An artificial intelligence early warning system and method

Abstract

The application discloses an artificial intelligence early warning system and method, wherein the method comprises the following steps: firstly, collecting equipment time sequence operation data, reconstructing operation state trajectory, identifying early degradation through trajectory curvature change and generating an early warning signal; in response to the early warning signal, collecting multi-device operation phase, constructing a phase difference matrix, combining time sequence causality to determine the fault source device; relying on the degradation conduction directed graph, the fault conduction path is deduced in a forward direction, the node failure trend is predicted, and the corresponding disposal strategy is output. At the same time, the real fault history data is used to trace back and mine implicit features, and the conduction network parameters are iteratively updated to realize model self-optimization. The application can realize early degradation identification, accurate fault tracing, conduction path prediction and self-adaptive upgrading of the linkage equipment through the collaborative closed-loop operation of multiple technical links, and meet the early warning needs of reliable continuous operation of industrial coupled equipment.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for industrial equipment, specifically to an artificial intelligence early warning system and method. Background Technology

[0002] In industrial continuous production scenarios, multi-unit interconnected equipment such as multi-stage pump sets, series fan groups, and chain-driven production lines generally exhibit strong coupling relationships at the mechanical, transmission, and control levels. The gradual, latent degradation of a single piece of equipment can be transmitted to related equipment through the coupling link, easily triggering cascading failures and unplanned downtime accidents. Existing conventional early warning methods mostly rely on fixed threshold criteria, time-frequency domain feature analysis, or ordinary data-driven models. These methods generally suffer from shortcomings such as over-reliance on significant changes in signal amplitude, inability to capture early signs of subtle distortions in trajectory morphology, inability to accurately distinguish the source from affected objects in multi-equipment coupled failures, and a lack of fault propagation path deduction and intelligent handling decision-making capabilities.

[0003] Existing fault early warning technologies can only achieve local single-point status alarms. They have not formed a full-link collaborative early warning system that includes early hidden degradation identification, precise location of coupled fault sources, fault propagation path prediction, operation and maintenance strategy output, and closed-loop self-optimization of operation data. This system cannot meet the actual needs of early detection, accurate source tracing, and proactive blocking of progressive and conductive faults in multi-machine coupled and linked equipment, becoming a core technical challenge restricting the continuous and reliable operation of industrial production lines.

[0004] In view of the above, this application is hereby submitted. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence early warning system and method to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides an artificial intelligence-based early warning method, comprising the following steps: S1: Collecting time-series operational data of interconnected equipment, constructing the equipment's operational trajectory, calculating the second derivative curvature of the trajectory, identifying early equipment degradation based on curvature mutations, and generating an early warning signal; S2: Responding to the early warning signal, collecting the operational phases of multiple interconnected devices, constructing a phase difference matrix, and determining the fault source device based on the offset of the phase difference matrix; S3: Based on a directed graph of degradation propagation, performing forward propagation deduction starting from the fault source device, determining the fault propagation path and expected failure time, and outputting a handling strategy; wherein, S1, S2, and S3 are triggered sequentially, are interdependent, and are executed in a closed loop; this method, through the sequential and coordinated execution of curvature mutation identification, phase difference matrix positioning, and directed graph deduction of degradation propagation, forms an integrated early warning closed loop of early identification, source positioning, and propagation prediction, which differs from existing single early warning methods and possesses significant and unique technical features.

[0007] Further, S1 includes the following steps: S11: Filtering and denoising the time-series running data; S12: Constructing a continuous time-series trajectory based on the denoised data; S13: Calculating the second derivative curvature of the continuous time-series trajectory, comparing the curvature with a dynamic threshold, and determining a curvature mutation when the curvature continuously exceeds the dynamic threshold; By improving the reliability of curvature calculation through filtering and denoising and constructing a continuous trajectory, and by using dynamic threshold comparison to achieve accurate early degradation identification, it can capture weak degradation signs that traditional methods cannot identify, and significantly advance the warning time.

[0008] Furthermore, the curvature is calculated by the ratio of the cross product modulus of the first and second derivatives of the trajectory to the cube of the modulus of the first derivative. The dynamic threshold is adaptively updated based on the curvature distribution under normal operating conditions. The dedicated curvature calculation formula accurately characterizes the trajectory bending change, and the dynamic threshold adaptively matches the equipment operating conditions, breaking through the technical bias of relying on amplitude mutation warnings in the industry and achieving effective identification of degradation at the very early stage.

[0009] Furthermore, the curvature change determination in S13 uses the curvature exceeding the dynamic threshold within three consecutive sliding time windows as the trigger condition; the continuous multi-window confirmation mechanism effectively suppresses transient curvature fluctuations caused by noise interference, avoids false alarms triggered by single-point anomalies, and significantly reduces the false alarm rate while ensuring early warning sensitivity, thereby improving the system's operational stability.

[0010] Further, S2 includes the following steps: S21: extract the operating phase angles of multiple devices; S22: construct a real-time phase difference matrix based on the operating phase angles; S23: calculate the offset between the real-time phase difference matrix and the reference phase difference matrix, quantify the degree of offset using the matrix norm, and identify the device with the largest offset and the curvature change first, followed by the phase offset, as the fault source device; based on the dual location of the fault source using the phase difference matrix and the temporal causal relationship, the root cause device and the passively affected device can be accurately distinguished, avoiding misjudgment and significantly improving the accuracy of fault tracing.

[0011] Further, S3 includes the following steps: S31: Load a pre-constructed directed graph of degradation propagation, which includes device nodes, component nodes, parameter nodes, and directed edges representing degradation propagation relationships. The weights of the directed edges include propagation probability and average propagation time; S32: Determine the fault propagation path by tracing forward along the directed edges starting from the fault source device; S33: Calculate the expected failure time of each node based on the propagation path and generate corresponding handling strategies. The three-level hierarchical directed graph structure finely depicts the fault propagation link, and combined with weight information, it realizes quantitative prediction of remaining lifetime, transforming early warning information into an executable handling plan, forming a complete decision-making closed loop.

[0012] Furthermore, the handling strategies in S33 include load reduction operation, switching to backup equipment, and cutting off coupling paths; providing clear and executable proactive handling measures, transforming early warning signals into actual control actions, effectively blocking the transmission and spread of faults, and avoiding chain-like shutdown accidents.

[0013] Furthermore, it also includes step S4: after the equipment experiences a real shutdown fault, it automatically traces back the time-series operation data, phase signal data, and vibration characteristic data from the 72 hours prior to the fault, identifies early latent features, and updates the edge weights of the deteriorated directed graph; based on the real fault data, it realizes closed-loop iterative optimization of the model, continuously improves the accuracy of transmission inference, and gradually enhances the system's ability to identify similar hidden faults.

[0014] Furthermore, the system continuously monitors the operating status of the equipment. When a real equipment shutdown is detected, step S4 is triggered. If the shutdown time is within the planned maintenance window and the work order is marked as planned maintenance, it is determined to be a planned shutdown and no backtracking is triggered. If the shutdown is unplanned or associated with an alarm, it is determined to be a real fault shutdown.

[0015] An artificial intelligence early warning system includes a data acquisition module, an early degradation identification module, a fault source location module, a transmission and deduction handling module, and a self-optimization and iteration module; The data acquisition module is used to collect the timing operation data of the linked equipment and the operation phase data of multiple devices; The early degradation identification module is used to filter and denoise the time-series running data, construct a continuous time-series trajectory, calculate the second derivative curvature of the trajectory, and generate an early warning signal based on curvature abrupt changes. The fault source location module is used to extract the operating phase angles of multiple devices in response to early warning signals, construct a phase difference matrix and calculate the matrix offset, and determine the fault source device by combining the temporal causal relationship. The transmission simulation and handling module is used to load the directed graph of degradation transmission, deduce the fault transmission path from the fault source equipment, calculate the expected failure time of each node and output the handling strategy; The self-optimizing iterative module is used to identify early latent features by tracing back historical feature data after a real equipment downtime failure, update the edge weights of the directed graph of degradation propagation, and realize the system's adaptive iterative optimization. This system adopts a modular architecture to correspond to each core step of the method. Each module works in coordination and timing, and fully realizes the functions of early degradation perception, accurate fault source location, propagation path prediction and adaptive self-optimization. The overall operational reliability, early warning timeliness and fault handling are significantly improved.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention abandons the traditional early warning judgment mode that relies on changes in signal amplitude and energy. Instead, it identifies early degradation by relying on changes in the geometric shape of the equipment's operating trajectory. It can capture latent degradation features where the amplitude does not fluctuate significantly but the trajectory shape undergoes subtle distortions. This breaks through the inherent technical limitations of the industry and enables ultra-early perception of progressive equipment degradation, solving the problems of lagging early warning recognition and difficulty in detecting weak faults in traditional methods from the source.

[0017] 2. This invention constructs a multi-device phase difference matrix combined with temporal causal logic to identify the source of faults, accurately distinguishing between the actively deteriorating source device and passively affected related devices in a coupled and interconnected system, eliminating the need for manual experience to troubleshoot and locate faults one by one. Simultaneously, it performs forward path deduction through a directed graph of degradation propagation in a multi-level node architecture, predicting the fault propagation path and development trend, and matching and generating appropriate handling methods to suit the on-site operating conditions. This achieves automated fault tracing and proactive fault propagation blocking, overcoming the shortcomings of existing technologies such as vague location and lack of deductive decision-making capabilities.

[0018] 3. This invention integrates the timing, causal dependence, and collaborative operation of various stages, including trajectory curvature change detection, phase difference matrix positioning, degradation propagation deduction, and fault data self-optimization, to form a complete closed-loop early warning system. By leveraging real fault data to backtrack and uncover hidden sensitive features, and adaptively updating propagation correlation parameters, the early warning system can effectively adapt to equipment aging and changes in operating conditions. The various technical means mutually empower each other to form synergistic gains, significantly improving overall operational stability and adaptability, and effectively avoiding the risks of cascading failures and unplanned downtime in industrial production lines. Attached Figure Description

[0019] Figure 1 A flowchart of an artificial intelligence early warning method; Figure 2 This is a block diagram illustrating the principle of an artificial intelligence early warning system. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 Please see Figures 1 to 2An artificial intelligence-based early warning system and method are proposed, applicable to multi-equipment coupled operation scenarios such as multi-level transmission production lines, pump group linkage systems, fan groups, and continuous chemical production units. This method forms a closed-loop early warning decision-making system through the sequential coordinated triggering of three core components: early identification of curvature abrupt changes, source location through phase difference matrix, and directed graph deduction of degradation propagation.

[0022] The artificial intelligence early warning method described in this embodiment operates on the factory edge computing node, connecting to the equipment sensor acquisition unit upstream and the equipment control system and maintenance terminal downstream. It can collect various types of time-series data such as vibration, temperature, current, speed, phase, and pressure in real time, and continuously analyze the equipment operating status. This embodiment uses a typical industrial scenario as an example for complete explanation. The application scenario is the main shaft pressing equipment and chain drive conveyor system in the motor assembly line of an automotive parts manufacturing plant. The two are rigidly coupled by mechanical structure to form a linked equipment group. Among them, the main shaft bearing is a key component that is prone to deterioration, and the transmission sprocket and conveyor chain are related components. The progressive wear of the bearing will cause a small distortion in the running trajectory. The distortion will further cause a phase shift in the transmission system, and the shift, in turn, will aggravate the bearing wear, forming a closed-loop deterioration transmission path.

[0023] This method is divided into three core stages: early degradation identification, fault source localization, and propagation and handling. These three stages are triggered sequentially, with the output of each stage serving as the basis for the next. They are independent yet interdependent, forming an inseparable early warning decision chain. The early degradation identification stage captures subtle anomalies based on curvature changes in the equipment's operating trajectory. The fault source localization stage identifies the root cause based on the phase difference matrix of multiple devices. The propagation and handling stage predicts the fault propagation path and assesses remaining lifespan based on the directed graph of degradation propagation, ultimately outputting an executable handling strategy. The system also includes a self-optimization stage, which corrects model parameters in reverse after a real fault occurs, continuously improving early warning accuracy.

[0024] Step S1: Detailed implementation of the early degradation identification step: Step S1 is the initiation stage of the entire early warning method. It is responsible for collecting the time-series operating data of the linked equipment, constructing an operating trajectory that reflects the true state of the equipment, and identifying early degradation and generating an early warning signal by calculating the curvature of the second derivative of the trajectory. The core objective of this step is to complete the anomaly capture at a stage where equipment degradation is extremely weak and traditional monitoring indicators show no significant changes, providing triggering conditions for subsequent location and deduction. In the early stage of progressive degradation, conventional indicators such as vibration amplitude, temperature, and current will not show significant jumps, only subtle changes in operational stability, which cannot be identified by traditional monitoring methods based on thresholds, spectrum, and trends. Curvature can characterize the degree and rate of change of the trajectory bending, and the second derivative curvature is highly sensitive to small distortions in the trajectory shape. Therefore, this step is set up to achieve very early degradation identification.

[0025] Step S11: Filtering and denoising of timing data; S11 filters and denoises the collected raw timing data to eliminate invalid fluctuations caused by environmental interference, electromagnetic interference, and mechanical vibration, and retains the effective signal components that truly reflect the operating status of the equipment.

[0026] Signal denoising is achieved using a wavelet packet decomposition and reconstruction algorithm. The original data is decomposed into multiple frequency band components according to the signal frequency distribution, eliminating high-frequency components dominated by noise, and reconstructing the signal from the effective components. Data sources include vibration sensors, temperature sensors, and current sensors, with a fixed acquisition time interval to ensure temporal continuity.

[0027] In the scenario of monitoring the spindle bearing on a motor assembly line, time-series data of spindle vibration acceleration is collected. This data includes interference signals from motor electromagnetic noise, workshop floor vibration, and assembly impact. Wavelet packet decomposition is used to remove interference, preserving the true vibration trajectory information generated by the contact between the bearing rolling elements and the raceway. After wavelet packet denoising, the signal-to-noise ratio increases from 15 dB to 28 dB, noise energy is reduced by 95%, and the subtle bearing degradation characteristics are effectively separated from background noise.

[0028] The unique technical approach of this solution lies in the strong binding of filtering and denoising with subsequent curvature calculation. The goal of denoising is not to improve the clarity of signal amplitude, but to preserve the subtle changes in trajectory shape, effectively improving the stability of subsequent curvature change identification, reducing invalid calculations caused by noise, and improving the usability of weak and degraded signals.

[0029] Step S12: Construct a continuous time-series trajectory based on denoised data; S12 uses the denoised time-series data as a basis to construct a continuous running trajectory with time as the horizontal axis and equipment state characteristics as the vertical axis, so that the changes in equipment state are presented in the form of a visual curve, providing a unified data carrier for curvature calculation.

[0030] A time-delay embedding method is used to reconstruct a one-dimensional time series into a high-dimensional phase space. Based on phase space reconstruction theory, an embedding dimension of 6 and a time delay of 8 sampling points are selected to map the original sequence into a point cloud in a six-dimensional phase space. Cubic spline interpolation is used to smooth the discrete trajectory, forming a continuously differentiable state trajectory curve. After interpolation, 150 continuous trajectory points are obtained within each sliding time window. A fixed-length sliding window is used to extract time-series data, with data points arranged in the order of sampling time. Missing points are filled in by interpolation, forming a continuous and smooth state trajectory curve.

[0031] In the scenario of monitoring the spindle bearings on a motor assembly line, the vibration data of the spindle bearings, after noise reduction, are used to generate a continuous vibration trajectory based on a time series. The trajectory is generally stable and periodic, with no obvious abrupt changes, consistent with the characteristics of healthy equipment operation. However, it contains subtle distortions that can be identified by curvature. In the healthy state, the trajectory shape is approximately a closed limit cycle. After entering the early deterioration stage, the trajectory gradually contracts inward or expands outward, and the shape undergoes a detectable change. However, the absolute position of the trajectory points changes very little, with amplitude changes of less than one percent.

[0032] The technical solution is that the trajectory construction is based on the principle of keeping the small distortion unchanged, without using excessive smoothing, and retaining the subtle curvature changes that can reflect early degradation, thereby improving the trajectory's ability to characterize minor degradation and providing a reliable basis for subsequent curvature change detection.

[0033] Step S13: Second derivative curvature calculation and curvature mutation judgment; S13 calculates the second derivative curvature of the continuous time-series trajectory, compares the curvature with the dynamic threshold, and judges the curvature mutation when the curvature continuously exceeds the threshold, generating an early warning signal.

[0034] On a continuous time-series trajectory, the curvature at each trajectory point is calculated from the first and second derivatives of the trajectory. The first derivative represents the velocity of the trajectory at that point, and the second derivative represents the acceleration of the trajectory at that point. The curvature is calculated as the ratio of the modulus of the cross product of the first and second derivatives to the cube of the modulus of the first derivative.

[0035] In practical calculations, discrete difference approximation is used instead of continuous derivatives. For each trajectory point, the first derivative is approximated by the difference between the current point and its adjacent points, and the second derivative is approximated by the second-order central difference between the current point and its adjacent points.

[0036] The geometric meaning of curvature is that when the trajectory is a straight line, the curvature is zero; when the trajectory is curved, the curvature is greater than zero; and the greater the degree of curvature, the greater the curvature value.

[0037] The dynamic threshold is adaptively updated based on the curvature distribution under normal operating conditions. In the initial stage of healthy operation, curvature samples are collected over one thousand time windows to construct a curvature distribution model. Assuming the curvature follows a normal distribution, the mean and standard deviation are calculated. The upper boundary of the dynamic threshold is set as the product of the mean, the threshold coefficient, and the standard deviation. The threshold coefficient is selected as four based on the sensitivity requirements of the industrial environment.

[0038] As the device operates, the dynamic threshold is updated using an exponentially weighted moving average. The curvature sample at the current moment is used in the calculation of the updated mean, and the learning rate ranges from 0.05 to 0.2; in this implementation, 0.1 is used.

[0039] The trigger condition for determining curvature abrupt changes is that the average curvature calculated in each of three consecutive sliding time windows exceeds the upper boundary of the dynamic threshold. Employing a continuous multi-window confirmation mechanism instead of single-window triggering effectively suppresses transient curvature fluctuations caused by noise interference.

[0040] In practical applications at the bearing press-fitting station of a motor assembly line, the curvature value of the vibration signal is stable between 0.15 and 0.25 under normal operating conditions, with the upper boundary of the dynamic threshold set at 0.32. When the bearing enters the early wear stage, the curvature value gradually rises to between 0.35 and 0.45, but the vibration amplitude remains at 1.2 mm / s, lower than the traditional alarm threshold of 2.0 mm / s. An early warning is triggered when the curvature exceeds 0.32 for three consecutive windows, and the system outputs a level one warning signal.

[0041] This technical solution uses the curvature of the second derivative as the sole indicator for early degradation identification. Combined with adaptive dynamic thresholds and continuous window judgment rules, it can complete the identification at an earlier stage of equipment degradation, extending the warning lead time by about 120 to 168 hours, reducing the probability of false triggering, and breaking through the industry's technical bias that warnings can only be issued by relying on amplitude changes.

[0042] As can be seen from S1, the process of integrating filtering and denoising, trajectory construction, curvature calculation, dynamic threshold, and continuous window determination into a unified early degradation identification process, with all steps serving the detection of weak trajectory distortion, can identify equipment degradation trends earlier, reduce missed detections and false detections, improve the reliability of early warning, and provide accurate triggering conditions for subsequent positioning and inference.

[0043] Example 2 like Figure 1 and 2 As shown, the specific implementation of step S2: fault source location step is as follows: Step S2 is initiated upon receiving the early warning signal from S1. It collects the operating phases of multiple interconnected devices, constructs a phase difference matrix, and identifies the faulty device through offset calculation and temporal causality determination. The core objective of this step is to distinguish between the truly deteriorated source device and related devices passively following the fluctuations. In a multi-device interconnected system, an anomaly in one device can trigger synchronous fluctuations in other devices. Relying solely on a single device anomaly cannot determine the root cause of the fault, leading to misjudgments in maintenance. Phase characterizes the synchronicity of device operation, and the phase difference matrix quantifies the consistency of operation among multiple devices; therefore, this step is designed to achieve accurate source tracing.

[0044] Step S21: Extract phase angles; Step S21 extracts the operating phase angles of multiple linked devices under the same time reference, providing basic data for the construction of the phase difference matrix.

[0045] The phase angle extraction methods vary depending on the equipment type. For equipment with encoders, the phase angle is extracted from the encoder signal. The encoder outputs 1024 pulses per revolution, achieving a phase angle resolution of 0.35 degrees. For motor equipment without encoders, the phase angle is estimated by monitoring the zero-crossing points of the three-phase current. The current signal is filtered, and the rising edge zero-crossing point of each phase current is detected. The time interval between adjacent zero-crossing points corresponds to 180 electrical degrees. For equipment equipped only with vibration sensors, the vibration signal is subjected to a Fourier transform to extract the rotational frequency component, and the instantaneous phase is calculated using a Hilbert transform.

[0046] In the press-fitting station of the motor assembly line, the spindle motor is equipped with an incremental encoder, outputting 1024 pulses per revolution with a phase angle resolution of 0.35 degrees. The conveyor sprocket uses a vibration signal fundamental frequency phase extraction method to extract the 2.5 Hz phase component from the vibration acceleration signal. The phase angle of all equipment is uniformly based on its own zero-degree reference point, and the sampling time synchronization is achieved through the precise time protocol of industrial Ethernet, with a synchronization accuracy better than one millisecond.

[0047] The technical solution is that the phase extraction is strictly aligned with the early warning trigger time, and high-precision acquisition is only started after the early warning, which reduces the normal computing overhead, improves the data accuracy in the tracing stage, and ensures that the phase difference matrix truly reflects the synchronous changes under abnormal conditions.

[0048] Step S22: Constructing the real-time phase difference matrix; S22 constructs a real-time phase difference matrix based on the phase angles of multiple devices, where the matrix elements represent the instantaneous phase difference values ​​between the devices.

[0049] Suppose there are K interconnected devices in the system. Then the phase difference matrix is ​​a square matrix of K rows and K columns. The element in the i-th row and j-th column of the matrix represents the phase difference between device i and device j. The phase difference is calculated using the cyclic interpolation method, mapping the difference to the interval between 0 and 360 degrees.

[0050] In the initial stage of steady-state operation of the equipment, collect data for a sufficient duration, such as 24 hours, calculate the phase difference at each moment, and take the average value of each element as the reference phase difference matrix.

[0051] The real-time phase difference matrix is ​​calculated using a sliding time window method. Each time window is, for example, ten seconds long. Within the window, the average phase angle of each device is calculated, and then the phase difference matrix is ​​constructed. The window sliding step is one second.

[0052] In the press-fitting station and conveyor chain drive system, let the phase angle of the press-fitting station be θA and the phase angle of the conveyor sprocket be θB. Then the phase difference matrix is ​​a second-order square matrix, where the elements are zero and θA minus θB and θB minus θA. Under normal and stable conditions, the phase difference is stable at around 180 degrees, and the oscillation amplitude does not exceed three degrees.

[0053] The technical solution is that the phase difference matrix is ​​only constructed after the warning is triggered and does not run all the time, which reduces the system's computing power consumption, improves the correspondence between the matrix and the fault state, and makes the change of offset more accurately reflect the fault transmission relationship.

[0054] Step S23: Offset calculation and fault source device determination; S23 calculates the offset between the real-time phase difference matrix and the reference matrix, uses the matrix norm to quantify the degree of offset, and determines the fault source device by combining the temporal causal relationship.

[0055] The offset calculation uses the Frobenius norm to quantify the overall difference between the real-time matrix and the reference matrix. This norm simultaneously considers the differences of all elements in the matrix, and can sensitively reflect changes in the phase relationship between any pair of devices. When the offset exceeds a preset threshold, it indicates that the system has experienced a significant phase consistency shift, which may indicate a coupling fault.

[0056] The fault source is located using a temporal causal determination method. For each device, the time of its curvature change and the start time of its phase shift are recorded. The determination logic is: the device with the largest shift, and whose curvature change occurs first and phase shift occurs later, is determined to be the fault source device.

[0057] In the press-fitting station and conveyor chain drive system, curvature abrupt change detection indicates that the curvature abrupt change at the bearing press-fitting station occurs 1.2 seconds before the curvature abrupt change of the conveyor sprocket. The offset of the phase difference matrix jumps from the reference value of 0.12 to 0.98. Based on the dual judgment conditions of maximum offset and time priority, the system outputs that the fault source equipment is the bearing at the press-fitting station.

[0058] This technical solution combines matrix norm offset quantization with temporal causality judgment to achieve precise fault source location, improve source identification accuracy, reduce the probability of false positives, and effectively distinguish between actively degraded equipment and passively affected equipment. The source location accuracy is increased from approximately 70% to over 95%, and the average location time is reduced from several hours to several minutes.

[0059] As can be seen from S2, by using the warning signal as the sole triggering condition, integrating phase extraction, matrix construction, offset quantization, and temporal causal judgment into a fault tracing process, the root cause of the fault can be quickly identified, the location time can be shortened, the targeted nature of maintenance can be improved, and ineffective repairs can be avoided.

[0060] Example 3 like Figure 1 and 2 As shown, the specific implementation of step S3: the transmission deduction and processing output step is as follows: Step S3 starts with the faulty source equipment and performs forward deduction based on the directed graph of degradation propagation to determine the fault propagation path, the expected failure time of each node, and output the corresponding handling strategy. The core objective of this step is to predict the scope and trend of fault spread and provide an actionable proactive handling plan. Simply completing early warning and location cannot stop the fault from spreading; the fault will propagate step by step along equipment, components, and parameters, eventually causing a chain reaction of shutdowns. The directed graph of degradation propagation can clearly express the propagation relationship, propagation probability, and propagation time; therefore, this step is set up to achieve proactive blocking.

[0061] Step S31: Load the deterioration propagation directed graph; S31 loads the pre-constructed deterioration propagation directed graph, which contains device nodes, component nodes, parameter nodes, directed edges of propagation relationships and edge weights.

[0062] The directed graph is constructed using a combination of expert knowledge and historical data, specifically comprising a three-level node system. Equipment-level nodes represent the physical equipment in the system, including pumps, fans, motors, compressors, and drive sprockets. Component-level nodes represent the key components of the equipment, including bearings, gears, seals, couplings, and belts. Parameter-level nodes represent monitorable operating parameters, including vibration amplitude, temperature, current, pressure, flow rate, and phase difference.

[0063] Directed edges represent degradation propagation relationships, pointing from the source node to the target node, indicating that degradation of the source node will lead to degradation of the target node. Propagation allows cross-layer connections, forming a complete causal network. The weight of a directed edge includes two values: propagation probability and average propagation time. The propagation probability ranges from zero to one, representing the probability that degradation of the source node will propagate to the target node. The average propagation time represents the time interval from the occurrence of degradation in the source node to the appearance of a detectable change in the target node.

[0064] In a bearing drive sprocket system, the directed graph contains the following nodes and edges. Equipment-level nodes include the bearing press-fit station and the conveyor sprocket. Component-level nodes include the bearing inner ring, bearing outer ring, rolling elements, sprocket tooth surface, and chain links. Parameter-level nodes include vibration amplitude, vibration curvature, phase difference, and temperature. Directed edges include: the bearing inner ring pointing to the vibration curvature, with a propagation probability of 0.95 and an average propagation time of 24 hours; the bearing outer ring pointing to the vibration amplitude, with a propagation probability of 0.85 and an average propagation time of 48 hours; the vibration curvature pointing to the phase difference, with a propagation probability of 0.75 and an average propagation time of 72 hours; and the phase difference pointing to the sprocket tooth surface, with a propagation probability of 0.70 and an average propagation time of 120 hours.

[0065] The technical solution here is to use a three-level hierarchical directed graph to finely depict the cross-layer transmission relationship, improve the clarity of the transmission path, enhance the reliability of the deduction, and fully express the degradation and diffusion law of complex coupled systems.

[0066] Step S32: Forward propagation path deduction; Starting from the fault source device, S32 deduces forward along the directed edge, traversing all possible propagation branches to determine the final propagation path.

[0067] Starting with the fault source device located in step S2, a forward deduction is performed along the direction of the directed edges. The deduction process uses a breadth-first search algorithm to expand the propagation path layer by layer.

[0068] For each current node, obtain the set of successor nodes pointed to by all its outgoing edges. For each successor node, calculate the cumulative propagation probability and cumulative propagation time from the source node to that successor node. The cumulative propagation probability is the product of the propagation probabilities of all edges on the path, and the cumulative propagation time is the sum of the average propagation times of all edges on the path.

[0069] When multiple paths can reach the same node, the one with the highest cumulative propagation probability is selected as the primary propagation path, and the others are considered alternative paths. In a system with multiple paths coexisting, after the primary path is blocked, alternative paths may become the new dominant propagation paths, so they should all be recorded.

[0070] The simulation terminates when the parameter layer node is reached and the parameter of that node corresponds to a critical value that can trigger an early warning, or when the device layer node is reached and the system boundary has been exceeded, thus no longer having any further impact.

[0071] The forward deduction process, with the bearing as the starting node, is as follows: Starting from the bearing press-fitting station at the equipment level node, the vibration propagates through the directed edge to the inner ring of the bearing at the component level node, with a cumulative probability of 1.0 and a cumulative time of zero. From the inner ring of the bearing, the vibration propagates to the vibration curvature node and the vibration amplitude node respectively. The cumulative probability via the curvature path is 0.95, with a cumulative time of 24 hours; the cumulative probability via the amplitude path is 0.85, with a cumulative time of 48 hours. From the vibration curvature node, the vibration propagates to the phase difference offset node, with a cumulative probability of 0.95 multiplied by 0.75 equaling 0.7125 and a cumulative time of 24 plus 72 equaling 96 hours. From the phase difference offset node, the vibration propagates to the conveyor sprocket node, with a cumulative probability of 0.7125 multiplied by 0.70 equaling 0.49875 and a cumulative time of 96 plus 120 equaling 216 hours.

[0072] The simulation results show that, originating from the bearing press-fitting station, the primary transmission path is: bearing inner ring wear leading to vibration curvature, then to phase difference shift, and finally to the conveyor sprocket. The conveyor sprocket will begin to be affected approximately four days later. The secondary transmission path is: bearing outer ring wear leading to vibration amplitude, then to temperature rise, and finally to machine shutdown. A chain-like shutdown accident may occur approximately seven days later.

[0073] The unique technical approach of this solution lies in its directional extrapolation starting from the actual source of the fault, without traversing the entire graph, thereby improving extrapolation efficiency, reducing interference from invalid paths, and making the propagation prediction more closely match the actual fault development process.

[0074] Step S33: Failure time calculation and handling strategy generation; S33 calculates the expected failure time of each node based on the transmission path and edge weight, and generates the corresponding handling strategy in combination with the equipment operating status.

[0075] The estimated failure time is calculated by taking the minimum cumulative propagation time of all end nodes. When there are multiple end nodes, the minimum value is taken as the predicted value of the remaining safe operating time of the system, that is, the node that first reaches the warning threshold is regarded as the weakest link.

[0076] The response strategy is generated based on the fault source device and the predicted propagation path. The maintenance priority is ranked as follows: the source device is treated before the affected devices, the propagation path with high accumulation probability is treated first, and the path with short accumulation time is treated first.

[0077] The response strategies fall into three categories.

[0078] The first category is reduced-load operation. For source equipment, the operating load is reduced to slow down the rate of degradation, and the transmission time is extended after the load is reduced. The original operating load is denoted as L0, and the load after the load is reduced is β multiplied by L0, where β is less than one. Therefore, the extended failure time is the original failure time divided by β.

[0079] The second category is switching to backup equipment. For systems with redundant configurations, the load of the faulty device is switched to the backup equipment.

[0080] The third category involves cutting off the coupling path. For decoupling relationships between devices, the transmission path is cut off by disconnecting the coupling, closing the valve, or other methods.

[0081] For the transmission fault warning at the bearing press-fitting station, the handling strategy is as follows: reduce the load by decreasing the spindle speed from 300 rpm to 200 rpm, β is two-thirds, and the remaining life after extension is 180 hours; at the same time, coordinate the maintenance plan to arrange bearing replacement within 72 hours; the coupling path is cut off by setting up a buffer zone in front of the conveyor sprocket to add a flexible link.

[0082] The unique technical approach of this solution lies in binding the quantified failure time with directly executable actions, thereby improving the targeting of the response, reducing the risk of chain-like downtime, and achieving a complete closed loop from early warning to maintenance.

[0083] As shown in step S3, by integrating propagation reasoning, failure prediction, and handling output into a unified decision-making process based on a three-level directed graph, the propagation reasoning, failure prediction, and handling output can effectively prevent the spread of faults, extend the remaining life of equipment, and improve the continuous operation capability of the production line. Chain-related downtime accidents are reduced by approximately 60%, and unplanned downtime is reduced by approximately 40%.

[0084] Step S4: Detailed implementation of the model self-optimization iterative step: Step S4 is the system self-optimization phase, which automatically starts after a real equipment downtime fault. It traces back to pre-fault characteristic data, identifies early latent features, updates the edge weights of the deterioration propagation directed graph, and improves subsequent early warning accuracy. Fixed models cannot adapt to long-term effects such as equipment aging, changes in operating conditions, and load fluctuations, and early warning accuracy will gradually decrease. By using real fault data to correct the model, the system can continuously learn and gradually optimize. Specific technical methods are as follows: The system continuously monitors the operating status of the equipment. When a genuine equipment shutdown is detected, step S4 is triggered. To distinguish between planned shutdowns and fault shutdowns, the system accesses work order information from the production management system: if the shutdown time is within the planned maintenance window and the work order is marked as planned maintenance, it is determined to be a planned shutdown and no backtracking is triggered; if the shutdown is unplanned or associated with an alarm, it is determined to be a genuine fault shutdown.

[0085] The system automatically extracts complete historical data from the 72 hours preceding the shutdown time. The retrospective window length can be adjusted according to the equipment type; a 24-hour window is used for rapidly degrading equipment, and a 168-hour window is used for slowly degrading equipment. The extracted data types include time-series operational data, phase signal data, and vibration characteristic data, i.e., all features used in steps S1 to S3.

[0086] For each feature dimension, calculate its rate of change within the backtracking window, which is the difference between the mean of the second half of the window and the mean of the first half of the window, divided by the standard deviation. Mark the five features with the largest rates of change as sensitive features.

[0087] Based on the identified sensitive features and fault source devices, update the propagation probability weights and average propagation times of relevant edges in the directed graph.

[0088] The transmission probability is updated using an exponentially weighted average, meaning the new transmission probability equals the difference between one and the learning rate multiplied by the original transmission probability, plus the learning rate multiplied by the actual transmission indicator. The actual transmission indicator is either zero or one, and the learning rate ranges from 0.1 to 0.3; in this implementation, it is 0.2.

[0089] The average propagation time is updated by multiplying the difference between the new and the learning rate by the original average propagation time, and then adding the learning rate multiplied by the actual propagation time difference. The learning rate ranges from 0.1 to 0.3, and this implementation uses 0.2. For edges with high confidence, such as those derived from standardized propagation parameters provided by the equipment manufacturer, the learning rate is smaller; for edges whose initial values ​​are estimated based on expert experience, the learning rate is larger.

[0090] In a scenario where the main spindle bearing of a motor assembly line eventually causes a shutdown failure, the system traces back all relevant data from the 72 hours prior to the failure, identifies early latent features that were previously overlooked, and updates the weight and time parameters of the bearing degradation propagation to the sprocket in the directed graph. This allows for earlier warnings, more accurate location, and more reliable predictions when the same type of degradation occurs again.

[0091] As shown in S4, by using real faults as the sole supervisory signal, feature mining and weight updates are automatically completed without relying on manual annotation, improving the model's adaptability, reducing manual maintenance costs, and making the system more sensitive and accurate with use. The initial monthly warning accuracy was 75%. After six iterations of backtracking from real faults, the monthly warning accuracy increased to 95%, and the false alarm rate decreased from 15% to 3%.

[0092] In summary, this technical solution strictly combines four parts—early curvature mutation identification, phase difference matrix tracing, directed graph propagation deduction, and fault backtracking self-optimization—according to their temporal dependencies. Each step provides a unique trigger condition for the next, and they do not operate independently, are not reversed in order, or are simply superimposed. The synergistic relationship among the three is as follows: Curvature change detection serves as a trigger engine, capturing the earliest signs of early device degradation when the signal amplitude is normal. It generates an early warning signal and wakes up the phase difference analysis module, thus avoiding the waste of computing resources caused by the phase difference matrix running at high frequency throughout the day.

[0093] Phase difference matrix analysis, as a localization engine, utilizes the timestamp information of curvature abrupt changes, combined with the order and degree of phase shift, to accurately distinguish between the fault source equipment and passively affected equipment, providing an accurate starting point for directed graph deduction.

[0094] The degradation propagation directed graph serves as a decision engine, starting from the source device to perform multi-level forward deduction, predict the fault propagation path, the expected failure time of each node, and generate specific blocking strategies.

[0095] The core of this invention lies in the causal relationship in which the three technical means are triggered sequentially, depend on each other, and are executed in a closed loop, with none of them being dispensable.

[0096] The technical advantages of this invention are as follows: the early warning lead time is 120 to 336 hours earlier than the traditional threshold method; the source location accuracy is greater than 95%; the false alarm rate is controlled within 3%; chain-related downtime accidents are reduced by approximately 60%; and unplanned downtime is reduced by approximately 40%. In the first month of system deployment, the early warning accuracy was approximately 75%. After fault backtracking and iteration, the early warning accuracy improved to 95% in the twelfth month.

[0097] Based on the classification and technical framework of publicly available documents, this technical solution breaks through the limitations of traditional monitoring approaches. It combines a series of innovative methods, such as time-series trajectory curvature, phase difference matrix tracing, three-level directed graph reasoning, continuous window determination, dynamic threshold updating, time-series causal localization, and backtracking self-optimization updating, to form a complete and indivisible technical solution. This solution enables efficient perception and proactive handling of gradual degradation, coupled transmission, weak anomalies, and hidden faults. The overall technical effect is significantly better than existing publicly available technical methods.

[0098] The foregoing description is a detailed account of specific embodiments of the present invention, used to illustrate the technical solution of the present invention and not to limit it. The scope of protection of the present invention is determined by the claims.

Claims

1. An artificial intelligence-based early warning method, characterized in that: Includes the following steps: S1: Collect time-series operation data of the linked equipment, construct the equipment operation status trajectory, calculate the second derivative curvature of the trajectory, identify early equipment degradation based on curvature abrupt changes, and generate early warning signals; S2: In response to the warning signal, collect the operating phases of multiple devices in the linkage device, construct a phase difference matrix, and determine the fault source device based on the offset of the phase difference matrix; S3: Based on the directed graph of degradation propagation, perform forward propagation simulation starting from the fault source equipment to determine the fault propagation path and the expected failure time, and output the handling strategy; S1, S2, and S3 are triggered sequentially, depend on each other, and execute in a closed loop.

2. The artificial intelligence early warning method according to claim 1, characterized in that: S1 includes the following steps: S11: Perform filtering and noise reduction processing on the time-series running data; S12: Construct a continuous time-series trajectory based on the denoised data; S13: Calculate the second derivative curvature of the continuous time-series trajectory, compare the curvature with a dynamic threshold, and determine the curvature abrupt change when the curvature continuously exceeds the dynamic threshold.

3. The artificial intelligence early warning method according to claim 2, characterized in that: The curvature is calculated by the ratio of the cross product of the first and second derivatives of the trajectory to the cube of the first derivative's modulus. The dynamic threshold is adaptively updated based on the curvature distribution under normal operating conditions of the device.

4. The artificial intelligence early warning method according to claim 3, characterized in that: The curvature mutation determination in S13 uses the curvature exceeding the dynamic threshold within three consecutive sliding time windows as the trigger condition.

5. The artificial intelligence early warning method according to claim 1, characterized in that: S2 includes the following steps: S21: Extract the operating phase angles of multiple devices; S22: Construct a real-time phase difference matrix based on the running phase angle; S23: Calculate the offset between the real-time phase difference matrix and the reference phase difference matrix, quantify the degree of offset using the matrix norm, and identify the device with the largest offset and the curvature change first, followed by the phase offset, as the source of the fault.

6. The artificial intelligence early warning method according to claim 1, characterized in that: S3 includes the following steps: S31: Load the pre-built directed graph of degradation propagation. The directed graph contains device nodes, component nodes, parameter nodes, and directed edges representing degradation propagation relationships. The weights of the directed edges include propagation probability and average propagation time. S32: Starting from the fault source device, deduce the fault propagation path in a forward direction along the directed edge; S33: Calculate the estimated failure time of each node based on the transmission path and generate the corresponding handling strategy.

7. The artificial intelligence early warning method according to claim 6, characterized in that: The handling strategies in S33 include load reduction operation, switching to backup equipment, and cutting off coupling paths.

8. The artificial intelligence early warning method according to claim 1, characterized in that: It also includes step S4: After the equipment experiences a real shutdown fault, it automatically traces back the time-series operation data, phase signal data, and vibration characteristic data from the 72 hours prior to the fault, identifies early latent characteristics, and updates the edge weights of the deterioration transmission directed graph.

9. The artificial intelligence early warning method according to claim 8, characterized in that: The system continuously monitors the operating status of the equipment. When a real equipment shutdown is detected, step S4 is triggered. If the shutdown time is within the planned maintenance window and the work order is marked as planned maintenance, it is determined to be a planned shutdown and no backtracking is triggered. If the shutdown is unplanned or associated with an alarm, it is determined to be a real fault shutdown.

10. An artificial intelligence early warning system, characterized in that: It includes a data acquisition module, an early degradation identification module, a fault source location module, a transmission and deduction handling module, and a self-optimization and iteration module; The data acquisition module is used to collect the timing operation data of the linked equipment and the operation phase data of multiple devices; The early degradation identification module is used to filter and denoise the time-series running data, construct a continuous time-series trajectory, calculate the second derivative curvature of the trajectory, and generate an early warning signal based on curvature abrupt changes. The fault source location module is used to extract the operating phase angles of multiple devices in response to early warning signals, construct a phase difference matrix and calculate the matrix offset, and determine the fault source device by combining the temporal causal relationship. The transmission simulation and handling module is used to load the deterioration transmission directed graph, deduce the fault transmission path from the fault source equipment, calculate the expected failure time of each node and output the handling strategy; The self-optimizing iterative module is used to identify early latent features by tracing back historical feature data after a real equipment downtime failure, update the edge weights of the deterioration propagation directed graph, and realize the system's adaptive iterative optimization.

Images