A method for remote diagnosis and maintenance support of electromechanical equipment

Abstract

The application discloses a kind of electromechanical equipment remote diagnosis and maintenance support method, it is related to electromechanical equipment remote operation and maintenance technical field.The method includes: diagnostic model generates diagnostic conclusion simultaneously output diagnostic reasoning chain, record complete reasoning path from sensor feature deviation to final diagnostic conclusion;When remote expert guides maintenance through augmented reality terminal, automatically extract verification result information from conversation voice and structure;After structured verification data and each reasoning node in diagnostic reasoning chain are associated and matched, along reasoning path, node by node executes cause-effect traceability comparison, when diagnostic conclusion is inconsistent with actual fault, backtrack layer by layer to locate the reasoning node where deviation occurs and deviation type;After deviation record is aggregated and is statistically triggered directional correction, the model parameters of corresponding reasoning node are executed differentiating correction.The application establishes the closed loop of each link check and directional correction from maintenance verification to diagnostic reasoning chain, realizes that diagnostic model continues self-evolution with the increase of use times.

Description

Technical Field

[0001] This invention relates to the field of remote operation and maintenance technology for electromechanical equipment, and more specifically, to a method for remote diagnosis and maintenance support of electromechanical equipment. Background Technology

[0002] Stamping production lines are a core production link in industries such as automobiles and home appliances. The presses and other electromechanical equipment within these lines are high-value and complex, involving multi-domain coupling of electrical, transmission, and hydraulic systems. Downtime due to malfunctions can result in significant losses. To ensure stable equipment operation, a common maintenance solution is based on sensor monitoring and remote diagnostic platforms: sensors deployed at the equipment collect and upload operational data; the remote diagnostic platform outputs fault conclusions based on diagnostic models; and then, maintenance work orders are issued, with remote experts guiding on-site personnel through augmented reality terminals to complete troubleshooting and repairs.

[0003] Chinese patent CN121680348A discloses a remote fault diagnosis method for stamping equipment based on the Internet of Things (IoT). It constructs a three-dimensional data acquisition framework encompassing equipment, process, and environment, and combines edge computing and adaptive feature transfer algorithms to achieve cross-condition fault diagnosis, then sends the diagnostic results to the field terminal. Chinese patent CN201821617065.0 discloses a remote operation and maintenance service system for a large closed-loop mechanical press. This system compares the real-time operating parameters of the equipment with normal range values ​​using a PLC controller, sends out-of-range parameters to a remote operation and maintenance server for fault diagnosis, and pushes maintenance guidance to the field.

[0004] The above scheme reflects the general approach of existing remote diagnostic and maintenance support technologies: the diagnostic system outputs fault conclusions based on sensor data, and the maintenance support system guides on-site handling based on these conclusions. The output of the fault diagnosis system flows to the maintenance support system, but information such as the actual faulty component, fault type, and severity confirmed through disassembly and measurement during the maintenance process is not fed back to the diagnostic system.

[0005] As a result, the diagnostic model lacks real feedback from maintenance practice. When the diagnostic conclusion is biased, it is impossible to determine the specific link in feature extraction, component mapping or fault classification where the bias occurred, and the model optimization lacks a link-level basis. Summary of the Invention

[0006] To overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide a remote diagnosis and maintenance support method for electromechanical equipment, which aims to solve the following problems: In the process of remote diagnosis and maintenance support for electromechanical equipment such as presses in stamping production lines, the diagnosis system lacks step-by-step feedback information from the maintenance process. When the diagnosis conclusion deviates, it is impossible to determine the specific link in the reasoning chain where the deviation occurs, making it difficult to make targeted corrections to the model.

[0007] According to one aspect of the present invention, a method for remote diagnostics and maintenance support of electromechanical equipment is provided, comprising the following steps: S1. When the remote diagnostic platform detects an abnormality in the press machine's electromechanical equipment in the stamping production line and triggers a diagnostic task, the diagnostic model outputs a diagnostic reasoning chain while generating a diagnostic conclusion.

[0008] The diagnostic reasoning chain records the complete reasoning path from sensor feature deviation to the final diagnostic conclusion in the form of a directed acyclic graph. It makes the internal reasoning process of the diagnostic model explicit into a structured record that can be traced node by node, so that the input data, output conclusion and local confidence of each reasoning link can be accessed and verified independently.

[0009] The diagnostic inference chain contains multiple inference nodes. Each inference node records the source of its input data, the output conclusion, the local confidence level, and the identifier of the causal association rule on which it is based.

[0010] Furthermore, the reasoning nodes of the diagnostic reasoning chain include: Feature deviation node: Records the feature value extracted from the original sensor data and the deviation of the feature value from the normal baseline value. The feature value includes the change in vibration spectrum energy, the change in current waveform slope, the change in pressure pulsation amplitude, or the brake angle drift. The feature-component mapping node records the reasoning process and mapping rules used to map the feature deviation to candidate faulty components of the press. The fault mode matching node records the process of matching fault modes within the range of candidate faulty components and the matching degree of each fault mode. The conclusion output node records the final diagnostic conclusion and the overall confidence level.

[0011] Furthermore, the causal association rule identifier points to the corresponding rule in the preset causal association rule base, which stores the mapping relationship between feature deviation type and candidate faulty parts of the press, as well as the initial weight value of each mapping relationship.

[0012] By establishing a pre-built causal association rule base, the prior knowledge of domain experts on the causal relationship between equipment failure characteristics and components is solidified in the system in the form of rules and weights, providing an initial judgment basis for the reasoning process of feature-component mapping nodes.

[0013] S2. During the process of remote experts guiding on-site maintenance personnel to troubleshoot and repair the press machine's electromechanical equipment through augmented reality terminals, the audio stream of the remote guidance session is acquired in real time and converted into a text sequence.

[0014] The text sequence is subjected to instruction intent recognition to determine the operation intent category corresponding to the remote expert instruction. The target component name is extracted from the instruction text and mapped to the standard component identifier. The verification result information is extracted from the voice feedback text of the on-site maintenance personnel within a preset time window after the instruction is issued.

[0015] The extracted verification information is organized into structured verification data. Each piece of structured verification data includes the target component identifier, verification type, and verification result value.

[0016] The above steps complete the collection and structuring of verification information during the natural interaction of the maintenance guidance session, without requiring maintenance personnel to perform additional data entry operations, so that each remote maintenance support process can generate actual fault confirmation data for diagnostic verification.

[0017] Furthermore, the categories of operational intent include instructing the opening of a specified part of the press, instructing the measurement of a specified physical quantity, instructing the visual inspection of the status of a specified component, and instructing the execution of a specified functional test, wherein the specified physical quantity includes the thickness of the clutch friction plate, the brake clearance, the slider guide clearance, or the main motor bearing clearance. The extraction of verification result information adopts the corresponding extraction template according to the operation intention category. The verification result information includes the measured value and its unit, the observed component status description or the functional test result value.

[0018] S3. Associate and match the structured verification data with the inference nodes in the diagnostic inference chain. Starting from the conclusion output node of the diagnostic inference chain, perform causal tracing comparison node by node along the inference path from bottom to top.

[0019] When the diagnostic conclusion output by the conclusion output node is inconsistent with the actual fault determined based on the structured verification data, the inference node where the deviation first occurs is traced back layer by layer, and the deviation type is determined based on the location of the deviation.

[0020] By comparing each node from bottom to top along the reasoning chain, the location of diagnostic deviations is refined from judging the correctness of the overall conclusion to identifying specific links in the reasoning chain, which can distinguish whether the deviation originates from feature extraction, component mapping or fault classification.

[0021] The types of deviations include: Classification layer deviation indicates that the component mapping is correct but the fault classification selection is incorrect; Feature attribution layer bias indicates that the mapping relationship between features and components is itself incorrect; Quantification layer bias indicates a deviation in the judgment of severity.

[0022] Furthermore, the method for associating and matching structured verification data with inference nodes is as follows: The target component identifier in the structured verification data is matched with the output component identifier of the feature and component mapping node, and the physical quantity corresponding to the verification type is matched with the feature type extracted by the feature deviation node. After a successful match, a verification label is attached to the corresponding inference node.

[0023] The above two-level matching establishes a one-to-one correspondence between maintenance verification information and corresponding nodes in the inference chain, enabling subsequent comparisons to determine which inference stage the deviation occurred in.

[0024] Furthermore, the process of backtracking layer by layer to locate the inference node where the deviation first occurs includes: When the fault type output by the conclusion output node is inconsistent with the actual fault type, backtrack to the fault mode matching node to determine whether the actual fault type is in the candidate fault mode list output by that node. If the actual fault type is in the candidate list but is not selected as the final conclusion, the judgment bias occurs at the conclusion output node and is identified as classification layer bias. If the actual fault type is not in the candidate list, the process continues to backtrack to the feature and component mapping node to determine whether the candidate fault component output by the node is consistent with the actual fault component. If they are consistent, the deviation is determined to occur at the fault mode matching node; if they are inconsistent, the deviation is determined to occur at the feature and component mapping node and is identified as a feature attribution layer deviation.

[0025] Furthermore, when a deviation occurs at a feature-to-component mapping node, the system traces back to the feature deviation node to determine whether a pre-defined causal relationship exists between each deviating feature and the actual faulty component. If it exists, the type of deviation is determined to be the feature-component mapping weight deviation, indicating that the causal association rule exists but the weight allocation fails to correctly integrate the contributions of each feature; If all significant deviation features are not directly causally related to the actual faulty component, then the deviation type is determined to be feature extraction deviation, indicating that the feature deviation node misjudges normal fluctuations or noise as abnormal features, and the problem occurs in the feature extraction stage.

[0026] Furthermore, for diagnostic samples where the diagnostic conclusions output by the conclusion output nodes are consistent with the actual faults, the logical consistency between the local confidence and the final confidence of each inference node in the diagnostic inference chain is verified. When the absolute value of the difference between the final confidence level and the geometric mean of the local confidence levels of each node exceeds a preset threshold, the diagnostic sample is marked as a confidence inconsistency sample.

[0027] This verification process identifies situations where the diagnostic conclusion is correct but the confidence level at a certain node in the reasoning process is abnormal, thus preventing such samples from being directly used for positive model reinforcement and masking potential biases in the reasoning process.

[0028] S4. Store the deviation records generated by each causal tracing comparison into the deviation database. The deviation records include the deviation type, the inference node identifier where the deviation occurred, the original output conclusion, and the target output conclusion determined based on the verification information.

[0029] The verification information and the comparison results of the diagnostic reasoning chain during the maintenance process are continuously accumulated in a unified format to form a structured deviation sample set that can be classified and statistically analyzed according to deviation type and reasoning node.

[0030] S5. Aggregate and statistically analyze the deviation records in the deviation database according to the deviation type and the inference node identifier. When the cumulative number of similar deviation records of the same inference node reaches a preset number threshold, and the proportion of the number of similar deviation records to the total number of diagnoses of the inference node exceeds a preset proportion threshold, trigger the targeted correction process for the inference node.

[0031] By setting dual trigger conditions of quantity and proportion, correction is only initiated after the accumulated biased samples reach statistical significance, thus avoiding unnecessary disturbances to model parameters due to individual random deviations.

[0032] S6. Based on the deviation type corresponding to the deviation record that triggered the correction, perform differentiated targeted correction on the model parameters of the corresponding inference node in the diagnostic inference chain. Specifically, to address classification layer bias, contrastive learning sample pairs are constructed to adjust the output weights of the fault classification boundary; To address the bias in feature attribution, adjust the mapping weights between features and components, or add working condition constraints to the mapping rules; To address the bias in the quantification layer, the severity mapping function is regressed and calibrated using measured verification values ​​as labels.

[0033] The aforementioned differentiated targeted correction only adjusts the parameters of the specific inference steps where systematic deviations occur in the system, while keeping other parts of the diagnostic model unchanged, thus achieving precise localized optimization.

[0034] Furthermore, the differentiated orientation correction also includes: To address the bias in feature attribution, a rule conflict report is generated that includes the original features and component mapping rules and conflict evidence. After expert confirmation, the mapping weights are redistributed or the operating condition constraints are supplemented. To address feature extraction bias, the original sensor data fragments of the biased samples are traced back, and the anomaly detection threshold or enhanced filtering parameters of the feature deviation nodes are adaptively adjusted.

[0035] Furthermore, the method also includes a confidence calibration step: using samples that are determined to be diagnostically accurate in step S3, the original confidence values ​​output by the diagnostic model are divided into multiple intervals, the actual diagnostic accuracy is calculated in each interval, and a calibration mapping function is constructed to map the original confidence to the actual accuracy of the corresponding interval.

[0036] By calibrating the confidence level based on historical accuracy, the confidence level output by the diagnostic model approaches the actual diagnostic accuracy of the corresponding confidence level interval, making the confidence level a practical reference for users.

[0037] The technical effects and advantages of this invention are reflected in the following aspects: First, the deviation location of the diagnostic conclusion has the precision of the link level. By synchronously outputting the diagnostic reasoning chain that records the complete reasoning path during diagnosis, the input data, output conclusion, and local confidence of each reasoning node are explicitly recorded. When the diagnostic conclusion is inconsistent with the actual fault, the comparison and verification are carried out layer by layer along the reasoning chain from the conclusion node to the feature node. It can be determined that the deviation occurs at the feature extraction node, the feature and component mapping node, or the fault mode matching node. The deviation types include classification layer deviation, feature attribution layer deviation, and quantization layer deviation.

[0038] Second, the diagnostic model's corrections are targeted at specific stages. Deviation records are aggregated and statistically analyzed by deviation type and node type. Once a trigger condition is met, targeted corrections are performed only on inference nodes with systematic deviations. This includes adjusting the boundary weights of the classification scoring function, adjusting the mapping weights between features and components, adding operational constraints to the mapping rules, and performing regression calibration on the severity mapping function, rather than retraining the entire model.

[0039] Third, the diagnostic model possesses the ability to continuously self-evolve with increasing usage. Automatically extracting maintenance verification information from remote guidance conversations does not increase the operational burden on on-site personnel. The comparison results between the verification information and the diagnostic inference chain are continuously accumulated. After the deviation records are aggregated and triggered for correction, the corresponding inference links of the model are optimized in a targeted manner, enabling the diagnostic model to gradually improve its inference accuracy and confidence during the long-term operation of the press on the stamping production line. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the overall implementation of the method of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1 As attached Figure 1 The method for remote diagnosis and maintenance support of electromechanical equipment shown is applied to the remote operation and maintenance scenario of press electromechanical equipment in a stamping production line.

[0043] The press is equipped with angle encoders, vibration sensors, current sensors, and pressure sensors. Data from each sensor is uploaded to a remote diagnostic platform in real time via an edge computing gateway. On-site maintenance personnel are equipped with augmented reality glasses, enabling them to establish audio and video conversations with remote experts. The augmented reality glasses capture on-site audio and video in real time and upload them to the platform.

[0044] The remote diagnostic platform is equipped with diagnostic models, a causal relationship rule base, a fault determination rule base, and a baseline feature base, specifically: The diagnostic model consists of four sub-models arranged sequentially, with each sub-model corresponding to a reasoning node in the diagnostic reasoning chain. The causal association rule base stores the mapping relationship between feature deviation type and candidate faulty parts of the press, as well as the initial weight value of each mapping relationship; The fault determination rule base stores the determination thresholds corresponding to each fault type. For example, if the remaining thickness of the friction plate is less than 1.5 mm, it is determined to be excessive wear. The baseline feature library stores the normal fluctuation range and standard deviation of each feature, which are statistically derived from the historical normal operation data of the equipment.

[0045] The following section provides a detailed explanation of the complete implementation process of this method, using the typical failure scenario of excessive wear of the clutch friction plate in a press as an example.

[0046] S1. When the remote diagnostic platform detects an abnormality in the pressure machine and triggers a diagnostic task, the diagnostic model outputs a diagnostic reasoning chain while generating a diagnostic conclusion.

[0047] The platform continuously analyzes operational data from various sensors on the press. When it detects multiple abnormal signs, such as a drift in the average braking angle from the normal baseline, prolonged braking response time, and increased engagement impact vibration amplitude, the platform determines that the equipment is in an abnormal state and triggers a diagnostic task.

[0048] After the diagnostic task is triggered, the diagnostic model begins to perform inference. The four sub-models sequentially complete feature deviation calculation, component mapping, fault mode matching, and conclusion output. Each sub-model generates a corresponding inference node after completing the inference.

[0049] First, the feature deviation sub-model retrieves raw data from the angle encoder, vibration sensor, and current sensor from the data buffer. It then extracts feature values ​​according to preset rules, including changes in vibration spectrum energy, current waveform slope, pressure pulsation amplitude, and braking angle drift. This sub-model then retrieves the normal fluctuation range and standard deviation corresponding to each feature from the baseline feature library, compares the current feature value with the normal baseline value one by one, and calculates the deviation of each feature.

[0050] After the calculation is completed, the sub-model generates a feature deviation node, and the node records the following information: The input data source is the address and length of the sensor's raw data in the data buffer; The output conclusion is the specific numerical value of the deviation of each feature; The local confidence level is assessed by the sub-model based on the relationship between the eigenvalue deviation magnitude and the historical standard deviation; The causal association rule used is identified by the number of the feature extraction rule called by the sub-model; The node type is identified as feature deviation.

[0051] In a specific example, the output conclusion of the feature deviation node record is: braking angle deviation of 0.8 degrees, response time deviation of 65 milliseconds, vibration spectrum energy increase of 42%, local confidence level of 0.95, and the rule on which it is based is identified as R001.

[0052] Next, the feature-to-component mapping sub-model receives the aforementioned feature deviations as input and retrieves the corresponding mapping rules from the causal association rule base. Each mapping rule defines a mapping relationship between a specific set of feature deviations and a candidate faulty component, assigning different weights to each feature within the rule. The sub-model calculates a score for each candidate component, obtained by multiplying each feature deviation by its initial weight value in the rule and then summing the results. The sub-model selects the component with the highest score as output and generates a feature-to-component mapping node.

[0053] This node records: The input data source is the node identifiers of the feature-deviation nodes; The output conclusion is the component identifier of the candidate faulty component; The local confidence score is obtained by normalizing the difference between the highest and second-highest scores. The causal association rule used is identified by the number of the mapping rule; The node type is identified as a component mapping.

[0054] In a specific example, the sub-model retrieved mapping rule R017, which points to the candidate faulty component, the clutch and brake assembly. This rule involves three features and their weights: brake angle deviation (weight 0.4), response time deviation (weight 0.35), and vibration amplitude increase (weight 0.25). The clutch and brake assembly scored the highest among all candidate components and was output as the candidate faulty component, with a local confidence level of 0.88.

[0055] Then, the fault mode matching sub-model takes the candidate faulty component identifier as input and queries the fault mode library for all fault modes corresponding to that component. For each fault mode, the sub-model calculates the cosine similarity between the standard feature deviation vector of the fault mode and the current actual feature deviation vector, using the cosine similarity as the matching degree. The sub-model sorts the fault modes from high to low matching degrees, selects the top three fault modes with the highest matching degrees to form a candidate fault mode list, and generates a fault mode matching node.

[0056] This node records: The input data source is the node identifiers of the feature-component mapping nodes; The output results are a list of candidate failure modes and the matching degree values ​​of each mode; The local confidence score is the highest matching score in the candidate list. The causal association rule used as the basis is identified by the matching rule number; The node type is identified as pattern matching.

[0057] Continuing with the previous example, the fault modes corresponding to the clutch and brake assembly include excessive wear of the friction plates, fatigue of the return spring, and oil contamination on the friction plate surface. Matching calculations yielded a matching degree of 0.82 for excessive wear of the friction plates, 0.31 for fatigue of the return spring, and 0.15 for oil contamination on the friction plate surface. These three fault modes and their respective matching degrees constitute the candidate fault mode list output, with a local confidence level of 0.82.

[0058] Finally, the conclusion output sub-model takes the candidate fault mode list as input and performs a comprehensive score on each candidate fault mode in the list. The scoring considers both the matching degree of the fault mode and its frequency of occurrence in historical diagnostic records. The sub-model selects the fault mode with the highest score as the final diagnostic conclusion, transforms the score value, outputs it as the overall confidence score, and generates a conclusion output node.

[0059] This node records: The input data source is the node identifier of the fault mode matching node; The output conclusions are the final fault type and the overall confidence level. Local confidence level equals overall confidence level; The causal association rule identifier used is the same as the fault mode matching node. The node type is identified as the conclusion output.

[0060] In this example, the conclusion output sub-model selects excessive wear of the friction plate as the final diagnostic conclusion, with an overall confidence level of 0.78.

[0061] The four inference nodes are interconnected in the order of their generation via the input data source field, forming a directed acyclic graph (DAG). This DAG constitutes the diagnostic inference chain for this diagnostic event. The platform assigns a unique identifier to this diagnosis and stores this identifier in association with the entire inference chain.

[0062] As another implementation, the diagnostic model can also be implemented using a single end-to-end neural network structure, such as a Transformer encoder, with multi-sensor time-series data as input and directly outputting the fault type and confidence level. In this structure, the platform extracts intermediate representations from each layer of the model to generate inference nodes during forward inference. The attention weights corresponding to the features of each sensor are extracted from the attention layer, and features with weights greater than a preset threshold are identified as significant features to generate feature deviation nodes. The hidden layer vectors are extracted from the intermediate coding layer and mapped to the component identifiers of the candidate faulty components by the auxiliary decoder to generate feature-component mapping nodes; The top three fault modes with the highest softmax scores are extracted from the classification layer to form a candidate list and generate fault mode matching nodes. The final classification result and confidence level are extracted from the output layer to generate conclusion output nodes. Each node also records the input source, local confidence level, and rule identifier, forming a complete diagnostic reasoning chain.

[0063] S2. After the diagnostic report is generated, the remote diagnostic platform issues a repair work order, which includes an identifier for the current diagnostic event. On-site repair personnel, wearing augmented reality (AR) devices, establish an audio-visual session with a remote expert. During session initialization, the AR device extracts the diagnostic event identifier from the work order and uploads it to the platform. The platform then binds this identifier to the current session to ensure that subsequent verification information is associated with the correct diagnostic event.

[0064] During remote expert-guided repairs, the augmented reality terminal captures and uploads the audio stream of the conversation in real time. The platform calls a speech recognition service to convert the audio stream into a text sequence. For each remote expert instruction in the text sequence, the platform uses a pre-trained intent classification model to identify its operational intent category. Operational intent categories include instructing to open a specified part, instructing to measure a specified physical quantity, instructing to visually inspect the status of a specified component, and instructing to perform a specified functional test. Specified physical quantities include clutch friction plate thickness, brake clearance, slider guide clearance, or main motor bearing clearance.

[0065] After identifying the type of operation intent, the platform extracts the name of the target component from the instruction text. The platform maintains a component tree for the press equipment, which organizes all serviceable components of the press in a hierarchical structure. The platform uses a maximum length matching algorithm based on this component tree to map the extracted component names to the corresponding standard component identifiers.

[0066] Within a preset time window after the command is issued, the platform extracts verification result information from the voice feedback text of the on-site maintenance personnel, using the corresponding extraction template according to the operation intent category. After extraction, the platform combines the target component identifier, verification type, verification result value, and diagnostic event identifier into a structured verification data set.

[0067] In a specific example, a remote expert issues the instruction, "Remove the clutch end cover and measure the remaining thickness of the friction plate with calipers." The intent classification model identifies this as measuring a specified physical quantity, specifically friction plate thickness. The platform extracts "friction plate" from the instruction and maps it to the standard part identifier P023 via a part tree. Approximately 25 seconds after the instruction is issued, the on-site maintenance personnel report, "Measured, 0.8 mm remaining." The platform uses a measurement template to extract the value 0.8 and the unit millimeters from this feedback. The target part identifier P023, the verification type thickness measurement, the verification result value 0.8 mm, and the diagnostic event identifier are combined into structured verification data.

[0068] The length of the preset time window can be dynamically adjusted according to the type of operation intent. The platform maintains a correspondence between intent types and window lengths: 60 seconds for opening a specified part, 25 seconds for measuring a specified physical quantity, 15 seconds for visually inspecting the status of a specified component, and 20 seconds for performing a specified function test. After intent recognition is completed, the platform looks up the correspondence based on the recognition results and sets the length of the time window for this extraction.

[0069] S3. The platform uses the diagnostic event identifiers carried in the structured verification data to retrieve the corresponding diagnostic reasoning chain, and then performs association matching, actual fault determination, causal source comparison and deviation location.

[0070] Association matching is performed in two levels: At the first level, the target component identifier and features in the structured verification data are compared with the component identifiers output by the component mapping node. If the target component is a sub-component of the output component in the component tree, it is considered a successful match. For example, if the target component P023 is a clutch friction plate and the output component of the mapping node is the clutch and brake assembly, a successful match is determined because P023 is a sub-node of the clutch and brake assembly in the component tree.

[0071] The second level involves matching the physical quantity corresponding to the verification type with the feature type extracted from the feature deviation node. The platform has a pre-set physical quantity-feature lookup table, which stores the correspondence between measurable physical quantities and sensor features reflecting changes in those quantities. For example, changes in friction pad thickness are reflected by changes in vibration spectrum energy, and changes in brake clearance are reflected by brake angle offset. After both levels of matching are successful, a verification label is attached to the corresponding feature deviation node, with the label content being the verification result value.

[0072] After completing the association matching, the platform determines the actual fault. The platform checks each rule in the fault judgment rule base one by one, compares the verification result value with the triggering condition of the rule, and the fault type corresponding to the rule whose triggering condition is met is the actual fault type. At the same time, the actual faulty component and the actual severity are determined.

[0073] Continuing with the previous example, the verification result is 0.8 mm of remaining friction plate thickness. The rule stored in the fault determination rule base is: if the remaining thickness is less than 1.5 mm but greater than 0.5 mm, it is determined to be excessive wear of the friction plate, with the corresponding severity being severe wear. 0.8 mm triggers this rule, the platform determines the actual fault type to be excessive wear of the friction plate, the actual faulty component to be the clutch friction plate, and the actual severity to be severe wear.

[0074] Once the actual fault is identified, the platform begins performing a causal source comparison. First, it compares whether the fault type output by the conclusion output node matches the actual fault type.

[0075] In this example, the diagnosis is also that the friction plate is excessively worn, and since both are consistent, this diagnosis is marked as accurate.

[0076] After marking the diagnosis as accurate, the platform further verifies the consistency of confidence levels within the diagnostic inference chain. The local confidence levels of the four inference nodes are denoted as follows: , , , ,in The platform calculates the geometric mean of the local confidence scores of the first three nodes, representing the overall confidence score of the output node. ,Right now Then calculate the deviation metric. for and The absolute value of the difference .

[0077] Will Deviation from preset threshold Comparison: like The diagnostic sample is marked as a confidence inconsistency sample. For the inference node involved, its weight in subsequent confidence fusion calculations is temporarily reduced until the node type to which the node belongs is corrected and restored to its original weight. like If the confidence level is consistent internally, the sample is considered a positive sample for subsequent confidence level calibration.

[0078] Substitute the specific values ​​from this example: , , , ,calculate , .set up ,because The confidence level of this diagnosis was consistent internally.

[0079] In addition to the consistency ratio of the diagnostic conclusions mentioned above, the platform also independently determines whether there is a quantification layer bias in this diagnosis. The platform compares the severity predicted by the diagnostic model with the actual severity determined based on the validation results, classifying severity into three levels: slight wear, moderate wear, and severe wear. In this example, the diagnostic model predicted moderate wear, while the actual wear was severe. This discrepancy indicates a quantification layer bias, with the bias occurring in the severity mapping portion of the conclusion output node.

[0080] The above describes the handling procedure when the diagnostic conclusion matches the actual fault. The following describes the handling procedure when the diagnostic conclusion does not match the actual fault.

[0081] In other diagnostic events, the diagnostic conclusion may differ from the actual fault. Suppose that in another diagnosis, the diagnostic model outputs the conclusion of return spring fatigue with an overall confidence level of 0.71. Following the same process, the maintenance verification information measures the remaining thickness of the friction plate to be 0.9 mm. The platform determines the actual fault type as excessive wear of the friction plate, and the actual faulty component as the clutch friction plate, based on the fault determination rules. The fault type output by the conclusion output node is inconsistent with the actual fault type, and the platform initiates a layer-by-layer backtracking process.

[0082] First, we trace back to the fault mode matching node. The candidate fault mode list output by this node shows: excessive friction plate wear (matching score 0.82), return spring fatigue (matching score 0.31), and friction plate surface oil contamination (matching score 0.15). The actual fault type, excessive friction plate wear, is in this candidate list, but the conclusion output node did not select it as the final conclusion. Therefore, the deviation type is determined to be classification layer deviation, and the deviation link is the conclusion output node. The original output conclusion is return spring fatigue, and the target output conclusion is excessive friction plate wear.

[0083] In another scenario, if the actual fault type is not in the candidate fault mode list, the process continues backtracking to the feature and component mapping node, comparing whether the candidate fault component output by that node matches the actual fault component: If the components are identical, it is determined to be a fault mode omission deviation, and the deviation link is the fault mode matching node; If the components are inconsistent, it is determined to be a feature attribution layer deviation, and the deviation link is the feature-component mapping node.

[0084] For cases identified as feature attribution layer bias, further backtracking to the feature deviation node is required for refined differentiation. Extract all features identified as significantly biased to form a set. The criterion for significant deviation is that the deviation exceeds twice the standard deviation of the normal fluctuation range of that feature stored in the baseline feature library. Let the set be... The total number of features is Query the causal relationship rule base, from A subset is formed by selecting features that have a predetermined causal relationship with the actual faulty component. Subset The number of features is Calculate the coverage of associated features Set a coverage threshold. :like If it is determined to be a deviation in the mapping weights between features and components; This was determined to be a feature extraction bias.

[0085] To illustrate the coverage calculation process, a specific example is given below: Suppose that in a certain feature attribution layer deviation determination, four significant deviation features are extracted from the feature deviation node, forming a set S = {braking angle deviation, response time deviation, vibration amplitude increase, and current waveform slope change}. A query of the causal relationship rule base revealed that three features—braking angle deviation, response time deviation, and increased vibration amplitude—are causally correlated with the actual faulty components, the clutch and brake assembly, forming a subset. , .calculate ,set up ,because This is determined to be a deviation in the mapping weights between features and components. In another scenario, if only the increase in vibration amplitude is causally related to a feature and the actual faulty component, then... ,but ,because This was determined to be a feature extraction bias.

[0086] As an alternative implementation, the aforementioned layer-by-layer backtracking process can also be replaced by a parallel comparison method. The platform simultaneously compares the output conclusions of the fault mode matching node, the feature-to-part mapping node, and the feature deviation node with the actual values.

[0087] If the candidate fault mode list contains actual fault types, it is determined to be a classification layer bias. Otherwise, if the candidate faulty component matches the actual faulty component, it is determined to be a fault mode omission deviation; Otherwise, it is judged as a feature attribution layer bias.

[0088] Further subdivision of the feature attribution layer deviation still employs the aforementioned method of determining associated feature coverage. This method reduces the number of backtracking steps and is suitable for scenarios with long inference chains or those requiring rapid deviation localization. After deviation localization is completed, subsequent processing follows the same procedure as the main process.

[0089] S4. After each causal attribution comparison, the platform generates a deviation record and stores it in the deviation database. Each deviation record includes the deviation type, the inference node identifier where the deviation occurred, the original output conclusion, and the target output conclusion determined based on the verification information. For samples with accurate diagnosis and no quantification-level deviation, no deviation record is generated, but a positive sample label record is generated for subsequent confidence calibration. Positive sample label records and deviation records are stored in different partitions of the deviation database and are distinguished by different record type identifiers.

[0090] Taking the classification layer deviation scenario above as an example, the generated deviation record is as follows: the deviation type is classification layer deviation, the inference node where the deviation occurs is the conclusion output, the original output conclusion is return spring fatigue, and the target output conclusion is excessive wear of the friction plate. For the case in the aforementioned main example where the diagnosis is accurate but there is a quantization layer deviation, the generated deviation record is as follows: the deviation type is quantization layer deviation, the inference node where the deviation occurs is the conclusion output, the original output conclusion is moderate wear, and the target output conclusion is severe wear.

[0091] S5. The platform groups and statistically analyzes the accumulated deviation records in the deviation database according to two dimensions: deviation type and node type identifier. For any group, the cumulative number of deviations within that group is denoted as... The total number of diagnostics performed on this node type since the system began is [number missing]. The platform sets a quantity threshold. and proportional threshold ,when and When this occurs, it indicates that the node has exhibited a statistically significant systematic deviation in this type of deviation, triggering a targeted correction process for that node type.

[0092] Taking a specific aggregation-triggered process as an example: the classification layer bias records of the conclusion output node are accumulated one by one in the early stages of system operation. Assume... The value is 10. The value is 15%, representing the cumulative number of deviations for that group when the 10th classification layer deviation record is generated. The threshold has been reached. At this point, the total number of diagnoses for this node type has been reached. , deviation ratio If the proportion exceeds the threshold of 15%, and both conditions are met simultaneously, the platform triggers a classification layer bias-oriented correction process for the conclusion output node.

[0093] S6. After triggering the directional correction, the platform performs differentiated directional correction on the parameters of the corresponding sub-model according to the deviation type. The correction methods corresponding to each deviation type are explained below.

[0094] The corresponding classification layer bias is corrected by adjusting the classification scoring function of the output sub-model. Let the feature vector of a biased sample be... The correct failure mode is The fault mode of the original error output is The classification scoring function is Define the marginal loss function: ; in This is a preset marginal value. The platform aims to minimize... For the target scoring function The parameters are updated using gradient descent. After the update, the scores of correct patterns increase while the scores of incorrect patterns decrease, thereby correcting the classification ranking boundary.

[0095] The correction targets the weight matrix within the feature-part mapping sub-model to address the deviation in feature-part mapping weights. Let the feature deviation vector be... Candidate component score vector Depend on and Multiply to obtain. Calculate the cross-entropy loss using the actual faulty components identified in the biased samples as the target labels, and determine the relationship between this loss and the target. The gradient, along the negative gradient direction, is... An update is performed, with the update step size controlled by a preset learning rate. After the correction is complete, the updated weight value replaces the weight value stored in the causal association rule base for that mapping rule, and serves as the weight value for that rule in subsequent diagnostics.

[0096] Corresponding to feature extraction bias, the correction target is the anomaly detection threshold of the feature deviation sub-model. The deviation amount of each misclassified feature in the biased samples is extracted, and the current detection threshold is set to... The deviation of the misjudged point is The new threshold Set as Add a tiny positive quantity, for example, take The threshold is set at 5% to ensure that the adjusted threshold is higher than the misjudgment point, thus avoiding normal fluctuations of the same magnitude from being misjudged as abnormal again in subsequent diagnoses.

[0097] The corresponding quantification layer deviation is corrected using the severity mapping function. Let the comprehensive deviation index be... This is a weighted sum of deviations from each feature, with the weights of each feature obtained from a causal association rule base. The mapping function is in linear form, let... This represents the severity level numerical value. and For mapping parameters, i.e. The platform collects multiple deviation samples that trigger corrections, with each sample providing a pair of combined deviation values. And the actual severity level numerical value The least squares method was used to fit these samples and solve for the parameters. and The updated value. When the number of biased samples is small, a single-step gradient update method is used, updating the value of a single sample. and Find the loss pairs separately and The partial derivatives are updated one step along the negative direction of the partial derivatives with a preset learning rate.

[0098] For any omissions or deviations in the corresponding fault modes, the platform generates a rule conflict report. This report includes the identifier and content of the original mapping rule, the feature deviation mode in the current deviation sample, the omitted correct fault mode, and the verified correct component identifier. The report is submitted to the expert review interface. After expert confirmation, the correspondence between the feature deviation mode and the correct fault mode is added to the fault mode library, or the existing matching rules are adjusted.

[0099] This method also includes a confidence calibration step, which is executed independently and in parallel with the aforementioned orientation correction process. The platform continuously collects positive samples that are determined to be diagnostically accurate by S3, and extracts the original confidence values ​​output by the diagnostic model from each sample. The range of the original confidence values ​​is divided into multiple consecutive left-closed and right-open intervals. One specific division method is 10 equal-width intervals, each with a width of 0.1.

[0100] For each interval, the total number of samples falling into that interval and the number of accurately diagnosed samples within that interval are counted to calculate the actual diagnostic accuracy for that interval. If the cumulative number of samples in an interval is zero, the calibration output value for that interval temporarily uses the actual accuracy of the adjacent interval. Based on this, the platform establishes a calibration mapping table, and when the sample size is insufficient in the initial stage of system operation, the original confidence value is directly output for each interval.

[0101] In each subsequent diagnosis, after the diagnostic model outputs the original confidence value, it first determines the interval in which the value falls, looks up the actual accuracy corresponding to the interval from the mapping table, and uses the obtained accuracy as the output of the calibrated confidence, thus completing the conversion from the original confidence to the calibrated confidence.

[0102] The steps S1 to S6 and the confidence calibration together constitute a self-evolving closed-loop system. After each remote diagnostic and maintenance support task is completed, the diagnostic reasoning chain is fully recorded, and the verification information naturally generated during the maintenance process is automatically extracted and structured. Causal tracing and comparison can accurately identify which reasoning stage the deviation occurred in and what type of deviation it belongs to.

[0103] Deviation records are aggregated and statistically analyzed. When a deviation becomes statistically significant, targeted correction is triggered, adjusting parameters only for the erroneous inference steps. As the system continues to run, the accuracy of the diagnostic model improves step by step in feature extraction, component mapping, fault classification, and severity assessment. The output confidence level also gets closer and closer to the true accuracy, achieving continuous self-evolution of the remote diagnostic capability for the press equipment of the stamping production line.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., 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 remote diagnostics and maintenance support of electromechanical equipment, characterized in that, Includes the following steps: S1. When the remote diagnostic platform detects an abnormality in the press in the stamping production line and triggers a diagnostic task... The diagnostic model outputs a diagnostic reasoning chain while generating a diagnostic conclusion; The diagnostic reasoning chain records the complete reasoning path from sensor feature deviation to the final diagnostic conclusion in the form of a directed acyclic graph; The diagnostic inference chain contains multiple inference nodes. Each inference node records the source of its input data, the output conclusion, the local confidence level, and the causal association rule identifier on which it is based. S2. During the process of remote experts guiding on-site maintenance personnel to troubleshoot and repair the press using augmented reality terminals. Acquire the audio stream of the remote guidance session in real time and convert it into a text sequence; The text sequence is subjected to instruction intent recognition to determine the operation intent category corresponding to the remote expert instruction; Extract the target component name from the instruction text and map it to the standard component identifier; The verification result information is extracted from the voice feedback text of the on-site maintenance personnel within a preset time window after the instruction is issued. The extracted verification information is organized into structured verification data, and each piece of structured verification data includes the target component identifier, verification type, and verification result value; S3. Associate and match the structured verification data with the inference nodes in the diagnostic inference chain; Starting from the conclusion output node of the diagnostic reasoning chain, perform causal tracing and comparison node by node along the reasoning path from bottom to top; When the diagnostic conclusion output by the conclusion output node is inconsistent with the actual fault determined based on the structured verification data; Backtrack layer by layer to locate the inference node where the deviation first occurs, and determine the type of deviation based on the location where the deviation occurs; The types of deviations include: Classification layer deviation indicates that the component mapping is correct but the fault classification selection is incorrect; Feature attribution layer bias indicates that the mapping relationship between features and components is itself incorrect; Quantification layer bias indicates a deviation in the severity assessment; S4. Store the deviation records generated from each causal tracing comparison into the deviation database. The deviation record includes the deviation type, the inference node identifier where the deviation occurred, the original output conclusion, and the target output conclusion determined based on the verification information; S5. Aggregate and statistically analyze the deviation records in the deviation database according to the deviation type and inference node identifier; When the cumulative number of similar deviation records of the same inference node reaches a preset number threshold, and the proportion of the number of similar deviation records to the total number of diagnoses of the inference node exceeds a preset proportion threshold; Trigger a targeted correction process for this inference node; S6. Based on the deviation type corresponding to the deviation record that triggered the correction, perform differentiated directional correction on the model parameters of the corresponding inference node in the diagnostic inference chain; Specifically, to address classification layer bias, contrastive learning sample pairs are constructed to adjust the output weights of the fault classification boundary; To address the bias in feature attribution, adjust the mapping weights between features and components, or add working condition constraints to the mapping rules; To address the bias in the quantification layer, the severity mapping function is regressed and calibrated using measured verification values ​​as labels.

2. The method for remote diagnostics and maintenance support of electromechanical equipment according to claim 1, characterized in that, The reasoning nodes of the diagnostic reasoning chain mentioned in step S1 include: Feature deviation node: Records the feature value extracted from the original sensor data and the deviation of the feature value from the normal baseline value. The feature value includes the change in vibration spectrum energy, the change in current waveform slope, the change in pressure pulsation amplitude, or the brake angle drift. The feature-component mapping node records the reasoning process and mapping rules used to map the feature deviation to candidate faulty components of the press. The fault mode matching node records the process of matching fault modes within the range of candidate faulty components and the matching degree of each fault mode. The conclusion output node records the final diagnostic conclusion and the overall confidence level.

3. The method for remote diagnostics and maintenance support of electromechanical equipment according to claim 1, characterized in that, The operational intent categories mentioned in step S2 include instructing to open a specified part of the press, instructing to measure a specified physical quantity, instructing to visually inspect the status of a specified component, and instructing to perform a specified functional test; The specified physical quantities include clutch friction plate thickness, brake clearance, slider guide clearance, or main motor bearing clearance. The extraction of verification result information adopts the corresponding extraction template according to the operation intention category. The verification result information includes the measured value and its unit, the observed component status description or the functional test result value.

4. The method for remote diagnosis and maintenance support of electromechanical equipment according to claim 1, characterized in that, The method for associating and matching structured verification data with inference nodes in step S3 is as follows: Match the target component identifiers and features in the structured verification data with the output component identifiers of the component mapping nodes; And match the physical quantity corresponding to the verification type with the feature type extracted from the feature deviation node; Once a match is successful, a verification tag is attached to the corresponding inference node.

5. The method for remote diagnosis and maintenance support of electromechanical equipment according to claim 1 or 4, characterized in that, The process of backtracking layer by layer to locate the inference node where the deviation first occurs, as described in step S3, includes: When the fault type output by the conclusion output node is inconsistent with the actual fault type, backtrack to the fault mode matching node to determine whether the actual fault type is in the candidate fault mode list output by that node. If the actual fault type is in the candidate list but is not selected as the final conclusion, the judgment bias occurs at the conclusion output node and is identified as classification layer bias. If the actual fault type is not in the candidate list, continue backtracking to the feature-component mapping node to determine whether the candidate faulty component output by that node matches the actual faulty component: If they match, the discrepancy is determined to occur at the fault mode matching node; If there is a discrepancy, the deviation is determined to occur at the feature-component mapping node and is identified as a feature attribution layer deviation.

6. The method for remote diagnosis and maintenance support of electromechanical equipment according to claim 5, characterized in that, When a deviation occurs at the feature-to-component mapping node, the process further backtracks to the feature deviation node to determine whether a pre-defined causal relationship exists between each deviating feature and the actual faulty component. If it exists, then the deviation type is determined to be feature-to-component mapping weight deviation; If all significant deviation features have no direct causal relationship with the actual faulty component, then the deviation type is determined to be feature extraction deviation.

7. The method for remote diagnostics and maintenance support of electromechanical equipment according to claim 1, characterized in that, Step S3 also includes: For diagnostic samples where the diagnostic conclusions output by the conclusion output node are consistent with the actual faults, verify the logical consistency between the local confidence and the final confidence of each inference node in the diagnostic inference chain. When the absolute value of the difference between the final confidence level and the geometric mean of the local confidence levels of each node exceeds a preset threshold, the diagnostic sample is marked as a confidence inconsistency sample.

8. The method for remote diagnosis and maintenance support of electromechanical equipment according to claim 1, characterized in that, The differential orientation correction in step S6 also includes: To address the bias in feature attribution, a rule conflict report is generated that includes the original features and component mapping rules and conflict evidence. After expert confirmation, the mapping weights are redistributed or the operating condition constraints are supplemented. To address feature extraction bias, the original sensor data fragments of the biased samples are traced back, and the anomaly detection threshold or enhanced filtering parameters of the feature deviation nodes are adaptively adjusted.

9. The method for remote diagnostics and maintenance support of electromechanical equipment according to claim 1, characterized in that, The method further includes a confidence calibration step: Using the samples that were determined to be diagnostically accurate in step S3, the original confidence values ​​output by the diagnostic model were divided into multiple intervals, and the actual diagnostic accuracy was calculated in each interval. A calibration mapping function is constructed to map the original confidence level to the actual accuracy of the corresponding interval.

10. The method for remote diagnosis and maintenance support of electromechanical equipment according to claim 2, characterized in that, The causal association rule identifier mentioned in step S1 points to the corresponding rule in the preset causal association rule base; The causal association rule base stores the mapping relationship between feature deviation type and candidate faulty parts of the press, as well as the initial weight value of each mapping relationship.

Images