AI-based Fault Root Cause Analysis Methods and Systems
By constructing an acyclic dynamic causal topology graph and using counterfactual virtual zeroing intervention, the problem of causal topology misjudgment in power equipment was solved, achieving accurate and stable location of fault root causes and improving the accuracy and anti-interference capability of fault assessment.
Patent Information
- Application Number
- CN202611366199.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-04
- Publication Date
- 2026-11-17
AI Technical Summary
In power equipment and multi-axis transmission machinery, existing technologies struggle to construct acyclic causal topologies while considering the physical propagation delay constraints of the transmission medium, leading to misjudgments of fault causes. Furthermore, it is difficult to block pseudo-causal loops under dynamic operating conditions, affecting the accuracy and stability of fault root cause analysis.
By acquiring multi-channel sensing time-series signals, a global multidimensional state tensor is constructed, the conditional transfer entropy and physical propagation delay lower limit between measurement points are calculated, an acyclic dynamic directed causal topology graph is generated, and a counterfactual virtual zeroing intervention is performed using a graph neural network model to calculate the fault root cause marginal contribution index and comprehensive root cause confidence, and output closed-loop adjustment parameters.
It effectively filters out false correlation edges, reduces the masking of source faults by downstream anomalies, improves the accuracy and stability of fault root cause localization, reduces dynamic operating condition interference, and ensures accurate identification of fault propagation paths and causal reasoning.
Smart Images

Figure CN123115869A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an AI-based method and system for fault root cause analysis, belonging to the field of industrial equipment fault prediction and health management technology. Background Technology
[0002] In the operation of power equipment and multi-axis transmission machinery, multi-channel sensors are usually deployed in key components to collect time-series signals such as vibration acceleration, drive current, oil supply pressure and end cover temperature. Then, neural networks or graph topology models are used to identify abnormal equipment status and infer potential fault causes. This is a common technical method in industrial equipment health status monitoring and fault root cause analysis. As the electromechanical coupling between various components of the transmission system increases, the abnormal characteristics caused by early minor damage are prone to change during transmission and spread to multiple sensing channels.
[0003] Under unsteady operating conditions such as frequent speed adjustments or dynamic load switching, localized weak anomalies can propagate to other parts along fluid pipelines, mechanical connection surfaces, and electrical circuits. Affected by factors such as mechanical damping losses, fluid flow resistance, and the heat capacity of the metal casing, different physical quantities will experience varying degrees of delay and attenuation during propagation. Measurement points located downstream and subjected to strong mechanical impact or frictional heating often exhibit higher signal fluctuation amplitudes and alarm frequencies than the source measurement points where the initial damage occurred. Conventional models based on statistical correlation analysis tend to identify downstream measurement points with more pronounced anomalies as the dominant root cause, leading to misjudgments of the source fault cause. Simply using conventional methods such as increasing the number of network layers, expanding the timing window, or adjusting the correlation threshold, without considering the actual constraints on the internal physical structure and energy flow paths of the equipment, still makes it difficult to eliminate timing co-currents caused by load switching. Linear interference is a concern. Furthermore, for reverse paths blocked by fluid check valves or statistically associated signals that do not conform to the heat conduction delay law, the model may still misjudge them as causal connections, leading to false correlations and closed-loop circuits in the generated causal topology that violate physical laws. In existing technologies, some causal analysis methods attempt to introduce temporal attention or shallow intervention mechanisms for auxiliary judgment. For example, Chinese invention patent application CN121207601A discloses an industrial equipment and its fault root cause diagnosis method and system, which uses an attention mechanism to construct a causal graph and perform static reference replacement on abnormal data. However, because its causal analysis process is still constrained by data statistical laws and does not deeply integrate the extreme value constraints of physical medium propagation delay such as mechanical wave velocity, fluid pressure wave, and structural thermal resistance, it is still difficult to block the pseudo-causal circuits generated by associated resonance under sudden changes in operating conditions.
[0004] Therefore, the technical problem to be solved by this invention is how to construct an acyclic causal topology under the constraint of physical propagation delay of transmission medium, quantify the marginal contribution of each measuring point to the abnormal state of the system by performing causal intervention deduction on candidate abnormal nodes, and then make a judgment and output closed-loop protection control parameters in combination with the confidence difference under dynamic working conditions, so as to reduce the masking of the source fault by the abnormal downstream performance and the interference caused by false correlation. Summary of the Invention
[0005] To address the problems in the background technology, the technical solution of the present invention is as follows: an AI-based fault root cause analysis method, comprising the following steps: Step S1: Obtain multi-channel sensor timing signals and discrete event alarms within the controlled equipment operation window; calculate the state deviation sequence based on the steady-state benchmark; and construct a global multi-dimensional state tensor by combining discrete event alarms. Step S2: Calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix, remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, and generate an acyclic dynamic directed causal topology graph. Step S3: Input the global multidimensional state tensor and the acyclic dynamic directed causal topology graph into the neural network model to extrapolate the baseline reconstruction error; sequentially implement counterfactual virtual zeroing intervention on the candidate nodes, cut off the downstream propagation of the intervened nodes and extrapolate the increase in the state reconstruction error of the non-intervention nodes, and calculate the fault root cause marginal contribution index of each candidate node. Step S4: Determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index; determine the target fault root cause entity based on the descending order of the comprehensive root cause confidence and the preset separation threshold, and output the closed-loop adjustment parameters for driving oil supply flow compensation and speed limiting load reduction.
[0006] Preferably, step S2 includes: step S21, within a preset time delay range, solving for the measurement point using the Gaussian kernel density estimation method. With measuring points The joint probability density distribution between the points is calculated under different discrete time delay conditions. State deviation sequence towards measurement point The information transfer entropy gain of the state deviation sequence flow; step S22, taking the maximum conditional transfer entropy under each discrete time delay as the measurement point. Pointing to the measuring point The initial correlation edge strength is used to construct an initial directed correlation matrix containing the temporal asymmetric dependencies between measurement points across the entire domain.
[0007] Preferably, step S2 further includes: step S23, obtaining the physical topology connectivity paths between the components of the controlled equipment, and establishing a physical propagation delay constraint matrix that includes the lower limit of mechanical vibration wave velocity propagation delay, the lower limit of fluid medium pressure wave propagation delay, and the lower limit of metal structure thermal resistance conduction hysteresis delay; in the initial directed correlation matrix, when the measuring point With measuring points When the time delay between the two is less than the corresponding lower limit of the physical propagation delay of the medium, or the maximum conditional transfer entropy does not reach the preset information transmission threshold, or the correlation direction between the two violates the flow prohibition of the fluid check valve, the measuring point will be... Pointing to the measuring point The weights of the directed edges are reset to zero, generating an acyclic dynamic directed causal topological graph that satisfies the properties of a directed acyclic topology.
[0008] Preferably, step S3 includes: step S31, inputting the weighted adjacency matrix of the acyclic dynamic directed causal topology graph and the global multidimensional state tensor into the spatiotemporal graph convolutional inference network, aggregating the spatial neighborhood features and temporal evolution features of the nodes, outputting the full system reconstruction prediction state under the no-intervention state, and calculating the baseline mean square reconstruction error of the full system within the operating window.
[0009] Preferably, step S3 further includes: step S32, sequentially selecting candidate nodes to be evaluated, forcibly assigning the state deviation sequence input to the candidate node to zero, and forcibly setting all outgoing edge weights of the candidate node in the acyclic dynamic directed causal topology graph to zero, thereby blocking the forward influence propagation of the candidate node to downstream successor nodes; re-inputting the global multidimensional state tensor and topology structure after intervention into the graph neural network model, deduce the counterfactual state prediction values of the non-intervention nodes in the whole system other than the intervened candidate node, and calculate the increase in mean square reconstruction error of the other non-intervention nodes.
[0010] Preferably, step S3 further includes: step S33, retaining the value when the increase in mean square reconstruction error is greater than zero, otherwise taking the value as zero; dividing the increase in mean square reconstruction error corresponding to each candidate node by the sum of the increase in mean square reconstruction error of all candidate nodes in the whole system and the sum of the preset zero-prevention smoothing constant, and calculating the normalized fault root cause marginal contribution index of each candidate node.
[0011] Preferably, step S4 includes: step S41, determining the operating condition sensitivity components based on the real-time collected operating speed and load factor of the controlled equipment, and normalizing each component to construct a dynamic sensitivity weight vector; step S42, multiplying the dynamic sensitivity weight corresponding to each measuring point with the fault root cause marginal contribution index of that measuring point, and calculating the comprehensive root cause confidence of each candidate node; sorting the comprehensive root cause confidence of all candidate nodes in descending order, and when the difference between the largest confidence in the first order and the second largest confidence in the second order reaches a preset separation threshold, determining the measuring point corresponding to the first order confidence as the dominant fault root cause entity, wherein the preset separation threshold ranges from 0.20 to 0.30.
[0012] Preferably, the multi-channel sensing timing signals include the stator current of the drive motor, the radial vibration acceleration of the bearing housing, the supply pressure of the lubrication pipe, and the temperature of the bearing end cover.
[0013] Preferably, the closed-loop control parameters include oil flow compensation and pressure boosting control parameters issued to the variable frequency drive unit of the lubricating hydraulic pump, and speed reduction and load reduction torque limiting commands issued to the spindle power drive unit.
[0014] An AI-based fault root cause analysis system, which implements an AI-based fault root cause analysis method, includes: The system includes a state tensor construction module, a causal graph generation module, a marginal contribution deduction module, and a closed-loop decision arbitration module. Among them, the state tensor construction module is used to acquire multi-channel sensor time-series signals and discrete event alarms within the operating window of the controlled equipment, calculate the state deviation sequence based on the steady-state benchmark, and construct a global multi-dimensional state tensor in combination with discrete event alarms. The causal graph generation module is used to calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; and perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix to remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, thereby generating an acyclic dynamic directed causal topology graph. The marginal contribution deduction module is used to deduce the baseline reconstruction error of the input graph neural network model from the global multidimensional state tensor and the acyclic dynamic directed causal topology graph; it sequentially implements counterfactual virtual zeroing intervention on candidate nodes, cuts off the downstream propagation of the intervened nodes and deduces the increase in the state reconstruction error of non-intervention nodes, and calculates the fault root cause marginal contribution index of each candidate node. The closed-loop decision arbitration module is used to determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index; based on the descending order of the comprehensive root cause confidence and the preset separation threshold, the target fault root cause entity is determined, and the closed-loop adjustment parameters for driving fuel supply flow compensation and speed limiting load reduction are output.
[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In AI-based fault root cause analysis, constraining causal relationships by physical propagation delay and energy flow direction can filter out false correlation edges that do not conform to actual propagation laws, and avoid attributing obvious abnormalities such as severe downstream vibration and temperature rise to early weak upstream damage. This can reduce spurious causal loops that are prone to occur in statistical correlation analysis, making the formed fault propagation relationship more consistent with the actual physical process inside the equipment, and providing a reliable basis for accurate identification of fault transmission paths and subsequent causal reasoning.
[0016] 2. By implementing counterfactual virtual zeroing intervention on candidate nodes and measuring the actual impact of each measurement point on the abnormal state of the system with the fault root cause marginal contribution index, the interference of time collinearity caused by dynamic operating conditions and load switching on the judgment results can be reduced, and the fault root cause can be avoided by judging the fault root cause solely based on the magnitude of local abnormality. This makes it easier to identify the weak source damage that is masked by severe downstream abnormalities, and improves the accuracy and timeliness of early fault root cause location under complex operating conditions.
[0017] 3. By combining the current operating speed and load to determine the dynamic sensitivity weight vector, and using the comprehensive root cause confidence to distinguish candidate nodes, the root cause judgment can be adjusted according to the actual working conditions. When multiple measurement points are disturbed at the same time or are in critical working conditions, it can reduce the misjudgment caused by transient measurement noise and occasional sensor contact failure, and maintain a clear distinction between the main candidate root cause and the secondary candidate root cause, thereby improving the stability and anti-interference ability of fault root cause judgment. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the fault location process using physical delay pruning and counterfactual causal inference in this invention. Figure 2 This is a structural diagram of the fault root cause analysis system module for multi-level data flow according to the present invention.
[0019] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.
[0021] An AI-based method for fault root cause analysis includes the following steps: Step S1: Obtain multi-channel sensor timing signals and discrete event alarms within the controlled equipment operation window; calculate the state deviation sequence based on the steady-state benchmark; and construct a global multi-dimensional state tensor by combining discrete event alarms. Step S2: Calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix, remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, and generate an acyclic dynamic directed causal topology graph. Step S3: Input the global multidimensional state tensor and the acyclic dynamic directed causal topology graph into the neural network model to extrapolate the baseline reconstruction error; sequentially implement counterfactual virtual zeroing intervention on the candidate nodes, cut off the downstream propagation of the intervened nodes and extrapolate the increase in the state reconstruction error of the non-intervention nodes, and calculate the fault root cause marginal contribution index of each candidate node. Step S4: Determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index; determine the target fault root cause entity based on the descending order of the comprehensive root cause confidence and the preset separation threshold, and output the closed-loop adjustment parameters for driving oil supply flow compensation and speed limiting load reduction.
[0022] Preferably, step S2 includes: step S21, within a preset time delay range, solving for the measurement point using the Gaussian kernel density estimation method. With measuring points The joint probability density distribution between the points is calculated under different discrete time delay conditions. State deviation sequence towards measurement point The information transfer entropy gain of the state deviation sequence flow; step S22, taking the maximum conditional transfer entropy under each discrete time delay as the measurement point. Pointing to the measuring point The initial correlation edge strength is used to construct an initial directed correlation matrix containing the temporal asymmetric dependencies between measurement points across the entire domain.
[0023] Preferably, step S2 further includes: step S23, obtaining the physical topology connectivity paths between the components of the controlled equipment, and establishing a physical propagation delay constraint matrix that includes the lower limit of mechanical vibration wave velocity propagation delay, the lower limit of fluid medium pressure wave propagation delay, and the lower limit of metal structure thermal resistance conduction hysteresis delay; in the initial directed correlation matrix, when the measuring point With measuring points When the time delay between the two is less than the corresponding lower limit of the physical propagation delay of the medium, or the maximum conditional transfer entropy does not reach the preset information transmission threshold, or the correlation direction between the two violates the flow prohibition of the fluid check valve, the measuring point will be... Pointing to the measuring point The weights of the directed edges are reset to zero, generating an acyclic dynamic directed causal topological graph that satisfies the properties of a directed acyclic topology.
[0024] Preferably, step S3 includes: step S31, inputting the weighted adjacency matrix of the acyclic dynamic directed causal topology graph and the global multidimensional state tensor into the spatiotemporal graph convolutional inference network, aggregating the spatial neighborhood features and temporal evolution features of the nodes, outputting the full system reconstruction prediction state under the no-intervention state, and calculating the baseline mean square reconstruction error of the full system within the operating window.
[0025] Preferably, step S3 further includes: step S32, sequentially selecting candidate nodes to be evaluated, forcibly assigning the state deviation sequence input to the candidate node to zero, and forcibly setting all outgoing edge weights of the candidate node in the acyclic dynamic directed causal topology graph to zero, thereby blocking the forward influence propagation of the candidate node to downstream successor nodes; re-inputting the global multidimensional state tensor and topology structure after intervention into the graph neural network model, deduce the counterfactual state prediction values of the non-intervention nodes in the whole system other than the intervened candidate node, and calculate the increase in mean square reconstruction error of the other non-intervention nodes.
[0026] Preferably, step S3 further includes: step S33, retaining the value when the increase in mean square reconstruction error is greater than zero, otherwise taking the value as zero; dividing the increase in mean square reconstruction error corresponding to each candidate node by the sum of the increase in mean square reconstruction error of all candidate nodes in the whole system and the sum of the preset zero-prevention smoothing constant, and calculating the normalized fault root cause marginal contribution index of each candidate node.
[0027] Preferably, step S4 includes: step S41, determining the operating condition sensitivity components based on the real-time collected operating speed and load factor of the controlled equipment, and normalizing each component to construct a dynamic sensitivity weight vector; step S42, multiplying the dynamic sensitivity weight corresponding to each measuring point with the fault root cause marginal contribution index of that measuring point, and calculating the comprehensive root cause confidence of each candidate node; sorting the comprehensive root cause confidence of all candidate nodes in descending order, and when the difference between the largest confidence in the first order and the second largest confidence in the second order reaches a preset separation threshold, determining the measuring point corresponding to the first order confidence as the dominant fault root cause entity, wherein the preset separation threshold ranges from 0.20 to 0.30.
[0028] Preferably, the multi-channel sensing timing signals include the stator current of the drive motor, the radial vibration acceleration of the bearing housing, the supply pressure of the lubrication pipe, and the temperature of the bearing end cover.
[0029] Preferably, the closed-loop control parameters include oil flow compensation and pressure boosting control parameters issued to the variable frequency drive unit of the lubricating hydraulic pump, and speed reduction and load reduction torque limiting commands issued to the spindle power drive unit.
[0030] An AI-based fault root cause analysis system, which implements an AI-based fault root cause analysis method, includes: The system includes a state tensor construction module, a causal graph generation module, a marginal contribution deduction module, and a closed-loop decision arbitration module. Among them, the state tensor construction module is used to acquire multi-channel sensor time-series signals and discrete event alarms within the operating window of the controlled equipment, calculate the state deviation sequence based on the steady-state benchmark, and construct a global multi-dimensional state tensor in combination with discrete event alarms. The causal graph generation module is used to calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; and perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix to remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, thereby generating an acyclic dynamic directed causal topology graph. The marginal contribution deduction module is used to deduce the baseline reconstruction error of the input graph neural network model from the global multidimensional state tensor and the acyclic dynamic directed causal topology graph; it sequentially implements counterfactual virtual zeroing intervention on candidate nodes, cuts off the downstream propagation of the intervened nodes and deduces the increase in the state reconstruction error of non-intervention nodes, and calculates the fault root cause marginal contribution index of each candidate node. The closed-loop decision arbitration module is used to determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index; based on the descending order of the comprehensive root cause confidence and the preset separation threshold, the target fault root cause entity is determined, and the closed-loop adjustment parameters for driving fuel supply flow compensation and speed limiting load reduction are output.
[0031] Example 1: When the transmission chain system of a large CNC rotary equipment is running, the state perception network deploys four physical quantity sensing nodes for the drive motor, coupling, main bearing housing, and lubrication hydraulic station, respectively. Corresponding to the effective value of the three-phase stator current of the drive motor, node The root mean square value of the radial vibration acceleration corresponding to the main bearing housing, node Corresponding to the main pipeline pressure of the lubrication circuit, node Corresponding to the infrared temperature measurement value of the main bearing end cover; the system is pre-calibrated to set the steady-state mean and standard deviation of each measuring point under the reference operating conditions as follows: A and A, m / s² and m / s², MPa and MPa, and ℃ and ℃; When the spindle is operating in high-speed, light-load mode and the oil supply pressure in the lubrication circuit decreases, the multi-channel sensor interface uses a sampling period of s continuous acquisition includes The original time-series signal sequence with discrete sampling points It acquires discrete event alarms within the same running window, using the alarm occurrence time as the event reference for timing alignment of each channel, where the measurement points... At any moment to The original measurement sequence is A, measuring point The original measurement sequence is m / s², measuring point The original measurement sequence is MPa, measuring point The original measurement sequence is ℃.
[0032] The state tensor construction module is based on the expression Calculate the dimensionless transient state deviation sequence of each measuring point on a continuous discrete time axis, where For measuring points At any moment The original physical quantity measurement value, For measuring points The steady-state baseline mean, For measuring points The steady-state baseline standard deviation; the calculated node The state deviation sequence is ,node The state deviation sequence is ,node The state deviation sequence is ,node The state deviation sequence is After aligning the four state deviation sequences according to the occurrence time of discrete event alarms, the sequences are combined to construct a global multidimensional state tensor. During this state evolution process, nodes The radial vibration of the main bearing housing, as characterized, rapidly increased to a state deviation amplitude of 9 at the fifth sampling point due to the rupture of the lubricating oil film, while the source node... The pressure state deviation is -4.4, exhibiting a strong coupled state where the amplitude of secondary vibration energy release is higher than the amplitude of the initial induced pressure drop.
[0033] The causal graph generation module solves for the joint probability density distribution between the state deviation sequences of each measuring point within a sliding time window using the Gaussian kernel density estimation method, and calculates the time delay between any two measuring points. Conditional transitive entropy under certain conditions ,in, Denotes the joint probability density function. Indicates the discrete time delay order; simultaneously retrieves the lower limit of the propagation time delay of the pressure wave in the fluid medium. Lower limit of mechanical vibration wave velocity transmission delay and the lower limit of thermal resistance conduction hysteresis delay in metal structures Physical propagation delay constraint matrix ; within the preset information transmission threshold Perform spatiotemporal boundary filtering and physical pruning under bit constraints for path ,exist The calculated conditional propagation entropy is 0.78, which is higher than the preset information propagation threshold and meets the lower limit of time delay, thus establishing the directed edge weights. For path ,exist The calculated conditional transitive entropy is 0.42, establishing the directed edge weights. For path ,exist The calculated conditional transfer entropy is 0.35, which satisfies the lower limit of the thermal resistance conduction hysteresis delay in metal structures, thus establishing the directed edge weights. For the reverse path Based on the unidirectional flow blocking characteristic of the fluid pipeline check valve, the directed edge weight is forcibly reset to zero, thus establishing... After physical pruning, the retained directed edges are checked according to the propagation direction of the physical topology connection paths of each component of the controlled equipment. If a retained edge causes the propagation path to point back to its upstream node, the weight of the edge is reset to zero until the retained edge no longer forms a directed closed loop. This process eliminates false associations that violate the extrema of the medium transmission timing and the direction of energy transfer, generating an acyclic dynamic directed causal topology graph. and its corresponding weighted adjacency matrix .
[0034] The marginal contribution deduction module uses an acyclic dynamic directed causal topology graph. With global multidimensional state tensor By inputting a spatiotemporal graph convolutional inference network, the baseline mean square reconstruction error of the entire system under no-intervention conditions is derived. Subsequently, counterfactual virtual zeroing intervention is applied to each candidate node to be evaluated in turn, forcibly cutting off all outgoing edge weights of the selected candidate node in the weighted adjacency matrix and forcibly assigning its input state deviation sequence to zero, and deduce the increase in state reconstruction error of the remaining non-intervention nodes in the entire system within the time window. ,in, For nodes When the node is in the zero-intervention state At any moment The counterfactual state prediction value, The state prediction is the value under the uninterrupted baseline state. During the deduction process of the spatiotemporal graph convolutional inference network, the baseline reconstructed state representation model reconstructs the current actual abnormal state based on the topology of each node. When the source fault node is subjected to counterfactual virtual zeroing intervention and its outward path is blocked, the downstream subordinate nodes lose the physical support of the upstream front-end driving input, causing the graph neural network's prediction output for the downstream nodes to approach zero deviation from the baseline. At this time, the actual state collected by each sensor channel of the physical equipment is still in an objective abnormal state of severe vibration and heat. The deviation between the predicted output and the actual measured state is thus amplified sharply, resulting in a significant jump in the increase of the state reconstruction error of the uninterrupted node. Therefore, the intervention node is determined to be the fault source that induces the cascading anomaly downstream. For the node at the end of the transmission... With nodes When zeroing intervention is implemented, the increase in reconstruction error of non-intervention nodes obtained from the deduction are as follows: and For nodes located at the vibration bearing center When implementing a zeroing intervention, it cuts off its downstream nodes. With nodes The positive input, the derived increase in reconstruction error is For nodes located at the source When zeroing intervention is implemented, the increase in reconstruction error obtained from the deduction reaches [amount] because the downstream nodes of the entire topology lose the preceding driving input. According to the expression Calculate the marginal contribution index of the root cause of failure for each candidate node, where To pre-determine the zero smoothing constant, the marginal contribution index of the root cause of the fault at each measuring point was calculated as follows: , , as well as .
[0035] The closed-loop decision arbitration module retrieves the dynamic sensitivity weight vector that matches the current high-speed light-load operating conditions. According to the expression The calculated root cause confidence levels for each measuring point are as follows: , , as well as ,in For measuring points The corresponding dynamic sensitivity weights and satisfy The candidate nodes are obtained by sorting the comprehensive root cause confidence in descending order. And the candidate nodes for the second-order sort are The confidence difference between the preferred candidate node and the second-best candidate node is calculated as follows: The difference has not yet reached the preset separation threshold set in the range of 0.20 to 0.30. Therefore, the comprehensive root cause confidence of each candidate node continues to be updated according to the same sliding time window; when the first-order ranked candidate nodes remain at [value missing], [the remaining value missing]. The candidate nodes of the second-order sort remain as When the confidence difference between the two reaches a preset separation threshold, the node is determined. Assuming the target is the root cause of the fault, the control system immediately outputs a pressure boosting control parameter to the variable frequency drive unit of the lubricating hydraulic pump, increasing the oil supply flow by 18%. At the same time, it outputs a speed reduction and load reduction torque limit command to the spindle power drive unit, reducing the upper limit of the spindle motor output torque by 30%, to suppress dry friction and burning of the main bearing.
[0036] Example 2: The current large-scale CNC rotary equipment transmission chain testing system is deployed in an industrial operating environment containing electromagnetic interference and mechanical vibration. The test objects include variable frequency drive motors, flexible couplings, heavy-duty main bearing housings, and servo lubrication hydraulic stations. The system is equipped with four physical quantity sensing nodes, among which nodes... A current sensor with a range of 0 to 100A and a resolution of 0.01A is used to acquire the effective value of the three-phase stator current of the drive motor. The root mean square value of the radial vibration acceleration of the main bearing housing was acquired using a piezoelectric accelerometer with a measurement range of -50 to +50 m / s² and a frequency response range of 0.1 to 10 kHz. A piezoresistive pressure sensor with a range of 0 to 2.5 MPa and a measurement accuracy of 0.1% FS is used to collect the pressure of the main pipeline supplying lubricating oil. An infrared sensor with a measurement range of 0 to 150℃ and a temperature resolution of 0.1℃ was used to collect the temperature of the main bearing end cover. Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency harmonic interference at a frequency of 50Hz were superimposed in the signal transmission link of the current and vibration channels. The system was set with steady-state reference mean and standard deviation as follows: A and A, m / s² and m / s² MPa and MPa, and ℃ and ℃, sampling period s, the sliding observation window contains A series of continuous sampling points.
[0037] The test system consisted of a test group using the method of this invention and nine control groups. The test group employed a complete processing flow including global multidimensional state tensor calculation, physical pruning based on the physical propagation delay constraint matrix, counterfactual virtual zeroing intervention, and arbitration of dynamic sensitivity weights and preset separation thresholds. Control group 1 removed the physical propagation delay constraint matrix. The physical pruning steps of the control group directly construct the correlation matrix based on the conditional transitive entropy; control group 2 removes the counterfactual virtual zeroing intervention and the calculation of the fault root cause marginal contribution index, and directly uses the absolute value of the maximum transient deviation of each measuring point as the anomaly assessment weight; control group 3 uses a fixed average weight vector. Replace the dynamic sensitivity weight vector; control groups 4 and 5 respectively use preset separation thresholds The preset separation thresholds were set to 0.12 and 0.38; control groups 6 and 7 used preset separation thresholds of 0.20 and 0.30, respectively; control group 8 used a low-intensity oil supply pressure decay condition, with the oil pressure slightly decreasing to 0.48 MPa; control group 9 used a high-intensity sudden pressure loss condition, with the oil pressure suddenly decreasing to 0.32 MPa. For each group, the peak value of the state deviation, the number of false correlation edges removed, the marginal contribution index of the fault root cause, the confidence difference, the judgment result of the target fault root cause entity, the judgment accuracy, the control response delay, and the end cover temperature rise suppression amount were recorded during the process of increased bearing vibration caused by abnormal lubrication pressure.
[0038] In the test group with high-speed light-load mode and superimposed 20dB noise, the node With nodes Peak state deviation The causal graph generation module utilizes the physical propagation delay constraint matrix [8.9, -4.3]. Physically prune the initial directed incidence matrix to remove reverse edges that violate the flow prohibition of the one-way valve. In addition, spurious edges that violated the lower limit of thermal resistance conduction hysteresis delay in metal structures were removed, totaling 3 spurious edges. Subsequently, counterfactual virtual zeroing intervention was implemented on candidate nodes, and the source node was... The corresponding increase in mean square reconstruction error is The marginal contribution index of the root cause of the failure was calculated. Combined with dynamic sensitivity weights Subsequently, its overall root cause confidence level was: At the 5th sampling point after the disturbance, the first-order candidate node With second-order candidate nodes The confidence difference was 0.1367. At this moment, the difference had not yet reached the preset separation threshold set in the range of 0.20 to 0.30, and the system continued to update the comprehensive root cause confidence with the sliding observation window; after the subsequent confidence difference reached the preset separation threshold, the node was finally determined. As the target fault root cause entity, the accuracy rate of the group's full-time judgment was 98.6%, the control response delay was 25.4ms, after the target fault root cause entity was determined, the lubrication hydraulic pump frequency conversion drive unit was output with 18% oil supply flow compensation and pressure boost control parameters, and the spindle power drive unit was output with 30% speed reduction and load reduction torque limit command, and the end cover temperature rise suppression was 8.4℃.
[0039] After removing physical pruning, the peak state deviation of control group 1 The value is [9.1, -4.4], the number of spurious edges pruned is 0, and the node... The marginal contribution index of the root cause of the failure decreased to 0.312, and the confidence difference between the first-order and second-order candidate nodes was 0.0315. Due to the fact that the bidirectional false correlation edges and pseudo-causal loops formed by statistical association are still retained in the correlation matrix, the downstream vibration nodes with larger energy release amplitudes... The entity identified as the root cause of the target failure was determined to have an accuracy rate of 61.2%, a control response delay of 58.7 ms, and an end cap temperature rise suppression of 3.2℃. Control group 2 retained physical pruning but removed the counterfactual virtual zeroing intervention, and its peak state deviation was... The value is [8.8, -4.2]. Three spurious correlation edges are still removed. Since the fault root cause marginal contribution index is no longer calculated, its corresponding value is not obtained. After directly using the absolute value of the maximum transient deviation at each measuring point for anomaly assessment, the confidence difference is only 0.012, and the vibration peak deviation... Make the node The entity identified as the root cause of the target failure had an accuracy rate of 43.5%, a control response delay of 64.2 ms, and an end cap temperature rise suppression of 1.8℃. It failed to provide effective oil supply compensation for the abnormal lubrication pressure at the source. Control group 3 retained physical pruning and counterfactual virtual zeroing interventions, using only a fixed average weight vector. Its peak state deviation was... For the region [9, -4.3], remove 3 spurious edges. The confidence difference after applying fixed weights is 0.669; the final decision node is determined by the confidence difference of 0.086. The target fault root cause entity was identified with an accuracy rate of 82.4%, a control response delay of 38.6 ms, and an end cover temperature rise suppression of 6.1℃.
[0040] When performing boundary tests on the preset separation threshold, control group 4 will... Set to 0.12, its peak state deviation Given [8.9, -4.3], remove 3 spurious edges. The confidence difference is 0.1367, which is 0.669. Since 0.12 is below the lower limit of 0.20, the order of first-order and second-order candidate nodes is prone to jump when high-frequency measurement noise is superimposed at a single point; the final decision node of this group The target fault root cause entity was identified with an accuracy rate of 78.4%, a control response delay of 22.1 ms, and an end cap temperature rise suppression of 5.3℃. The control group 5... Set to 0.38, peak state deviation Given [9, -4.4], remove 3 spurious edges. The confidence difference was 0.669, and the confidence difference was 0.1367. Since 0.38 is higher than the upper limit of 0.30, the confidence difference in the early stage of transient disturbance cannot cross this threshold. The judgment result remains undecided, the judgment accuracy rate is 52%, the control response delay is extended to 88.6ms, and the end cover temperature rise suppression is 2.1℃.
[0041] Control group 6 used a preset separation threshold of 0.20, and its state deviation peak value was... Given [8.9, -4.3], remove 3 spurious edges. The confidence difference for the 5th sampling point was 0.1367, so no root cause determination was output at that moment. As pressure decay accumulated in subsequent sampling periods, the confidence difference for the 7th sampling period increased to 0.235, and the node was determined after crossing the preset separation threshold of 0.20. For the target fault root cause entity, the full-time judgment accuracy was 96.2%, the control response delay was 26.8ms, the end cap temperature rise suppression was 8.1℃, and the control group 7 used a preset separation threshold of 0.30, whose peak state deviation was... For the value [9.1, -4.4], also remove 3 spurious edges. The confidence difference was 0.669 at the first sampling point and 0.1367 at the fifth sampling point. By the ninth sampling period, the confidence difference had increased to 0.312, reaching the preset separation threshold of 0.30, at which point the node was determined. As the target fault root cause entity, the full-time judgment accuracy rate is 94.8%, the control response delay is 29.5ms, and the end cover temperature rise suppression is 7.9℃. Thus, under the two boundary values of 0.20 and 0.30, the control response delay is kept within 30ms.
[0042] The lower limit of the preset separation threshold range of 0.20 is determined based on the maximum confidence jitter extreme value when a multi-channel sensor network is superimposed with white noise at a signal-to-noise ratio of 20dB. Below this value, concurrent and occasional disturbances at multiple measurement points will cause frequent changes in the ranking of first-order and second-order candidate nodes. The upper limit of 0.30 is determined based on the maximum physical response energy difference between the dominant node and the accompanying node in the early weak damage evolution stage. Above this value, the confidence difference corresponding to the early weak degradation is difficult to exceed the decision threshold in the fault latency stage, and the output delay of the control system exceeds 80ms. Therefore, the preset separation threshold is set in the range of 0.20 to 0.30.
[0043] In the oil supply pressure disturbance intensity test, the oil pressure of control group 8 slightly decreased to 0.48 MPa, node With nodes Peak state deviation For the value [3.2, -2.4], physical pruning removed 3 spurious edges, and the calculated values were... The confidence difference between first-order and second-order candidate nodes is 0.1042; the final node is determined over the entire time period. The target fault root cause entity was identified with an accuracy rate of 97.1%, a control response delay of 24.8 ms, and an end cap temperature rise suppression of 4.6℃. Control group 9 experienced a sudden drop in oil pressure to 0.32 MPa, corresponding to a peak state deviation. Increased to [14.6, -5.6], physical pruning still removes 3 spurious edges. The confidence level increased to 0.742, and the confidence difference increased to 0.1825; the final determination node for the entire time series was achieved. The target fault root cause entity was identified with an accuracy rate of 99.2%, the control response delay was reduced to 21.6ms, and the end cover temperature rise suppression was increased to 11.2℃.
[0044] Example 3: The root cause analysis system for transmission chain faults in large-scale CNC rotary equipment consists of an industrial edge computing unit, a multi-channel synchronous data acquisition board, and a sensor network. The controlled physical entities include the spindle drive motor, high-rigidity diaphragm coupling, main bearing housing, and lubrication hydraulic pump station. The sensor network is configured with four physical quantity sensing nodes, where each node... Corresponding to the effective value of the three-phase stator current of the drive motor, node The root mean square value of the radial vibration acceleration corresponding to the main bearing housing, node Corresponding to the main pipeline pressure of the lubrication circuit, node Corresponding infrared temperature measurement value of the main bearing end cover; the data acquisition board uses a sampling period. s synchronously samples and buffers the timing signal sequences of each channel. The sliding time window contains Each discrete sampling point is calculated based on the state deviation formula. Extract the dimensionless state deviation sequence and combine it to form a global multidimensional state tensor. The steady-state reference parameters are stored in non-volatile memory and set. A, A, m / s² m / s², MPa MPa, and ℃ ℃.
[0045] When constructing an acyclic dynamic directed causal topological graph, the causal graph generation module discretizes the joint probability density distribution of the state deviation sequence using the Gaussian kernel density estimation method, with the kernel function window width parameter... Based on statistical smoothing relationship Adaptive selection, where The standard deviation of the sample sequence within the sliding window; time lag. In the set of discrete integers Use inner traversal to obtain values and calculate the conditional transitive entropy between any two measurement points. The system employs a sliding time window with five sampling points to update the state in real time and perform causal inference. Gaussian kernel density estimation uses 200 pre-cached continuous discrete sampling points as rolling baseline data. During the probability density integral solution, a statistical space is established based on these 200 sampling points. Density mapping and conditional probability inference are performed on the transient state of the local five sampling points for joint probability distribution calculation of the three variable dimensions. This avoids the underdetermined sample and singular probability density phenomena that occur when directly using a single five-point window for density estimation.
[0046] Physical propagation delay constraint matrix The lower limit of propagation delay corresponding to each physical topology connectivity path is stored. The main pipeline for lubrication supply uses ISO VG46 anti-wear hydraulic oil, and the pipeline length is 2.4m; the fluid bulk elastic modulus is... GPa, oil density kg / m³, fluid medium pressure wave propagation velocity according to Based on the m / s calculation, the theoretical propagation time of the pressure wave in the main pipeline section is 1.83 ms. When the pressure wave propagates from the upstream pressure node to the main bearing side along the physical connection path, it also passes through the hydraulic servo valve with an inherent response lag of 15 ms. Therefore, the total time delay from the occurrence of upstream pressure change to the earliest detectable pressure response at the downstream measuring point is 16.83 ms. Taking this earliest pressure response time as the lower limit of the propagation time delay of the fluid medium pressure wave in this connection path, the corresponding sampling period is... s is determined after rounding up. .
[0047] The main bearing housing is made of HT250 gray cast iron with a wall thickness of 35mm. The material has a thermal conductivity of 48W / (m·K) and a thermal diffusivity of [missing value]. The heat generated by friction of the bearing outer ring is conducted along the metal-to-metal contact path of the bearing housing, main bearing housing, and end cover to the infrared thermography location. The thermal resistance conduction lag time from the start of frictional heat generation to the appearance of a detectable initial temperature rise at the infrared thermography point on the end cover is measured to be 23.5 ms. Based on this, the lower limit of the thermal resistance conduction lag time of the metal structure is determined. The 23.5ms corresponds to the time when the initial micro-area temperature rise of the thin-walled mounting edge inside the bearing end cover reaches the detection threshold of the infrared temperature sensor, and is not based on the overall thermal equilibrium of the cast iron housing. In the initial stage of frictional heating of the bearing outer ring, micron-level surface micro-protrusions tear to form contact flash temperature. Heat is conducted to the end cover through the metal contact area between the bearing housing and the end cover. The high-sensitivity infrared temperature probe detects an initial infrared radiation increment exceeding the sensor noise floor at 23.5ms, making the detectable response of the initial thermal anomaly within the microsecond to millisecond time range.
[0048] For the mechanical propagation path, the mechanical vibration propagates from the main bearing housing through the mechanical physical topology of the controlled equipment to the main spindle drive motor side. The propagation time of the mechanical vibration wave along this path is measured to be 8.5 ms. Based on this, the lower limit of the mechanical vibration wave velocity transmission delay is determined. The causal graph generation module uses a preset information transmission threshold. bit, when the maximum conditional propagation entropy between measurement points is greater than 0.15 and the time delay satisfies The corresponding directed connection edges are retained; for reverse paths that violate the flow prohibition of the one-way valve, the weights of the directed edges are reset to zero and determined. After physical pruning, an acyclic dynamic directed causal topological graph and its weighted adjacency matrix are generated. .
[0049] The marginal contribution inference module embeds a spatiotemporal graph convolutional inference network. This network consists of two spatiotemporal graph convolutional blocks and a fully connected reconstruction output layer. Each spatiotemporal graph convolutional block is composed of a Chebyshev graph convolutional layer in the spatial direction and a one-dimensional dilated causal convolutional layer in the temporal direction. The feature hidden channel dimension is set to 32, the temporal convolutional kernel size is set to 3, and the activation function is a modified linear unit. The network input is a global multidimensional state tensor. The weighted adjacency matrix of an acyclic dynamic directed causal topological graph It aggregates neighborhood node features in the spatial direction, extracts temporal evolution features on the time axis, and outputs the predicted state tensor of the whole system reconstruction under no-intervention conditions. The baseline mean square reconstruction error is obtained by the following formula: The polynomial expansion truncation order of the spatial Chebyshev graph convolution is set to 2, and the local second-order neighborhood node features are aggregated by normalizing the graph Laplacian matrix. The one-dimensional dilated causal convolution in the temporal direction has a hole expansion rate of 1 in the first layer spatiotemporal graph convolution block and a hole expansion rate of 2 in the second layer. Combined with a temporal convolution kernel of size 3, the causal receptive field on the time axis covers 5 discrete sampling periods.
[0050] In the counterfactual virtual zeroing intervention process, candidate nodes are selected sequentially. The input tensor will be the first The state deviation sequence is forcibly assigned to a zero vector to construct the intervention tensor. Simultaneously, the weighted adjacency matrix will be... The Middle All outgoing edge weights of the row are forcibly reset to zero, and an intervention topology matrix is constructed. The global multidimensional state tensor and topology after intervention are re-input into the spatiotemporal graph convolutional inference network, which outputs the counterfactual state predictions of the remaining non-intervention nodes, and the increase in mean square reconstruction error is calculated using the following formula: Thus we obtain , , as well as Then perform non-negativity processing and normalization according to the following formula: The calculated marginal contribution index of the root cause of failure for each candidate node is as follows: , , as well as .
[0051] The closed-loop decision arbitration module obtains the per-unit value of the current spindle operating speed. Per-unit value of motor output torque In this embodiment, the spindle speed is 85% of the rated speed, and the load torque is 30% of the rated value, that is... , Based on the continuous mapping relationship of operating conditions, the current component is: The vibration component is The pressure component is Temperature component After normalization, the dynamic sensitivity weight vector is obtained. Combining the marginal contribution index of root causes of failure, according to Calculate the comprehensive root cause confidence score to obtain , , as well as .
[0052] The constant reference terms 0.1, 0.15, and 0.2 in the above sensitivity components are taken from the background noise power ratio and minimum sensitivity basis of each sensing channel under the reference state of motor no-load rotation; the proportional gain coefficients 0.3 and 0.2 are obtained from the spindle transmission dynamics calibration and are used to characterize the coupling relationship between the linear increase of motor stator current with load torque and the increase of bearing housing radial vibration with the square root of spindle centrifugal speed; the negative load influence coefficient 0.1 in the pressure component is used to characterize the fluid dynamic response when the shear thinning effect of lubricating oil film weakens under light load conditions. All coefficients are dimensionless calibrated by quasi-static step loading test on the transmission bench.
[0053] After sorting the root cause confidence in descending order, the first-order candidate nodes are: The second-order candidate node is The difference between the two within the current sliding time window is The preset separation threshold is determined within the range of 0.20 to 0.30; this embodiment uses the lower limit. According to the original continuous mapping relationship, the value obtained by linearly scaling 0.20 with the torque per unit value is 0.1. This value is lower than the lower limit of the preset separation threshold. Therefore, the decision threshold is still limited to 0.20 during actual arbitration. The current confidence difference of 0.1235 has not yet reached this threshold. The closed-loop decision arbitration module continues to update the comprehensive root cause confidence of each candidate node with the sliding time window. When the subsequent first-order candidate nodes remain at... Second-order candidate nodes are maintained as When the difference between the two reaches 0.20, the node is determined. The target failure root cause entity.
[0054] After identifying the root cause of the target fault, the closed-loop decision arbitration module sends an oil supply flow compensation boost control parameter of 18% to the variable frequency drive unit of the lubrication hydraulic pump via the programmable logic controller. At the same time, it sends a speed reduction and load reduction torque limit command to the spindle power drive unit to reduce the upper limit of the spindle motor output torque by 30%, so that the main bearing can be restored to the hydrodynamic lubrication state before plastic scratches appear on the surface.
[0055] Example 4: This example combines Figures 1 to 2 This section describes AI-based methods and systems for fault root cause analysis, such as... Figure 1As shown, firstly, step S1 is executed to acquire multi-channel sensor timing signals and discrete event alarms within the controlled equipment's operating window. Based on the steady-state baseline, the state deviation sequence is calculated, and a global multidimensional state tensor is constructed by combining the discrete event alarms. Then, step S2 is executed to calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix. The lower limit of the physical medium propagation delay for each measurement point is obtained to construct a physical propagation delay constraint matrix. Based on the physical propagation delay constraint matrix, the initial directed correlation matrix is physically pruned, removing edges that violate the lower delay limit and unidirectional flow paths, generating an acyclic dynamic directed causal topology graph. Finally, step S3 is executed. The global multidimensional state tensor and the acyclic dynamic directed causal topology graph are input into the graph neural network model to extrapolate the baseline reconstruction error. Counterfactual virtual zeroing intervention is then applied to candidate nodes in sequence to cut off the downstream propagation of the intervened nodes and extrapolate the increase in state reconstruction error of non-interventioned nodes. The marginal contribution index of the root cause of each candidate node is calculated. Finally, step S4 is executed to determine the dynamic sensitivity weight vector based on the current operating conditions. The comprehensive root cause confidence of each candidate node is calculated by weighting it with the marginal contribution index of the root cause of each candidate node. The target root cause entity is determined based on the descending order of the comprehensive root cause confidence and the preset separation threshold. The closed-loop adjustment parameters for driving fuel supply flow compensation and speed limiting and load reduction are output.
[0056] like Figure 2 As shown, the system provided by this invention includes a state tensor construction module, a causal graph generation module, a marginal contribution deduction module, and a closed-loop decision arbitration module for mutual signal and data interaction. The state tensor construction module acquires multi-channel sensor time-series signals and discrete event alarms, constructs and outputs a global multidimensional state tensor to the causal graph generation module and the marginal contribution deduction module through time-series deviation calculation. The causal graph generation module establishes an initial directed correlation matrix based on the global multidimensional state tensor and performs pruning filtering by retrieving the physical propagation delay constraint matrix to generate... The module outputs an acyclic dynamic directed causal topology graph to the marginal contribution inference module. The marginal contribution inference module inputs the global multidimensional state tensor and the acyclic dynamic directed causal topology graph into the graph neural network model, sequentially performs counterfactual virtual zeroing intervention on the candidate nodes, and infers and calculates the fault root cause marginal contribution index of each candidate node. The closed-loop decision arbitration module receives the fault root cause marginal contribution index and calculates the comprehensive root cause confidence of each candidate node by combining it with the dynamic sensitivity weight vector. After sorting in descending order and determining the target fault root cause entity by the separation threshold, the closed-loop adjustment parameters are output.
[0057] Example 5: When the CNC spindle drive system is deployed and initially debugged on-site, the state-aware network establishes a steady-state baseline self-calibration channel for sensor zero-point temperature drift and installation mechanical clearance deviation caused by long-term operation; during the spindle's no-load preheating cycle, the data acquisition board continuously acquires the physical quantity measurement sequence of each channel from low speed to rated speed at a sampling interval of 0.01s. The calculation unit calculates the time-series mean of each measuring point in the fault-free steady-state operation stage based on the sliding statistical window to obtain the benchmark mean. A, m / s² MPa and ℃, simultaneously calculate the sample standard deviation to obtain the benchmark standard deviation. A, m / s² MPa and When the ambient temperature exceeds 15℃ or the viscosity of the lubricating medium decreases, the system triggers the baseline reconstruction process, uses the sliding mean filter to remove low-frequency drift components, and writes the updated steady-state benchmark into the non-volatile storage register, providing an updated timing benchmark for the construction of the global multidimensional state tensor.
[0058] When the spindle enters a variable-speed, heavy-load cutting condition and is subjected to external vibration disturbances, the closed-loop decision arbitration module determines a preset separation threshold within the range of 0.20 to 0.30 according to the current operating conditions. In the rated heavy-load cutting section, the preset separation threshold is 0.20; in the high-speed, light-load finishing section, it is 0.30. When the lubrication circuit becomes blocked, the pressure node... The state bias is generated, and the confidence level of the first-order integrated root cause is... With second-order comprehensive root cause confidence When the difference reaches the currently used preset separation threshold, the node is determined. As the target fault root cause entity, the closed-loop decision arbitration module sends oil supply flow compensation and pressure boosting control parameters to the variable frequency drive unit of the lubrication hydraulic pump, driving the variable frequency oil supply pump to increase the main pipeline oil supply servo flow in a stepwise manner with an 18% flow increment; at the same time, it sends a speed reduction and load reduction torque limit command to the spindle power drive unit, reducing the upper limit of the spindle motor output torque by 30%, and executing spindle speed limiting and load reduction control until the state deviation amplitude of each measuring point in the global multidimensional state tensor falls back to within the steady-state tolerance band.
[0059] When the 18% incremental step size of the oil supply flow is calibrated based on the friction loss of the lubrication hydraulic pipeline with a length of 2.4m and an inner diameter of 12mm and the frequency modulation response curve of the servo pump, oil supply flow compensation is performed according to this step size. Within 0.2s, the pressure at the oil inlet of the main bearing housing can be restored to above 0.58MPa, and a hydrodynamic oil film with a thickness greater than 3.5μm can be rebuilt. The 30% torque reduction in the speed reduction and load reduction torque limiting command is determined based on the frictional heat generation balance calculation of the main bearing contact surface. After this limitation is implemented, the frictional heat generation rate is reduced by 51%, and the natural convection heat dissipation rate of the housing is higher than the frictional heat generation rate, which is used to limit the further aggravation of the main bearing temperature rise and avoid bearing surface seizure failure.
[0060] Example 6: When the transmission chain system of a large CNC rotary equipment is performing heavy-load cutting, there is strong electromagnetic interference and mechanical vibration on site. When the terminals of the sensors deployed in the state sensing network are transiently loosened or subjected to electromagnetic spike interference, high-frequency abnormal disturbances will be generated in a single sensing channel. The steady-state reference of each measuring point remains unchanged. A, A, m / s² m / s², MPa MPa, and ℃ ℃; When the radial vibration measuring point of the main bearing housing When a momentary poor contact occurs and is superimposed with a sudden high-frequency white noise pulse with an amplitude of 12 m / s², the stator current measuring point of the drive motor... Lubrication pipeline supply pressure measuring point and bearing end cover temperature measuring points The values remain at the steady-state baseline, with the corresponding measured values being: A, MPa and ℃, the state tensor construction module during the sampling period s, number of consecutive sampling points Calculate the state deviation of each measuring point under the given conditions, and obtain , , ,node The transient pulse deviation is A global multidimensional state tensor is constructed from the state deviation sequences of each measuring point. .
[0061] The causal graph generation module calculates the conditional transfer entropy between each measurement point, due to the nodes When an isolated pulse occurs, the lubrication pipeline supply pressure measuring point Maintaining a steady-state baseline, with time delay Under these conditions, the calculated conditional propagation entropy did not reach the preset information propagation threshold. bit; simultaneously, this isolated measurement pulse does not generate a corresponding associated response along the mechanical and thermal resistance transmission path of the device, therefore the weight of the relevant directed edge is determined as , and After physical pruning, the nodes In the causal topology, it remains an isolated node, generating an acyclic dynamic directed causal topology graph and its weighted adjacency matrix. .
[0062] The marginal contribution derivation module will use the global multidimensional state tensor with weighted adjacency matrix Input the spatiotemporal graph convolutional inference network to obtain the baseline mean square reconstruction error of the entire system. Subsequently, counterfactual virtual zeroing intervention was applied to each candidate node in turn, when dealing with isolated abnormal nodes. When the input state is forcibly assigned to zero and all outgoing edge weights are reset to zero, since there are no slave nodes in the topology... Pointing to node , and The forward propagation path of the nodes was observed, and the state reconstruction trajectories of the remaining non-intervention nodes did not deviate within the observation time window, thus obtaining the node's... The corresponding increase in state reconstruction error ,node , and The corresponding increase in state reconstruction error is also close to zero; after normalization calculation, the node The marginal contribution index of the root cause of failure is Combined with dynamic sensitivity weight vector Afterwards, the overall root cause confidence level of each measuring point approached zero.
[0063] The closed-loop decision arbitration module sorts the comprehensive root cause confidence scores in descending order. If the difference between the first-order and second-order comprehensive root cause confidence scores does not reach the preset separation threshold set within the range of 0.20 to 0.30, the decision threshold used in this embodiment is... Under this condition, node The severe pulse was determined to be a hardware failure of the isolated sensing channel, rather than a physical fault in the mechanical transmission chain; the system outputs a sensor measurement link verification prompt signal, and does not output oil supply flow compensation boost control parameters or speed reduction and load reduction torque limit commands, and the spindle drive system maintains continuous and stable operation.
[0064] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.
Claims
1. An AI-based method for fault root cause analysis, characterized in that, Includes the following steps: Step S1: Obtain multi-channel sensor timing signals and discrete event alarms within the controlled equipment operation window; calculate the state deviation sequence based on the steady-state benchmark; and construct a global multi-dimensional state tensor by combining discrete event alarms. Step S2: Calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix, remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, and generate an acyclic dynamic directed causal topology graph. Step S3: Input the global multidimensional state tensor and the acyclic dynamic directed causal topology graph into the neural network model to extrapolate the baseline reconstruction error; sequentially implement counterfactual virtual zeroing intervention on the candidate nodes, cut off the downstream propagation of the intervened nodes and extrapolate the increase in the state reconstruction error of the non-intervention nodes, and calculate the fault root cause marginal contribution index of each candidate node. Step S4: Determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index. Based on the descending order of the comprehensive root cause confidence and the preset separation threshold, the target fault root cause entity is determined, and the closed-loop adjustment parameters for driving fuel supply flow compensation and speed limiting load reduction are output.
2. The AI-based fault root cause analysis method according to claim 1, characterized in that, Step S2 includes: Step S21, within the preset time delay range, solving for the measurement point using the Gaussian kernel density estimation method. With measuring points The joint probability density distribution between the points is calculated under different discrete time delay conditions. State deviation sequence towards measurement point The information transfer entropy gain of the state deviation sequence flow; step S22, taking the maximum conditional transfer entropy under each discrete time delay as the measurement point. Pointing to the measuring point The initial correlation edge strength is used to construct an initial directed correlation matrix containing the temporal asymmetric dependencies between measurement points across the entire domain.
3. The AI-based fault root cause analysis method according to claim 1, characterized in that, Step S2 further includes: Step S23, obtaining the physical topology connectivity paths between the components of the controlled equipment, and establishing a physical propagation delay constraint matrix that includes the lower limit of mechanical vibration wave velocity propagation delay, the lower limit of fluid medium pressure wave propagation delay, and the lower limit of metal structure thermal resistance conduction hysteresis delay; in the initial directed correlation matrix, when the measuring point With measuring points When the time delay between the two is less than the corresponding lower limit of the physical propagation delay of the medium, or the maximum conditional transfer entropy does not reach the preset information transmission threshold, or the correlation direction between the two violates the flow prohibition of the fluid check valve, the measuring point will be... Pointing to the measuring point The weights of the directed edges are reset to zero, generating an acyclic dynamic directed causal topological graph that satisfies the directed acyclic topological property.
4. The AI-based fault root cause analysis method according to claim 1, characterized in that, Step S3 includes: Step S31, inputting the weighted adjacency matrix of the acyclic dynamic directed causal topology graph and the global multidimensional state tensor into the spatiotemporal graph convolutional inference network, aggregating the spatial neighborhood features and temporal evolution features of the nodes, outputting the full system reconstruction prediction state under the no-intervention state, and calculating the baseline mean square reconstruction error of the full system within the operating window.
5. The AI-based fault root cause analysis method according to claim 1, characterized in that, Step S3 further includes: Step S32, sequentially selecting candidate nodes to be evaluated, forcibly assigning the state deviation sequence input to the candidate nodes to zero, and forcibly setting all outgoing edge weights of the candidate nodes in the acyclic dynamic directed causal topology graph to zero, blocking the forward influence propagation of the candidate nodes to downstream successor nodes; re-inputting the global multidimensional state tensor and topology structure after intervention into the graph neural network model, deduce the counterfactual state prediction values of the non-intervention nodes in the whole system except for the intervened candidate nodes, and calculate the increase in mean square reconstruction error of the other non-intervention nodes.
6. The AI-based fault root cause analysis method according to claim 5, characterized in that, Step S3 further includes: Step S33, when the increase in mean square reconstruction error is greater than zero, retain its value, otherwise take the value as zero; divide the increase in mean square reconstruction error corresponding to each candidate node by the sum of the increase in mean square reconstruction error of all candidate nodes in the whole system and the sum of the preset zero-prevention smoothing constant, and calculate the normalized fault root cause marginal contribution index of each candidate node.
7. The AI-based fault root cause analysis method according to claim 1, characterized in that, Step S4 includes: Step S41, determining the operating condition sensitivity components based on the real-time collected operating speed and load factor of the controlled equipment, and normalizing each component to construct a dynamic sensitivity weight vector; Step S42, multiplying the dynamic sensitivity weight corresponding to each measuring point with the fault root cause marginal contribution index of the measuring point to calculate the comprehensive root cause confidence of each candidate node; sorting the comprehensive root cause confidence of all candidate nodes in descending order, and when the difference between the largest confidence in the first order and the second largest confidence in the second order reaches a preset separation threshold, determining the measuring point corresponding to the first order confidence as the dominant fault root cause entity, wherein the preset separation threshold ranges from 0.20 to 0.
30.
8. The AI-based fault root cause analysis method according to claim 1, characterized in that, The multi-channel sensing timing signals include the stator current of the drive motor, the radial vibration acceleration of the bearing housing, the supply pressure of the lubrication pipe, and the temperature of the bearing end cover.
9. The AI-based fault root cause analysis method according to claim 1, characterized in that, The closed-loop control parameters include oil flow compensation and pressure boosting control parameters sent to the variable frequency drive unit of the lubrication hydraulic pump, and speed reduction and load reduction torque limiting commands sent to the spindle power drive unit.
10. An AI-based fault root cause analysis system, used to implement the AI-based fault root cause analysis method of claim 1, characterized in that, include: The system includes a state tensor construction module, a causal graph generation module, a marginal contribution deduction module, and a closed-loop decision arbitration module. Among them, the state tensor construction module is used to acquire multi-channel sensor time-series signals and discrete event alarms within the operating window of the controlled equipment, calculate the state deviation sequence based on the steady-state benchmark, and construct a global multi-dimensional state tensor in combination with discrete event alarms. The causal graph generation module is used to calculate the conditional transfer entropy between measurement points based on the global multidimensional state tensor to construct an initial directed correlation matrix; obtain the lower limit of the physical medium propagation delay for each measurement point to construct a physical propagation delay constraint matrix; and perform physical pruning on the initial directed correlation matrix based on the physical propagation delay constraint matrix to remove the associated edges that violate the lower limit of the delay and the unidirectional flow path, thereby generating an acyclic dynamic directed causal topology graph. The marginal contribution deduction module is used to deduce the baseline reconstruction error of the input graph neural network model from the global multidimensional state tensor and the acyclic dynamic directed causal topology graph; it sequentially implements counterfactual virtual zeroing intervention on candidate nodes, cuts off the downstream propagation of the intervened nodes and deduces the increase in the state reconstruction error of non-intervention nodes, and calculates the fault root cause marginal contribution index of each candidate node. The closed-loop decision arbitration module is used to determine the dynamic sensitivity weight vector based on the current operating conditions, and calculate the comprehensive root cause confidence of each candidate node by weighting it with the fault root cause marginal contribution index; based on the descending order of the comprehensive root cause confidence and the preset separation threshold, the target fault root cause entity is determined, and the closed-loop adjustment parameters for driving fuel supply flow compensation and speed limiting load reduction are output.
Citation Information
Patent Citations
Industrial equipment and fault root cause diagnosis method and system thereof
CN121207601A
