 2550 within 1 second, and the reference resistance calibration value... If the value is 1000, then the transient differential characteristic of aerodynamic drag calculated by the first-order difference is... for The highly quantified integrated thermodynamic anomaly characteristics and aerodynamic flow resistance transient differential characteristics obtained from the calculation are then fully encapsulated into data packets and sent to the upper-level diagnostic server via industrial Ethernet for use by the subsequent spectrum characterization activation module.

[0055] Through the aforementioned technical solutions, the multi-source physical field decoupling module can provide more accurate and robust anomaly features. Specifically, by quantifying the spatial dispersion and the degree of deviation from the uniform temperature, the system can comprehensively capture temperature anomalies inside the catalytic bed, whether it is local overheating or overall temperature drift, thus avoiding misjudgments or missed judgments that may be caused by a single temperature index. At the same time, the time-domain sliding filter denoising introduced by the transient differential features of aerodynamic flow resistance effectively suppresses sensor noise and transient disturbances, ensuring the accuracy of the flow resistance change trend; the anti-singularity protection mechanism avoids numerical instability in calculations under extreme conditions (such as low flow rates), significantly improving the reliability of feature extraction. These optimized and refined features provide high-quality input for subsequent spectral characterization activation, cross-stage time-delay inference, and root cause prediction and diagnosis modules, greatly improving the entire intelligent diagnostic system's ability to identify and diagnose early and hidden faults in multi-stage catalytic coupling hazardous waste treatment processes, thereby effectively reducing false alarm rates and missed alarm rates, and ensuring the safe and stable operation of the equipment.

[0056] In a preferred embodiment of the present invention, in the spectrum characterization activation module: the activation intensity probability value of the i-th level characterization node is calculated by a monotonically nonlinear mapping function with probability saturation characteristics, the value range of which strictly converges to the interval [0,1); the input of the independent variable of the mapping function is the weighted sum of the dimensionless comprehensive thermodynamic anomaly characteristics and the aerodynamic flow resistance transient differential characteristics after taking the absolute value; wherein, taking the absolute value of the aerodynamic flow resistance transient differential characteristics aims to construct a bilateral response mechanism to equally capture the catalytic bed blockage fault characterized by the increase in flow resistance and the catalytic bed collapse and flow deviation fault characterized by the sudden drop in flow resistance. Preferably, a bilateral response mode is used to capture bidirectional flow resistance anomalies, and the comprehensive thermodynamic anomaly characteristics and the aerodynamic flow resistance transient differential characteristics are independently and in parallel with a weighted relationship; the activation intensity probability value of the i-th level characterization node The calculation formula is: In the formula, Let represent the activation intensity probability value of the i-th level characterization node. It is dimensionless and its value range is restricted to [0,1). This represents the preset comprehensive thermal anomaly calibration threshold, expressed in Kelvin squares. This represents the preset flow resistance differential sensitivity coefficient, in seconds; This indicates that the absolute value of the transient differential characteristic of aerodynamic flow resistance is taken to capture the blockage caused by the increase in flow resistance and the collapse and flow deviation fault caused by the sudden drop in flow resistance on both sides.

[0057] The two-sided response model for capturing bidirectional flow resistance anomalies means that the system can identify and respond to abnormal changes in fluid resistance (aerodynamic flow resistance) in both directions (increasing or decreasing) above and below the normal value. Traditional fault diagnosis may only focus on increased flow resistance (e.g., blockage, coking), but sudden drops in flow resistance (e.g., catalyst bed collapse, bypass leakage, equipment structural damage) are equally serious fault indicators. The two-sided response model ensures effective detection of both situations. This can be achieved by taking the absolute value of the transient differential characteristic of aerodynamic flow resistance, ensuring that its positive or negative change contributes to the anomaly intensity; alternatively, two independent thresholds can be designed, one for detecting increased flow resistance and the other for detecting decreased flow resistance, and then the results are logically ORed. The integrated thermodynamic anomaly characteristics and the transient differential characteristic of aerodynamic flow resistance are independently weighted in parallel, indicating that these two different types of characteristics contribute their anomaly information independently when calculating the activation intensity probability value, and are combined in a weighted manner, rather than being serially connected or mutually dependent. Independent parallel operation means that thermodynamic anomalies and aerodynamic resistance anomalies can occur and be detected independently, and their contributions to the final activation intensity are adjustable (through weighting). This allows the system to respond promptly when either feature exhibits a significant anomaly, even if the other feature behaves normally. This can be achieved by dividing the two feature terms by their respective thresholds or sensitivity coefficients and then adding them directly, followed by a nonlinear function mapping; alternatively, a parallel layer structure in a neural network can be used to process the two features separately, and then the outputs can be weighted and summed. The activation intensity probability value of the i-th level characterizing node. The calculation formula transforms the processed integrated thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic resistance into a unified, dimensionless probability value between 0 and 1, used to quantify the degree of anomaly at the current node. A higher probability value indicates a greater likelihood of node failure. The introduction of the exponential function 1-exp(-x) ensures that the probability value increases slowly when the anomaly characteristic value is small, and rapidly approaches 1 when the anomaly characteristic value is large, exhibiting good nonlinear mapping characteristics and effectively distinguishing between minor and severe anomalies. This can be achieved through direct numerical calculation according to the formula, where... and Provided by the aforementioned module, and These are preset parameters; alternatively, in the software implementation, mathematical library functions can be used for calculation. This represents taking the absolute value of the transient differential characteristic of aerodynamic flow resistance. It should be noted that the directed acyclic graph of this invention is mainly used for explicit modeling of downstream thermodynamic and aerodynamic mass flow propagation. For the reverse back pressure phenomenon (pressure backlash propagating at the speed of sound) caused by severe downstream blockage in multi-stage systems under actual operating conditions, this system does not rely on the reverse propagation of the graph, but instead constructs a bilateral response mechanism through this absolute value calculation. The surge in back pressure caused by downstream blockage will be instantly reflected in the inlet and outlet pressure difference of this stage, making the collapse and flow deviation faults (characterized by a sudden drop in flow resistance)... ), or the increased flow resistance is characterized by blockage at this stage or a surge in downstream reverse back pressure ( All of these can equivalently trigger the activation intensity probability value of this level. This mechanism allows inverse physical anomalies to be instantly converted into self-locking faults at the current level, perfecting the fault capture loop across the entire topology without altering the underlying directed graph structure. This can be achieved through computation... Previously, for This can be achieved by performing absolute value operations; alternatively, two independent detection branches can be set up, one for detection and the other for general detection. One branch detects positive anomalies, the other detects negative anomalies, and then the anomaly intensities of the two branches are combined.

[0058] To ensure the activation strength probability value The comprehensive thermal anomaly calibration threshold in the formula can accurately reflect the equipment's fault sensitivity. With the differential sensitivity coefficient of flow resistance Offline calibration can be performed through standard industrial data-driven processes, without relying on subjective experience.

[0059] The specific calibration method is as follows: First, construct a historical multiphysics dataset of the target multi-stage catalytic converter. This dataset must include labeled 'fault-free healthy sample sets' and 'known fault sample sets' (such as typical catalytic bed blockage or flow deviation events recorded in historical logs). Second, maximize the activation intensity probability values ​​of the two types of samples after mapping. The statistical discrimination (e.g., using Fisher's discrimination criterion or cross-entropy loss function) is used as the objective function. Finally, particle swarm optimization (PSO) or genetic algorithm (GA) is used to perform iterative optimization calculations on the historical dataset to obtain the converged result. and The optimal value. This calibration process is a standard optimization method in this field, ensuring the adaptability and feasibility of the mapping function on different process specifications and equipment.

[0060] The scheme in this application receives the comprehensive thermodynamic anomaly features from the multi-source physical field decoupling module through the spectrum characterization activation module. Transient differential characteristics of aerodynamic flow resistance To overcome the limitations of unidirectional flow resistance anomaly capture, this module specifically introduces a two-sided response mode, which utilizes the transient differential characteristics of aerodynamic flow resistance. Absolute values ​​are taken to ensure that the degree of anomaly, whether due to an increase or a sudden drop in flow resistance, can be effectively quantified. Subsequently, the processed comprehensive thermodynamic anomaly characteristics are analyzed. and the transient differential characteristics of aerodynamic flow resistance in absolute form The features are combined using an independent, parallel, weighted relationship. This combination method allows two different types of anomalous features to contribute their information independently, and these features are then combined using pre-defined weights (by...). and The weighted anomaly intensities are flexibly adjusted to reflect their relative importance to node failures. Finally, a nonlinear exponential function 1-exp(-x) maps the weighted anomaly intensity to an activation intensity probability value between 0 and 1. This nonlinear mapping makes the system sensitive to minor anomalies while rapidly providing high-probability indications for severe anomalies, thus providing accurate and robust input for subsequent cross-stage time-delay inference and root cause prediction and diagnosis. This mechanism ensures that the system can comprehensively and accurately assess the operating status of each node in the multi-stage catalytic converter, effectively identify various types of potential faults, and significantly improve the accuracy and reliability of fault diagnosis.

[0061] As a specific implementation method, a particular catalytic node in a multi-stage catalytic coupling hazardous waste treatment system can be considered. A multi-source physical field decoupling module monitors and calculates the comprehensive thermodynamic anomaly characteristics of this node in real time. For example, 50 Kelvin squared, and the transient differential characteristic of aerodynamic flow resistance. For example, a value of -0.5 seconds to the power of negative one could indicate a sudden drop in flow resistance. After receiving these characteristic values, the spectral characterization activation module first analyzes the transient differential characteristics of the aerodynamic flow resistance. Taking the absolute value, we get The system detects anomalies in flow resistance by calculating the negative first power of seconds. Then, it uses a preset comprehensive thermal anomaly calibration threshold. For example, 100 Kelvin squared, and a preset flow resistance differential sensitivity coefficient. For example, 5 seconds. Then, according to... The calculation is performed using the formula. Substituting the specific values, we obtain... This calculation result indicates that the activation probability of this node is very high, strongly suggesting the presence of a potential fault.

[0062] Through the aforementioned technical solution, the system can employ a two-sided response mode to comprehensively capture anomalies in both directions: increased flow resistance (such as blockage) and sudden decrease in flow resistance (such as collapse and flow deviation), significantly improving the sensitivity and coverage of fault detection in multi-stage catalytic converters. Simultaneously, by fusing thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic flow resistance in an independent and parallel weighted relationship, the system can flexibly weigh the contribution of different types of anomalies to fault diagnosis, avoiding the risk of a single feature dominating or missing crucial information. This precise and comprehensive feature fusion mechanism, combined with nonlinear activation intensity probability value calculation, enables the system to more accurately and robustly assess the operating status of each node, effectively addressing the shortcomings of traditional methods in terms of comprehensiveness, accuracy, and robustness, and providing a more reliable foundation for subsequent fault reasoning and diagnosis.

[0063] like Figure 2 As shown in the preferred embodiment of the present invention, in the cross-level time-delay inference module: for any i-th level target node and its predecessor parent node set in the directed acyclic knowledge graph, the aerodynamic time delay between the predecessor node and the target node is determined by the ratio of the effective physical volume between the two nodes to the actual gas volume flow rate of the predecessor node after smoothing by the moving average and constrained by the small positive flow rate lower limit, so as to reflect the pure delay time of gas transmission between levels; the occurrence probability of the i-th level target fault node is calculated by the probability constraint transfer function, which linearly weights and fuses the activation intensity probability value of the current level characterization node with the occurrence probability of the predecessor parent node set after alignment by aerodynamic time delay, and further performs nonlinear probability space compression; the coupling weight and directed causal transmission weight of the current level characterization node participating in the weighted fusion are strictly limited to non-negative real numbers, and satisfy the local normalization constraint condition of summing to one within the local topology structure they point to, so as to prevent probability divergence and saturation effect in the recursive inference process from the network topology level.

[0064] Preferably, aerodynamic time delays are calculated based on the time-domain averaged gas dynamic states of the inter-level physical geometry and predecessor nodes, and non-negative aggregation weights that satisfy local normalization constraints are used to prevent the recursive probability of the graph network from diverging and saturating; for any i-th level target node in the directed acyclic knowledge graph, let... Let j be the set of all predecessor parent nodes pointing to node i, and j be the aerodynamic time delay between predecessor node j and target node i. The probability of occurrence of node i The calculation formula is: In the formula, This represents the aerodynamic time delay between the precursor node j and the target node i, in seconds. This represents the effective physical volume between two nodes, expressed in cubic meters. , and Let represent the absolute static pressure, space average temperature, and gas mass flow rate of the precursor node j, respectively, in Pascals, Kelvin, and kilograms per second. Let represent the probability of the i-th level target fault node occurring at time t, which is dimensionless; This represents the coupling weight of the characterization node at this level, and is dimensionless. This indicates that the knowledge graph contains weights that propagate causality, and these weights are dimensionless; all of the above weights are strictly limited to non-negative real numbers and satisfy normalization constraints. The boundary conditions are specified as follows: when node i is the starting source node, i.e., the set... When the summation term is empty, the summation term is always equal to 0.

[0065] Aerodynamic time delay refers to the time required for gas to travel from one catalytic node to the next. In multi-stage catalytic coupling hazardous waste treatment systems, different stages are typically connected by pipes or channels. The flow velocity and transport time of gas in these connecting structures are affected by physical geometry and gas dynamics. Accurate calculation of aerodynamic time delay is crucial for understanding fault propagation paths and times. One calculation method is based on fluid dynamics principles, determining time delay by measuring or estimating the effective volume of the connecting pipes and combining it with the average gas velocity within them. Another method utilizes tracer gas experiments or computational fluid dynamics (CFD) simulations to obtain time delay data under different operating conditions and establish corresponding empirical models or lookup tables. Aggregation weights are used to balance the activation intensity of the current stage node with the fault propagation probability from the predecessor node when calculating the failure probability of the target node. These weights are set to non-negative values ​​to ensure the physical meaning of the probability, i.e., that fault propagation does not cancel each other out. Local normalization constraints ensure that the total probability contribution does not accumulate or decay indefinitely during each aggregation, effectively preventing probability divergence or saturation problems that may occur during recursive probability calculations in knowledge graph networks, thus maintaining the stability and effectiveness of the diagnostic model. The aerodynamic time delay between the predecessor node j and the target node i. The calculation formula is used to quantify the time required for gas to travel from the predecessor node j to the target node i. Among them, This represents the effective volume of the physical space (such as a pipe) connecting the two nodes. It is a fixed parameter that reflects the physical structure. , and These represent the absolute static pressure, spatial average temperature, and gas mass flow rate of the precursor node j at time t, respectively. These are dynamically changing physical quantities that reflect the instantaneous state of the gas at node j. Using these parameters, the average density and velocity of the gas in the connecting pipe can be calculated, and thus the transmission time can be estimated. The moving average filter operator is used to smooth flow data and reduce the impact of instantaneous fluctuations on time delay calculations. The anti-singularity flow lower limit protection function avoids calculation instability or meaningless results caused by the denominator approaching zero when the flow rate is extremely low. In actual control logic, when the equipment is shut down or under extremely low load conditions, causing the actual flow rate to be lower than the limit, the protection function prevents such calculations from occurring. At that time, the system will simultaneously trigger the "graph propagation freeze mechanism" at the software layer, temporarily freezing the weights of cross-level directed causal transmission. Logical values ​​are set to zero (i.e., the physical transmission delay is considered infinite), thus preventing the propagation of false faults caused by static gas pressure holding until the actual flow rate returns to normal. The probability of node i occurring... The formula is used to calculate the danger evolution state of target node i at time t. It is a nonlinear activation function based on the hyperbolic tangent (tanh). It is important to note that, due to the local normalization constraint (weights sum to 1), the numerical value mapped by this function strictly converges to... In this system, this value does not represent a traditional statistical probability, but rather a "standardized nonlinear hazard index" characterizing the equipment approaching its shutdown extreme. The full-scale hazard state of the system is defined as tanh(1). This space compression mechanism plays a natural role in preventing overflow and limiting amplitude at the algorithm topology level. The formula contains two parts: the first part This represents the activation strength probability value of node i at this level. The contribution to the probability of failure is represented by the coupling weight of the node at this level. Weighted. Part Two Let represent the probability contribution of fault propagation from all predecessor parent nodes j, where The precursor node j takes aerodynamic time delay into account. The probability of subsequent failures is determined by the directed causal transit weights. Weighted aggregation. This weighted aggregation method, combined with time-delay alignment, enables accurate modeling of the causal relationships of fault propagation.

[0066] Coupling weights of local representation nodes participating in weighted aggregation With directed causal transit weights The specific methods for obtaining these weights are as follows: These weights reflect the system's physical understanding of the relative impact of 'self-deterioration of the fault at this level' and 'cross-level propagation of the upstream fault'. Before system deployment, the objective transfer probability of a precursor node fault (such as dust escape) leading to a target node fault (such as pore blockage) is assessed by extracting the catalytic reaction kinetic model from the equipment design phase or by consulting the fault tree analysis (FTA) prior report of similar process units. After normalizing this objective transfer probability, it is directly used as the directed causal transfer weight. The input values; then the local normalization constraints (i.e. Automatically derive and lock the coupling weights of the current level's representation nodes. This allocation mechanism strongly anchors the abstract graph weights to the actual process mechanism parameters.

[0067] The proposed scheme accurately infers the probability of occurrence of a target fault node by precisely calculating the inter-stage aerodynamic time delay and employing aggregation weights with local normalization constraints. Specifically, the system first dynamically calculates the aerodynamic time delay required for gas to travel from the precursor node to the target node based on the physical geometry connecting different catalytic nodes and the gas dynamic state at the precursor node. This time delay calculation considers the actual physical process of gas flow, enabling accurate modeling of the time dimension of fault propagation. Subsequently, when calculating the fault occurrence probability of target node i, the system not only considers the activation intensity probability value of the node itself but also aggregates the fault occurrence probabilities of all precursor parent nodes j after the corresponding aerodynamic time delay. This aggregation process introduces coupling weights representing the nodes at this stage. and directed causal transitive weights Weighting is applied. These weights are strictly limited to non-negative real numbers and satisfy local normalization constraints, i.e. This constraint mechanism ensures that during recursive probability calculations in the knowledge graph, the propagation of fault probabilities neither accumulates infinitely leading to divergence nor decays excessively leading to saturation, thus guaranteeing the stability and reliability of the diagnostic results. In this way, the system can organically combine instantaneous physical field data with the time-delay effect of cross-stage fault propagation, achieving accurate and timely diagnosis of faults in multi-stage catalytically coupled hazardous waste treatment systems.

[0068] The following example illustrates this concept. Suppose a three-stage catalytic reactor in series, where the first stage (node ​​1) is connected to the second stage (node ​​2) via a pipe, and the second stage is in turn connected to the third stage (node ​​3) via another pipe. The question is how to calculate the probability of failure at node 2 of the second stage. At that time, the system will first calculate the aerodynamic time delay between the first-level node 1 and the second-level node 2. This can be determined by consulting the design drawings to obtain the effective physical volume of the connecting pipes. (Assuming it is) (and combined with the real-time gas mass flow rate of the first-level node 1) (Assuming it is) absolute static pressure (Assuming it is) ) and space average temperature (Assuming it is) (Given gas constant) for Substituting into the formula, the actual gas volumetric flow rate is derived as follows: Thus, the aerodynamic time delay can be accurately determined. Second.

[0069] After obtaining the aerodynamic time delay, the system will utilize the activation intensity probability value of the second-level node 2 itself. (Assuming the current value is 0.5), and the probability of failure of the first-level node 1 at time t-2.0. To calculate Specifically, the pre-defined coupling weights of the current-level representation nodes. The weight is 0.4, indicating a directed causal transitivity. The value is 0.6 to satisfy the normalization constraint. For the initial source node 1, since the predecessor set is empty, if its activation strength 2 seconds ago is... If it is 0.8, then The system then substitutes these values ​​into the transfer function of node 2, obtaining... By employing this step-by-step, strictly time-delay aligned computational approach that incorporates physical evolution data, the system can accurately track the dynamic propagation of faults in a multi-stage catalytic system, thereby providing a more precise probability of fault occurrence.

[0070] Through the above technical solution, this application effectively solves the problems of inaccurate and delayed diagnosis caused by the time delay of fault propagation in multi-stage catalytic coupling hazardous waste treatment systems. By accurately calculating the inter-stage aerodynamic time delay based on physical geometric space and real-time gas dynamics, the system can time-align the fault information of the precursor nodes, ensuring that the aggregated fault probabilities consider events that are truly causally related and occur at the correct time points. Furthermore, the use of non-negative aggregation weights that satisfy local normalization constraints effectively avoids the divergence or saturation phenomena that may occur when recursively calculating fault probabilities in knowledge graph networks, significantly improving the stability and reliability of the diagnostic model. This method enables the system to more accurately capture the dynamic propagation process of faults, improves the prediction accuracy of the probability of occurrence of target fault nodes, thereby providing more timely and reliable fault diagnosis for hazardous waste treatment equipment, reducing the false diagnosis rate and false negative rate, and ensuring the safe and stable operation of the equipment.

[0071] like Figure 3 As shown, in a preferred embodiment of the present invention, in the root cause prediction and diagnosis module: the system locks the node with the highest intrinsic activation intensity probability value as the root cause fault node, extracts its aggregation occurrence probability and spatial average temperature, and calculates the final remaining safe operating life through the following multi-branch safety logic: constructing an inverse analytical operator of the nonlinear mapping space, inversely restoring the aggregation occurrence probability with a limited value range to an unbounded linear degradation expansion domain, and obtaining the first-order time difference of the characteristic trajectory of the linear degradation expansion domain to obtain the degradation derivative of the mapping space, which represents the true evolution rate of the current fault; at the same time, the upper limit of the remaining theoretical life is determined based on the maximum theoretical design life and the accumulated operating time.

[0072] Preferably, a physical circuit breaker mechanism with immediate direct access and the highest priority is introduced, along with a health condition protection mechanism that is strictly constrained by the low-risk range and includes the current level of thermal penalty. The nonlinear probability is mapped to the linear degradation expansion domain using an inverse hyperbolic tangent function, and a first-order difference linear extrapolation is performed. Specifically, this includes: the system traverses all nodes and locks the intrinsic activation strength probability value. The largest node is identified as the root cause failure node, and the aggregation probability of this root cause node is extracted. and space average temperature Used for lifetime calculation; defines the degenerate derivative of the mapping space. And define the upper limit of the remaining theoretical lifetime after deducting the running time. Its final remaining safe operating life Multi-branch safety logic is used for computation.

[0073] Condition 1: When the probability of aggregation occurs Greater than or equal to the preset safe downtime tolerance probability threshold Or the root cause of the spatial average temperature of the nodes Greater than or equal to the ultimate failure temperature When this occurs, the physical circuit breaker mechanism is triggered, forcing the output to show that the final remaining lifespan is zero. .

[0074] Condition 2: When the circuit breaker mechanism is not triggered, and the probability of aggregation occurs... Less than or equal to the preset upper limit of health probability Simultaneously, when the degenerate derivative of the mapping space is less than or equal to a preset small positive threshold for determining the stationarity of the derivative, i.e. If the device is determined to be in a healthy state with no deterioration trend under the constraint of a low-risk range, the backup mechanism is triggered, and the remaining theoretical life limit, which is discounted by the local thermal penalty factor, is directly output as the original life sequence.

[0075] Condition 3: If the above two conditions are not met, the remaining degradation time is calculated based on the difference between the safe shutdown tolerance probability threshold and the current aggregation probability in the linear degradation expansion domain, combined with the degradation derivative of the mapping space constrained by the conservative benchmark lower limit. The original lifetime sequence is then calculated by combining the upper limit of the remaining theoretical lifetime and the local thermal penalty factor.

[0076] Ultimately, only the original lifetime sequences calculated for those satisfying condition two or three are considered. Apply a first-order inertial smoothing filter to output the final remaining safe operating life. .

[0077] Among them, the local thermal penalty factor characterizes the life loss based on the exponential thermal stress accelerated degradation mechanism, and its loss intensity increases nonlinearly with the magnitude of the approach of the root node spatial average temperature to the ultimate failure mean temperature.

[0078] The preferred method for calculating the remaining lifetime under condition two is as follows: The preferred method for calculating the remaining lifetime under condition three is as follows: In the formula, The maximum theoretical design life is expressed in hours. This represents the cumulative running time, in hours. and Both represent remaining lifespan, in hours; Let f be the downtime tolerance probability threshold, which is dimensionless and satisfies ; To predetermine the upper limit of the probability of health, which is dimensionless and satisfies... ; The threshold value for a small positive number is set to determine if the derivative is stationary, and the unit is the negative first power of seconds; This is the degenerate derivative of the mapping space, expressed in negative first power of seconds; This is a conservative baseline degradation rate lower limit, expressed in negative first power of seconds, and satisfies... ; This is a unit conversion factor, with a value of 1 / 3600, where the unit is hours per second; exponent term Characterizes the local thermal penalty discount based on the exponential thermal stress accelerated degradation model; Characterizes the bidirectional temperature deviation amplitude of the root cause node, in Kelvin; and These are the design rated average temperature and the ultimate failure average temperature, respectively, both in Kelvin. The acceleration coefficient for material thermal sensitivity is dimensionless.

[0079] During the system initialization and parameter tuning phases, strict constraints are imposed on the engineering configuration: the 'limit failure mean temperature' in the local thermal penalty factor formula. It must be absolutely consistent with the extreme damage uniform temperature threshold that triggers the physical melting mechanism in "Condition One". This cross-level parameter binding design ensures a tight logical loop—that is, when the physical operating condition approaches or exceeds... When this happens, the system will be intercepted by 'Condition One' with absolute priority and forced to output a lifetime of zero, thereby completely blocking the physical contradiction risk of a negative denominator or abnormal reverse growth of the exponent caused by exceeding the temperature limit in the formula of Condition Two or Condition Three at the underlying logic level.

[0080] The physical shutdown mechanism aims to address extreme hazardous conditions faced by equipment, such as when the probability of failure or key physical parameters (e.g., temperature) reach critical safety thresholds, by immediately implementing forced shutdown measures to prevent catastrophic accidents. This can be achieved by directly cutting off the equipment's power supply through hardware circuitry or by sending an emergency shutdown signal to the equipment control system via software instructions, ensuring the equipment is placed in a safe state as quickly as possible. Another implementation method is that when any preset hazardous condition is detected, the system immediately assesses the remaining safe operating life as zero, thereby forcing a shutdown at the decision-making level. The health condition baseline mechanism provides a conservative life assessment based on theoretical life and current thermal stress correction when the equipment is in a low-risk, healthy operating state without significant deterioration. Its purpose is to avoid misjudging life due to overdiagnosis when the equipment is actually healthy, while reflecting the microscopic impact of the actual operating environment on life by considering local thermal penalties. To avoid excessive double penalties caused by temperature parameters in system simulations, this invention strictly decouples two temperature deviations in the physical time scale and parameter tuning: the comprehensive thermodynamic anomaly features in the front-end feature extraction module. The module quantifies "transient thermal imbalance states," focusing on immediate alarm responses to sudden changes in the state of spectral nodes. The exponential local thermal penalty factor in this module, however, characterizes "long-term material thermal stress accumulation loss" based on the Arrhenius physical mechanism. The former's fusion weight is highly sensitive to transient changes, while the latter's is a material thermal sensitivity acceleration factor. It needs to be calibrated as a smoothly evolving microscopic physical constant. This cross-scale decoupling design ensures the agility of spectrum state capture and the smoothness of the remaining lifetime assessment trajectory. This mechanism can establish a baseline lifetime model based on historical operating data and equipment design parameters, and, combined with real-time monitored temperature data, reduce the baseline lifetime using a preset temperature sensitivity coefficient. The inverse hyperbolic tangent function is used to map the nonlinear, bounded (typically between 0 and 1) probability of failure to a broader, approximately linear degradation expansion domain. This mapping helps transform the nonlinear change in probability into a quantity that is easier to analyze and extrapolate linearly, thereby improving the accuracy of lifetime prediction. The first-order difference linear extrapolation method uses the degradation derivative of the mapping space, i.e., the rate of change of probability after mapping by the inverse hyperbolic tangent function, to predict the trend of failure probability over a future period. By calculating the degradation rate at the current moment and assuming that this rate remains relatively stable in the short term, the time required to reach the preset shutdown threshold can be linearly inferred, thereby estimating the remaining lifetime. Traverse all nodes and lock the intrinsic activation strength probability value. The largest node is identified as the root cause of the failure. This step aims to identify the most likely source of system failure from the knowledge graph. By comparing the intrinsic activation strength probability values ​​of all nodes, the system can focus on the most active and anomalous node, identifying it as the root cause. The aggregate occurrence probability of this root cause node is then extracted. and space average temperature For lifetime calculation, after identifying the root cause failure node, the system needs to obtain the node's key state parameters, including its failure probability after inference aggregation in the knowledge graph and its real-time spatial average temperature. The degenerate derivative of the mapping space is defined. This definition will aggregate the probability of root cause failure nodes. First, a mapping is performed using the inverse hyperbolic tangent function, and then its derivative with respect to time is calculated. The upper limit of the remaining theoretical lifetime after deducting the already run time is defined. This definition calculates the maximum remaining operating time of a device under ideal, fault-free conditions, starting from the current moment. Multi-branch safety logic is a decision framework that uses different calculation paths to determine the remaining safe operating life of the device based on different operating states and risk levels. First-order inertial smoothing filtering is used to smooth the calculated raw lifetime sequence. Smoothing is performed to eliminate transient fluctuations and noise, thus extending the remaining safe operating life of the final output. More stable and reliable.

[0081] All extreme values, upper and lower limit constraints, and decision thresholds involved in the multi-branch security logic of this module have clear objective engineering basis and standard acquisition methods, completely eliminating the risk of arbitrary boundary settings for extreme values.

[0082] Design life and limiting temperature acquisition: Maximum theoretical design Directly retrieve the 'rated design guaranteed working hours' from the catalyst manufacturer's technical specifications; ultimate failure uniform temperature With the material's thermal sensitivity acceleration coefficient The data were obtained by referring to tables based on the critical phase transition temperature handbook for catalyst thermal sintering and the basic data from the Arrhenius accelerated aging experiment.

[0083] During the system initialization and parameter tuning phases, strict constraints are imposed on the engineering configuration: the 'limit failure mean temperature' in the local thermal penalty factor formula. It must be absolutely consistent with the extreme temperature threshold that triggers the physical melting mechanism in Condition 1. This cross-level parameter binding design ensures a tight logical loop, completely eliminating the physical contradictions in the formulas of Condition 2 or Condition 3 that might cause the denominator to be negative or the exponent to grow abnormally in the opposite direction due to excessive temperature.

[0084] Conservative degradation floor and stationarity determination: Lower limit of conservative baseline degradation rate With the threshold of small positive numbers for determining the stationarity of the derivative The calibration is based on historical time-series data of the equipment during its full-load, healthy, and stable operation phase, and the degradation derivative of the mapping space is calculated. Statistical analysis was performed to extract the measurement noise and the basis of small fluctuations. The confidence limit is taken as a parameter input.

[0085] Multi-level probability safety threshold acquisition: Upper limit of health probability With the safe downtime tolerance probability threshold The value is hard-mapped based on the Hazard and Operability Analysis (HAZOP) and Safety Integrity Level (SIL) risk matrix of the hazardous waste treatment process workshop. For example, the value is... Anchored to the green health belt boundary of the risk matrix, Anchored to the high-risk critical zone boundary of interlocking actions. Furthermore, in the software deployment of actual industrial control systems, the determination of derivative stationarity in condition two ( The system software layer is additionally configured with a state hysteresis interval or a time delay confirmation mechanism (such as requiring conditions to be met continuously for a specific period before performing state switching) to prevent the derivative from fluctuating at high frequency due to environmental noise near the critical threshold, thereby ensuring the stability and robustness of the underlying multi-branch lifetime extrapolation state machine.

[0086] The root cause prediction and diagnosis module of this application, based on the multi-source physical field decoupling module, the spectral characterization activation module, and the cross-level time delay inference module that identify potential faults and locate root cause fault nodes, further provides a robust and secure remaining lifetime assessment mechanism. This module first locks the intrinsic activation intensity probability value by traversing all nodes. The largest node is designated as the root cause failure node, ensuring that lifetime assessment is based on the most critical failure source. The system then extracts the aggregate occurrence probability of this root cause node. and space average temperature As the core input for lifetime calculation, in order to accurately capture the nonlinear characteristics of the equipment degradation process and perform effective linear extrapolation, the system introduces the inverse hyperbolic tangent function to map the nonlinear failure probability to the linear degradation expansion domain, and defines the degradation derivative in the mapping space based on this. Meanwhile, taking into account the inherent design life of the equipment and its already operational time, the upper limit of the remaining theoretical lifespan was calculated. In actual life assessment, this module employs multi-branch safety logic to address operating conditions with varying risk levels. This includes considering the probability of root cause failure nodes occurring. or space average temperature When the preset danger threshold is reached, the physical circuit breaker mechanism is immediately triggered, forcibly ending the remaining safe operating life. Setting it to zero ensures the highest priority for safety, preventing the device from continuing to operate under extremely dangerous conditions. If the circuit breaker mechanism is not triggered, and the device is in a low-risk, healthy state with no signs of deterioration (i.e., the probability of failure is below the upper limit of the healthy probability), then the device will be safe to operate. And degenerate derivative If the condition is stable, the health condition protection mechanism will be activated, outputting an original lifetime sequence that takes into account local thermal penalty correction. This allows for a conservative and reasonable lifespan assessment even under healthy operating conditions, avoiding overdiagnosis. For general degradation conditions falling between the two mentioned above, the system utilizes the degenerate derivative of the mapping space. Perform first-order difference linear extrapolation, combined with the shutdown tolerance probability threshold. The original lifetime sequence considering thermal penalty was calculated. Finally, a first-order inertial smoothing filter is applied to the calculated raw lifetime sequence to eliminate instantaneous fluctuations and output a stable and reliable final remaining safe operating lifetime. Through this hierarchical, multi-mechanism diagnostic and prediction method, the root cause prediction and diagnosis module of this application can effectively combine the fault reasoning capability of knowledge graphs with real-time physical field data. It can not only accurately identify the root cause of the fault, but also provide a remaining life assessment that takes into account safety, accuracy and real-time performance under different operating conditions, which significantly improves the reliability and safety of hazardous waste treatment equipment operation.

[0087] As one specific implementation, the root cause prediction and diagnostic module can be deployed on the edge computing unit or cloud server of an industrial control system (ICS). This module first receives the failure probability of each node from the cross-level time-delay inference module and the average spatial temperature of each node from the multi-source physics decoupling module. When calculating the remaining safe operating life, the system sets a downtime tolerance probability threshold. The upper limit of the health probability is 0.70. It is 0.1; The catalyst bed material is designed to withstand a maximum temperature of 800 Kelvin, with a uniform temperature. The material's thermal sensitivity acceleration factor is 500 Kelvin. It is 0.05.

[0088] Assuming a certain node is currently identified as the root cause, its spatial average temperature Increased to 600 Kelvin, and the equipment has been running for [time]. 10,000 hours (preset maximum design life) The theoretical upper limit is calculated to be 50,000 hours. (hours). Under normal degradation conditions (triggering condition three), if the aggregation of this root cause node occurs with a probability of... Reaching 0.50 (between 0.1 and 0.70), the system maps it to the linear domain using the inverse hyperbolic tangent function, i.e. And lock the target shutdown mapping value. Assume that the degenerate derivative of its mapping space is calculated in real time at this point. for (Higher than the set stable noise level) The system first calculates the remaining degradation time using a unit conversion factor. Dividing the distance difference (0.867-0.549)=0.318 by the degenerate derivative, we get (1 / 3600). Hours. Subsequently, the system extracts the local thermal penalty factor based on the current temperature. Combining the two, and since the extrapolated time did not exceed the upper limit, the original lifetime sequence was calculated. Hours. Finally, to ensure the output lifetime prediction is stable, the system uses a first-order low-pass filter with a time constant of 5 seconds. The process is performed to obtain the final, real-time updated remaining safe operating life. It is then sent to the safety interlock interface.

[0089] It should be noted that in the actual software deployment of industrial control systems, the determination of a small positive threshold for determining the stationarity of the derivative (i.e.) The system additionally configures a state hysteresis interval or a time-delay confirmation mechanism at the software logic layer (e.g., requiring conditions to be continuously met for a specific time period before performing a state switch). This is intended to prevent the system from frequently switching between the health protection mechanism (condition two) and the degradation extrapolation mechanism (condition three) when the degradation derivative fluctuates at a high frequency due to environmental noise near the critical threshold, thereby ensuring the stability and robustness of the underlying lifetime extrapolation state machine.

[0090] Through the above technical solutions, the root cause prediction and diagnosis module of this application can effectively solve the technical problem of how to accurately and safely assess the remaining lifespan of hazardous waste treatment equipment during operation. The introduced physical fuse mechanism ensures immediate forced shutdown in the face of sudden high-risk failures, maximizing the safety of equipment and personnel. The health condition baseline mechanism avoids misjudgments and unnecessary shutdowns caused by over-diagnosis when the equipment is actually healthy, improving equipment utilization. Simultaneously, by mapping nonlinear probabilities to the linear degradation expansion domain using the inverse hyperbolic tangent function and performing first-order difference linear extrapolation, the prediction of equipment degradation trends becomes more accurate and stable, especially when the failure probability is in the medium-to-high risk range, providing more refined lifespan predictions. The application of multi-branch safety logic allows the system to flexibly switch lifespan assessment strategies according to different operating states and risk levels, thereby achieving the best balance between safety, accuracy, and economy, significantly improving the operational reliability and intelligence level of hazardous waste treatment equipment.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge graph-based intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment, characterized in that, include: The multi-source physical field decoupling module is used to collect physical field data of each level of nodes in a multi-stage catalytic device in real time, and to decouple and calculate the comprehensive thermodynamic anomaly characteristics and transient differential characteristics of aerodynamic flow resistance of each level of node. The graph representation activation module is used to fuse comprehensive thermodynamic anomaly features with transient differential features of aerodynamic flow resistance, and calculate the activation intensity probability value of each representation node in the directed acyclic knowledge graph. The cross-level time delay reasoning module is used to extract the topological pointing relationship between nodes in the knowledge graph, calculate the aerodynamic time delay between adjacent nodes, and perform weighted aggregation calculation on the time delay alignment probability of the predecessor parent node set along the directed edge of the graph, and output the occurrence probability of each target fault node. The root cause prediction and diagnosis module is used to locate the root cause fault node based on the activation intensity probability value of each node, and calculate the remaining safe operating life based on the occurrence probability of the root cause node and its thermodynamic state, combined with the equipment's operating time.

2. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 1, characterized in that, In the multi-source physical field decoupling module: The comprehensive thermodynamic anomaly features include the quantification of the degree of deviation between spatial dispersion and uniform temperature; the aerodynamic flow resistance transient differential features introduce time-domain sliding filter denoising and anti-singularity protection. For any i-th stage catalytic node, the real-time acquired physical field data includes: the gas mass flow rate at time t, the absolute pressure difference between the inlet and outlet of the catalytic bed at time t, the absolute static pressure of the catalytic bed at time t, and the temperatures of the N thermocouples located on the grid node at time t. The comprehensive thermodynamic anomaly characteristics are a linear weighted combination of a temperature spatial dispersion penalty term and a mean temperature bidirectional deviation penalty term; The temperature spatial dispersion penalty term is used to quantify the degree of fluctuation and dispersion of the temperature of each thermocouple relative to the average temperature of the space at this level, and the average temperature bidirectional deviation penalty term is used to quantify the absolute deviation of the average temperature of the space at this level from the preset rated design average temperature of the catalytic bed, so as to achieve isomorphic measurement of bidirectional deviation between overheating and low temperature.

3. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 2, characterized in that, The transient differential characteristic of the aerodynamic flow resistance is characterized as the relative rate of change of the aerodynamic flow resistance over time after time-domain sliding filter denoising. The pneumatic flow resistance is dynamically equivalent based on the ratio of the absolute pressure difference between the inlet and outlet of the catalytic bed to the actual gas volume flow rate. The actual gas volume flow rate is derived by the gas mass flow rate, absolute static pressure, and average temperature of the space through an ideal gas state mechanism. In the derivation process, a small positive flow rate lower limit constraint is applied to the gas mass flow rate to eliminate the division-to-zero singularity calculation collapse caused by the flow rate approaching zero under low frequency or shutdown conditions of the equipment.

4. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 3, characterized in that, In the map representation activation module: The activation intensity probability value of the i-th level characterization node is calculated by a monotonic nonlinear mapping function with probability saturation characteristics. The range of the nonlinear mapping function is strictly converged to the interval [0,1). The input of the mapping function is a weighted sum of the comprehensive thermodynamic anomaly characteristics after dimensionless processing and the transient differential characteristics of aerodynamic flow resistance after taking the absolute value.

5. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 4, characterized in that, The absolute value of the transient differential characteristic of aerodynamic flow resistance is taken to construct a two-sided response mechanism to equally capture catalytic bed blockage faults characterized by increased flow resistance and catalytic bed collapse and flow deviation faults characterized by sudden decrease in flow resistance.

6. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 4, characterized in that, In the cross-level time-delay inference module: For any i-th level target node and its predecessor parent node set in a directed acyclic knowledge graph, the aerodynamic time delay between the predecessor node and the target node is determined by the ratio of the effective physical volume between the two nodes to the actual gas volume flow rate of the predecessor node after smoothing by the moving average and constrained by the small positive flow rate lower limit, so as to reflect the pure time delay of gas transmission between stages. The occurrence probability of the i-th level target fault node is calculated by the probability constraint transfer function. This transfer function linearly weights and fuses the activation intensity probability value of the current level characterization node with the occurrence probability of the predecessor parent node set after aerodynamic time delay alignment, and further performs nonlinear probability space compression. The coupling weights and directed causal propagation weights of the local representation nodes participating in the weighted fusion are strictly limited to non-negative real numbers, and satisfy the local normalization constraint condition that sums to one within the local topology they point to, so as to prevent probability divergence and saturation effects in the recursive deduction process from the network topology level.

7. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 6, characterized in that, In the root cause prediction and diagnosis module: The system identifies the node with the highest intrinsic activation intensity probability as the root cause failure node, extracts its aggregation probability and spatial average temperature, and determines the upper limit of the remaining theoretical lifespan of the equipment based on the maximum theoretical design life and the cumulative operating time.

8. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 7, characterized in that, The root cause prediction and diagnosis module constructs an inverse analytical operator for the nonlinear mapping space, which inversely restores the aggregation occurrence probability with a limited value range to an unbounded linear degenerate expansion domain, and calculates the first-order time difference of the characteristic trajectory of the linear degenerate expansion domain to obtain the degenerate derivative of the mapping space, which represents the true evolution rate of the current fault.

9. The intelligent fault diagnosis system for multi-level catalytic coupling hazardous waste treatment based on knowledge graphs according to claim 8, characterized in that, The root cause prediction and diagnosis module calculates the remaining safe operating lifetime through the following multi-branch safety logic: Condition 1: When the probability of aggregation is greater than or equal to the preset safe shutdown tolerance probability threshold, or when the root node space average temperature is greater than or equal to the ultimate destruction average temperature, the physical fuse mechanism is triggered, and the final remaining lifetime is forced to be zero. Condition 2: When the circuit breaker mechanism is not triggered, and the probability of aggregation is less than or equal to the preset health probability limit. Meanwhile, when the degenerate derivative of the mapping space is less than or equal to the preset threshold of a small positive number indicating that the derivative is stable, the device is determined to be in a healthy state with no deterioration trend constrained by the low-risk range, triggering the safety net mechanism and directly outputting the upper limit of the remaining theoretical lifespan, which is corrected by the discount of the local heat penalty factor, as the original lifespan sequence. Condition 3: If the above two conditions are not met, the remaining degradation time is calculated based on the difference between the safe shutdown tolerance probability threshold and the current aggregation probability in the linear degradation expansion domain, combined with the degradation derivative of the mapping space constrained by the conservative benchmark lower limit, and the original lifetime sequence is calculated by combining the upper limit of the remaining theoretical lifetime and the local thermal penalty factor. Finally, a first-order inertial smoothing filter is applied only to the original lifetime sequence calculated under conditions two or three to output the final remaining safe operating lifetime. Among them, the local thermal penalty factor characterizes the life loss based on the exponential thermal stress accelerated degradation mechanism, and its loss intensity increases nonlinearly with the magnitude of the approach of the root node spatial average temperature to the ultimate failure mean temperature.