A multi-parameter condition monitoring and fault early warning system for prefabricated substations based on the Industrial Internet

Abstract

This invention discloses a multi-parameter status monitoring and fault early warning system for transformer substations based on the Industrial Internet, belonging to the field of intelligent operation and maintenance technology for industrial equipment. It includes a data acquisition and quality quantification unit, used to collect monitoring data from sensors and quantify the quality of the monitoring data in real time to generate a data quality vector; a dynamic weight optimization unit, used to receive the data quality vector and the fault posterior probability distribution output by the causal reasoning unit, and construct the current state of the decision-making environment based on the data quality vector and the fault posterior probability distribution, and then output an evidence weight vector through a built-in reinforcement learning agent; and a causal reasoning unit, used to receive the evidence weight vector and use the evidence weight vector to weight the conditional probability table of the built-in dynamic Bayesian network model, thereby updating and outputting the fault posterior probability distribution. This invention significantly improves the diagnostic system's resistance to data noise and interference.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of industrial equipment, specifically to a multi-parameter status monitoring and fault early warning system for prefabricated transformer substations based on the Industrial Internet. Background Technology

[0002] Current monitoring methods for prefabricated substations rely on fixed thresholds or static diagnostic models. When processing multi-source, heterogeneous, and time-varying monitoring data, these methods suffer from diagnostic ambiguity and high false alarm rates due to data quality fluctuations, external operating condition interference, and fault mode coupling. Existing expert knowledge is difficult to encode into fixed logical rules, limiting the model's ability to handle complex and unknown scenarios. To address this issue, a dynamic monitoring and early warning system that can simulate a multi-expert consultation mechanism and adaptively adjust the weight of diagnostic evidence is proposed.

[0003] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-parameter status monitoring and fault early warning system for prefabricated transformer substations based on the Industrial Internet, so as to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes:

[0006] The data acquisition and quality quantification unit is used to collect monitoring data from sensors and quantify the quality of the monitoring data in real time to generate a data quality vector.

[0007] The dynamic weight optimization unit is used to receive the data quality vector and the fault posterior probability distribution output by the causal reasoning unit, and to construct the current state of the decision environment based on the data quality vector and the fault posterior probability distribution, and then output the evidence weight vector through the built-in reinforcement learning agent.

[0008] The causal reasoning unit is used to receive the evidence weight vector and use the evidence weight vector to weight the conditional probability table of the built-in dynamic Bayesian network model, thereby updating and outputting the fault posterior probability distribution.

[0009] The diagnostic decision output unit is used to make judgments based on the posterior probability distribution of the fault and output early warning information when the preset early warning triggering conditions are met.

[0010] Preferably, the data acquisition and quality quantification unit generates a data quality vector through the following steps: first, estimating the noise level of the monitoring data and calculating the normalized standard deviation; then, combining the noise level and the normalized standard deviation, constructing a data quality index through a preset composite function to generate a data quality vector.

[0011] Preferably, the state of the reinforcement learning agent is formed by concatenating a data quality vector and a fault posterior probability distribution; the action performed by the reinforcement learning agent based on the state is to generate a new evidence weight vector.

[0012] Preferably, the dynamic weight optimization unit uses a reward function to guide the learning process of the reinforcement learning agent;

[0013] The reward function is generated by linearly combining the evidence convergence and diagnostic ambiguity.

[0014] Preferably, the evidence convergence is calculated based on Shannon entropy in information theory to measure the concentration of the posterior probability distribution of the fault.

[0015] Preferably, the diagnostic ambiguity is a count-type index, obtained by counting competing fault hypotheses whose probability values ​​are higher than a preset confidence threshold in the fault posterior probability distribution.

[0016] Preferably, the dynamic Bayesian network model of the causal reasoning unit has a topology structure that is pre-defined based on domain expert knowledge, and the initial conditional probability table is obtained by learning from historical datasets.

[0017] Preferably, the preset warning trigger condition is activated when both of the following conditions are met simultaneously:

[0018] The posterior probability of the fault consistently exceeds the preset alarm threshold over a given period of time;

[0019] And the diagnostic ambiguity is below the preset decision threshold.

[0020] Preferably, the early warning information includes: the most likely fault type determined from the fault posterior probability distribution, the confidence level corresponding to the most likely fault type, and the key evidence chain for forming a diagnostic conclusion.

[0021] Preferably, the key evidence chain is obtained by identifying the component with the highest weight in the evidence weight vector.

[0022] This invention provides an improved multi-parameter status monitoring and fault early warning system for prefabricated transformer substations based on the Industrial Internet, which has the following improvements and advantages compared with the prior art:

[0023] 1. The system's data acquisition and quality quantification unit can construct composite data quality indicators based on the noise level and volatility of the monitored signals, thereby generating a standardized data quality vector. This design enables the system to actively evaluate the validity and stability of each source of evidence instead of passively receiving all data. The subsequent dynamic weight optimization unit receives this quality vector and uses it as a key input to the decision-making environment, allowing subsequent diagnostic reasoning to intelligently adjust the influence of each sensor's data in the decision-making process based on the data quality, significantly improving the diagnostic system's ability to resist data noise and interference.

[0024] 2. A self-optimizing closed-loop diagnostic control logic was constructed; the reinforcement learning agent in the dynamic weight optimization unit and the dynamic Bayesian network model in the causal reasoning unit form a tight feedback loop; the fault posterior probability distribution output by the causal reasoning unit is not only a reflection of the current diagnostic state, but also fed back to the agent as part of the environmental state to guide the generation of the evidence weight vector at the next moment; this mechanism of linking diagnostic output with weight optimization enables the system to continuously optimize its evidence acceptance strategy according to the evolution of diagnostic conclusions, realizing a multi-expert consultation-style decision fusion, breaking the limitations of the traditional static model weight fixation;

[0025] 3. The scheme employs an ingenious reward and decision-making mechanism. The dynamic weight optimization unit guides the reinforcement learning process through a reward function that combines evidence convergence and diagnostic ambiguity. This reward function, based on the principle of information entropy, encourages diagnostic conclusions to converge towards a single deterministic fault while penalizing ambiguous states where multiple competing faults coexist. This design provides a clear and effective guide for the agent to learn the optimal weighted strategy. The final diagnostic decision output unit adopts a dual triggering condition of stable posterior probability exceedance and controlled diagnostic ambiguity, ensuring that warning information is issued only when the system has high confidence in the fault judgment and there is no obvious ambiguity. This greatly improves the reliability and prudence of the warning and effectively suppresses false alarms.

[0026] 4. This invention significantly enhances the transparency and traceability of diagnostic conclusions; the dynamic Bayesian network model of the causal reasoning unit has its topology pre-set from domain expert knowledge, ensuring the physical conformity of the model's causal logic; its initial parameters are learned from historical data, ensuring the objectivity and truthfulness of probabilistic reasoning; more importantly, the warning information output by the system not only includes the most likely fault type and its confidence level, but also provides a key chain of evidence; this chain of evidence is obtained by identifying the component with the highest weight in the evidence weight vector, intuitively revealing the core sensor information on which the current diagnostic conclusion depends, providing maintenance personnel with clear traceability basis, enhancing their trust in the system's conclusions, and providing efficient guidance for subsequent fault review and troubleshooting. Attached Figure Description

[0027] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0028] Figure 1 This is a flowchart of the multi-parameter status monitoring and fault early warning system for prefabricated transformers based on the Industrial Internet of Things, according to the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0030] Example 1

[0031] Please see Figure 1 This invention provides a multi-parameter status monitoring and fault early warning system for prefabricated transformer substations based on the Industrial Internet, comprising:

[0032] The data acquisition and quality quantification unit is used to collect monitoring data from sensors and quantify the quality of the monitoring data in real time to generate a data quality vector.

[0033] The dynamic weight optimization unit is used to receive the data quality vector and the fault posterior probability distribution output by the causal reasoning unit, and to construct the current state of the decision environment based on the data quality vector and the fault posterior probability distribution, and then output the evidence weight vector through the built-in reinforcement learning agent.

[0034] The causal reasoning unit is used to receive the evidence weight vector and use the evidence weight vector to weight the conditional probability table of the built-in dynamic Bayesian network model, thereby updating and outputting the fault posterior probability distribution.

[0035] The diagnostic decision output unit is used to make judgments based on the posterior probability distribution of the fault and output early warning information when the preset early warning triggering conditions are met.

[0036] This invention provides a multi-parameter status monitoring and fault early warning system for prefabricated transformer substations based on the Industrial Internet;

[0037] The data acquisition and quality quantification unit aims to collect sensor monitoring data from different physical dimensions of the transformer substation and to evaluate its effectiveness in real time and quantitatively, generating a standardized data quality vector; this unit receives the raw signal streams from each sensor in real time.

[0038] The dynamic weight optimization unit aims to intelligently adjust the influence of each sensor's data on diagnostic decisions based on the overall state of the current system. This unit receives the data quality vector output by the data acquisition and quality quantification unit, as well as the fault posterior probability distribution fed back by the subsequent causal reasoning unit. These two together constitute the environmental state information required for the reinforcement learning agent to make decisions. Based on this state, the agent outputs an evidence weight vector, which is passed to the causal reasoning unit to guide its diagnostic process.

[0039] The causal reasoning unit aims to simulate the diagnostic logic of domain experts, inferring the probability of various potential failure modes based on weighted evidence. This unit receives the evidence weight vector output by the dynamic weight optimization unit and uses this vector to dynamically weight the conditional probability table of the internally preset dynamic Bayesian network model. The weighting operation adjusts the influence of each observation evidence on the failure hypothesis. This unit executes the standard Bayesian inference algorithm to update the posterior probability of all failure hypothesis nodes in the network, thereby generating a posterior probability vector reflecting the current system failure confidence distribution. This vector is fed back to the dynamic weight optimization unit, thus forming a closed-loop control loop for policy optimization.

[0040] The diagnostic decision output unit aims to make a final fault warning judgment based on the reasoning results. This unit receives the fault posterior probability distribution generated by the causal reasoning unit and makes continuous judgments based on a set of preset warning triggering conditions. Once the conditions are met, this unit generates and outputs warning information, which includes the specific fault type, confidence level, and key evidence chain.

[0041] This embodiment constructs a closed-loop adaptive monitoring and early warning framework through the collaborative work of the aforementioned units. The system can evaluate the quality of input data in real time and dynamically optimize the weights of each piece of evidence in the diagnostic model using reinforcement learning, thereby realizing a decision fusion mechanism for multi-expert consultation. This design overcomes the problems of diagnostic ambiguity and high false alarm rates that traditional fixed threshold or static models are prone to when facing data quality fluctuations and operating condition interference. It achieves accurate evaluation of the operating status of the transformer substation and improves the robustness of the early warning, significantly improving the accuracy and reliability of the diagnosis.

[0042] The data acquisition and quality quantification unit generates a data quality vector through the following steps: First, it estimates the noise level of the monitored data and calculates the normalized standard deviation; then, it combines the noise level and the normalized standard deviation to construct a data quality index through a preset composite function to generate a data quality vector.

[0043] In this embodiment, the method by which the data acquisition and quality quantization unit generates the data quality vector is further defined; the core of this process lies in constructing data quality indicators by combining the noise level and volatility of the signal through a composite function.

[0044] To further clarify, for the monitoring data stream of any sensor, the system processes it through a preset high-pass filter; this filter is designed to separate high-frequency noise components from effective signal components in the data. The working principle is to allow only signal components with frequencies higher than a certain preset cutoff frequency to pass through, thereby estimating the noise level in the signal.

[0045] The cutoff frequency is set based on the physical characteristics of the signal. Taking the monitoring of electrical signals from a transformer substation with a power frequency of 50Hz as an example, its effective signal and harmonic components are mainly concentrated in the low-frequency band below several hundred hertz. Therefore, the cutoff frequency of the high-pass filter can be set to... This effectively filters out low-frequency effective signals and allows only higher-frequency noise components to pass through, thereby enabling accurate estimation of noise levels.

[0046] At the same time, the system calculates the normalized standard deviation of the original signal of the sensor within the moving time window; the normalized standard deviation here is defined as the ratio of the standard deviation of the signal within the time window to the reference standard deviation under the operating condition. This indicator aims to quantify the dispersion or volatility of the signal itself and reflect the stability of the information carried.

[0047] The system uses a pre-defined composite function to integrate the two indicators mentioned above. This composite function takes the estimated noise level and the calculated normalized standard deviation as input and outputs a single, normalized data quality indicator. The construction of this composite function is as follows: ;

[0048] in, This refers to the data quality of sensor j; It is an estimated noise level obtained after processing by a high-pass filter; It is the maximum noise reference value preset by the system, which represents the upper limit of acceptable noise. The value can be set according to the statistical distribution of noise level in the historical operating data of the equipment, for example, by taking the 99th percentile value or referring to the relevant equipment factory test standards. It is the normalized standard deviation calculated using the aforementioned method; It is the balance coefficient; the balance coefficient here These are adjustable parameters set based on prior knowledge. The values ​​are differentiated according to the noise sensitivity of different types of sensors. The purpose is to optimize the accuracy of quality assessment. The specific values ​​can be optimized and calibrated for different sensor types through offline experiments so that the final quality assessment results are most consistent with expert judgment. Exponential function;

[0049] To further clarify, The calibration process can be specifically described as follows: In the offline historical dataset, select a segment of sensor signal that has been manually labeled with data quality levels by domain experts, such as excellent, good, and poor; minimize the data quality index output by the model. The classification error between the classification level and the expert-labeled level is the optimization objective. Methods such as grid search are used to optimize this error. Find the optimal value within the interval. Value; for example, for a temperature sensor with relatively stable noise characteristics, its The value can be set to For vibration sensors where the signal itself is highly volatile, its The value can be set to This is to increase the weight of signal volatility assessment;

[0050] In system implementation, to ensure the robustness of the model, the input parameters need to be validated. For example, the maximum noise reference value... Perform checks to ensure the values ​​are greater than zero, in order to avoid division by zero errors during calculations; at the same time, perform outlier detection on all sensor input data to prevent extreme noise or signal interruptions from impacting the quality assessment.

[0051] In addition, the system is equipped with a global diagnostic confidence monitoring mechanism. When the data quality vector output by the data acquisition and quality quantification unit indicates that the data quality of most key evidence is lower than the preset minimum availability threshold for a period of time, the system will pause the output of specific fault warnings and instead output system status information indicating that the data quality is too low and the diagnostic results are unreliable. This avoids making misleading decisions based on invalid evidence and ensures the safety and robustness of the system in extremely harsh signal environments.

[0052] The minimum availability threshold can be calibrated based on historical experience data, for example, set as the minimum average data quality level required to ensure 90% diagnostic accuracy in historical data;

[0053] This specific implementation provides a clear and quantifiable technical path for data quality assessment; by simultaneously considering the absolute noise level and relative volatility of the signal, it can more comprehensively and accurately reflect the true validity of the monitoring data; this composite assessment mechanism avoids the one-sidedness brought about by a single indicator; this method enables the subsequent weight optimization unit to obtain a more reliable decision basis, thereby improving the foundation for the diagnostic accuracy of the entire system.

[0054] The state of a reinforcement learning agent is formed by concatenating a data quality vector and a fault posterior probability distribution; the action performed by the reinforcement learning agent based on the state is to generate a new evidence weight vector.

[0055] In this embodiment, the working mechanism of the reinforcement learning agent built into the dynamic weight optimization unit is further clarified;

[0056] The state of the agent is defined as a composite vector composed of two parts: the first part is the data quality vector output by the data acquisition and quality quantification unit at the current moment; the second part is the fault posterior probability distribution vector output by the causal reasoning unit at the current moment; the two are combined to form an environmental state vector that can comprehensively describe the reliability of external evidence and the internal diagnostic status of the current system.

[0057] The agent's action is defined as generating a new evidence weight vector. After receiving the composite state vector, the policy network inside the agent calculates and outputs an action vector. This action vector is then processed by a Softmax activation function layer. The Softmax function converts the input vector into a probability distribution, ensuring that all elements in the final output evidence weight vector are positive and that the sum of all elements is strictly equal to 1. This processing method ensures that the generated weight vector is mathematically sound and can be directly used for subsequent probability weighting calculations.

[0058] By concatenating data quality and failure probability into a state, reinforcement learning agents can make decisions in a more complete information space; this allows the agents to comprehensively consider the reliability of evidence and the tendency of diagnosis, thereby learning more complex weighted strategies; and the design of its action space ensures the effectiveness and standardization of output weights, providing a guarantee for stable connection with causal reasoning units.

[0059] The reinforcement learning agent in this embodiment can be trained using a proximal policy optimization algorithm, which has high sample efficiency while ensuring learning stability. The internal policy network can be constructed as a three-layer fully connected neural network containing an input layer, a hidden layer with 256 neurons, and an output layer. The hidden layer uses a modified linear unit as the activation function to effectively learn the highly nonlinear mapping relationship from state to action.

[0060] Example 2

[0061] The dynamic weight optimization unit uses a reward function to guide the learning process of the reinforcement learning agent.

[0062] The reward function is generated by linearly combining the evidence convergence and diagnostic ambiguity;

[0063] Evidence convergence is calculated based on Shannon entropy in information theory to measure the degree of concentration of the posterior probability distribution of the fault.

[0064] Diagnostic ambiguity is a count-type index, obtained by counting competing fault hypotheses whose probability values ​​are higher than a preset confidence threshold in the fault posterior probability distribution;

[0065] In this embodiment, in order to effectively guide the learning process of the reinforcement learning agent, a specific reward function is designed and introduced into the dynamic weight optimization unit; the reward function is generated by linearly combining the evidence convergence degree representing the convergence of the diagnostic conclusion and the diagnostic ambiguity degree representing the ambiguity.

[0066] The reward function is defined as follows:

[0067] ;

[0068] in, This refers to the reward value given to the agent at time t; It is the convergence of evidence at time t, and its calculation result comes from the formula described later; It is the diagnostic ambiguity at time t, and the calculation result is also derived from the formula described later; and These are coefficients used to adjust the convergence reward and ambiguity penalty, respectively. Both are positive values, and their proportions are set based on the system's sensitivity preference for diagnostic certainty and ambiguity. For example, in scenarios with extremely low tolerance for false negatives, they can be appropriately increased. Compared to The proportion of [something] is used to incentivize agents to form concentrated diagnostic conclusions more quickly;

[0069] For example, in a scenario where avoiding underreporting is the primary objective, the system needs to quickly generate centralized diagnostic conclusions. In this case, rewards can outweigh penalties, such as setting... , Conversely, in a scenario where preventing false alarms is the primary objective, the system needs to make more cautious judgments. In this case, the penalty for ambiguous states can be increased, such as by setting... , ;

[0070] Evidence convergence Its purpose is to measure the concentration of the posterior probability distribution of faults; the calculation is based on Shannon entropy in information theory; the formula is:

[0071] ;

[0072] in, It is the posterior probability of the fault hypothesis at time t, output by the causal reasoning unit; It represents the total number of fault hypotheses; this formula is the result of transforming and normalizing the standard Xiangnon entropy; when diagnostic conclusions are highly concentrated on a specific fault... The value will be close to 1; conversely, if the probability distribution of each fault is uniform, The value will be close to 0; Sum all fault assumptions; Logarithmic function with base N;

[0073] Diagnostic ambiguity The purpose is to quantify the number of competing failure hypotheses with high confidence; it is a count-type indicator, calculated as follows:

[0074] ;

[0075] in, It is an indicator function that has a value of 1 when its internal condition is true, and 0 otherwise; It is the posterior probability of fault i, and its source is the same as above; This is a preset confidence threshold; this threshold can be determined by analyzing the receiver operating characteristic curve on a historical dataset, selecting the probability value corresponding to the maximum Youden index, thereby achieving the optimal balance between distinguishing meaningful fault indications and random probability fluctuations; this formula measures the degree of ambiguity in diagnosis by calculating the number of faults with posterior probabilities exceeding this threshold.

[0076] This reward mechanism successfully transforms high-level, difficult-to-optimize diagnostic goals into scalable, optimizable reward signals; evidence convergence. The introduction of this concept, utilizing the mathematical principle of information entropy, precisely rewards weight adjustment actions that make diagnostic conclusions more certain; diagnostic ambiguity... The introduction of this directly penalizes ambiguous states that lead to multiple fault conclusions; the combination of the two provides a clear and effective learning guide for reinforcement learning agents, enabling them to autonomously learn the optimal dynamic evidence weighting strategy, thereby naturally achieving the ultimate goal of improving diagnostic certainty and reducing ambiguity in the process of maximizing long-term cumulative rewards.

[0077] Example 3

[0078] The dynamic Bayesian network model of the causal reasoning unit has a topology that is pre-defined based on domain expert knowledge, and the initial conditional probability table is obtained by learning from historical datasets.

[0079] In this embodiment, the construction method of the core model of the causal reasoning unit—the dynamic Bayesian network—is described in detail;

[0080] The topology of this DBN, namely the nodes in the network and the directed connections between the nodes, is pre-defined based on the prior knowledge of domain experts; this setting process is an abstract modeling of the fault mechanism of the transformer substation.

[0081] As an example, for the specific fault mode of main transformer winding overheating, the corresponding DBN topology can be set as follows: the winding temperature sensor node, load current sensor node, and cooling fan operation status node, which represent observational evidence, are set as parent nodes, and the winding overheating fault node, which represents the fault hypothesis, is set as its child node; the parent nodes can also be configured with dependencies based on physical mechanisms. For example, a directed edge can be drawn from the load current sensor node to the winding temperature sensor node to indicate that an increase in load current will lead to an increase in winding temperature; in this way, a local network structure that can clearly reflect physical causal relationships is constructed.

[0082] The initial conditional probability table of the network is obtained through offline learning. The conditional probability table defines the probability that a child node will take different states given the state of its parent node. The learning process utilizes a large amount of historical dataset and employs standard machine learning algorithms such as the expectation-maximization algorithm for parameter estimation. The EM algorithm can effectively handle the situation where there are missing values ​​in the dataset. Through iterative expectation E-steps and maximization M-steps, it eventually converges to a set of optimal probability parameters.

[0083] By combining expert knowledge and data-driven methods to construct the DBN model, the advantages of both aspects are complemented. The network topology structure pre-set by expert knowledge ensures that the causal logic of the model conforms to physical reality. Meanwhile, the initial CPT learned from historical data enables the model's probability parameters to objectively reflect the statistical characteristics in real operating data. This construction method not only improves the initial accuracy of the model, but also provides a solid and reasonable foundation for subsequent online dynamic weighting and inference.

[0084] The preset warning trigger condition is activated when both of the following conditions are met:

[0085] The posterior probability of the fault consistently exceeds the preset alarm threshold over a given period of time;

[0086] And the diagnostic ambiguity is lower than the preset decision threshold;

[0087] In this embodiment, the preset warning triggering condition used by the diagnostic decision output unit to trigger the warning is specifically defined as two logical conditions that must be met simultaneously.

[0088] The first condition is that the failure probability consistently exceeds the limit; this condition requires that the posterior probability of a specific failure k is... If, over a sustained period of time, the value consistently and stably exceeds a preset alarm threshold... This alarm threshold The numerical value is set based on the criterion of controlling the precision and recall of the alert within the target range. For example, it can be set as the minimum probability threshold corresponding to achieving a 95% recall rate on the validation set. Continuous and stable judgment can be used to filter out short-term, spike-like probability fluctuations.

[0089] The duration of this period is further quantified as continuous. One data sampling point; The value can be set according to the system's sampling frequency and tolerance to interference; for example, if the system's data sampling frequency is 10Hz, that is, 10 points per second, it can be set to... This means that the posterior probability of a specific fault must consistently exceed the alarm threshold for 5 consecutive seconds. This effectively filters out probability spikes that are caused by transient disturbances and last for less than 5 seconds.

[0090] The second condition is that the overall diagnostic ambiguity of the system is under control; this condition requires that the overall diagnostic ambiguity of the system be controlled. Below the preset decision threshold This decision threshold The setting is intended to ensure the uniqueness and clarity of the warning. Its value is usually set to a small integer, such as 2, which means that the system only confirms the warning condition when there is no or only one major competing failure assumption.

[0091] This dual-condition triggering mechanism greatly improves the reliability and prudence of the early warning; a single probability over-limit condition is easily triggered by noise or transient interference, leading to false alarms; while increasing the constraint on diagnostic ambiguity ensures that the early warning is issued when the diagnostic conclusion is highly convergent, effectively suppressing false alarms caused by diagnostic ambiguity; this design makes the final early warning output of higher quality.

[0092] Example 4

[0093] The warning information includes: the most likely fault type determined from the fault posterior probability distribution, the confidence level corresponding to the most likely fault type, and the key chain of evidence for forming a diagnostic conclusion;

[0094] The key chain of evidence is obtained by identifying the component with the highest weight in the evidence weight vector;

[0095] In this embodiment, the content included in the output warning information by the diagnostic decision output unit, and the generation method of the key evidence chain therein, are further explained;

[0096] When the warning triggering conditions defined above are met, the output warning information is a structured data packet, which includes at least the following three items:

[0097] Most likely failure type: This is the failure hypothesis with the highest posterior probability value selected from the final failure posterior probability distribution as the most likely failure type;

[0098] Confidence level: This value is the posterior probability value corresponding to the most likely fault type, which directly quantifies the system's confidence in the diagnostic conclusion;

[0099] Key chain of evidence: The purpose is to provide a basis for the diagnosis and explain what information the system uses to make its judgment;

[0100] The key evidence chain is generated by directly combining the evidence weight vector output by the dynamic weight optimization unit. Related; the system will identify the evidence weight vector at the current moment. In the vector, the components with the highest weight values ​​are identified. Since each component of the vector corresponds to a sensor data source or a specific observational evidence, the sensor or evidence corresponding to the component with the highest weight is considered to be the key evidence that contributes the most to the diagnosis process. Arranging the names or identifiers of these key evidences in descending order of weight constitutes a key evidence chain that can be used as a reference by maintenance personnel.

[0101] This implementation transforms early warning information from inexplicable conclusions into transparent and traceable ones. By providing the most likely fault type and its confidence level, it offers clear diagnostic directions. More importantly, by identifying and providing key evidence chains, it reveals the reasoning behind the diagnostic conclusion, enhancing operators' confidence in the system's conclusions. Maintenance personnel can quickly locate the relevant sensors or equipment parts for verification based on the evidence chains, which not only verifies the correctness of the early warning but also significantly improves the efficiency of subsequent fault investigation.

[0102] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet, characterized in that: include: The data acquisition and quality quantification unit is used to collect monitoring data from sensors and quantify the quality of the monitoring data in real time to generate a data quality vector. The dynamic weight optimization unit is used to receive the data quality vector and the fault posterior probability distribution output by the causal reasoning unit, and to construct the current state of the decision environment based on the data quality vector and the fault posterior probability distribution, and then output the evidence weight vector through the built-in reinforcement learning agent. The causal reasoning unit is used to receive the evidence weight vector and use the evidence weight vector to weight the conditional probability table of the built-in dynamic Bayesian network model, thereby updating and outputting the fault posterior probability distribution. The diagnostic decision output unit is used to make judgments based on the posterior probability distribution of the fault and output early warning information when the preset early warning triggering conditions are met. The data acquisition and quality quantification unit generates a data quality vector through the following steps: First, it estimates the noise level of the monitored data and calculates the normalized standard deviation; then, it combines the noise level and the normalized standard deviation to construct a data quality index using a preset composite function to generate the data quality vector; the formula for calculating the composite function is as follows: ； in, Refers to sensors Data quality; It is an estimated noise level obtained after processing by a high-pass filter; This is the system's preset maximum noise reference value; It is the calculated normalized standard deviation, defined as the ratio of the standard deviation of the signal within the time window to the reference standard deviation under the operating condition. It is the balance coefficient; Exponential function; The state of the reinforcement learning agent is formed by concatenating a data quality vector and a fault posterior probability distribution; the action performed by the reinforcement learning agent based on the state is to generate a new evidence weight vector.

2. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 1, characterized in that, The dynamic weight optimization unit uses a reward function to guide the learning process of the reinforcement learning agent. The reward function is generated by linearly combining the evidence convergence and diagnostic ambiguity.

3. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 2, characterized in that, The evidence convergence is calculated based on Shannon entropy in information theory to measure the degree of concentration of the posterior probability distribution of the fault.

4. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 2, characterized in that, The diagnostic ambiguity is a count-type index, obtained by counting competing fault hypotheses whose probability values ​​in the fault posterior probability distribution are higher than a preset confidence threshold.

5. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 1, characterized in that, The dynamic Bayesian network model of the causal reasoning unit has a topology that is pre-defined based on domain expert knowledge, and the initial conditional probability table is obtained by learning from historical datasets.

6. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 1, characterized in that, The preset warning trigger condition is activated when both of the following conditions are met: The posterior probability of the fault consistently exceeds the preset alarm threshold over a given period of time; And the diagnostic ambiguity is below the preset decision threshold.

7. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 1, characterized in that, The early warning information includes: the most likely fault type determined from the fault posterior probability distribution, the confidence level corresponding to the most likely fault type, and the key evidence chain that forms the diagnostic conclusion.

8. The multi-parameter status monitoring and fault early warning system for prefabricated substations based on the Industrial Internet as described in claim 7, characterized in that, The key chain of evidence is obtained by identifying the component with the highest weight in the evidence weight vector.

Images