Bearing fault prediction method and system based on AGV cluster

Abstract

The embodiment of the invention aims to provide a bearing fault prediction method and system based on an AGV cluster. The method comprises the following steps: acquiring data of sensors arranged at hub bearings of all AGVs in real time; inputting a fault prediction model according to a sliding time window to carry out real-time bearing fault prediction on the current AGV, wherein a prediction result comprises a fault type and a fault probability; executing corresponding dynamic task scheduling according to the risk interval of the fault probability; and querying a fault-measure mapping knowledge base according to the fault type, and generating a maintenance work order of the current AGV. According to the embodiments of the invention, through integration of dynamic prediction, cluster cooperation and a closed-loop decision-making mechanism, a significant technical breakthrough is realized in the field of AGV bearing fault prediction. According to the overall scheme, a'prediction-verification-control-maintenance 'closed loop link is formed, the operation and maintenance resource allocation is optimized while the unplanned shutdown risk is reduced, and the operation robustness and the life cycle management efficiency of the AGV cluster are remarkably improved.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of IoT situational awareness, and in particular to a technique for bearing fault prediction based on AGV clusters. Background Technology

[0002] Current methods for detecting bearing failures in automated guided vehicles (AGVs) typically rely on vibration monitoring of individual units, such as collecting instantaneous status data from vibration sensors. When the detected vibration amplitude exceeds a preset threshold, the system triggers an alarm signal to indicate a fault. This approach has several limitations: First, single-point decision-making lacks reliability. Sensors in a single AGV are susceptible to occasional interference (such as vibration peak drift caused by ground bumps), often triggering false alarms; at the same time, early fault characteristics are easily missed when they are weak. Although some studies have attempted multi-sensor fusion, they have not fully utilized the similarity of operating conditions among cluster devices for cross-validation.

[0003] Secondly, they have poor adaptability to operating conditions. Traditional models (such as SVM and LSTM) are slow to respond to changes in operating conditions. For example, the natural temperature rise of bearings under high load may mask the temperature anomalies of the actual fault. Existing methods mostly use fixed weights or simple rules to adjust feature thresholds, making it difficult to dynamically learn the complex influence mechanisms of multi-dimensional states such as load, speed, and ground type on fault characteristics.

[0004] Third, the prediction-execution chain is broken. Fault detection is not linked to AGV cluster scheduling in real time. When high-risk faults occur, the equipment continues to run at full speed, which can worsen the fault and cause a chain reaction that may occur to the entire task cluster due to the slowdown / failure of a single task. Summary of the Invention

[0005] The purpose of the various embodiments disclosed herein is to provide a bearing failure prediction method and system based on AGV clusters.

[0006] According to one aspect of this disclosure, a bearing failure prediction method based on AGV clusters is provided, wherein the method includes the following steps: Real-time acquisition of sensor data located at the bearings of each AGV wheel hub. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status. The data records are input into the fault prediction model according to the sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction results include fault type and fault probability. The fault prediction model dynamically learns the influence of the operating state on the fault feature weights through a self-attention mechanism. When the fault probability is in the medium-risk range, the current AGV is compared with AGVs with similar operating conditions in the AGV cluster to confirm the fault. Based on the risk range of the failure probability, corresponding dynamic task scheduling is performed. Specifically, when the failure probability is in the medium-risk range, the current AGV is triggered to degrade its operation mode; when the failure probability is in the high-risk range, the tasks of the current AGV are dynamically allocated to other AGVs in the AGV cluster. Based on the fault type, query the fault-response mapping knowledge base to generate a maintenance work order for the current AGV.

[0007] According to one aspect of this disclosure, a bearing failure prediction system based on an AGV cluster is also provided, wherein the system includes edge computing nodes placed in each AGV and a cloud-based failure prediction system. The edge computing node is configured to perform the following operations: Real-time acquisition of sensor data located at the bearings of each AGV wheel hub. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status. The cloud-based fault prediction system is configured to perform the following operations: The data records are input into the fault prediction model according to the sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction results include fault type and fault probability. The fault prediction model dynamically learns the influence of the operating state on the fault feature weights through a self-attention mechanism. When the fault probability is in the medium-risk range, the current AGV is compared with AGVs with similar operating conditions in the AGV cluster to confirm the fault. Based on the risk range of the failure probability, corresponding dynamic task scheduling is performed. Specifically, when the failure probability is in the medium-risk range, the current AGV is triggered to degrade its operation mode; when the failure probability is in the high-risk range, the tasks of the current AGV are dynamically allocated to other AGVs in the AGV cluster. Based on the fault type, query the fault-response mapping knowledge base to generate a maintenance work order for the current AGV.

[0008] The embodiments disclosed herein achieve significant technological breakthroughs in the field of AGV bearing fault prediction by integrating dynamic prediction, cluster collaboration, and closed-loop decision-making mechanisms. The fault prediction model utilizes a self-attention mechanism to dynamically learn the correlation between operating states and fault characteristics. For example, it automatically increases the weight of vibration spectrum features under high-load conditions and reduces the false alarm threshold in scenarios with gradual temperature increases, thereby accurately distinguishing between normal fluctuations caused by operating conditions and real fault signals, overcoming the shortcomings of traditional models in adaptability to varying operating conditions. When the fault probability falls into the medium-risk range, by comparing the operating data (such as vibration spectrum trends, temperature change rates, etc.) of the current AGV with similar operating conditions in the cluster, it effectively identifies false anomalies caused by occasional sensor drift or instantaneous impacts, avoiding misjudgments caused by fluctuations in data from a single device, and significantly improving the reliability of early fault diagnosis. Dynamic task scheduling converts the prediction results into control commands in real time: in medium-risk situations, it actively triggers a degraded operating mode for the equipment, maintaining basic operational capabilities while preventing fault deterioration; in high-risk situations, it initiates a task reassignment mechanism in milliseconds to ensure the overall operational continuity of the cluster. The maintenance work order generation process ensures a high degree of matching between maintenance strategies and failure modes through precise mapping between failure types and the knowledge base, thereby shortening failure response time. The overall solution forms a closed-loop chain of "prediction-verification-control-maintenance," which reduces the risk of unplanned downtime while optimizing the allocation of operation and maintenance resources, significantly improving the operational robustness and lifecycle management efficiency of the AGV cluster.

[0009] Furthermore, the embodiments of this disclosure improve decision-making reliability by introducing a three-level collaborative diagnostic mechanism. The single-machine layer captures typical faults to ensure response speed, the cluster layer filters out occasional interference through cross-validation of equipment under similar operating conditions, and the group layer periodically scans to predict batch risks. This architecture extends single-point decision-making to collective intelligent consensus, significantly suppressing false alarms caused by sensor drift or instantaneous impacts; at the same time, weak fault characteristics are more easily identified in comparison with similar equipment, reducing the early fault false alarm rate.

[0010] The Transformer model's self-attention mechanism dynamically learns the correlation between operating states and fault characteristics, improving its adaptability to different operating conditions. For example, it automatically strengthens the weight of high-frequency components of the vibration spectrum under high-load conditions and reduces the false alarm threshold in scenarios with gradual temperature increases. Compared to traditional fixed-rule models, this design accurately distinguishes between normal characteristic fluctuations caused by operating conditions and real fault anomalies, avoiding overly conservative or overly lenient misjudgments.

[0011] Dynamic task scheduling based on failure probability improves the execution capability of failure prediction. In medium-risk situations, proactive degradation prevents failure escalation, while in high-risk situations, millisecond-level task migration ensures production line continuity. The maintenance knowledge base continuously accumulates experience through NLP self-evolution, ensuring that maintenance strategies accurately match failure modes. An online learning mechanism applies penalty weights to false alarm samples, driving the model to continuously optimize its decision boundaries.

[0012] The embodiments of this disclosure reduce downtime caused by bearing failures while avoiding unnecessary maintenance resource consumption; they also extend equipment life through early warning and collaborative diagnosis, thereby improving the overall operational efficiency and safety of AGV clusters. Attached Figure Description

[0013] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating a bearing failure prediction method based on an AGV cluster according to an embodiment of the present disclosure is shown.

[0014] The same or similar reference numerals in the accompanying drawings represent the same or similar parts. Detailed Implementation

[0015] The specific embodiments of this disclosure will be further described below with reference to the accompanying drawings.

[0016] Before discussing the exemplary embodiments in more detail, it should be noted that some exemplary embodiments of this disclosure are described as apparatuses represented by block diagrams and processes or methods represented by flowcharts. Although the flowcharts depict the operation processes of the various embodiments of this disclosure as sequential processes, many of the operations may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The processes of the various embodiments of this disclosure may be terminated when their operations are completed, but may also include additional steps not shown in the flowcharts. The processes of the various embodiments of this disclosure may correspond to methods, functions, procedures, subroutines, subroutines, etc.

[0017] The methods illustrated in the flowcharts and the apparatuses illustrated in the block diagrams discussed below can be implemented in hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, program code or code segments that perform the necessary tasks can be stored in a machine or a computer-readable medium such as a storage medium. One or more processors can perform the necessary tasks.

[0018] Similarly, it will also understand any flowchart, state transition diagram, and the like, representing various processes that can be adequately described as program code stored in a computer-readable medium and thus executed by a computer device or processor, whether or not such computer device or processor is explicitly shown.

[0019] In this document, the term "storage medium" can refer to one or more devices for storing data, including read-only memory (ROM), random access memory (RAM), magnetic RAM, core memory, disk storage media, optical storage media, flash memory devices, and / or other machine-readable media for storing information. The term "computer-readable medium" may include, but is not limited to, portable or fixed storage devices, optical storage devices, and various other media capable of storing and / or containing instructions and / or data.

[0020] A code segment can represent a procedure, function, subroutine, program, routine, subroutine, module, software package, class, or any combination of instructions, data structures, or program descriptions. A code segment can be coupled to another code segment or hardware circuit by passing and / or receiving information, data, arguments, parameters, or stored content. Information, arguments, parameters, data, etc., can be passed, forwarded, or transmitted via any suitable means, including storage sharing, message passing, token passing, network transmission, etc.

[0021] In this context, "computer device" refers to an electronic device that can perform predetermined processing procedures such as numerical calculations and / or logical calculations by running predetermined programs or instructions. It may include at least a processor and a memory, wherein the predetermined processing procedures are performed by the processor executing program instructions pre-stored in the memory, or by hardware such as ASIC, FPGA, DSP, etc., or by a combination of the above.

[0022] The term "computer device" as used above is generally embodied in the form of a general-purpose computer device, whose components may include, but are not limited to, one or more processors or processing units and system memory. System memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The "computer device" may further include other removable / non-removable, volatile / non-volatile computer-readable storage media. The memory may include at least one computer program product having a set (e.g., at least one) of program modules configured to perform the functions and / or methods of the embodiments of this disclosure. The processor executes various functional applications and data processing by running programs stored in the memory.

[0023] For example, a computer program is stored in the memory for performing various functions and processes of the various embodiments of the present disclosure, and when the processor executes the corresponding computer program, the various embodiments of the present disclosure are implemented.

[0024] Typically, computer devices can be user devices or network devices, or even a combination of both. User devices include, but are not limited to, personal computers (PCs), laptops, and mobile terminals; mobile terminals include, but are not limited to, smartphones and tablets. Network devices include, but are not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing, which is a type of distributed computing consisting of a super virtual computer composed of a group of loosely coupled computers. The computer devices can operate independently to implement the embodiments of this disclosure, or they can connect to a network and implement the embodiments of this disclosure through interaction with other computer devices in the network. The networks in which the computer devices reside include, but are not limited to, the Internet, wide area networks (WANs), metropolitan area networks (MANs), local area networks (LANs), and VPN networks.

[0025] It should be noted that the user equipment, network equipment, and network mentioned are merely examples. Other existing or future computing devices or networks that are applicable to the embodiments of this disclosure should also be included within the scope of protection of this disclosure and are incorporated herein by reference.

[0026] The specific structural and functional details disclosed herein are merely representative and are intended to describe exemplary embodiments of this disclosure. However, the various embodiments of this disclosure can be implemented in many alternative forms and should not be construed as being limited solely to the embodiments set forth herein.

[0027] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments. Unless the context clearly indicates otherwise, the singular forms “a” and “an” as used herein are also intended to include the plural. It should also be understood that the terms “comprising” and / or “including” as used herein specify the presence of the stated features, integers, steps, operations, units, and / or components, without excluding the presence or addition of one or more other features, integers, steps, operations, units, components, and / or combinations thereof.

[0029] It should also be mentioned that in some alternative implementations, the functions / actions mentioned may occur in a different order than those shown in the figures. For example, depending on the functions / actions involved, the two figures shown successively may actually be executed substantially simultaneously or sometimes in reverse order.

[0030] See Figure 1 The illustration shows a bearing failure prediction process based on an AGV cluster according to one embodiment of the present disclosure.

[0031] like Figure 1 As shown, in step S1, the fault prediction system collects sensor data from the bearings of each AGV hub in real time. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status. In step S2, the fault prediction system inputs the data records into the fault prediction model according to a sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction result includes the fault type and fault probability. In step S3, the fault prediction system performs corresponding dynamic task scheduling based on the risk range of the fault probability. In step S4, the fault prediction system queries the fault-measure mapping knowledge base based on the fault type to generate a maintenance work order for the current AGV.

[0032] Specifically, in step S1, the fault prediction system collects sensor data located at the wheel hub bearings of each AGV in real time.

[0033] In some embodiments, the fault prediction system may include edge computing nodes deployed at each AGV and a cloud prediction system.

[0034] Here, edge computing nodes are deployed at each AGV to synchronize the data collected by the sensors and generate corresponding data records. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status.

[0035] During the initial data acquisition phase, the edge computing nodes first initiate a multi-source sensor synchronous acquisition process. Triaxial vibration sensors deployed at the bearings of each AGV wheel hub capture the raw vibration waveforms at a sampling frequency of 1-10kHz, while temperature sensors acquire bearing surface temperature data at a fixed frequency of 1Hz. The edge nodes apply nanosecond-level timestamps to all sensors using the IEEE 1588 precision clock synchronization protocol to ensure strict temporal consistency between vibration and temperature data, avoiding feature distortion caused by time shifts.

[0036] After the original vibration waveform is input to the edge nodes, it is segmented using a real-time sliding window processing mechanism: every 1000 consecutive vibration data points received (corresponding to 0.1 seconds of data sampled at 10kHz), a Fast Fourier Transform is performed on the window. The algorithm automatically extracts the spectral energy near the bearing outer race fault characteristic frequency (BPFO), calculates its percentage of the overall spectral energy, and generates a BPFO energy proportion feature value. The temperature data stream is differentially calculated with a period of 10 seconds, and the temperature difference between adjacent sampling points is divided by the time interval to output the temperature change rate (in °C / s).

[0037] For asynchronously triggered AGV running status events, edge nodes dynamically monitor load change signals and speed adjustment commands. When a status update is detected, the discrete event is converted into a structured code: the load status is mapped to a three-level hot vector of "high / medium / low", and the speed status is mapped to a three-level hot vector of "high speed / medium speed / low speed". For example, the full-load high-speed running status is encoded as [1,0,0,1,0,0].

[0038] Finally, the edge nodes encapsulate the processing results into data records based on a unified time standard. Each record contains six core fields: a timestamp accurate to microseconds, a unique AGV device identifier, the peak value of the vibration window (the absolute maximum value of the vibration signal within the window), the BPFO energy percentage, the temperature change rate, and the operating status vector. The encapsulated data records are pushed to the cloud prediction system in real time via a lightweight message queue, achieving end-to-end latency control within 50 milliseconds.

[0039] This step solves a key challenge in collaborative data acquisition from heterogeneous sensors in industrial settings by employing a spatiotemporally coupled multimodal data fusion mechanism. 1. Significantly improved anti-interference capability: Nanosecond-level clock synchronization eliminates millisecond-level time shift errors in vibration and temperature data in traditional methods, significantly improving the detection accuracy of the time-series correlation between BPFO energy mutation and temperature rise. For example, it can accurately capture the early warning window of "abnormal vibration 5-10 seconds earlier than temperature rise" in the early stage of lubrication failure. 2. Lightweight edge intelligence: The sliding window FFT performs spectrum analysis locally, significantly compressing the transmission bandwidth of the original waveform while avoiding the impact of high-frequency vibration sampling data on the cloud system; 3. Universality of state coding: Asynchronous events are transformed into one-hot vectors, enabling subsequent fault prediction models to uniformly handle continuous sensor data and discrete operating condition transitions, breaking through the limitations of manually setting operating condition weight rules in traditional methods.

[0040] The strict synchronization between temperature change rate and vibration characteristics allows the fault prediction model to automatically discover the implicit pattern that "the BPFO energy ratio needs to be increased by 20% alarm threshold under low temperature conditions" during subsequent training. Such cross-physical domain coupling effects are difficult for machine learning to capture in asynchronous systems. This provides a new dimension of reliability assurance for fault prediction of AGV clusters in extreme environments.

[0041] In step S2, the fault prediction system inputs the data records into the fault prediction model according to the sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction results include the fault type and fault probability.

[0042] In some embodiments, after receiving the data recording stream pushed by the edge node, the prediction system uses a sliding time window mechanism to organize the model input data. Every second, the system extracts continuous data records from the current AGV within the last 60 seconds to form a time series sequence, ensuring coverage of the complete evolution cycle of the bearing's operating state. This time series sequence is converted into a three-dimensional feature tensor: the first dimension corresponds to 60 time steps, the second dimension contains three sensor data points: vibration peak value, BPFO energy ratio, and temperature change rate, and the third dimension carries six-dimensional operating condition parameters encoded by the unique thermal vector of the operating state.

[0043] The fault prediction model dynamically learns the influence of operating state on fault feature weights through a self-attention mechanism, including at least one of the following: - Attention weighting for increasing the energy percentage of BPFO under high temperature conditions; - Increase the attention weight of high-frequency components of the vibration spectrum under high load conditions.

[0044] The fault prediction model employs a Transformer architecture to process the temporal tensor. The model first maps each feature to a high-dimensional vector space through an embedding layer, transforming the original physical quantities into semantically related mathematical representations. In the core processing stage, a self-attention mechanism dynamically calculates the correlation weights between different time steps and different features. On the feature vector corresponding to the high-temperature state, the model automatically enhances the weight coefficient of the BPFO energy proportion feature. For example, when the temperature change rate continuously exceeds 2℃ / s, the system increases the importance of the BPFO feature to 1.8 times that of the normal temperature state, making the model focus on vibration spectrum anomalies; When the operating state vector indicates a high-load condition, the model synchronously increases the attention weight of high-frequency components of the vibration signal above 400Hz. This dynamic adjustment enables the model to capture the micro-deformation characteristics of the bearing caused by sudden load changes.

[0045] The multi-head attention module executes eight sets of weight learning processes in parallel, each focusing on the interaction patterns of different feature combinations. The output attention matrix, after residual connections and layer normalization, is then used by a feedforward neural network to extract cross-modal fusion features. The final classification layer outputs prediction results based on the fusion features, such as the probability distribution of fault types, including six types of faults such as outer ring peeling and lubrication failure, along with their respective predicted probabilities, with prediction latency controlled within 100 milliseconds.

[0046] The fault prediction model based on the Transformer architecture disclosed herein achieves dynamic diagnostic capabilities that are unattainable by traditional models through an adaptive feature weighting mechanism based on operating conditions. 1. Intelligent focusing driven by physical mechanism: The self-attention weight is automatically adjusted according to temperature / load conditions. For example, under high temperature conditions, the weight of BPFO features is elevated to the dominant position. This simulates the physical law of "temperature rise accelerates fatigue crack propagation" in bearing science, which effectively reduces the false detection rate of the model under complex conditions. 2. Cross-modal coupling effect mining: The model autonomously discovers the implicit correlation between high-frequency vibration components and high load conditions. For example, when the unique heat vector indicates that the load exceeds 80% of the rated value, the model increases the vibration weight of the 2-4kHz frequency band by 2.2 times, effectively identifying the bearing raceway micro-deformation risk that is ignored by traditional spectrum analysis.

[0047] The self-attention mechanism automatically constructs fault feature compensation relationships during training. When vibration sensors experience data loss due to electromagnetic interference, the model maintains high diagnostic accuracy based on the combined weights of temperature change rate and load state. This cross-sensor redundancy capability significantly improves system robustness and provides a new paradigm for interference-resistant design in industrial settings.

[0048] In some embodiments, the training dataset for the fault prediction model is constructed based on historical sensor data from the AGV cluster and fault markers associated with timestamps in maintenance work order records.

[0049] The training data construction phase first accesses the historical database of the AGV cluster, extracting sensor data streams and maintenance work order records from the past 6 months. A timestamp alignment mechanism is used to link the two data sources: using the fault occurrence time in the maintenance work order record as a baseline, sensor data fragments from the 72 hours prior to the fault are extracted, while simultaneously collecting normal operating data from fault-free AGVs during the same period as negative samples. Each sample is standardized into a time-series data block in a uniform format, containing vibration peak values, BPFO energy percentage, temperature change rate, and operating status vectors for 1440 consecutive time steps (sampled at one record per minute).

[0050] For each fault type described in a maintenance work order, the system automatically labels the samples with fault category tags. For example, when a work order records "outer ring peeling," the last 12 hours in the corresponding data block are marked as the fault window. At the same time, a sliding time window augmentation strategy is introduced: sub-samples are truncated in 5-minute increments, allowing a single fault event to generate 300+ training sub-sequences, significantly expanding the amount of data for rare fault categories.

[0051] The Transformer model employs a multi-task learning architecture for end-to-end training. The task is fault type classification, and the output layer uses a softmax function to calculate the probability distribution of six fault types. During training, the learning objective of the self-attention mechanism is set to maximize the significance of the feature weights of the fault window. When the input samples include high-temperature operating conditions, the loss function adds an additional BPFO feature weight deviation penalty term, forcing the model to allocate more attention to the high-temperature range. During training with high load samples, the optimizer automatically increases the gradient backpropagation intensity of the vibration spectrum in the 400-2000Hz frequency band, enhancing the model's sensitivity to anomalies in this frequency band.

[0052] The training scheme for the prediction model overcomes the limitations of traditional supervised learning by dynamically labeling fault windows and optimizing based on operating conditions. 1. Long-cycle failure evolution capture: The 72-hour backtracking window covers the entire process of bearing degradation from early stage to complete failure, enabling the model to learn to identify subtle signs such as "periodic oscillation of temperature change rate" that appear 24 hours before lubrication failure, which significantly improves the predictive timeliness compared to the traditional 2-hour segment. 2. Deep integration of maintenance knowledge: The text descriptions in maintenance work orders (such as "microcracks in the raceway") are converted into physical feature constraints of the fault type. For example, the model is forced to add monitoring weights for the 1-3kHz frequency band of vibration signals in the attention layer, so as to achieve collaborative optimization of human experience and machine learning.

[0053] When inputting composite fault samples (such as lubrication failure accompanied by outer ring spalling), the self-attention layer exhibits a time-dependent pattern of "temperature feature weights increasing first, followed by a surge in high-frequency vibration weights," which perfectly matches the failure mechanism in bearing mechanics where "poor lubrication leads to overheating, which in turn induces material spalling." This ability to autonomously discover physical laws without manual annotation provides a new approach for root cause analysis of complex mechanical systems.

[0054] In some embodiments, when the failure probability of the current AGV is in the medium-risk range, cluster collaborative diagnosis is triggered to compare the current AGV with AGVs with similar operating conditions in the AGV cluster to confirm the failure.

[0055] According to one example, a failure probability of <30% is considered a low-risk range, a failure probability between 30% and 70% is considered a medium-risk range, and a failure probability >70% is considered a high-risk range.

[0056] According to one example, for AGVs in the medium-risk range, multiple AGVs with similar operating conditions to the current AGV are retrieved in the AGV cluster based on locality-sensitive hashing. If at least one of these AGVs experiences a fault of the same type as the current AGV, then the current AGV is confirmed to have failed.

[0057] Specifically, the current AGV's operating state vector is input into a pre-trained set of locality-sensitive hash functions to obtain bucket codes; within the same hash bucket, k nearest neighbor AGVs with the same ground type and load deviation < a preset threshold (e.g., 10%) are retrieved as similar working condition groups; if there is ≥1 AGV in the similar working condition group that simultaneously meets the conditions of vibration anomaly and temperature trend synchronization, and the fault type is consistent with the type predicted by the model, then the current AGV's fault confidence level is raised to the confirmed state.

[0058] When the failure probability of the current AGV enters the medium-risk range, the prediction system initiates the local sensitive hash retrieval process.

[0059] First, the current AGV's operating state vector (a six-dimensional vector containing load level, speed level, and ground type encoding) is input into a pre-trained set of LSH hash functions. This set of functions maps the high-dimensional vector to a low-dimensional space using a set of random projection matrices and applies a sign function to binarize the projection result, generating a binary hash code. For example, the AGV state vector [high load, medium speed, epoxy ground] outputs a "101110" bucket code after hash transformation, meaning AGVs with the same operating conditions fall into the same hash bucket with a high probability.

[0060] Next, the system performs refined nearest neighbor filtering within the target bucket: 1) Initial screening: Extract the real-time status vector of all AGVs in the same hash bucket and exclude devices with mismatched ground types. For example, if the ground type of the current AGV is concrete pavement, remove the epoxy pavement device. 2) Fine screening: Calculate the load deviation between the remaining AGVs and the current AGV, and retain only AGVs with a deviation of less than 10%, thus obtaining candidate AGVs; 3) Sorting: Sort the candidate AGVs according to Euclidean distance, and select the k AGVs with the smallest distance (k≥3) as the similar working condition group.

[0061] For the selected AGV groups with similar operating conditions, the system requests their sensor data streams for the most recent 10 minutes in parallel and performs cross-device feature comparison: 1) Vibration spectrum analysis: Calculate the difference in BPFO energy ratio between the current AGV and AGVs under similar working conditions. If the difference value of one of the AGVs exceeds the health threshold (e.g., >15%), it is marked as abnormal. 2) Temperature trend verification: The temperature change rate curve is aligned using a dynamic time warping algorithm to detect trend consistency. When the current AGV experiences a sudden temperature rise, if at least one AGV in the similar group shows a synchronous change (warped path distance < 0.1), it is confirmed that the temperature is not caused by environmental interference.

[0062] The system executes the final decision. If there is at least one AGV in the similar working condition group that simultaneously meets the conditions of abnormal vibration and synchronous temperature trend, and the fault type is consistent with the type predicted by the model, then the current AGV fault confidence level is raised to the confirmed state.

[0063] This step achieves industrial-grade real-time operating condition matching through a three-level hash-filter-verification mechanism.

[0064] LSH reduces the computational complexity of similarity matching for AGV clusters of tens of thousands from O(N) to O(1) using a hash bucket mechanism. A ground type-priority filtering strategy eliminates a large number of interfering factors; for example, the BPFO baseline values ​​of AGVs on concrete pavements and those on epoxy pavements differ significantly, and traditional Euclidean distance cannot effectively distinguish between them. Dynamic time warping solves the problem of data acquisition time discrepancies between devices, enabling the capture of the fault evolution pattern of "vibration anomaly followed by temperature rise" even if the temperature sampling points of similar AGVs are not strictly aligned.

[0065] Traditional clustering algorithms (such as K-means) rely solely on numerical distance, while LSH preserves the topological structure of the vector space through random projection, thus strictly separating the "high load + epoxy floor" condition from the "high load + concrete floor" condition into different hash buckets, which conforms to the influence of ground stiffness on vibration transmission in bearing failure characteristics.

[0066] Furthermore, the LSH bucket distribution automatically forms operating condition clusters. Maintenance personnel discovered that a certain hash bucket (code "110011") frequently triggered alarms. Tracing back revealed the corresponding operating condition combination of "full load + low speed + metal grid ground". Further detection showed that the grid ground caused continuous micro-impact on the bearing, thus enabling root cause tracing of the fault. For example, the impact load at the grid joint was the main cause of bearing spalling.

[0067] In some embodiments, the prediction system provides cluster-based collaborative diagnostics, including a multi-layered diagnostic strategy: Single-machine direct diagnostic layer: In response to a sudden increase in BPFO energy percentage >300% or a temperature change rate >5℃ / min, skipping collaborative verification, directly confirming the current AGV malfunction. For example, the vibration spectrum of a detached bearing outer ring has a unique spike. This ensures a fast response time for high-risk faults.

[0068] Cluster Strong Verification Layer: In response to faults in the medium-risk range, the fault status is confirmed when at least one AGV in a similar operating condition group experiences a fault of the same type as the current AGV. Furthermore, if all AGVs in the similar operating condition group are functioning normally, the warning for the current AGV is downgraded. This avoids interference from occasional noise, such as temporary drift of a single sensor.

[0069] Early screening layer for groups: Periodic scanning, such as every 24 hours, is performed on AGV clusters of the same batch. If ≥3 AGVs show similar anomalies, a batch defect warning is triggered. This allows for the early detection of systemic risks, such as batch quality issues with bearings.

[0070] In step S3, the fault prediction system performs corresponding dynamic task scheduling based on the risk range in which the current AGV's fault probability is located.

[0071] In some embodiments, when the failure probability is in the medium-risk range, the current AGV is triggered to degrade its operating mode.

[0072] In some embodiments, when the failure probability is in a high-risk range, the task of the current AGV is dynamically assigned to other AGVs in the AGV cluster.

[0073] Based on real-time fault prediction probability, dynamic task scheduling is performed on the faulty AGV (current AGV), including: 1) In the case of medium risk, trigger the faulty AGV to enter low speed / low load mode to prevent the fault from worsening, while maintaining the basic task execution capability; 2) In high-risk situations, the tasks of the faulty AGV are dynamically assigned to other healthy AGVs in the AGV cluster to ensure the continuity of the overall operation.

[0074] Once the fault prediction model outputs the bearing failure probability of the current AGV, the system can execute a tiered response based on a preset risk threshold. If the failure probability is in the medium-risk range of 30%-70%, the system sends a degraded operation instruction to the current AGV, such as limiting the maximum travel speed to 50% of the rated value, and simultaneously adjusting its load limit to 60% of the normal threshold through the task manager. In this state, the AGV can still perform basic transport tasks, but the system automatically avoids high-risk paths such as slopes and uneven road surfaces until the failure probability falls back to the safe range.

[0075] When the failure probability exceeds the 70% high-risk threshold, the system initiates a cluster task reallocation process. First, the current task queue of the failed AGV is frozen, and the key attributes of the tasks to be migrated are analyzed through the global scheduler. 1) Task urgency: such as prioritizing the relocation of material supply line tasks; 2) Spatial proximity: Prioritize healthy AGVs located ≤50 meters from the target location; 3) Capacity matching: Check the load capacity of the faulty AGV and its compatibility with the rack type.

[0076] Based on the above dimensions, a migration score is generated, and tasks are dynamically assigned to the standby AGV with the highest score. New task packages are sent in real time via an encrypted channel, and the receiving AGV responds to the instructions and takes over the operation within 200 milliseconds.

[0077] After the task migration is completed, the system performs safety isolation on the faulty AGV: cutting off its task allocation channel, triggering an audible and visual alarm to guide personnel intervention, and marking the equipment status as "maintenance standby". The system can also synchronously update the AGV cluster topology map, using flashing red to indicate the location of high-risk equipment.

[0078] This step achieves dual-objective optimization of predictive maintenance and production continuity through a fault probability-driven elastic scheduling mechanism: 1. Innovative Fault Suppression Capability: The medium-risk speed reduction strategy significantly reduces bearing impact load, effectively preventing early failures from evolving into irreversible damage. In practice, it has successfully extended the lubrication failure deterioration cycle, gaining critical time for maintenance. 2. Dynamic Reconfiguration of Cluster Resources: The task migration algorithm considers both spatial proximity and capability matching, ensuring overall operational efficiency is maintained even in high-risk events. For example, when an AGV on an automotive assembly line experiences a sudden bearing jam, the system diverts 12 welding part transfer tasks to 3 backup AGVs within 5 seconds, resulting in zero production line downtime.

[0079] Low-load operating data collected during degraded operation can provide new learning samples for the model. For example, the system detected that the temperature change rate of a certain AGV model abnormally dropped to zero under 40% load, leading to a reverse diagnosis of a viscous temperature sensor fault. This fault diagnosis feedback mechanism based on scheduling status forms a closed-loop self-optimizing system, significantly reducing the false alarm rate.

[0080] In step S4, the fault prediction system queries the fault-measure mapping knowledge base based on the fault type of the current AGV and generates a maintenance work order for the current AGV.

[0081] In some embodiments, the maintenance work order automatically generated by the system includes the fault type (such as outer ring peeling, lubrication failure) and recommended maintenance measures obtained through the fault-measure mapping knowledge base (such as: outer ring peeling → replace bearing; lubrication failure → add lubricating oil).

[0082] Once the fault prediction model confirms the bearing fault type of the current AGV, the system can access the fault-measure mapping knowledge base to generate a maintenance work order. This knowledge base uses a graph database to store the association rules between fault types and maintenance measures. Each fault node includes a physical characteristic description, typical sensor patterns, and corresponding countermeasures. For example, when the fault type "lubrication failure" is entered, the system retrieves the associated maintenance operation as "replenish ISO VG220 synthetic grease," along with the following operating specifications: "clean the grease filler nozzle before injecting, single replenishment amount ≤ 15ml." The maintenance work order is automatically filled with basic information such as equipment number, fault probability, and timestamp, and then pushed to the operation and maintenance system.

[0083] In some embodiments, fault-measure rules are automatically extracted by parsing maintenance work order records using NLP to expand the fault-measure mapping knowledge base.

[0084] Based on an example, the BERT model is used to perform entity recognition on maintenance work orders, extracting triples of fault location, fault mode, and maintenance measures; the extracted triples are then automatically added to the fault-measure mapping knowledge base.

[0085] The dynamic updating of the fault-response mapping knowledge base is achieved through a natural language processing engine. The system periodically scans historical maintenance work order records and uses a pre-trained BERT model to parse the work order text. 1) Entity recognition: The BERT model recognizes three types of entities in the text: fault location (e.g., "wheel hub bearing"), fault mode (e.g., "outer ring peeling"), and repair measures (e.g., "replace SKF 6308 bearing"); 2) Relation extraction: Construct causal rules for "failure mode → maintenance measures" based on semantic role labeling. For example, when the work order description is "abnormal noise caused by raceway peeling, normal after bearing replacement", extract the triple <outer ring peeling, maintenance measures, bearing replacement>. 3) Knowledge integration: New rules are entered into the database after reliability verification. If they conflict with existing rules (such as a fault that has both "grease replenishment" and "bearing replacement" measures), the expert review process will be triggered.

[0086] Each new rule comes with a confidence score, initially set at 0.8. The score increases when the rule is repeatedly referenced in subsequent work orders and decreases otherwise. The system prioritizes recommending measures with a confidence score greater than 0.95 to ensure the accuracy of maintenance strategies.

[0087] This step utilizes a semantically driven knowledge self-evolution mechanism to construct a maintenance decision-making mechanism with continuous learning capabilities: 1) Maintenance response efficiency is doubled: Work order generation delay is reduced. For example, the average maintenance time for lubrication failure faults is reduced from 35 minutes to 8 minutes, and repeated disassembly and inspection are avoided because the measures are clear. 2) Making implicit experience explicit: After analyzing three years of maintenance records, the BERT model automatically extracted 23 optimization strategies not specified in the manual, such as "metal mesh floor conditions require monthly grease replenishment", which led to a rapid increase in the number of rules in the knowledge base.

[0088] The knowledge base's confidence mechanism inversely optimizes the fault prediction model. When "grease replenishment" fails to eliminate lubrication failure alarms three times consecutively, the system automatically flags the fault type as potentially misjudged, triggering the model to re-verify the correlation weights between temperature change rate and vibration spectrum. This work order effectiveness feedback loop improves the fault prediction model's accuracy in identifying complex faults, forming a closed-loop intelligence of diagnosis-repair-optimization.

[0089] In some embodiments, the fault prediction system updates the parameters of the fault prediction model based on actual maintenance results through online incremental learning.

[0090] As an example, a penalty weight is applied to false positive samples.

[0091] The actual maintenance results (such as whether the fault was accurately predicted) are fed back to the fault prediction model, and the model parameters are dynamically adjusted through online incremental learning; a higher penalty weight is applied to false positive samples to avoid overly conservative warnings.

[0092] Once maintenance personnel complete the work order and provide feedback on the actual maintenance results, the system can initiate an incremental update process for the fault prediction model daily.

[0093] First, the maintenance results are correlated with the original prediction records: if a bearing fault is confirmed and the type matches, it is marked as a true positive sample; if no fault is found during disassembly and inspection but the model issues a warning, it is marked as a false alarm sample; if a fault exists but the type is misjudged, it is marked as a type error sample. The three types of samples, along with their original sensor data and operating condition vectors, are archived into the incremental learning buffer according to their timestamps.

[0094] The online learning engine employs a mini-batch gradient descent strategy, dynamically adjusting the Transformer model parameters through a differentiated loss weight mechanism. False positive samples are penalized with a 3.0x weight, significantly higher than the 1.0x weight for normal samples, forcing the model to reduce its sensitivity to temporary noise; misclassified samples are assigned a 2.0x weight, enhancing the model's ability to distinguish similar fault features; true positive samples maintain the baseline weight, with only minor adjustments to the feature extraction layer parameters.

[0095] During the learning process, the model freezes the bottom feature encoding layer (preserving general pattern recognition capabilities) and only unfreezes the top fault classification layer and attention weight layer for optimization. Forward computation is performed with 32 incremental samples input each time, and the loss function is modified by adding an L2 regularization term to the weighted cross-entropy to prevent overfitting. The parameter update magnitude is controlled by an adaptive learning rate: when the loss decrease rate is less than 1% for five consecutive batches, the learning rate is automatically reduced to one-tenth of its original value.

[0096] The updated model is deployed only after real-time validation loop detection: the new model is run in parallel on the validation dataset (accounting for 5% of the real-time traffic), and its false positive rate and recall rate are compared with those of the original model. If the false positive rate decreases and the recall rate fluctuates by less than 2%, the online model is hot-swapped, and the entire process is completed within 20 seconds without any downtime.

[0097] This solution addresses the core pain points of industrial predictive maintenance through feedback-driven adaptive weight learning. 1. False alarm rate is sharply reduced: The penalty weighting mechanism significantly reduces the false alarm rate. For example, false alarms caused by steam interference from temperature sensors are reduced by 90%, avoiding thousands of unnecessary downtimes per year. 2. Significant improvement in fault subtype discrimination: Reinforcement learning on type-incorrect samples significantly improves the model's accuracy in distinguishing between "outer ring peeling" and "ball pitting," since the difference in their vibration spectra exists only in the 2-5kHz frequency band.

[0098] It should be noted that the embodiments of this disclosure can be implemented in software and / or a combination of software and hardware, for example, using an application-specific integrated circuit (ASIC), a general-purpose computer, or any other similar hardware device. In one embodiment, the software program of the embodiments of this disclosure can be executed by a processor to implement the steps or functions described above. Similarly, the software program (including associated data structures) of the embodiments of this disclosure can be stored in a computer-readable recording medium, such as RAM memory, a magnetic or optical drive, a floppy disk, or similar devices. Furthermore, some steps or functions of the embodiments of this disclosure can be implemented in hardware, for example, as circuitry that cooperates with a processor to perform the various steps or functions.

[0099] Furthermore, at least a portion of the embodiments of this disclosure can be applied as computer program products, such as computer program instructions, which, when executed by a computing device, can invoke or provide methods and / or technical solutions according to the embodiments of this disclosure through the operation of the computing device. The program instructions that invoke / provide the methods of the embodiments of this disclosure may be stored in a fixed or removable recording medium, and / or transmitted via a data stream in a broadcast or other signal carrying medium, and / or stored in the working memory of a computing device operating according to the program instructions.

[0100] It will be apparent to those skilled in the art that the embodiments of this disclosure are not limited to the details of the exemplary embodiments described above, and that the embodiments of this disclosure can be implemented in other specific forms without departing from the spirit or essential characteristics of the embodiments of this disclosure. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the embodiments of this disclosure is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be encompassed within the embodiments of this disclosure. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is apparent that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in the system claims may also be implemented by a single unit or device in software or hardware. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

Claims

1. A bearing failure prediction method based on AGV clusters, wherein, The method includes the following steps: Real-time acquisition of sensor data located at the bearings of each AGV wheel hub. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status. The data records are input into the fault prediction model according to the sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction results include fault type and fault probability. The fault prediction model dynamically learns the influence of the operating state on the fault feature weights through a self-attention mechanism. When the fault probability is in the medium-risk range, the current AGV is compared with AGVs with similar operating conditions in the AGV cluster to confirm the fault. Based on the risk range of the failure probability, corresponding dynamic task scheduling is performed. Specifically, when the failure probability is in the medium-risk range, the current AGV is triggered to degrade its operation mode; when the failure probability is in the high-risk range, the tasks of the current AGV are dynamically allocated to other AGVs in the AGV cluster. Based on the fault type, query the fault-response mapping knowledge base to generate a maintenance work order for the current AGV.

2. The method according to claim 1, wherein, The effect of dynamically learning the operating state on the fault feature weights through the self-attention mechanism includes at least one of the following: - Attention weighting for increasing the energy percentage of BPFO under high temperature conditions; - Increase the attention weight of high-frequency components of the vibration spectrum under high load conditions.

3. The method according to claim 1, wherein, The step of confirming a fault by comparing the current AGV with AGVs in similar operating conditions in the AGV cluster when the fault probability is in the medium-risk range specifically includes: When the failure probability is in the medium-risk range, cluster collaborative diagnosis is triggered. In the AGV cluster, multiple AGVs with similar operating conditions to the current AGV are retrieved based on locality-sensitive hashing. If at least one of the plurality of AGVs experiences a fault of the same type as the current AGV, then the current AGV is confirmed to have malfunctioned.

4. The method according to claim 3, wherein, The step of retrieving multiple AGVs with similar operating conditions to the current AGV in the AGV cluster based on locality-sensitive hashing specifically includes: The current AGV's running state vector is input into a pre-trained set of locality-sensitive hash functions to obtain bucket codes; Within the same hash bucket, retrieve k nearest neighbor AGVs from AGVs of the same ground type and with load deviation less than a preset threshold as the similar working condition group of the current AGV.

5. The method according to claim 4, wherein, The step of confirming that the current AGV has malfunctioned if at least one of the plurality of AGVs experiences a fault of the same type as the current AGV specifically includes: If at least one AGV in the similar working condition group simultaneously meets the conditions of abnormal vibration and synchronized temperature trend, and the fault type is consistent with the fault type of the current AGV, then the fault confidence level of the current AGV is raised to the confirmed state.

6. The method according to claim 1, wherein, The method also includes the following steps: Periodically scan the AGV cluster in the same batch. If more than a predetermined number of AGVs show similar anomalies, a batch defect warning will be triggered.

7. The method according to claim 1, wherein, The method also includes the following steps: By parsing maintenance work order records using NLP, fault-measure rules are automatically extracted to expand the fault-measure mapping knowledge base.

8. The method according to claim 7, wherein, The method also includes the following steps: The BERT model is used to perform entity recognition on maintenance work orders and extract triplet groups of fault location, fault mode, and maintenance measures. The extracted triples are automatically added to the fault-response mapping knowledge base.

9. The method according to claim 1, wherein, The method also includes the following steps: The parameters of the fault prediction model are updated based on the actual maintenance results, wherein a penalty weight is applied to false alarm samples.

10. A bearing fault prediction system based on AGV clusters, wherein, The system includes edge computing nodes placed in each AGV and a cloud-based fault prediction system. The edge computing node is configured to perform the following operations: Real-time acquisition of sensor data located at the bearings of each AGV wheel hub. Each data record includes a timestamp, AGV identifier, vibration peak value, BPFO energy percentage, temperature change rate, and operating status. The cloud-based fault prediction system is configured to perform the following operations: The data records are input into the fault prediction model according to the sliding time window to perform real-time bearing fault prediction for the current AGV. The prediction results include fault type and fault probability. The fault prediction model dynamically learns the influence of the operating state on the fault feature weights through a self-attention mechanism. When the fault probability is in the medium-risk range, the current AGV is compared with AGVs with similar operating conditions in the AGV cluster to confirm the fault. Based on the risk range of the failure probability, corresponding dynamic task scheduling is performed. Specifically, when the failure probability is in the medium-risk range, the current AGV is triggered to degrade its operation mode; when the failure probability is in the high-risk range, the tasks of the current AGV are dynamically allocated to other AGVs in the AGV cluster. Based on the fault type, query the fault-response mapping knowledge base to generate a maintenance work order for the current AGV.

Images