An operation and maintenance intelligent agent evidence reasoning method

Abstract

The application discloses an operation and maintenance intelligent agent evidence reasoning method, comprising the following steps: an operation and maintenance intelligent agent acquires an industrial operation and maintenance task to be diagnosed, initializes an evidence graph and a candidate hypothesis set; a tool call is initiated and tool observation is received; abnormal information, entity information and structured discovery are extracted from the tool observation and written into the evidence graph, and evidence coverage is calculated; the confidence is reweighted and normalized according to the degree of coincidence between the tool observation and the candidate hypothesis, the corresponding confidence is reduced when a preset contradiction situation is hit, and the current investigation plan is continued when it is not hit; when a preset sufficient condition is met, a candidate answer is generated, the candidate answer is decomposed into claims and the supporting evidence is verified one by one, the claims without supporting evidence are deleted or weakened, and the final answer anchored to the tool observation is output. The application can improve the traceability and auditability of operation and maintenance conclusions, and reduce the risk of unfounded assertions.

Description

Technical Field

[0001] This invention relates to the technical field of industrial operation and maintenance intelligent agents, and in particular to an evidence reasoning method for operation and maintenance intelligent agents. Background Technology

[0002] Industrial operations and maintenance tasks typically require intelligent agents to perform multi-step evidence collection and causal reasoning across various heterogeneous tools, ultimately providing accurate and traceable conclusions. In continuous process industries, a single equipment failure that is not located in a timely manner can cascade and amplify within hours; if the root cause is incorrectly attributed, it may lead to unnecessary component replacements, masking the true defect and causing the same failure to recur. Therefore, industrial operations and maintenance agents must not only complete tasks correctly but also provide auditable conclusions, meaning that each attribution should be traceable to specific sensor readings, work order records, or knowledge base entries.

[0003] When general tool-calling agents are directly transferred to industrial operations and maintenance scenarios, a problem arises: conclusions lack supporting evidence. After multiple tool calls, the final conclusion often contains assertions without any observational support, making it impossible for operations and maintenance personnel to verify or trust them. Furthermore, when tool calls continue to fail and key evidence cannot be obtained, the agent either forces unfounded conclusions or refuses to respond altogether, neither of which meets the actual needs of industrial operations and maintenance for auditable conclusions. Therefore, an evidence-based reasoning method for operations and maintenance agents is needed that enables traceable conclusions and retains explicitly labeled claims that do not contradict the collected tool observations when tool calls fail. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the lack of observational evidence for the conclusions of operation and maintenance agents and the inability to provide auditable conclusions when tool calls fail. This invention provides an evidence reasoning method for operation and maintenance agents. This method constructs an evidence graph, manages candidate hypotheses, and verifies claims one by one during the answer synthesis stage. This ensures that each conclusion is anchored to specific tool observations as much as possible. When tool calls continue to fail, it retains claims based on domain knowledge that are explicitly labeled and do not contradict the collected tool observations.

[0005] The objective of this invention is achieved through the following technical solution: a method for evidence reasoning in an operation and maintenance intelligent agent, comprising the following steps:

[0006] S1: The operation and maintenance agent acquires the industrial operation and maintenance task to be diagnosed and initializes the evidence graph and the candidate hypothesis set; wherein, the evidence graph is used to record the anomaly information, entity information and structured findings extracted from tool observations, and the candidate hypothesis set is used to record several candidate hypotheses for the industrial operation and maintenance task and the confidence level of each candidate hypothesis, and the candidate hypotheses in the candidate hypothesis set are sorted by confidence level.

[0007] S2: The operation and maintenance intelligent agent initiates a tool call for the industrial operation and maintenance task and receives the tool observation returned by the tool call;

[0008] S3: The operation and maintenance intelligent agent adopts rule-based extraction, that is, it extracts abnormal information, entity information and structured findings from the observation of the tool and writes them into the evidence graph. The observed facts and the hypotheses to be verified are separated in the evidence graph, and the co-occurrence relationship and temporal relationship between entities are recorded.

[0009] S4: The operation and maintenance intelligent agent calculates the evidence coverage rate by comparing the number of collected evidences with the number of evidences required to complete the industrial operation and maintenance task based on the evidence map processed in step S3.

[0010] S5: The operation and maintenance agent reweights the confidence levels of each candidate hypothesis based on the degree of consistency between the tool observation and each candidate hypothesis in the candidate hypothesis set. This increases the confidence level of candidate hypotheses that are consistent with the tool observation and decreases the confidence level of candidate hypotheses that contradict the tool observation. The reweighted confidence levels are then normalized so that the sum of the confidence levels of all candidate hypotheses in the candidate hypothesis set is 1. The tool observation is then subjected to contradiction detection. If the tool observation hits any preset contradiction situation, the confidence level of the corresponding candidate hypothesis is reduced, and the operation and maintenance agent is guided to other candidate hypotheses in the candidate hypothesis set. If the tool observation does not hit any preset contradiction situation, no contradiction penalty is imposed. The confidence levels are still reweighted and normalized according to the degree of consistency, and the current operation and maintenance agent continues to execute the pre-set execution plan.

[0011] S6: The operation and maintenance intelligent agent repeats steps S2 to S5, updating the evidence graph, candidate hypothesis set, and evidence coverage rate each time, until the preset sufficient condition is met; when the preset sufficient condition is met, candidate answers are generated based on the final evidence graph and candidate hypothesis set;

[0012] S7: The operation and maintenance agent decomposes the candidate answer into several claims, verifies one by one whether each claim can find supporting evidence in the collected tool observations, deletes or weakens claims that cannot find supporting evidence, and combines the verified and retained claims into a final answer anchored to the tool observations and outputs it.

[0013] Furthermore, the operation and maintenance intelligent agent maintains the operating status during execution, and this operating status is represented by the following formula:

[0014] ;

[0015] In the formula, the subscript t represents the step number; This represents the running state at step t; For step t, a record is created to document the executed tool calls and the tool observations they return. The evidence graph at step t is used to record the anomaly information, entity information, and structured findings extracted from tool observations, as well as the co-occurrence relationships and temporal relationships between entities; Let t be the evidence coverage rate at step t; Let be the set of candidate hypotheses at step t, which includes several candidate hypotheses sorted by confidence level and the confidence level of each candidate hypothesis; The t-th step in the pre-defined execution plan for the operations and maintenance agent; where the evidence diagram... Evidence coverage is maintained by step S3. The candidate hypothesis set is obtained from step S4. Updated from step S5.

[0016] Furthermore, in step S3, the rule-based extraction includes a general extraction layer and a scenario-adaptive extraction layer; the general extraction layer performs abnormal keyword detection, key-value field extraction, entity co-occurrence identification, and temporal relationship identification; the scenario-adaptive extraction layer configures corresponding field extraction rules according to different operation and maintenance scenarios.

[0017] a. For industrial asset operation and maintenance scenarios, extract site fields, asset fields, sensor fields and fault mode fields from tool observations, and construct the attribution relationship between fault modes, assets and sites;

[0018] b. For cloud fault operation and maintenance scenarios, extract indicator fields, log fields, and link fields, as well as deployment name, service name, and container name from the tool observation, and identify fault injection signals.

[0019] Furthermore, in step S4, the evidence coverage rate is calculated according to the following formula:

[0020] ;

[0021] In the formula, the subscript t represents the step number; Let t be the evidence coverage rate at step t; It is the set of evidence collected up to step t that matches the evidence required for the task. The number of pieces of evidence in this set; For industrial operation and maintenance tasks to be diagnosed, To complete the task The required set of non-empty evidence, The number of pieces of evidence in this set; the evidence coverage rate is then indicated to the operations and maintenance agent. At that time, the evidence coverage rate The evidence coverage rate is categorized into three levels: low, medium, and high. The value range of the evidence coverage rate is 0 to 1. The first and second thresholds are both between 0 and 1, and the first threshold is less than the second threshold. When the evidence coverage rate is less than the first threshold, it is classified as low level; when it is not less than the first threshold and less than the second threshold, it is classified as medium level; and when it is not less than the second threshold, it is classified as high level.

[0022] Furthermore, in step S5, the preset contradictory situation includes at least one of the following situations:

[0023] a. The equipment status is reported as normal, but this conflicts with the current candidate hypothesis;

[0024] b. Key data was reported as non-existent;

[0025] c. The specified resource has no record in the knowledge base;

[0026] If at least one of the above situations is met, the confidence level of the corresponding candidate hypothesis is reduced only, without directly refuting the candidate hypothesis.

[0027] Furthermore, in step S5, the confidence levels of each candidate hypothesis are reweighted according to the following formula:

[0028] ;

[0029] In the formula, the subscript t represents the step number; For the tool observation at step t; For a candidate hypothesis in the set of candidate hypotheses, Candidate Hypothesis The confidence level at step t, Candidate Hypothesis The confidence level at step t+1 after reweighting; The summation is performed on the candidate hypotheses traversed, i.e., any candidate hypothesis in the candidate hypothesis set, and is related to the candidate hypothesis being examined. To distinguish, in Add a stroke to the top right corner and mark it as , Candidate Hypothesis The confidence level at step t; Candidate Hypothesis Observation with the tool at step t The degree of consistency, Candidate Hypothesis Observation with the tool at step t The degree of consistency; It is an exponential function with the natural constant as its base; the summation in the denominator is the sum of the results over the set of candidate hypotheses. All candidate hypotheses Summation; the candidate hypothesis set is a non-empty set, and the confidence score of each candidate hypothesis is a non-negative number. After normalization, the sum of the confidence scores is 1; where, the degree of agreement Calculate using the following formula:

[0030] ;

[0031] In the formula, The entity matching signal characterizes the candidate hypothesis. The entities involved are observed by the tool. The degree to which it appears in; As a field consistency signal, the characterization tool observes... Field values ​​and candidate hypotheses The degree of consistency between the predictions, both of which range from 0 to 1; This is a contradictory signal, when the tool observes The value is 1 when any preset contradiction is met, and 0 when no contradiction is met. , , All are non-negative preset weight coefficients, which are the weights of the entity matching signal, field consistency signal, and contradiction signal, respectively. When the preset contradiction situation is not hit, no contradiction penalty is applied, and reweighting and normalization are still completed based on the entity matching signal and field consistency signal.

[0032] Furthermore, in step S6, the preset sufficient condition is: the evidence coverage rate reaches the preset coverage level, and the latest received tool observation has covered the fields required to complete the industrial operation and maintenance task; when the industrial operation and maintenance task also includes additional filtering conditions, and the additional filtering conditions have not yet been met, even if the evidence coverage rate reaches the preset coverage level, no candidate answer is generated, but steps S2 to S5 are repeated to collect the tool observations required for the additional filtering conditions.

[0033] Furthermore, in step S7, after decomposing the candidate answer into several claims, the grounding degree of the candidate answer is calculated according to the following formula, and claims without supporting evidence are deleted or weakened when the grounding degree is lower than a preset grounding threshold:

[0034] ;

[0035] In the formula, These are candidate answers to be verified. as candidate answers The non-empty set of claims obtained by decomposition The number of claims in this set; as candidate answers The number of claims obtained from decomposition and supported by evidence found in the collected tool observations or evidence maps; as candidate answers The grounding degree, whose value ranges from 0 to 1, is the grounding degree. The closer to 1, the better the candidate answer. The higher the proportion of claims with supporting evidence, the better; the preset grounding threshold ranges from 0 to 1; when the grounding degree of candidate answers is lower than the preset grounding threshold, claims without supporting evidence are deleted or weakened, and the remaining claims are combined into the final answer.

[0036] Furthermore, step S7 also includes rollback processing: when the tool call in step S2 continues to return errors and key evidence cannot be obtained, claims that cannot be supported by evidence but can be derived based on domain knowledge are marked as domain knowledge-based claims and retained; retained domain knowledge-based claims meet the following two requirements:

[0037] a. In the final answer, use a qualifier indicating that the claim is based on domain knowledge as a prefix;

[0038] b. This claim does not contradict the observations already collected by the tools.

[0039] Furthermore, this method is implemented based on a kernel and adapter layered architecture: the kernel maintains the evidence graph, the candidate hypothesis set and its confidence function and evidence coverage read / write interface, as well as the claim verification logic of the candidate answer, and does not limit the specific set of tool call objects and field extraction rules; the adapter is responsible for tool call object enumeration, initiating tool calls and normalizing tool observation format, and injecting the field extraction rules used for rule-based extraction into the kernel according to the operation and maintenance scenario, as well as the preset weight coefficients used when calculating the degree of consistency between the tool observation and each candidate hypothesis.

[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0041] 1. By constructing evidence graphs through rule-based extraction and separating observed facts from hypotheses to be verified, each conclusion can be traced back to specific tool observations, thereby improving the auditability of the conclusions.

[0042] 2. Based on the degree of consistency between the tool observations and the candidate hypotheses, the reliability of the candidate hypotheses is reweighted and normalized, and in conjunction with contradiction detection, the operation and maintenance agent can dynamically converge to the candidate hypotheses that are consistent with the observations in multi-step reasoning.

[0043] 3. During the answer synthesis stage, candidate answers are broken down into several claims and verified one by one. Claims without supporting evidence are deleted or weakened, thereby avoiding the output of assertions without observational basis.

[0044] 4. When tool calls continue to fail and key evidence cannot be obtained, retain claims based on domain knowledge that do not contradict the observations collected by the tools in an explicit annotation manner, so that the operation and maintenance agent can still form an auditable final answer when the evidence is incomplete, instead of collapsing into an unfounded definitive answer. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0046] Figure 2 This is a schematic diagram illustrating the structure for extracting and writing evidence data into an evidence graph using a tool, and calculating the evidence coverage.

[0047] Figure 3 A flowchart illustrating the process of reweighting the reliability of candidate hypotheses, generating candidate answers, and verifying grounding. Detailed Implementation

[0048] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0049] like Figure 1 As shown in the figure, this embodiment discloses an evidence reasoning method for an operation and maintenance intelligent agent, including the following steps:

[0050] S1: The operation and maintenance agent acquires the industrial operation and maintenance task to be diagnosed and initializes the evidence graph and the candidate hypothesis set; wherein, the evidence graph is used to record the anomaly information, entity information and structured findings extracted from tool observations, and the candidate hypothesis set is used to record several candidate hypotheses for the industrial operation and maintenance task and the confidence level of each candidate hypothesis, and the candidate hypotheses in the candidate hypothesis set are sorted by confidence level.

[0051] S2: The operation and maintenance intelligent agent initiates a tool call for the industrial operation and maintenance task and receives the tool observation returned by the tool call;

[0052] S3: The operation and maintenance intelligent agent adopts rule-based extraction, that is, it extracts abnormal information, entity information and structured findings from the observation of the tool and writes them into the evidence graph. The observed facts and the hypotheses to be verified are separated in the evidence graph, and the co-occurrence relationship and temporal relationship between entities are recorded.

[0053] S4: The operation and maintenance intelligent agent calculates the evidence coverage rate by comparing the number of collected evidences with the number of evidences required to complete the industrial operation and maintenance task based on the evidence map processed in step S3.

[0054] S5: The operation and maintenance agent reweights the confidence levels of each candidate hypothesis based on the degree of consistency between the tool observation and each candidate hypothesis in the candidate hypothesis set. This increases the confidence level of candidate hypotheses that are consistent with the tool observation and decreases the confidence level of candidate hypotheses that contradict the tool observation. The reweighted confidence levels are then normalized so that the sum of the confidence levels of all candidate hypotheses in the candidate hypothesis set is 1. The tool observation is then subjected to contradiction detection. If the tool observation hits any preset contradiction situation, the confidence level of the corresponding candidate hypothesis is reduced, and the operation and maintenance agent is guided to other candidate hypotheses in the candidate hypothesis set. If the tool observation does not hit any preset contradiction situation, no contradiction penalty is imposed. The confidence levels are still reweighted and normalized according to the degree of consistency, and the current operation and maintenance agent continues to execute the pre-set execution plan.

[0055] S6: The operation and maintenance intelligent agent repeats steps S2 to S5, updating the evidence graph, candidate hypothesis set, and evidence coverage rate each time, until the preset sufficient condition is met; when the preset sufficient condition is met, candidate answers are generated based on the final evidence graph and candidate hypothesis set;

[0056] S7: The operation and maintenance agent decomposes the candidate answer into several claims, verifies one by one whether each claim can find supporting evidence in the collected tool observations, deletes or weakens claims that cannot find supporting evidence, and combines the verified and retained claims into a final answer anchored to the tool observations and outputs it.

[0057] In a specific example, the operation and maintenance agent maintains the running state during execution, which is represented by the following formula:

[0058] ;

[0059] In the formula, the subscript t represents the step number; This represents the running state at step t; For step t, a record is created to document the executed tool calls and the tool observations they return. The evidence graph at step t is used to record the anomaly information, entity information, and structured findings extracted from tool observations, as well as the co-occurrence relationships and temporal relationships between entities; Let t be the evidence coverage rate at step t; Let be the set of candidate hypotheses at step t, which includes several candidate hypotheses sorted by confidence level and the confidence level of each candidate hypothesis; The t-th step in the pre-defined execution plan for the operations and maintenance agent; where the evidence diagram... Evidence coverage is maintained by step S3. The candidate hypothesis set is obtained from step S4. Updated from step S5.

[0060] In a specific example, the rule-based extraction includes a general extraction layer and a scenario-adaptive extraction layer; the general extraction layer performs abnormal keyword detection, key-value field extraction, entity co-occurrence identification, and temporal relationship identification; the scenario-adaptive extraction layer configures corresponding field extraction rules according to different operation and maintenance scenarios.

[0061] a. For industrial asset operation and maintenance scenarios, extract site fields, asset fields, sensor fields and fault mode fields from tool observations, and construct the attribution relationship between fault modes, assets and sites;

[0062] b. For cloud fault operation and maintenance scenarios, extract indicator fields, log fields, and link fields, as well as deployment name, service name, and container name from the tool observation, and identify fault injection signals.

[0063] In a specific example, the evidence coverage rate is calculated using the following formula:

[0064] ;

[0065] In the formula, the subscript t represents the step number; Let t be the evidence coverage rate at step t; It is the set of evidence collected up to step t that matches the evidence required for the task. The number of pieces of evidence in this set; For industrial operation and maintenance tasks to be diagnosed, To complete the task The required set of non-empty evidence, The number of pieces of evidence in this set; the evidence coverage rate is then indicated to the operations and maintenance agent. At that time, the evidence coverage rate The evidence coverage rate is categorized into three levels: low, medium, and high. The value of the evidence coverage rate ranges from 0 to 1. The first and second preset thresholds are both between 0 and 1, with the first threshold being less than the second threshold. An evidence coverage rate less than the first threshold is classified as low, a rate not less than the first threshold and less than the second threshold is classified as medium, and a rate not less than the second threshold is classified as high.

[0066] In a specific example, the pre-defined contradictory situation includes at least one of the following situations:

[0067] a. The equipment status is reported as normal, but this conflicts with the current candidate hypothesis;

[0068] b. Key data was reported as non-existent;

[0069] c. The specified resource has no record in the knowledge base;

[0070] If at least one of the above situations is met, the confidence level of the corresponding candidate hypothesis is reduced only, without directly refuting the candidate hypothesis.

[0071] In a specific example, the confidence levels of each candidate hypothesis are reweighted using the following formula:

[0072] ;

[0073] In the formula, the subscript t represents the step number; For the tool observation at step t; For a candidate hypothesis in the set of candidate hypotheses, Candidate Hypothesis The confidence level at step t, Candidate Hypothesis The confidence level at step t+1 after reweighting; The summation is performed on the candidate hypotheses traversed, i.e., any candidate hypothesis in the candidate hypothesis set, and is related to the candidate hypothesis being examined. To distinguish, in Add a stroke to the top right corner and mark it as , Candidate Hypothesis The confidence level at step t; Candidate Hypothesis Observation with the tool at step t The degree of consistency, Candidate Hypothesis Observation with the tool at step t The degree of consistency; It is an exponential function with the natural constant as its base; the summation in the denominator is the sum of the results over the set of candidate hypotheses. All candidate hypotheses Summation; the candidate hypothesis set is a non-empty set, and the confidence score of each candidate hypothesis is a non-negative number. After normalization, the sum of the confidence scores is 1; where, the degree of agreement Calculate using the following formula:

[0074] ;

[0075] In the formula, The entity matching signal characterizes the candidate hypothesis. The entities involved are observed by the tool. The degree to which it appears in; As a field consistency signal, the characterization tool observes... Field values ​​and candidate hypotheses The degree of consistency between the predictions, both of which range from 0 to 1; This is a contradictory signal, when the tool observes The value is 1 when any preset contradiction is met, and 0 when no contradiction is met. , , All are non-negative preset weight coefficients, representing the weights of the entity matching signal, field consistency signal, and contradiction signal, respectively. No contradiction penalty is applied when a preset contradiction condition is not met; instead, reweighting and normalization are performed based on the entity matching signal and field consistency signal.

[0076] In a specific example, the preset sufficient condition is: the evidence coverage rate reaches the preset coverage level, and the latest received tool observation has covered the fields required to complete the industrial operation and maintenance task; when the industrial operation and maintenance task also includes additional filtering conditions, and the additional filtering conditions have not yet been met, even if the evidence coverage rate reaches the preset coverage level, no candidate answer is generated, but steps S2 to S5 are repeated to collect the tool observations required for the additional filtering conditions.

[0077] In a specific example, after decomposing the candidate answer into several claims, the grounding degree of the candidate answer is calculated according to the following formula, and claims without supporting evidence are deleted or weakened when the grounding degree is lower than a preset grounding threshold:

[0078] ;

[0079] In the formula, These are candidate answers to be verified. as candidate answers The non-empty set of claims obtained by decomposition The number of claims in this set; as candidate answers The number of claims obtained from decomposition and supported by evidence found in the collected tool observations or evidence maps; as candidate answers The grounding degree, whose value ranges from 0 to 1, is the grounding degree. The closer to 1, the better the candidate answer. The higher the proportion of claims supported by evidence, the better. The preset grounding threshold ranges from 0 to 1; when the grounding of a candidate answer is lower than the preset grounding threshold, claims without supporting evidence are deleted or weakened, and the remaining claims are combined into the final answer.

[0080] In a specific example, step S7 also includes rollback processing: when the tool call in step S2 continues to return errors and key evidence cannot be obtained, claims that cannot be supported by evidence but can be derived based on domain knowledge are marked as domain knowledge-based claims and retained; retained domain knowledge-based claims meet the following two requirements:

[0081] a. In the final answer, use a qualifier indicating that the claim is based on domain knowledge as a prefix;

[0082] b. This claim does not contradict the observations already collected by the tools.

[0083] Below, we will take the retrieval task in the industrial asset operation and maintenance scenario, which requires filtering fault modes by sensor availability, as an example to explain the above steps.

[0084] In step S1, the operations and maintenance agent acquires a diagnostic task, such as acquiring fault modes in an asset that can be monitored by available sensors, and initializes an empty evidence graph and a candidate hypothesis set containing initial candidate hypotheses. The evidence graph is used to record anomaly information, entity information, and structured findings extracted from tool observations, as well as co-occurrence relationships and temporal relationships between entities; the candidate hypothesis set is used to store candidate hypotheses sorted by normalized confidence level.

[0085] Figure 2 The tool observation at step t is shown. After the observation format is normalized by the tool, structured discoveries are obtained through rule-based extraction, forming evidence nodes e. i Then add the evidence diagram. The process is shown in the figure, where i is the evidence node number. In the figure, e1 is the "temperature over-threshold" anomaly, e2 is the "chiller unit" entity, e3 is the time "09:42", e4 is the "supply water temperature" sensor, and e5 is the "heat exchanger scaling" fault. The "observation co-occurrence" edge between e1 and e2 indicates that they appear in the same tool observation; the "observation time" edge between e2 and e3 indicates the time when the tool observation corresponding to this entity occurs; the "anomaly attribution" edge between e1 and e4 indicates that the anomaly is generated by the supply water temperature sensor; the "sensor attribution" edge between e2 and e4 indicates that the supply water temperature sensor belongs to the chiller unit; the "fault occurrence time" edge between e3 and e5 indicates the time when the heat exchanger scaling fault occurs; and the "fault and sensor" edge between e4 and e5 indicates that the supply water temperature sensor is used to monitor this fault. Figure 2 Evidence coverage shown below Based on the set of evidence collected up to step t that is relevant to task x Size and the set of evidence required to complete task x The size ratio is calculated, and three levels (low, medium, and high) are indicated.

[0086] In steps S2 to S5, the operations and maintenance agent sequentially initiates tool calls and receives tool observations: locating sites and assets, retrieving fault modes, and querying sensor lists. For each tool observation received, the agent extracts fields such as site, asset, fault mode, and sensor information into an evidence graph through rule-based extraction, calculates the evidence coverage rate, and re-weights and normalizes the confidence of candidate hypotheses based on the degree of consistency between the tool observations and candidate hypotheses. When a tool observation indicates that the equipment status is normal but conflicts with the current candidate hypothesis, or when key data is reported as non-existent, or when a specified resource is not recorded in the knowledge base, a contradiction detection is triggered, reducing the confidence of the corresponding candidate hypothesis and guiding the operations and maintenance agent to other candidate hypotheses.

[0087] The process of reweighting the reliability of candidate hypotheses, generating candidate answers, and verifying relevance is as follows: Figure 3 As shown. For the candidate hypothesis set For any candidate hypothesis u, calculate its correlation with current tool observations based on entity matching signals, field consistency signals, and contradiction signals. The degree of consistency, and determined by the confidence level. After exponential weighting and normalization, the following was obtained: If the pre-set sufficient conditions are not yet met, continue collecting new tool observations and evidence nodes; if the pre-set sufficient conditions are met, generate candidate answers y and split them into non-empty claim sets. Verify each supporting piece of evidence and calculate the grounding degree. Conflicting claims are deleted or weakened; when the tool call in step S2 continues to return errors and key evidence cannot be obtained, claims that can be derived based on domain knowledge and do not contradict the observations collected by the tools are explicitly marked and retained; claims with supporting evidence are retained and the source of the evidence is recorded. The combination of retained claims after processing constitutes the final answer y*.

[0088] The method described in this embodiment can be implemented based on a kernel and adapter layered architecture. The kernel maintains the evidence graph, the set of candidate hypotheses and their confidence functions, the read / write interface for evidence coverage, and the claim verification logic for candidate answers; the adapter is responsible for enumerating tool invocation objects, initiating tool invocations, normalizing tool observation formats, and injecting the field extraction rules used for rule-based extraction into the kernel according to the operation and maintenance scenario, as well as the preset weight coefficients used when calculating the degree of consistency between the tool observations and each candidate hypothesis. This layered architecture decouples the kernel from the specific set of tool invocation objects, and can be adapted to different industrial operation and maintenance scenarios by replacing or extending the adapter.

[0089] In step S6, the operation and maintenance agent repeats the above-mentioned data collection and update until the preset sufficient conditions are met. The preset sufficient conditions include evidence coverage reaching a preset coverage level and the latest tool observations covering the fields required for the task; if the task also includes additional filtering conditions such as filtering by sensor availability that have not yet been met, candidate answers are not generated even if the coverage rate meets the standard, but data collection continues.

[0090] In step S7, the operation and maintenance agent decomposes the candidate answers into several claims, and compares each claim with the structured findings in the evidence graph one by one: if the entity and value are consistent, it is determined to have supporting evidence and is retained; if they are inconsistent or contradictory, it is determined to have no supporting evidence and is deleted or weakened. When the fault sensor mapping tool continues to return errors and key mapping evidence cannot be obtained, claims that can be derived based on domain knowledge are retained with a prefix indicating that they are based on domain knowledge, and it is ensured that the claim does not contradict the observations of the collected tools, thereby forming an auditable conclusion with explicit annotation.

[0091] To verify the effectiveness of the method of this invention, the operation and maintenance agent using the method of this invention was compared with the plan-execution baseline method on four language model families across 113 executable scenarios of the industrial asset operation and maintenance benchmark AssetOpsBench. The results show that the data retrieval accuracy, reflecting the quality of the conclusion, is improved on each model family, increasing from 31.0% to 77.0% on Qwen3.5-9B and from 61.9% to 86.7% on Qwen3.5-35B-A3B; the general result validation rate increased from 52.2% to 80.5% on Qwen3.5-35B-A3B; and the illusion rate decreased from 17.7% to 1.8% on GPT-5.4 and from 67.3% to 22.1% on Qwen3.5-9B. These results demonstrate that the method of this invention significantly improves the traceability of multi-step reasoning conclusions.

[0092] Furthermore, the ablation experiment was validated on Qwen3.5-35B-A3B using the six-dimensional joint pass rate as an indicator: the joint pass rate of the complete method was 75.2%; after removing the evidence graph-related steps, it decreased to 58.4%; after removing candidate hypothesis management, it decreased to 56.6%; and after removing claim verification, it decreased to 54.9%, with the hallucination rate increasing from 6.2% to 11.5%. These results indicate that evidence graph construction, candidate hypothesis management, and claim verification all significantly contribute to the quality of the conclusions, with claim verification playing the most crucial role in suppressing unfounded assertions, thus validating the effectiveness of each step in the method of this invention.

[0093] The above description is only a preferred embodiment of 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 method for evidence reasoning in an operational intelligent agent, characterized in that, Includes the following steps: S1: The operation and maintenance agent acquires the industrial operation and maintenance task to be diagnosed and initializes the evidence graph and the candidate hypothesis set; wherein, the evidence graph is used to record the anomaly information, entity information and structured findings extracted from tool observations, and the candidate hypothesis set is used to record several candidate hypotheses for the industrial operation and maintenance task and the confidence level of each candidate hypothesis, and the candidate hypotheses in the candidate hypothesis set are sorted by confidence level. S2: The operation and maintenance intelligent agent initiates a tool call for the industrial operation and maintenance task and receives the tool observation returned by the tool call; S3: The operation and maintenance intelligent agent adopts rule-based extraction, that is, it extracts abnormal information, entity information and structured findings from the observation of the tool and writes them into the evidence graph. The observed facts and the hypotheses to be verified are separated in the evidence graph, and the co-occurrence relationship and temporal relationship between entities are recorded. S4: The operation and maintenance intelligent agent calculates the evidence coverage rate by comparing the number of collected evidences with the number of evidences required to complete the industrial operation and maintenance task based on the evidence map processed in step S3. S5: The operation and maintenance agent reweights the confidence levels of each candidate hypothesis based on the degree of consistency between the tool observation and each candidate hypothesis in the candidate hypothesis set. This increases the confidence level of candidate hypotheses that are consistent with the tool observation and decreases the confidence level of candidate hypotheses that contradict the tool observation. The reweighted confidence levels are then normalized so that the sum of the confidence levels of all candidate hypotheses in the candidate hypothesis set is 1. The tool observation is then subjected to contradiction detection. If the tool observation hits any preset contradiction situation, the confidence level of the corresponding candidate hypothesis is reduced, and the operation and maintenance agent is guided to other candidate hypotheses in the candidate hypothesis set. If the tool observation does not hit any preset contradiction situation, no contradiction penalty is imposed. The confidence levels are still reweighted and normalized according to the degree of consistency, and the current operation and maintenance agent continues to execute the pre-set execution plan. S6: The operation and maintenance intelligent agent repeats steps S2 to S5, updating the evidence graph, candidate hypothesis set, and evidence coverage rate each time, until the preset sufficient condition is met; when the preset sufficient condition is met, candidate answers are generated based on the final evidence graph and candidate hypothesis set; S7: The operation and maintenance agent decomposes the candidate answer into several claims, verifies one by one whether each claim can find supporting evidence in the collected tool observations, deletes or weakens claims that cannot find supporting evidence, and combines the verified and retained claims into a final answer anchored to the tool observations and outputs it.

2. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, The operation and maintenance intelligent agent maintains its operational status during execution, which is represented by the following formula: ； In the formula, the subscript t represents the step number; This represents the running state at step t; For step t, a record is created to document the executed tool calls and the tool observations they return. The evidence graph at step t is used to record the anomaly information, entity information, and structured findings extracted from tool observations, as well as the co-occurrence relationships and temporal relationships between entities; Let t be the evidence coverage rate at step t; Let be the set of candidate hypotheses at step t, which includes several candidate hypotheses sorted by confidence level and the confidence level of each candidate hypothesis; The t-th step in the pre-defined execution plan for the operations and maintenance agent; where the evidence diagram... Evidence coverage is maintained by step S3. The candidate hypothesis set is obtained from step S4. Updated from step S5.

3. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S3, the rule-based extraction includes a general extraction layer and a scenario-adaptive extraction layer; the general extraction layer performs abnormal keyword detection, key-value field extraction, entity co-occurrence identification, and temporal relationship identification; the scenario-adaptive extraction layer configures corresponding field extraction rules according to different operation and maintenance scenarios. a. For industrial asset operation and maintenance scenarios, extract site fields, asset fields, sensor fields and fault mode fields from tool observations, and construct the attribution relationship between fault modes, assets and sites; b. For cloud fault operation and maintenance scenarios, extract indicator fields, log fields, and link fields, as well as deployment name, service name, and container name from the tool observation, and identify fault injection signals.

4. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S4, the evidence coverage rate is calculated according to the following formula: ； In the formula, the subscript t represents the step number; Let t be the evidence coverage rate at step t; It is the set of evidence collected up to step t that matches the evidence required for the task. The number of pieces of evidence in this set; For industrial operation and maintenance tasks to be diagnosed, To complete the task The required set of non-empty evidence, The number of pieces of evidence in this set; the evidence coverage rate is then indicated to the operations and maintenance agent. At that time, the evidence coverage rate The evidence coverage rate is categorized into three levels: low, medium, and high. The value range of the evidence coverage rate is 0 to 1. The first and second thresholds are both between 0 and 1, and the first threshold is less than the second threshold. When the evidence coverage rate is less than the first threshold, it is classified as low level; when it is not less than the first threshold and less than the second threshold, it is classified as medium level; and when it is not less than the second threshold, it is classified as high level.

5. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S5, the preset contradictory situation includes at least one of the following situations: a. The equipment status is reported as normal, but this conflicts with the current candidate hypothesis; b. Key data was reported as non-existent; c. The specified resource has no record in the knowledge base; If at least one of the above situations is met, the confidence level of the corresponding candidate hypothesis is reduced only, without directly refuting the candidate hypothesis.

6. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S5, the confidence levels of each candidate hypothesis are reweighted according to the following formula: ； In the formula, the subscript t represents the step number; For the tool observation at step t; For a candidate hypothesis in the set of candidate hypotheses, Candidate Hypothesis The confidence level at step t, Candidate Hypothesis The confidence level at step t+1 after reweighting; The summation is performed on the candidate hypotheses traversed, i.e., any candidate hypothesis in the candidate hypothesis set, and is related to the candidate hypothesis being examined. To distinguish, in Add a stroke to the top right corner and mark it as , Candidate Hypothesis The confidence level at step t; Candidate Hypothesis Observation with the tool at step t The degree of consistency, Candidate Hypothesis Observation with the tool at step t The degree of consistency; It is an exponential function with the natural constant as its base; the summation in the denominator is the sum of the results over the set of candidate hypotheses. All candidate hypotheses Summation; the candidate hypothesis set is a non-empty set, and the confidence score of each candidate hypothesis is a non-negative number. After normalization, the sum of the confidence scores is 1; where, the degree of agreement Calculate using the following formula: ； In the formula, The entity matching signal characterizes the candidate hypothesis. The entities involved are observed using tools. The degree to which it appears in; As a field consistency signal, the characterization tool observes... Field values ​​and candidate hypotheses The degree of consistency between the predictions, both of which range from 0 to 1; This is a contradictory signal, when the tool observes The value is 1 when any preset contradiction is met, and 0 when no contradiction is met. , , All are non-negative preset weight coefficients, which are the weights of the entity matching signal, field consistency signal, and contradiction signal, respectively. When the preset contradiction situation is not hit, no contradiction penalty is applied, and reweighting and normalization are still completed based on the entity matching signal and field consistency signal.

7. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S6, the preset sufficient condition is: the evidence coverage rate reaches the preset coverage level, and the latest received tool observation has covered the fields required to complete the industrial operation and maintenance task; when the industrial operation and maintenance task also includes additional filtering conditions, and the additional filtering conditions have not yet been met, even if the evidence coverage rate reaches the preset coverage level, no candidate answer will be generated, but steps S2 to S5 will continue to be executed repeatedly to collect the tool observations required for the additional filtering conditions.

8. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, In step S7, after decomposing the candidate answers into several claims, the grounding degree of the candidate answers is calculated according to the following formula, and claims without supporting evidence are deleted or weakened when the grounding degree is lower than a preset grounding threshold: ； In the formula, These are candidate answers to be verified. as candidate answers The non-empty set of claims obtained by decomposition The number of claims in this set; as candidate answers The number of claims obtained from decomposition and supported by evidence found in the collected tool observations or evidence maps; as candidate answers The grounding degree, whose value ranges from 0 to 1, is the grounding degree. The closer to 1, the better the candidate answer. The higher the proportion of claims that find supporting evidence; The preset grounding threshold ranges from 0 to 1; when the grounding degree of a candidate answer is lower than the preset grounding threshold, claims without supporting evidence are deleted or weakened, and the remaining claims are combined into the final answer.

9. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, Step S7 also includes rollback processing: when the tool call in step S2 continues to return errors and key evidence cannot be obtained, claims that cannot be supported by evidence but can be derived based on domain knowledge are marked as domain knowledge-based claims and retained; retained domain knowledge-based claims meet the following two requirements: a. In the final answer, use a qualifier indicating that the claim is based on domain knowledge as a prefix; b. This claim does not contradict the observations already collected by the tools.

10. The evidence reasoning method for an operation and maintenance intelligent agent according to claim 1, characterized in that, This method is implemented based on a kernel and adapter layered architecture: the kernel maintains the evidence graph, the candidate hypothesis set and its confidence function and evidence coverage read / write interface, as well as the claim verification logic of the candidate answer, and does not limit the specific set of tool call objects and field extraction rules; the adapter is responsible for enumerating tool call objects, initiating tool calls and normalizing tool observation formats, and injecting the field extraction rules used for rule-based extraction into the kernel according to the operation and maintenance scenario, as well as the preset weight coefficients used when calculating the degree of consistency between the tool observation and each candidate hypothesis.

Images