A device failure root cause diagnosis system and method based on causal AI

Abstract

The application discloses a device fault root cause diagnosis system and method based on causal AI, and belongs to the technical field of device health management. The method comprises the following steps: acquiring real-time three-mode sensor data of device operation, and generating theoretical prediction data synchronized with the real-time three-mode sensor data based on a preset device mechanism model; generating a sparse companion source excitation vector according to the deviation between the real-time three-mode sensor data and the theoretical prediction data; initializing a companion state vector, and injecting the sparse companion source excitation vector into the companion state vector to form an initial companion state; performing reverse iterative propagation on the initial companion state a predetermined number of times based on a causal topology transpose matrix to generate a final companion gradient vector, wherein the gradient peak value of the final companion gradient vector indicates the deviation physical fault root cause. The application realizes low-delay and high-precision device fault root cause positioning by reducing the dimensionality of the reverse search of the causal diagram to deterministic algebraic operation.

Description

Technical Field

[0001] This application relates to the field of equipment health management technology, and in particular to a system and method for diagnosing the root causes of equipment failures based on causal AI. Background Technology

[0002] In modern industrial production, the health status of large and complex equipment (such as wind turbines and aircraft engines) directly affects production safety and economic efficiency. Traditional fault diagnosis methods often rely on correlation analysis of sensor data, triggering alarms when abnormal data is detected. However, such methods often only identify fault phenomena but cannot pinpoint the root physical causes that trigger a chain reaction, leading to delayed maintenance responses and high repair costs. In recent years, artificial intelligence technology based on causal reasoning has provided a new approach to root cause diagnosis, but existing technologies generally face fundamental technical contradictions: on the one hand, accurate causal graph back-reasoning (such as using Bayesian networks) requires sampling a huge joint probability distribution, and its computational complexity increases exponentially with the number of equipment nodes, resulting in extremely high inference latency, which cannot meet the real-time intervention requirements of industrial scenarios. Summary of the Invention

[0003] The first aspect of this application provides a method for diagnosing the root cause of equipment faults based on causal AI, aiming to solve the technical problem that the computational complexity of causal inference is too high in the prior art, making real-time root cause diagnosis impossible. The method acquires the deviation between real-time and theoretically predicted data, constructs a sparse adjoint source excitation, and utilizes a deterministic inverse iterative process based on the causal topological transpose matrix, which is mathematically equivalent to backpropagation of a causal graph, to reduce the complex probabilistic search problem to efficient linear algebraic operations. This mechanism can accurately backpropagate the deviation energy to the physical fault root cause with extremely low computational latency, thereby technically achieving high-speed, high-precision root cause localization of complex equipment faults.

[0004] Optionally, the step of acquiring real-time trimodal sensor data of the device operation and generating theoretical prediction data synchronized with the real-time trimodal sensor data based on a preset device mechanism model includes performing time alignment processing on asynchronous data streams generated by sensors from at least three different physical modes to generate the real-time trimodal sensor data; and substituting an estimated value of the device's internal node state at a previous moment into the device mechanism model for single-step solution to generate the theoretical prediction data.

[0005] Optionally, according to the method of claim 1, the step of generating a sparse adjoint source excitation vector based on a deviation between the real-time trimodal sensor data and the theoretical prediction data includes obtaining the vector difference between the real-time trimodal sensor data and the theoretical prediction data, and normalizing each dimension component of the vector difference to obtain a standard residual vector; wherein the normalization is based on a standard deviation of the corresponding sensor channel data under historical health conditions.

[0006] Optionally, generating a sparse adjoint source excitation vector further includes setting the components in the standard residual vector whose absolute values ​​are less than a preset noise threshold to zero in order to generate the sparse adjoint source excitation vector.

[0007] Optionally, the step of performing a predetermined number of reverse iterative propagations on the initial adjoint state based on a causal topological transpose matrix includes using the initial adjoint state as the input to the iterative propagation and performing a predetermined number of matrix-vector multiplication operations in a loop according to the maximum depth of a causal graph, wherein each matrix-vector multiplication operation is performed by left-multiplying the adjoint state vector output from the previous iteration using the causal topological transpose matrix.

[0008] Optionally, the method further includes, after generating the final adjoint gradient vector, using the gradient value corresponding to the physical root node in the final adjoint gradient vector as a pseudo-label gradient; and performing gradient descent optimization on the connection weights in a causal topology matrix corresponding to the causal topology transpose matrix based on the pseudo-label gradient, so as to achieve self-supervised updating of the causal topology matrix.

[0009] Optionally, after each matrix-vector multiplication operation, the result of the matrix-vector multiplication operation is further processed by applying a nonlinear activation function to eliminate negative gradient components.

[0010] The second aspect of this application provides a device fault root cause diagnosis system based on causal AI for performing the aforementioned method.

[0011] A third aspect of this application provides a computer-readable storage medium on which a computer program stored is executed by a processor to implement the aforementioned method.

[0012] The beneficial effect of this application's embodiments is that, by constructing a discrete adjoint state causal propagation mechanism, the NP-hard causal graph back-search problem is transformed into... The application employs sparse matrix multiplication with low complexity, reducing the diagnostic latency of fault roots from seconds to milliseconds, thus enabling real-time, high-precision root cause diagnosis of complex industrial equipment. Furthermore, by utilizing the byproducts of the adjoint gradient to perform self-supervised optimization of the causal topology, the system can adapt to equipment aging or unknown fault modes, significantly improving the robustness and long-term effectiveness of the diagnostic system. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a device fault root cause diagnosis method based on causal AI provided in an embodiment of this application;

[0015] Figure 2 This is an architectural block diagram of a device fault root cause diagnosis system based on causal AI provided in an embodiment of this application;

[0016] Figure 3 This is a schematic diagram illustrating the principle of generating sparse adjoint source excitation vectors based on noise thresholds using the standard residual vector provided in this application embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] This application provides a method for diagnosing the root cause of equipment faults based on causal AI. In a specific implementation, this method employs a discrete adjoint state causal propagation mechanism to transform the inverse probabilistic reasoning process on the physical causal graph of the equipment into a deterministic algebraic operation process based on the causal topological transpose matrix. This enables the inverse inference of the physical fault root cause of the deviation from the real-time deviation of multimodal sensors with extremely low computational latency. This method solves the technical problem in the prior art where traditional causal reasoning methods based on Bayesian networks cannot meet the needs of real-time industrial diagnosis when dealing with complex equipment due to the exponential increase in computational complexity. It achieves the beneficial effect of high-precision root cause localization of equipment faults within milliseconds.

[0020] The method flow of this application embodiment is described in detail below. The method can be executed by a device fault root cause diagnosis system deployed in an edge-cloud collaborative architecture. In a preferred implementation scenario, the device is a large wind turbine generator set.

[0021] S100: Acquire real-time trimodal sensor data of the device operation, and generate theoretical prediction data synchronized with the real-time trimodal sensor data based on a preset device mechanism model.

[0022] In a specific embodiment, this step aims to provide two data foundations that are strictly aligned in time and precisely correspond in the physical dimension for subsequent deviation calculations: one is measured data reflecting the current actual physical state of the equipment, and the other is theoretical state data calculated based on first principles, representing the equipment's expected state under the current operating conditions. The three-modal sensor data, originating from physical modes covering different frequency ranges and physical phenomena, can more comprehensively capture the equipment's operating status. For example, in a wind turbine monitoring scenario, the three-modal sensor data may include high-frequency vibration sensor data, low-frequency temperature sensor data, and ultra-high-frequency acoustic emission sensor data. Vibration data can reflect the dynamic characteristics of rotating components such as gears and bearings; temperature data can reflect the thermodynamic characteristics of components such as friction and lubrication; and acoustic emission data can capture early fault signals such as the propagation of microcracks within materials.

[0023] This step S100 can be broken down into two sub-steps, S110 and S120.

[0024] S110: Time alignment processing is performed on asynchronous data streams generated from sensors of at least three different physical modes to generate the real-time trimodal sensor data.

[0025] In a specific implementation, the timestamps of data generated by various sensors deployed on wind turbine generators (such as accelerometers, thermocouples, and acoustic emission sensors) are naturally asynchronous and non-uniform due to differences in their physical characteristics and sampling circuits. For example, vibration sensors may sample at high speeds at frequencies of several kilohertz (kHz), while temperature sensors may sample at only 1 Hz. Directly comparing data from different times without time alignment would introduce significant errors, rendering subsequent deviation calculations meaningless. Therefore, to construct an accurate snapshot of the device's state over time, the data alignment subunit in sub-step S110 performs strict time synchronization and interpolation operations. This data alignment subunit continuously reads raw analog data streams from different sensors from the system's data bus. After analog-to-digital conversion (ADC), the raw analog data streams are appended with a high-precision clock stamp provided by a hardware-level Time Sensitive Network (TSN) or Network Time Protocol (NTP).

[0026] For example, suppose at time point A state snapshot needs to be generated. The data alignment sub-unit first defines a very small time window, for example... Then, it retrieves all sensor data points that arrive within this time window. For vibration sensors, there may be multiple data points within this window, for example, at... Reading at and in Reading at For a temperature sensor, the most recent reading was probably at... place In order to obtain For precise time values, the data alignment sub-unit employs a cubic spline interpolation algorithm. This algorithm utilizes information from a data point and its neighbors to construct a piecewise cubic polynomial function. This function has continuous first and second derivatives at connection points, thus ensuring the smoothness of the interpolation curve. For vibration data, the data alignment sub-unit will be based on... , Construct a cubic spline function from the data points and their surrounding data points. Then calculate Get In The interpolated vibration value at a given time. For temperature data with a low sampling frequency, it will be based on... A spline function is constructed from multiple historical data points within a longer timeframe before and after it. and calculate By performing this operation on all sensor channels, the final output is a... Strictly time-aligned, dimension is (in The measured state vector (total number of sensors) For example, if there are 10 vibration sensors, 5 temperature sensors, and 3 acoustic emission sensors, then It is an 18-dimensional column vector, where each component represents a specific sensor. This is the interpolated and corrected physical measurement value at a precise time cross-section. This vector is stored in a circular buffer, awaiting retrieval in subsequent steps. This process ensures that the benchmark for subsequent comparisons is fair and accurate, forming the cornerstone of the entire diagnostic system's accuracy.

[0027] S120: Substitute the estimated state of the internal nodes of the device at the previous moment into the device mechanism model for single-step solution to generate the theoretical prediction data.

[0028] After obtaining the equipment Real state snapshot at any given moment Next, to determine whether this actual state is normal, a theoretical normal reference is needed. This reference is not a fixed threshold, but a theoretical value dynamically calculated based on the current operating conditions of the equipment (such as wind speed, load, and ambient temperature). The twin solver in sub-step S120 is responsible for this calculation. The twin solver runs on hardware deployed on the edge gateway of the equipment side, and its core is a preset equipment mechanism model. The equipment mechanism model is a set of differential equations that can describe the physical interaction laws between the various components of the equipment. For the gearbox of a wind turbine generator, this model can be a rigid-flexible coupled multibody dynamics model, which includes a mathematical description of physical processes such as gear meshing force, bearing support reaction force, oil film lubrication effect, and heat conduction.

[0029] For example, suppose in At that moment, the system already has a state estimate of all internal nodes of the equipment (including gears, bearings, drive shafts, etc.), which includes the angles, angular velocities, temperatures, etc. of each component. The twin solver will then use this... The state estimate at time 1 is used as the initial condition and substituted into the system of differential equations for the rigid-flexible coupled multibody dynamics. Simultaneously, it also obtains the current time from external environmental sensors. Operating parameters, such as wind speed and grid load, are used as boundary conditions for the equation system. Then, the twin solver employs a numerical integration algorithm, such as the fourth-order Runge-Kutta method, to perform single-step time integration on the equation system, with a step size of [missing value]. This integration process physically simulates the process from... Time's up At any given moment, the evolution of all physical states within the equipment under its current operating conditions. The solution yields the result of the equipment's... The theoretical state values ​​of all internal nodes at each time step are obtained. Finally, the twin solver maps these theoretical state values ​​of the internal nodes (such as the theoretical rotational speed and theoretical temperature of a specific bearing) to the corresponding physical quantities measured by the sensors using an observation model (usually a simple linear mapping matrix), thus obtaining a solution at each time step. Theoretical sensor output vector at time This vector With the measured state vector They have the exact same dimensions and physical meanings when ordered. For example, if The first component is the vibration measurement value of bearing No. 1, then... The first component is the theoretical vibration value that bearing No. 1 should produce under the current operating conditions, calculated based on the mechanistic model. This process generates... This constitutes a dynamic, condition-adaptive health baseline, providing a solid physical basis for subsequent accurate identification of abnormal deviations.

[0030] S200: Generate a sparse adjoint source excitation vector based on the deviation between the real-time three-modal sensor data and the theoretical prediction data.

[0031] The core task of this step is to transform the two high-dimensional vectors obtained in S100. and The differences are compared and transformed into a dimensionless energy signal with a high signal-to-noise ratio that can drive subsequent causal backpropagation. If the difference is used directly in the original physical units (e.g., a deviation is...), the difference would be significantly different. The other is If these are not properly matched, their importance cannot be fairly compared, nor can they be addressed in subsequent unified models. Therefore, this step constructs a co-source that accurately reflects the source and intensity of anomalies through normalization and sparsification.

[0032] This step S200 can be broken down into two sub-steps, S210 and S220.

[0033] S210: Obtain the vector difference between the real-time three-modal sensor data and the theoretical prediction data, and normalize each dimension component of the vector difference to obtain a standard residual vector.

[0034] In a specific implementation, the deviation calculation unit of sub-step S210 first performs a vector subtraction operation to obtain the original residual vector. Each component of this vector represents the absolute deviation between the measured and theoretical values ​​of a sensor channel. However, the physical meaning and fluctuation range of the deviations differ significantly between channels. For example, a small deviation in a vibration signal may indicate a serious fault, while a large deviation in a temperature signal may simply be normal operating fluctuation. To eliminate these differences in dimensions and scales and make the deviations of all channels comparable, the deviation calculation unit... Each component Normalization is performed. This normalization is based on a standard deviation of the sensor channel data corresponding to historical health states. The standard deviation During the equipment commissioning phase or over a long period (e.g., 100 hours) after confirming healthy operation, the statistical analysis of the first... The deviation between the measured values ​​and their theoretical predicted values ​​for each sensor channel is calculated using a time series. This represents the typical fluctuation amplitude of that channel under normal operating conditions. The normalization operation is as follows: .

[0035] For example, assume the historical deviation standard deviation of vibration sensor No. 1 for The historical standard deviation of temperature sensor No. 2 for At the current moment The calculated original deviations are respectively and Although the temperature deviation is larger in absolute terms, after normalization, the standard residuals are respectively... and The results showed that the deviation in the vibration channel (3 standard deviations) was actually more anomalous than the deviation in the temperature channel (2 standard deviations). This was achieved by analyzing all... Each channel performs this operation, ultimately generating a dimensionless standard residual vector. Each component of this vector fairly reflects the statistical significance of the corresponding sensor channel deviating from its normal fluctuation range, providing a unified benchmark for subsequent judgments.

[0036] S220: The components in the standard residual vector whose absolute value is less than a preset noise threshold are set to zero to generate the sparse adjoint source excitation vector.

[0037] The standard residual vector generated in S210 Even if the equipment is perfectly healthy, due to model imperfections and sensor noise, most components will still be small non-zero values. Using these noisy vectors directly as input for subsequent causal inference would significantly increase unnecessary computation and potentially overwhelm the inference results with noise, preventing focus on the true anomaly source. The purpose of sub-step S220 is to extract the truly high signal-to-noise ratio deviation signals indicating physical anomalies through a filtering process, forming a sparse input. This sparsity operation is implemented using a hard threshold filtering logic. Specifically, the accompanying source building unit traverses the standard residual vector. Each component And compare it with a preset noise threshold. Compare. If If the value is zero, the component will be forced to zero; otherwise, its original value will be preserved.

[0038] In a preferred embodiment, the noise threshold The method for setting the value is as follows: Under healthy operating conditions, collect dimensionless standard residual vectors for at least 1000 cycles. This creates a large dataset containing residual values ​​from all sensor channels and at all time points. Then, the probability distribution of the absolute values ​​of all residual components in this dataset is calculated, and the 99th percentile of this distribution is set as... For example, suppose that the result is obtained through statistical calculation. This means that under normal circumstances, the probability of a sensor channel exhibiting a deviation exceeding 2.5 times its standard deviation is only 1%. Therefore, any deviation exceeding this threshold can be considered with high confidence as an anomalous signal rather than random noise.

[0039] For example, suppose the input is a 5-dimensional standard residual vector. After passing through the noise threshold After comparison, the absolute values ​​of the 2nd, 3rd, and 5th components are all less than 2.5, and therefore are set to zero. The final generated sparse adjoint source excitation vector... This vector It is clearly pointed out that the current physical anomalies are mainly manifested in the first and fourth sensors, and the intensity and direction of the anomalies are represented by the magnitude and sign of the non-zero values. This sparse vector will serve as the source of subsequent accompanying state propagation, and its sparsity greatly improves the efficiency and robustness of subsequent calculations.

[0040] S300: Backpropagation of discrete adjoint states.

[0041] The purpose of this step is to efficiently and accurately trace the cause (i.e., the fault source) of the physical root node that triggered the effect (i.e., the deviation) identified in S200 at the leaf nodes of the sensor back along the physical causal path of the device. This step abandons the complex probabilistic sampling or graph search methods in traditional causal inference, and instead uses an algebraic operation that is mathematically equivalent but computationally highly efficient. Its core mechanism is to map the complex causal network into a linear system, and the backward propagation of the deviation is transformed into a forward evolution in an adjoint system. This evolution process can be completed through simple matrix-vector multiplication iterations.

[0042] This step S300 can be broken down into three sub-steps: S310, S320, and S330.

[0043] S310: Initialize a companion state vector and inject the sparse companion source excitation vector into the companion state vector to form an initial companion state.

[0044] In a specific implementation, the core accompanying causal reasoning engine first creates a dimension in memory. The accompanying state vector ,in This is the total number of nodes in the device's cause-effect graph (including physical component nodes and sensor nodes), with all its elements initialized to zero. This vector Each dimension index uniquely corresponds to a physical node in the causal graph. Then, the inference engine reads the dimension index generated by S220. ( Sparse adjoint source excitation vector (number of sensors) The inference engine, based on a pre-defined mapping relationship, will... Each non-zero component is injected into The corresponding sensor node index position in the vector.

[0045] For example, suppose the device cause-effect graph has 100 nodes ( ). Among them, the positions with node indices 90 and 93 correspond to the generation... The sensor for the first and fourth components of the vector. In the aforementioned example, So, after initialization dimensional adjoint state vector This will be a very sparse vector, with a value of 3.8 at the 90th position, -4.1 at the 93rd position, and zero values ​​at the other 98 positions. This is the initial state of the propagation process. Physically, this is equivalent to applying two virtual force sources to the sensor leaf nodes of the causal graph, their magnitude and direction determined by the statistical significance of the sensor bias. Subsequent iterative propagation calculates how these virtual forces travel backward along the causal chain and converge at the true source of the force (i.e., the fault source). The construction of this initial state sets the starting point for the entire deterministic reverse tracing process.

[0046] S320: Based on a causal topological transpose matrix, perform a predetermined number of backward iterative propagation on the initial adjoint state to generate a final adjoint gradient vector.

[0047] Step S320 is the core computational step in the propagation process. Its core data structure is a... Causal topological matrix and its transpose matrix The causal topology matrix It is constructed by analyzing the CAD drawings and dynamic system diagrams of the equipment during the system initialization phase (S000). It is an adjacency matrix, whose elements Represents a node State changes affect nodes The direct impact strength or propagation probability of the state is normalized to... Between. For example, if node It is the driving gear, the node. If it is the driven gear, then It will be a relatively large value close to 1; if the node It's a bearing, a node. It is the temperature sensor on the casing that is connected to it. This will be a relatively small value representing the heat transfer efficiency. The entire matrix This describes how a disturbance propagates forward along the physical path within a device. According to the adjoint state theory, the forward propagation of a disturbance in the original system is mathematically rigorously equivalent to the forward propagation of an adjoint source in the transposed system. Therefore, the reverse propagation of the deviation energy can be achieved by... The network is used for forward iteration to achieve accurate simulation.

[0048] The iterative process is as follows: The inference engine generates the initial adjoint state from S310. This serves as the input for step 0. Then, based on the maximum depth of the causal graph... (i.e., the longest path length from any root cause node to a sensor leaf node), execute The iteration. In the... The next iteration ( From 1 to In this process, the core algebraic iterative equations are executed: .in, This is a standard matrix-vector multiplication. The physical meaning of this multiplication operation is to multiply the first matrix by the second... The accompanying energy accumulated at each node during a step is distributed to their parent node according to the transposed causal topology (i.e., reverse causality). For example, if a node... In the The accompanying energy of the step is And in the original graph it is affected by the nodes and The impact (i.e.) and In this multiplication operation, The energy will follow and The proportions of (elements of the transpose matrix) are backtracked and accumulated to the nodes. and In the Within the energy of the step.

[0049] After matrix multiplication, a nonlinear activation function is applied, preferably the rectified linear unit (ReLU), i.e. Its function is to eliminate negative gradient components. In the physical scenario of causal diagnostics, we are usually concerned with the cumulative effect of fault energy. Negative gradients may represent certain inhibitory or compensatory effects, and introducing these effects in the early stages of root cause localization complicates the problem. The ReLU function ensures that only positive energy flows indicating fault contributions propagate in the network, enhancing the robustness of the algorithm and the interpretability of the results.

[0050] For example, suppose the causal graph depth The inference engine will perform the matrix multiplication-ReLU activation operation described above eight times in a loop. From Start by generating sequentially As the iterations proceed, the energy initially injected into the sensor's leaf nodes flows back towards the root of the causal graph, layer by layer, like a tide. After the first iteration, the energy converges to the sensor's directly upstream node; after the second iteration, the energy continues to propagate upstream, and so on. After several iterations, it is sufficient to ensure that energy has propagated from any leaf node to all possible root cause nodes. The final vector is obtained. This is the final adjoint gradient vector. Each component of this vector... Its numerical value precisely quantifies the first Each node serves as a potential root cause, contributing to the final observation of that specific set of sensor biases. The sensitivity of the contribution. The larger the value, the greater the suspicion that the node is the instigator.

[0051] S330: Perform an optimization operation on the final adjoint gradient vector to output node information indicating the root cause of physical faults.

[0052] The final adjoint gradient vector was obtained in S320. (Right now Following this, step S330's task is to identify the most likely source of the fault from this vector containing the suspicion levels of all nodes in the entire network, and make a final diagnostic decision. The decision unit first reads... Vector. Due to the energy focusing effect of accompanying propagation, if a single physical fault source exists, then in In the vector, the gradient value of the node corresponding to the physical source will be significantly higher than that of all other nodes, forming a sharp peak.

[0053] The decision-making unit performs a simple ArgMax optimization operation, that is, in Find the maximum value among all components and record the node index where the maximum value is located. Then, this maximum value... With a preset alarm threshold The alarm thresholds are compared. The configuration is also white-box. In a preferred embodiment, the configuration method is as follows: During the first 100 hours of equipment trial operation (fault-free confirmation), the system continuously runs and calculates the normal accompanying gradients caused by random operating condition disturbances and measurement noise. The maximum value of all final accompanying gradient vectors during this period is collected to form a gradient peak dataset under normal operating conditions. Then, the mean of this dataset is calculated. and standard deviation And set the alarm threshold to This statistically based setting ensures that only gradient peaks that significantly exceed the normal fluctuation range will trigger an alarm, greatly reducing the false alarm rate of the system.

[0054] if The system then determines that a real physical failure has occurred and indexes the node. and its corresponding gradient value Output (as confidence level). For example, the output might be: Fault Root Cause Diagnosis: Node 5 (High-Speed ​​Bearing), Confidence Level: 3.82. Maintenance personnel can then use this precise location information to directly inspect or replace the high-speed bearing, significantly reducing troubleshooting time. If If the system determines that the current sensor deviation exists, its energy level after tracing is insufficient to constitute a high-confidence internal physical fault, and it is likely merely caused by a relatively severe external environmental disturbance. In this case, the system will remain silent and will not issue an alarm, thereby effectively avoiding false alarms caused by external interference (such as sudden changes in wind speed).

[0055] In an optional embodiment of this application, the method further includes a self-supervised online optimization step to enable the system to adapt to device aging or environmental changes.

[0056] S400: Self-supervised causal topological dynamic evolution.

[0057] In one specific implementation, the final adjoint gradient vector generated in S300 is not only used for diagnosis, but also contains information about the causal topological matrix. Valuable information for accuracy. The self-supervised causal topology optimizer (M400) runs silently in the background, using this information to... Fine-tuning is then performed. The optimizer defines a loss function. The goal of this loss function is to ensure that when a known, historically confirmed fault mode is injected into the system, the peak of the resulting gradient points as accurately as possible to the known fault source. During the online run phase without real fault labels, the optimizer employs a pseudo-label self-supervised learning strategy. It will use the gradients of high-confidence root cause nodes diagnosed in S330... This is considered a pseudo-true value monitoring signal. Then, using the chain rule in calculus, the loss function is calculated. For causality matrix Each connection weight partial derivatives This partial derivative indicates how the weights should be fine-tuned to make the diagnostic results more accurate (i.e., to make the gradient of the root cause node more prominent). .

[0058] For example, the optimizer uses a momentum gradient descent algorithm to update the weights. The update rule is as follows: ,in It uses a small learning rate. This update process is continuous and takes small steps. Over time, if the system frequently attributes a certain sensor bias pattern to a specific root cause, the weights on the causal paths from that root cause to these sensors will be gradually increased, while the weights on other irrelevant paths will be decreased. This mechanism makes the causal topology matrix... It can learn and adapt to gradual degradation of equipment (e.g., wear of a component leading to an increasing impact on vibration) or nonlinear correlations not precisely described in the initial model. Ultimately, the optimizer generates a fine-tuned causal topology matrix. It periodically writes the data back into the memory of the core inference engine (M300), thus completing a closed-loop self-learning process, enabling the entire diagnostic system to evolve and become more accurate with use.

[0059] The technical solution of this application will be further described below through a system embodiment.

[0060] This application also provides a device fault root cause diagnosis system based on causal AI. The system can be used to execute the methods described in any of the foregoing embodiments. In a specific implementation, the system adopts an edge-cloud collaborative architecture and consists of the following four core functional modules:

[0061] The Edge Forward Mechanism Twin Module (M100) is a physical device deployed on the edge computing gateway of the monitored device. Its hardware includes a multi-channel synchronous analog-to-digital converter (ADC) and an embedded processor (such as an ARM Cortex-A series). This module is physically connected to the device's three-modal sensors and is responsible for executing the S100 steps. Its internal data alignment subunit uses hardware timestamps to synchronize and interpolate asynchronous data streams, while the twin solver runs a simplified device mechanism model in real time on the embedded processor, generating theoretical prediction data. The module's output, i.e., and It is passed to the next module through a high-speed internal shared memory area.

[0062] The Deviation Dimension Alignment and Accompanying Source Construction Module (M200) can be deployed on the same edge gateway as the M100, or on a higher-performance aggregation node. Logically, it consists of a deviation calculation unit and an accompanying source construction unit. This module is responsible for executing the S200 steps. It reads data from shared memory. and Normalization and sparsification processes are performed to finally generate sparse adjoint source activation vectors. The module's output is sent to the core inference engine via an internal bus (such as PCIe) or a low-latency network protocol (such as ZeroMQ).

[0063] The core accompanying causal inference engine (M300), the brain of the entire diagnostic system, is configured to run as a standalone process on an edge server or cloud server with at least one graphics processing unit (GPU) equipped with a Tensor Core. This module specifically utilizes the hardware's half-precision floating-point (FP16) matrix multiply-accumulate (MAC) unit to accelerate the backpropagation of the accompanying gradient. It is responsible for executing the S300 steps. Its internal graph-resident subunits will store the device's causal topology matrix... The matrix is ​​pre-compiled into a sparse matrix format optimized for GPU architecture (such as Compressed Sparse Row Format, CSR) and locked in the GPU's cache. The accompanying inference subunit is a finite state machine that detects new... When the vector arrives, it triggers a fixed-length, containing A CUDA kernel function for sparse matrix-vector multiplication (SpMV). Ultimately, the module outputs the node ID and confidence level indicating the root cause of the fault.

[0064] The self-supervised causal topology optimizer (M400) is typically deployed on a cloud server because it needs to process historical data over long time spans. This module is responsible for executing the S400 steps. It subscribes to the diagnostic results (including the final adjoint gradient vector) generated by M400 via a message queue (such as Kafka). Once sufficient data has been collected, it performs a computationally intensive gradient descent optimization task in the background on the causal topology matrix. Update the matrix. It will be serialized and pushed back to the M300 via a secure API to replace the old matrix in its memory.

[0065] These four modules work together to form a complete closed-loop system that goes from data acquisition to root cause diagnosis and model self-optimization, enabling rapid, accurate and adaptive root cause diagnosis of equipment faults.

[0066] This application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it can implement the steps in any of the foregoing method embodiments. The storage medium can be any type of non-volatile memory, such as a solid-state drive (SSD), a hard disk drive (HDD), or flash memory.

[0067] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0068] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

Claims

1. A method for diagnosing the root causes of equipment failures based on causal AI, characterized in that, include: Acquire real-time trimodal sensor data of the device operation, and generate theoretical prediction data synchronized with the real-time trimodal sensor data based on a preset device mechanism model; A sparse adjoint source excitation vector is generated based on a deviation between the real-time trimodal sensor data and the theoretical prediction data. Initialize an adjoint state vector and inject the sparse adjoint source excitation vector into the adjoint state vector to form an initial adjoint state; Based on a causal topological transpose matrix, a predetermined number of backward iterative propagations are performed on the initial adjoint state to generate a final adjoint gradient vector, wherein a gradient peak of the final adjoint gradient vector indicates a physical fault root cause of the deviation.

2. The method according to claim 1, characterized in that, The process of acquiring real-time trimodal sensor data of the device operation and generating theoretical prediction data synchronized with the real-time trimodal sensor data based on a preset device mechanism model includes: Asynchronous data streams from sensors of at least three different physical modes are time-aligned to generate the real-time trimodal sensor data. The estimated state of the internal nodes of the device at the previous moment is substituted into the device mechanism model for single-step solution to generate the theoretical prediction data.

3. The method according to claim 1, characterized in that, The step of generating a sparse adjoint source excitation vector based on a deviation between the real-time trimodal sensor data and the theoretical prediction data includes: The vector difference between the real-time three-modal sensor data and the theoretical prediction data is obtained, and each dimension component of the vector difference is normalized to obtain a standard residual vector; wherein, the normalization process is based on a standard deviation of the corresponding sensor channel data under historical health conditions.

4. The method according to claim 3, characterized in that, The generation of a sparse adjoint source activation vector further includes: The components in the standard residual vector whose absolute value is less than a preset noise threshold are set to zero to generate the sparse adjoint source excitation vector.

5. The method according to claim 1, characterized in that, The step of performing a predetermined number of backward iterative propagations on the initial adjoint state based on a causal topological transpose matrix includes: The initial adjoint state is used as the input for iterative propagation, and a predetermined number of matrix-vector multiplication operations are performed cyclically according to the maximum depth of a causal graph. Each matrix-vector multiplication operation is performed by left-multiplying the adjoint state vector output from the previous iteration using the causal topological transpose matrix.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: After generating the final adjoint gradient vector, the gradient value corresponding to the physical root cause node in the final adjoint gradient vector is used as a pseudo-label gradient. Based on the pseudo-label gradient, gradient descent optimization is performed on the connection weights in a causal topology matrix corresponding to the causal topology transpose matrix to achieve self-supervised updating of the causal topology matrix.

7. The method according to claim 5, characterized in that, After each of the matrix-vector multiplication operations, the following is also included: The result of the matrix-vector multiplication operation is processed by a nonlinear activation function to eliminate negative gradient components.

8. A device fault root cause diagnosis system based on causal AI, characterized in that, include: An edge-forward mechanism twin module is used to acquire real-time trimodal sensor data of device operation and generate theoretical prediction data synchronized with the real-time trimodal sensor data based on a preset device mechanism model. A deviation dimension alignment and adjoint source construction module is used to generate a sparse adjoint source excitation vector based on a deviation between the real-time trimodal sensor data and the theoretical prediction data. A core adjoint causal reasoning engine is used to initialize an adjoint state vector and inject the sparse adjoint source excitation vector into the adjoint state vector to form an initial adjoint state. Furthermore, the core adjoint causal inference engine is also used to perform a predetermined number of backward iterative propagation on the initial adjoint state based on a causal topological transpose matrix to generate a final adjoint gradient vector, wherein a gradient peak of the final adjoint gradient vector indicates a physical fault root cause of the deviation.

9. The system according to claim 8, characterized in that, The system also includes: A self-supervised causal topology optimizer is used to optimize the connection weights in a causal topology matrix corresponding to the causal transpose matrix by using the gradient value corresponding to the physical root cause node in the final adjoint gradient vector as a pseudo-label gradient after the core adjoint causal inference engine generates the final adjoint gradient vector, thereby achieving self-supervised updating of the causal topology matrix.

10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images