Device anomaly detection method and system based on multi-source heterogeneous data

Abstract

The present invention discloses a device anomaly detection method and system based on multi-source heterogeneous data. The method includes: obtaining real-time multi-source monitoring data streams of production equipment in a smart park, performing cross-modal data alignment processing on the real-time multi-source monitoring data streams, generating a target monitoring data set, performing dynamic feature extraction operations in each edge computing node, generating a multi-dimensional device status feature set based on the target monitoring data set, inputting the multi-dimensional device status feature set into a trained anomaly detection integration model, generating a device anomaly probability distribution through a multi-level feature fusion strategy, and determining the anomaly type and anomaly confidence of the target device according to the device anomaly probability distribution, generating a device maintenance instruction set based on the anomaly type and anomaly confidence, and sending the device maintenance instruction set to the device management terminal of the smart park to trigger an anomaly handling operation, thereby improving the automation level and decision reliability of the smart park equipment maintenance response.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of industrial Internet of Things, and specifically to a device anomaly detection method and system based on multi-source heterogeneous data. Background Art

[0002] With the continuous development of Industrial Internet of Things (IIoT) technology, anomaly monitoring and analysis of industrial equipment has become increasingly important. However, traditional technologies often rely on single-dimensional data collection and centralized processing architectures, achieving anomaly warnings through preset thresholds or analyzing single sensor data. These technologies lack a mechanism for spatiotemporal alignment of cross-modal data, leading to problems such as data spatiotemporal misalignment and lack of feature relevance when assessing equipment status. Furthermore, traditional feature extraction methods, often based on fixed sampling periods and static feature templates, struggle to adapt to dynamic changes in equipment operating status and the impact of environmental coupling, making them prone to misjudgments when dealing with complex operating conditions.

[0003] In the anomaly detection phase, existing solutions often rely on a single detection model to perform linear analysis on structured data. This fails to effectively capture the nonlinear correlation characteristics between multi-source data, and the accuracy of identifying complex anomalies is limited. Furthermore, traditional systems typically use a cloud-based centralized decision-making model, resulting in high data processing latency and a lack of fine-grained parameter guidance for maintenance instruction generation, leading to delayed anomaly responses and insufficient adaptability of maintenance operations. These shortcomings severely restrict the real-time, accuracy, and decision-making effectiveness of smart campus equipment health management. Improving the automation level and decision-making reliability of smart campus equipment maintenance responses has become a currently difficult problem to overcome. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for detecting equipment anomaly based on multi-source heterogeneous data, which are used to improve the automation level and decision reliability of maintenance response of smart park equipment.

[0005] In the first aspect, an embodiment of the present invention provides an equipment anomaly detection method based on multi-source heterogeneous data, which is applied to an equipment anomaly detection system. The method includes: obtaining real-time multi-source monitoring data streams of production equipment in a smart park, wherein the real-time multi-source monitoring data streams include equipment operating status data, environmental sensor data, and equipment operation log data; performing cross-modal data alignment processing on the real-time multi-source monitoring data streams to generate a target monitoring data set, wherein the cross-modal data alignment processing includes timestamp synchronization, data sampling frequency unification, and data format normalization; performing dynamic feature extraction operations in each edge computing node, and generating multi-dimensional equipment status features based on the target monitoring data set. A multi-dimensional device status feature set is collected, wherein the multi-dimensional device status feature set includes device operation status features, environment-related features, and data abnormal fluctuation features; the multi-dimensional device status feature set is input into the trained anomaly detection integrated model, and a device anomaly probability distribution is generated through a multi-level feature fusion strategy, and the anomaly type and anomaly confidence of the target device are determined according to the device anomaly probability distribution; a device maintenance instruction set is generated based on the anomaly type and the anomaly confidence, and the device maintenance instruction set includes an abnormal device identifier, maintenance priority, and maintenance operation parameters, and the device maintenance instruction set is sent to the device management terminal of the smart park to trigger an exception handling operation.

[0006] In a second aspect, an embodiment of the present invention provides a device anomaly detection system, comprising:

[0007] processor;

[0008] a storage device having a computer program stored thereon,

[0009] When the computer program is executed by the processor, the processor implements any of the device anomaly detection methods based on multi-source heterogeneous data.

[0010] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the device anomaly detection method based on multi-source heterogeneous data are implemented.

[0011] It can be seen that the embodiments of the present invention have the following beneficial effects: through the collaborative optimization of cross-modal data alignment and edge computing dynamic feature extraction, an end-to-end analysis architecture for multi-source heterogeneous monitoring data is constructed, achieving a dual improvement in the real-time and accuracy of equipment anomaly detection. Through a multi-level data integration mechanism with timestamp synchronization, frequency unification and format standardization, the problem of separation between multi-source data streams in temporal correlation and semantic consistency is effectively overcome, so that a dynamic coupling analysis is formed between the equipment operating status, environmental parameters and operation logs; the edge node performs dynamic feature extraction based on the real-time data stream, which not only captures the equipment's own operating characteristics, but also integrates environmental correlation effects and abnormal fluctuation patterns, and constructs a health portrait of the equipment in a multi-dimensional feature space; the anomaly detection integration model uses a multi-level feature fusion strategy to map the nonlinear correlation between heterogeneous features into a probabilistic anomaly distribution, breaking through the limitations of a single detection model's ability to represent complex anomaly patterns. The maintenance instruction set finally generated realizes closed-loop decision-making on anomaly location, priority determination and operating parameters, improving the automation level and decision reliability of the maintenance response of smart park equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flowchart of a device anomaly detection method based on multi-source heterogeneous data provided by an embodiment of the present invention.

[0013] Figure 2 A schematic diagram of the basic structure of a device anomaly detection system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0015] See also Figure 1 As shown in FIG, this figure is a flow chart of a device anomaly detection method based on multi-source heterogeneous data provided by an embodiment of the present invention, which can be applied to a device anomaly detection system. Figure 1 As shown, the method may include steps 101 to 105.

[0016] For ease of understanding, the embodiments of the present invention are described using a six-axis collaborative robot (an exemplary production equipment) in a new energy vehicle battery module assembly line in an automobile company's production line park (an exemplary smart park) as a specific application object.

[0017] Step 101: Acquire real-time multi-source monitoring data streams of production equipment in the smart park, where the real-time multi-source monitoring data streams include equipment operating status data, environmental sensor data, and equipment operation log data.

[0018] In the embodiment of the present invention, the real-time multi-source monitoring data stream refers to a comprehensive data set continuously captured through a heterogeneous sensor network and an industrial Internet of Things platform.

[0019] The equipment operating status data represents the real-time parameters of the mechanical performance of the six-axis collaborative robot, covering core operating parameters such as the dynamic response indicators of the joint transmission system, the mechanical feedback signal of the end effector, the motion trajectory deviation monitoring value, and the temperature rise rate of key components. These data are sampled at high frequency through high-precision embedded sensors. For example, the vibration acceleration sensor of the harmonic reducer collects the micro-oscillation waveform of the mechanical transmission system in ultra-high frequency mode, and the Hall effect sensor of the servo motor continuously records the current phase changes of the rotor winding.

[0020] Environmental sensing data includes physical environmental indicators such as air quality parameters, temperature and humidity gradient distribution, and electrostatic intensity field strength in the assembly workshop. Multi-spectral environmental sensors deployed around the production line form a network to achieve spatial continuity monitoring.

[0021] The equipment operation log data consists of a structured operation sequence generated by a programmable logic controller, which records in detail timing operation information such as robot arm coordinate positioning instructions, workpiece clamping status change events, and safety protection system trigger records. Each log is accompanied by a precise time stamp and equipment identity.

[0022] In practical applications, the data acquisition system uses redundant transmission protocols to ensure the integrity of multi-source data. For example, Industrial Ethernet and Time-Sensitive Networking (TSN) are used to transmit high-priority device status data, while wireless mesh networks are used to transmit environmental sensor data to optimize cabling costs. The real-time data flow is guaranteed by the cache queues of edge computing nodes. When network bandwidth fluctuates, data priority transmission strategies are automatically triggered to ensure low-latency transmission of critical operating parameters.

[0023] Step 102: performing cross-modal data alignment processing on the real-time multi-source monitoring data stream to generate a target monitoring data set, wherein the cross-modal data alignment processing includes timestamp synchronization, data sampling frequency unification, and data format normalization.

[0024] In this embodiment of the present invention, cross-modal data alignment aims to address the issue of inconsistent spatiotemporal references for heterogeneous multi-source data. Timestamp synchronization eliminates time reference offsets between devices through a precise clock synchronization protocol. Specifically, the IEEE 1588 Precision Time Protocol (PTP) is used to calibrate all data sources to the nanosecond level. A hardware timestamp module is deployed at the edge gateway to uniformly map the analog signal sampling moments of vibration sensors, the discrete acquisition cycles of environmental sensors, and the generation times of operation logs to a global timeline. Data sampling frequency unification requires dynamic resampling of data streams with different acquisition frequencies. For high-frequency device status data, anti-aliasing filtering combined with a downsampling algorithm is used to generate an equivalent signal matching the target frequency. For low-frequency environmental parameters, a continuous function model is constructed based on a cubic spline interpolation algorithm and then resampled to the target frequency. Data format normalization involves the unified representation and conversion of multimodal data. For example, the raw analog voltage signals of a robotic arm joint angle sensor are converted to radians in the International System of Standard Units (SI), the multi-dimensional temperature field data of an infrared thermal imager is reconstructed into a normalized temperature matrix, and the natural language instructions in the operation log are converted into structured control instruction encoding using a semantic parsing engine.

[0025] In actual application, the data quality verification module runs synchronously, identifies abnormal data points caused by sensor failure through the outlier detection algorithm, and repairs or removes data based on the historical data distribution model, ultimately generating a target monitoring data set with temporal and spatial consistency.

[0026] Step 103: Execute a dynamic feature extraction operation in each edge computing node to generate a multi-dimensional device status feature set based on the target monitoring data set, wherein the multi-dimensional device status feature set includes device operation status features, environment-related features, and data abnormal fluctuation features.

[0027] In this embodiment of the present invention, dynamic feature extraction leverages the real-time processing capabilities of the edge computing architecture to implement multi-level feature engineering at the data source. Equipment operating status features focus on quantitatively characterizing the microscopic behavior of mechanical systems. For example, a wavelet packet transform algorithm is used to extract frequency band energy entropy features from vibration signals, revealing subtle changes in the meshing state of harmonic reducer gears. A Kalman filter is used to estimate the state of servo motor current signals, generating dynamic impedance features that characterize winding health. The statistical distribution characteristics of the positioning error of the manipulator end-point are calculated based on an inverse kinematic model, reflecting the cumulative backlash effect of the transmission system. Environmental correlation features reveal the coupling relationship between equipment operating parameters and external environmental variables. For example, a time-varying correlation matrix is ​​constructed between the rate of change of workshop temperature and humidity and the heat dissipation efficiency of the motor, and a causal relationship model is analyzed between air particulate matter concentration and the wear rate of precision bearings. Data abnormal fluctuation features utilize an adaptive threshold detection algorithm to identify potential abnormal events. For example, a sliding window statistic is used to analyze the transient mutation pattern of the current signal, combined with a hidden Markov model (HMM) to probabilistically model the duration and intensity of abnormal events.

[0028] It can be understood that the above feature extraction process can introduce an online learning mechanism to automatically adjust the feature extraction strategy according to the dynamic changes in the equipment operation stage. For example, in the equipment startup stage, it focuses on transient response feature extraction, and in the steady-state operation stage, it strengthens the spectrum analysis of periodic features.

[0029] In the specific scenario of a new energy vehicle battery module assembly line, the operating status of a six-axis collaborative robot is directly related to battery pack assembly quality and production line safety. Taking the battery cell stacking process as an example, the dynamic feature extraction in step 103 specifically focuses on physical feature dimensions strongly related to battery module assembly. At the equipment operating status feature layer, the feature extraction module constructs a joint clamping force-displacement feature matrix based on the end-effector mechanical feedback signal during the battery cell gripping process. By real-time analyzing the pulsating waveform of the vacuum chuck pressure sensor and combining it with micron-level positioning data from the contact displacement probe, characteristic parameters such as the steady-state retention rate of the clamping force and the uniformity of the contact surface pressure distribution are extracted. These parameters are strongly correlated with the surface flatness of the lithium battery cell. When the clamping force fluctuation coefficient exceeds a threshold, it may indicate surface warpage or diaphragm wrinkling defects. Furthermore, for the six-axis robotic arm's motion trajectory characteristics at the welding station, the system focuses on analyzing the positioning repeatability of the TCP (Tool Center Point) in three-dimensional space. By calculating the Hausdorff distance between the actual and theoretical trajectories between adjacent weld points, a decay index representing the robotic arm's dynamic accuracy is generated. This indicator has a decisive influence on the welding quality of the battery module busbar. When it is detected that the repeated positioning error of a certain axis joint shows a nonlinear growth trend, it can provide early warning of wear failure of the harmonic reducer.

[0030] In terms of environmentally relevant features, due to the extreme sensitivity of battery module assembly to cleanliness, temperature, and humidity, feature engineering requires a deep integration of environmental sensor data and equipment operating parameters. For example, when the static field strength in the workshop exceeds 50kV / m during the dry season, the system dynamically activates the electrostatic protection correlation analysis model: by analyzing the charge transfer at the moment of contact between the robot arm's end effector and the battery housing, combined with the spatial gradient distribution data from the environmental humidity sensor, a time-varying relationship model between the static accumulation rate and the equipment's grounding impedance is constructed. When an abnormally prolonged static dissipation time constant is detected at a certain workstation, poor contact of the conductive brush or blind spots in the ion blower's coverage can be precisely located. Furthermore, to meet the strict dew point temperature requirements of the lithium battery filling process, the system calculates in real time the deviation between the ambient temperature and humidity sensor data and the set value of the sealed chamber dew point control system. Sliding window regression analysis reveals the quantitative impact coefficient of temperature fluctuations on filling accuracy, providing dynamic compensation parameters for subsequent anomaly detection.

[0031] At the data abnormal fluctuation feature layer, the unique process risk points of battery module assembly determine the targeted feature extraction strategy. For example, in the hot pressing and shaping process, the six-axis robot needs to press the battery cell group into the mold cavity with constant pressure. At this time, the system will establish a multi-physics field coupling monitoring model: by integrating the servo motor current ripple characteristics, the pressure head temperature gradient distribution data and the step response curve of the pressure sensor, a thermal-mechanical coupling state feature vector is constructed. When an abnormal increase in the current harmonic component during the pressure holding stage is detected, accompanied by a sudden increase in the local temperature, the mechanical jamming phenomenon caused by the mold cavity positioning deviation can be quickly identified. This type of composite feature is of great value in preventing the misalignment of the internal pole pieces of the battery module. Compared with the traditional single-dimensional threshold alarm method, the detection timeliness is significantly improved.

[0032] Step 104: Input the multi-dimensional device status feature set into the trained anomaly detection integrated model, generate a device anomaly probability distribution through a multi-level feature fusion strategy, and determine the anomaly type and anomaly confidence of the target device based on the device anomaly probability distribution.

[0033] In this embodiment of the present invention, the integrated anomaly detection model utilizes a heterogeneous model fusion architecture to achieve collaborative analysis of multi-source features. At the bottom layer of the model, a convolutional neural network (CNN) branch processes vibration spectrum image features, extracting local anomaly patterns in the frequency domain using multi-layer convolution kernels. A long short-term memory (LSTM) branch analyzes the dynamic evolution of temporal features to capture the gradual degradation of device performance. A graph neural network (GNN) branch models the topological relationships between device components to identify anomaly propagation paths.

[0034] Furthermore, the multi-level feature fusion strategy dynamically assigns feature weights through an attention mechanism. For example, it increases the weight coefficient of the mechanical feature branch in the mechanical overload warning scenario, and enhances the decision-making contribution of environmental-related features in the environmental interference scenario. The probability distribution of equipment anomalies synchronously outputs the probability values ​​of preset fault categories and the confidence scores of unknown anomaly types through a multi-task learning framework. The library of known fault categories covers typical industrial scenarios such as mechanical transmission failure, electrical system failure, and environmental interference anomalies. The anomaly confidence calculation fusion model predicts the similarity measurement results with historical fault examples, and uses evidence reasoning theory to quantify the credibility of decisions under uncertain conditions. In addition, the model reasoning process introduces an explainability enhancement module, which assists operation and maintenance personnel in understanding the basis for anomaly judgment through feature importance ranking and decision path visualization technology.

[0035] It's worth noting that the aforementioned integrated anomaly detection model requires special optimization for its ability to identify process-specific faults in battery module assembly scenarios. For example, in the battery pack bolt tightening process, a six-axis robot must tighten hundreds of joints according to a preset torque curve. The model employs a multimodal fusion detection strategy: the CNN branch focuses on analyzing the time-frequency characteristics of the servo motor current signal to capture harmonic distortion patterns caused by sudden torque changes. The LSTM network tracks the temporal correlation of angle deviations between the joints during the tightening process to identify progressive precision loss due to tool wear. Furthermore, the graph neural network models the device topology between the robot, torque gun, and vision positioning system. When multi-node coordination errors are detected that exceed tolerances, an assembly process chain anomaly alert is automatically triggered. For battery module sealing inspection, the model incorporates a transfer learning mechanism, transferring the characteristic patterns of historical helium leak rate violations to current vibration signal analysis. This ensures a high fault recall rate even in the early stages of new machine production.

[0036] In terms of specific fault identification, the system has established a fault knowledge graph covering the entire battery assembly process. For example, when abnormal vibration occurs during busbar laser welding, the model will simultaneously associate the following multi-dimensional features: the penetration monitoring data of the welding head optical sensor, the pressure pulsation characteristics of the cooling water circulation system, and the changes in the readings of the ambient oxygen concentration sensor. By calculating the influence weight of each feature node through the graph attention network, three typical faults can be accurately distinguished: if the high-frequency vibration characteristics are strongly correlated with the cooling water pressure fluctuations (weight > 0.7), it is judged as a cooling pipe blockage; if the vibration pattern changes synchronously with the penetration data and the oxygen concentration is abnormal, it is classified as a shielding gas supply failure; when the vibration energy is concentrated in the low and medium frequency bands and is accompanied by abnormal joint temperature gradients, it is identified as a reducer lubrication failure. This fine-grained diagnostic capability significantly shortens the mean fault location time.

[0037] Step 105: Generate a device maintenance instruction set based on the exception type and the exception confidence, the device maintenance instruction set including the abnormal device identification, maintenance priority and maintenance operation parameters, and send the device maintenance instruction set to the device management terminal of the smart park to trigger the exception handling operation.

[0038] In an embodiment of the present invention, the generation of equipment maintenance instruction sets follows the principles of hierarchical response and resource optimization. In detail, the maintenance priority determination module comprehensively considers the abnormality confidence, equipment criticality index and production plan status. For example, the highest priority response is immediately triggered for a high-confidence core equipment failure, while the low-confidence auxiliary equipment abnormality is scheduled for processing during non-production time periods. The abnormal equipment identification adopts a multi-level coding system that integrates the physical location of the equipment, the functional module number and the production line level information to ensure that the equipment positioning accuracy reaches the workstation level granularity. The maintenance operation parameter library is associated with the equipment digital twin model to automatically match the standard maintenance process, special tool requirement list and safe operation constraint conditions corresponding to the failure mode.

[0039] Furthermore, the instruction distribution system uses a publish-subscribe model to achieve multi-terminal collaboration. For example, detailed maintenance procedures can be pushed to augmented reality (AR) inspection terminals via the OPCUA protocol, while simultaneously sending spare parts requests to the material management system. Closed-loop monitoring is implemented throughout the maintenance process, with IoT devices verifying operator compliance in real time as operators perform key steps. For example, a torque wrench sensor can be used to confirm that bolt tightening torque meets process specifications. Maintenance result data is automatically fed back to the anomaly detection model for online parameter fine-tuning, continuously optimizing fault diagnosis accuracy.

[0040] For example, if the positioning accuracy of a battery cell stacking robot is degraded, the system performs multi-dimensional decision-making and reasoning. First, the system analyzes the root cause of the accuracy loss based on historical maintenance records. If the kinematic model parameter error increases linearly, a servo motor encoder calibration instruction is generated. If the joint clearance error exhibits a nonlinear, sudden change, a harmonic reducer wear inspection work order is triggered. If excessive ambient vibration is detected, a priority recommendation is made to level the equipment base. Battery assembly-specific knowledge is pre-installed in the maintenance operation parameter library. For example, when replacing the vacuum suction cup of a battery cell gripper, the system not only provides standard torque parameters but also links to the battery cell model change records for that workstation over the past three months, dynamically adjusting the recommended suction cup material (e.g., silicone suction cups are suitable for prismatic cells, while polyurethane suction cups are more suitable for soft-pack cells).

[0041] The system uses a dynamic weighted evaluation model to prioritize maintenance. For example, when multiple device alarms occur simultaneously, if bus transmission delays cause the clock deviation between the stacking robot and the visual positioning system to reach 10ms (potentially resulting in a positioning error of 0.1mm), the system automatically adjusts the response level based on the accuracy requirements of the current production batch (energy storage batteries allow a 0.3mm error, while power batteries must be controlled within 0.1mm). For the power battery production line, such an anomaly immediately triggers a yellow alert and calibration is scheduled; for the energy storage battery production line, it is marked as an observation-level event. This differentiated management strategy significantly improves maintenance resource utilization at critical workstations while reducing unnecessary downtime.

[0042] At the maintenance instruction execution level, the system deeply integrates the digital twin of the battery production line. When wear is detected on the guide shaft of the module end plate press-fitting equipment, the maintenance instruction not only includes the standard operating procedures for replacing the component, but also simultaneously retrieves the 3D point cloud model of the equipment, overlaying a heat map of the critical disassembly path on the AR terminal, specifically marking hazardous operation areas that may affect adjacent high-voltage wiring harnesses. For complex maintenance tasks requiring collaboration among multiple trades (such as laser welding head optical path calibration), the system automatically breaks down the task steps and simultaneously pushes them to the terminals of technicians with different specialties, including mechanical, optical, and electrical technicians, ensuring strict time and space matching for each link.

[0043] In terms of safety controls unique to battery module assembly, the maintenance command system is deeply integrated with process safety parameters. For example, if environmental sensors detect that VOC concentrations at a certain workstation are approaching 25% of the lower explosion limit, all maintenance operations involving that area will automatically implement three levels of protection: mandatory activation of the explosion-proof ventilation system, disabling of wireless communication devices, and limiting the use of intrinsically safe tools. The system also dynamically adjusts maintenance schedules, prioritizing the replacement of potentially sparking mechanical components and rescheduling risky processes during low-volume periods during shift handovers.

[0044] Designed in this way, the embodiment of the present invention significantly improves the timeliness and accuracy of industrial equipment anomaly detection by constructing a closed-loop system of multi-source data fusion analysis and intelligent decision-making: cross-modal data alignment processing effectively solves the spatiotemporal mismatch problem of multi-source heterogeneous data in industrial sites, ensuring the basic quality of monitoring data; the dynamic feature extraction mechanism fully mines the multi-dimensional characterization information of the equipment operation status, and enhances the ability to identify abnormal patterns; the anomaly detection integration model realizes robust diagnosis under complex working conditions through a multi-level feature fusion strategy, reducing the risk of false alarms and missed alarms; the generation of intelligent maintenance instructions realizes the automation of the entire process from fault diagnosis to disposal execution, greatly shortening equipment downtime, which can adapt to the strict requirements of smart parks on equipment reliability and production continuity, and provide reliable technical support for intelligent manufacturing scenarios.

[0045] In one embodiment, the step 102 of performing cross-modal data alignment on the real-time multi-source monitoring data stream to generate a target monitoring data set includes:

[0046] Step 1021: extracting a first original data stream from the device operating status data, where the first original data stream includes device temperature, vibration amplitude, and current fluctuation parameters.

[0047] In the application scenario of six-axis collaborative robots in the assembly line of new energy vehicle battery modules, the equipment temperature is specifically manifested as the surface temperature of the harmonic reducer housing, the temperature of the servo motor windings, and the temperature of the end effector contact surface. The servo motor winding temperature is acquired by an embedded platinum resistance temperature sensor at a sampling frequency of 50 times per second, and its numerical change directly reflects the motor load status. The vibration amplitude data comes from the three-axis accelerometer installed in each joint of the robot. For example, the axial vibration amplitude of the third joint needs to be monitored in real time during the battery cell grasping process to determine abnormal backlash of the transmission system. Its original signal contains a vibration waveform in the 0-5kHz frequency band. The current fluctuation parameter specifically refers to the effective value of the three-phase current of the servo drive and its ripple coefficient. For example, in the battery module screw locking process, when the torque gun reaches the preset torque threshold, the instantaneous fluctuation amplitude of the B-phase current exceeds the baseline value by 15%, triggering the overload protection mechanism. Taking the battery cell stacking station as an example, the specific composition of the first raw data stream includes: the winding temperature of the vacuum suction cup servo motor used to grasp the battery cells (sampling value 78.3°C), the RMS value of the vibration acceleration of the third-axis reducer (0.23g), and the C-phase current ripple coefficient of the servo driver driving the rotary joint (12.8%).

[0048] Step 1022: Extract a second raw data stream from the environmental sensor data, where the second raw data stream includes environmental temperature, humidity, and dust concentration.

[0049] In the battery module assembly workshop, ambient temperature is measured using a distributed array of infrared temperature sensors at a density of one monitoring point per square meter. The data is accurate to 0.1°C to match the sensitivity requirements of the thermal expansion coefficient of lithium battery cells. Humidity data is collected from capacitive humidity sensors located above the assembly stations. During the liquid filling process, a strict 45% ± 2% RH is maintained within the sealed chamber. The raw data consists of 12 sampling sequences per minute. Dust concentration is monitored in real time by a laser particle counter, with a particular focus on the number concentration of particles >5μm. During the module packaging process, the air purification system is triggered when concentrations exceed ISO Class 7 cleanliness standards. For example, at the busbar welding station, the second raw data stream includes parameters such as the welding area ambient temperature (23.5°C), local humidity (43.7% RH), and the concentration of particles 5-10μm in size (182 particles / m³) within a 0.5m³ volume. These data are collected simultaneously by an environmental sensor array mounted at the base of the welding robot arm.

[0050] Step 1023: extracting a third original data stream from the device operation log data, wherein the third original data stream includes device start and stop records, operation instruction sequences, and alarm history records.

[0051] Equipment start and stop records are accurate to millisecond timestamps, recording power state transitions for six-axis robots, emergency stop commands (event code E-Stop-002) during batch switching, and their release times. Operational command sequences include G-code execution records for the motion control system, such as the linear interpolation command G01X120.5Y-35.2Z78.3F1500 during the cell positioning phase, whose parameters are interpreted as the tool center point coordinates and feed rate. Alarm history records are categorized by severity, including items such as servo overload alarms (code AL501) and vision positioning timeouts (code AL312), along with snapshots of the device status at the time of occurrence. Taking the module hot pressing and shaping process as an example, the third raw data stream includes: the hot press linkage start timestamp (2023-07-15 14:23:17.532), the pressure control PID parameter adjustment instruction sequence (Kp=1.25, Ki=0.03, Kd=0.12), and three hydraulic pressure overlimit warnings (code AL607) that occurred in the last 24 hours.

[0052] Step 1024: Perform a timestamp synchronization operation on the first original data stream, the second original data stream, and the third original data stream to generate a time-aligned initial data set. The timestamp synchronization operation includes: segmenting each original data stream based on a preset time window and filling in the data of the missing time segments.

[0053] During data integration on the battery module assembly line, the preset time window is set to 100ms, corresponding to the minimum control cycle of the six-axis robot. For the servo motor temperature data stream (sampling frequency 50Hz), each time window contains five sampling points. Linear interpolation is used to fill in the second sampling point in the third window (timestamps 14:23:17.532-14:23:17.632), which is lost due to network latency. For the ambient humidity data stream (sampling frequency 0.2Hz), nearest neighbor interpolation is used within each time window. For example, in the window at 14:23:17.700, the humidity value of 43.2% RH at 14:23:17.600 is directly used. Discrete event data, such as alarm records, is matched to the corresponding window using timestamps. When multiple events occur in the same window, a time series queue is established. Taking the battery cell barcode scanning station as an example, the initial data set after timestamp synchronization in the 14:23:17.800 window includes: servo motor B phase current (12.3A), barcode scanning gun success signal (event ID: Scan-0987), and ambient dust concentration at that moment (155 particles / m³).

[0054] Step 1025: Perform sampling frequency uniform processing on the time-aligned initial data set to generate a target data set, wherein the sampling frequency uniform processing includes: downsampling the high-frequency data stream to a preset reference frequency, and interpolating and completing the low-frequency data stream to the preset reference frequency.

[0055] The preset reference frequency is set at 10 Hz to meet the response requirements of most control systems. The raw data stream of the joint vibration signal (sampling frequency 5 kHz) is downsampled 100 times using an anti-aliasing FIR filter, preserving the 0-50 Hz effective frequency band. For example, the third-axis vibration acceleration data is processed to output an effective value of 0.25g during each reference period (100 ms). The ambient temperature data stream (original sampling frequency 0.1 Hz) is upsampled to 10 Hz using cubic spline interpolation, generating a continuous temperature curve within the period 14:23:17.800-14:23:17.900. Discrete events in the operation command sequence are converted into 10 Hz pulse signals. For example, a high-level flag lasting 100 ms is generated within the time window of receiving a G01 command. Taking the welding process as an example, the target data set after unified sampling frequency includes: welding gun current (350 A ± 5 A) at 100 ms intervals, ambient temperature (24.1 → 24.2°C) obtained by synchronous interpolation, and safety door status signals resampled at 10 Hz.

[0056] Step 1026: Perform format normalization processing on the target data set to generate the target monitoring data set. The format normalization processing includes data unit conversion, data range normalization and outlier removal.

[0057] As you can understand, servo motor temperature data is converted from its original Fahrenheit value to standard Celsius, for example, converting a raw value of 172.3°F to 78.0°C. Vibration acceleration data is uniformly converted to g. The original voltage signal (2.34V) is calibrated using the sensitivity calibration formula a = V / 0.5 (V / g) to obtain 4.68g. Data range normalization maps each parameter to the [0, 1] interval. For example, the current parameter is linearly normalized based on the servo drive rated value (30A), converting the actual value of 18A to 0.6. Outlier rejection uses a dynamic thresholding method. When the instantaneous value of ambient dust concentration exceeds the 3σ range of the moving average, it is considered invalid data. For example, an outlier value of >500 particles / m³ for three consecutive cycles will be replaced by the median value. Taking the injection process as an example, the standardized target monitoring data set includes: vacuum degree (0.85atm→0.85), injection rate (12.3ml / s→0.615), and ambient humidity after eliminating abnormalities (45.2%RH→0.452).

[0058] In one embodiment, the step 103 of performing a dynamic feature extraction operation in each edge computing node to generate a multi-dimensional device state feature set based on the target monitoring data set includes:

[0059] Step 1031: extracting a subset of equipment operation status data from the target monitoring data set, performing time domain feature extraction on the subset of equipment operation status data, and generating equipment operation status features, wherein the time domain feature extraction includes mean calculation, variance calculation, and peak detection.

[0060] During the battery cell handling process, the 10-second sliding mean (14.2A) of the vacuum chuck servo motor current is calculated to characterize the steady-state load level, and the variance (0.45A²) reflects the torque fluctuations when grasping battery cells with different surface conditions. Peak detection identifies extreme vibration amplitudes during the robot arm's acceleration phase. For example, the vibration acceleration of the fourth axis at a 90° rotation angle instantaneously reaches 1.2g. For the module press-fit process, the peak-to-valley difference in pressure sensor data (82.3kN → 75.1kN) is extracted as an indicator of backlash accumulation. Taking the welding station as an example, the equipment operating status characteristics include the 30-second mean of laser power (980W), the variance of cooling water flow (0.12L / min²), and the current peak (1050A) at the moment the weld is completed.

[0061] Step 1032: extracting a subset of environmental sensor data from the target monitoring data set, performing environmental correlation analysis on the subset of environmental sensor data, and generating environmental correlation features. The environmental correlation analysis includes calculating the difference between the device temperature and the ambient temperature, and extracting the influence coefficient of humidity on device vibration.

[0062] For example, a time-varying difference curve between the servo motor winding temperature and the ambient temperature is calculated. When the difference exceeds 15°C, a heat dissipation anomaly warning is triggered. For example, during continuous press-fitting operations, the difference reaches 17.3°C. A correlation model between humidity and joint vibration RMS values ​​is established through multivariate regression analysis, demonstrating that every 10% increase in humidity leads to a 0.05g increase in fifth-axis vibration. During the injection process, the sensitivity coefficient of ambient temperature fluctuations to vacuum pump energy consumption is analyzed to be 0.38kW / °C. For the cleaning station, for example, environmental correlation features include the difference between the cleaning agent temperature and the ambient temperature (8.7°C), the transfer coefficient of air flow rate to nozzle pressure (0.12 bar / (m / s)), and the humidity attenuation factor on static elimination efficiency (-0.35% / RH).

[0063] Step 1033: extracting a subset of equipment operation log data from the target monitoring data set, performing abnormal fluctuation detection on the subset of equipment operation log data, and generating data abnormal fluctuation features, wherein the abnormal fluctuation detection includes detection of sudden changes in operation instruction frequency and association matching of alarm records with operating parameters.

[0064] For example, the moving standard deviation of the G-code instruction transmission frequency is monitored. Any standard deviation exceeding 3 times / second within 1 minute is flagged as an abnormal scheduling event. Correlation analysis is performed on the frequency of coordinate correction instructions preceding the visual positioning failure alarm (code AL312). An association rule is established that increases the alarm probability to 82% when correction instructions occur >5 times / second. During the module testing process, it was detected that the interval between three consecutive battery cell internal resistance test instructions suddenly changed from the standard 5 seconds to 3.2 seconds, triggering abnormal production cycle fluctuations. Taking the assembly line changeover phase as an example, abnormal data fluctuations include the coefficient of variation of fixture replacement instruction frequency (0.37), the correlation coefficient between vacuum failure alarms and suction cup wear (0.79), and the sudden change in temperature gradient before and after the program restart event (ΔT = 4.2°C).

[0065] Step 1034: Merge the device operating status features, the environment-related features, and the data abnormal fluctuation features in chronological order to generate the multi-dimensional device status feature set.

[0066] At the final inspection station for battery modules, the feature merging process aligns the average soldering temperature (956W), the environmental electrostatic impact factor (0.67), and the code scanning failure fluctuation index (0.83) at millisecond timestamps, forming a time series consisting of 28-dimensional feature vectors. Each feature vector corresponds to a complete description of the state of a specific process step. For example, the feature set at the moment of cell stacking completion includes: gripping force variance (0.12kN²), ambient temperature difference (6.4°C), and the number of positioning command retries (2). The feature set establishes cross-dimensional correlations through timeline matching. For example, if the phase difference between the servo current peak and the ambient humidity fluctuation reaches 200ms, it is marked as a potential condensation interference event. The resulting multi-dimensional device status feature set provides standardized input for subsequent anomaly detection models, ensuring strict consistency across all feature dimensions in both time and space domains.

[0067] In an optional embodiment, the trained anomaly detection integrated model is generated by training through the following steps:

[0068] Step 201: Acquire a historical multi-source monitoring data set, where the historical multi-source monitoring data set includes normal device data samples and abnormal device data samples marked with abnormal types.

[0069] In the application scenario of a six-axis collaborative robot on a new energy vehicle battery module assembly line, the historical multi-source monitoring dataset covers 12 consecutive months of production cycle data. The normal equipment data samples include over 5,000 hours of steady-state servo motor operation records, demonstrating winding temperature fluctuations of ≤±3°C and stable RMS joint vibration acceleration values ​​within the range of 0.15-0.25g. The abnormal equipment data samples are categorized and annotated by an expert system using seven known fault types, including harmonic reducer wear (type code F01) and encoder signal drift (F02). Each abnormal sample includes a slice of multimodal data from 30 minutes before the fault to the moment of shutdown. For example, the F01 sample includes temperature gradient data showing a sudden increase in the third-axis reducer temperature from 65°C to 89°C, frequency domain characteristics showing a 2.5kHz sideband energy increase to 3.2 times the baseline value in the vibration spectrum, and an axial clearance excess alarm log (code AL704) at the corresponding time.

[0070] Step 202: performing cross-modal data alignment processing on the historical multi-source monitoring dataset to generate a historical aligned dataset, and extracting a historical multi-dimensional device status feature set based on the historical aligned dataset.

[0071] Cross-modal data alignment utilizes the same timing synchronization protocol as the online monitoring system, aligning the servo motor current waveform data (sampling frequency 50kHz) with the trigger signal of the visual positioning system (millisecond-level events) onto a unified timeline. For example, in the case of abnormal cell positioning, the time deviation between the moment of the current surge (14:23:17.532) and the coordinate correction command sent by the visual system was calibrated to within ±5ms. The strategy for extracting historical multi-dimensional device status feature sets aligns with online feature engineering. Features extracted for F02 abnormal samples include the encoder's raw pulse count mean (1423 counts / cycle), the influence coefficient of ambient temperature on counting error (0.08 counts / °C), and the frequency of 10 consecutive zero-point resets (3 times / minute).

[0072] Step 203: Divide the historical multi-dimensional device status feature set into a training set and a validation set, and construct an initial anomaly detection model set, wherein the initial anomaly detection model set includes a first detection model based on an isolation forest, a second detection model based on a temporal convolutional network, and a feature fusion model based on an attention mechanism.

[0073] For example, the training set consists of 80% historical data samples, stratified by device serial number to ensure a balanced distribution of fault types. For example, F01 samples account for 12.7% of the training set, and F02 samples account for 9.3%. The validation set retains the remaining 20% ​​of the data and includes six fault subtypes that were not included in training to test the model's generalization capabilities. The first detection model, based on isolation forests, has a tree depth parameter of 64 and uses a random subspace method to select 30% of the feature dimensions for node splitting calculations. The second detection model, based on a temporal convolutional network, has an architecture consisting of five dilated convolutional layers with dilation coefficients in the sequence [1, 2, 4, 8, 16], and the number of output channels per layer is set to 128. The feature fusion model uses a multi-head attention mechanism with four independent attention heads to capture feature correlations between different fault modes.

[0074] Step 204: Train the initial anomaly detection model set through a cascade training strategy to obtain a target anomaly detection model set, wherein the cascade training strategy includes: using the training set to perform unsupervised training on the first detection model, using the output probability of the first detection model as the input feature of the second detection model, and combining the second detection model with the feature fusion model for supervised training.

[0075] During the unsupervised training of the first detection model, the isolation forest algorithm automatically identified outlier samples representing 5.7% of the training set, 83% of which were manually verified as unlabeled potential fault events. The second detection model's input feature dimensionality was expanded to the original number of features plus anomaly probabilities. For example, in the task of detecting abnormal cell clamping force, the input feature dimension increased from 28 to 29 (with the addition of the F01-class anomaly probability output by the isolation forest). During the joint training phase, an alternating optimization strategy was employed, with the output of the temporal convolutional network and the attention weights of the feature fusion model being synchronously updated via backpropagation. In the detection of F03-class faults (vacuum leaks in the end effector), the model's attention weight for the suction cup pressure fluctuation feature was increased to 0.63.

[0076] Step 205: Performance evaluation of the target anomaly detection model set is performed based on the validation set. When the anomaly type recognition accuracy exceeds a preset threshold, the target anomaly detection model set is used as a trained anomaly detection integrated model and deployed to an edge computing node.

[0077] Performance evaluation metrics include multi-classification accuracy (≥92%), unknown fault detection rate (≥85%), and false alarm rate (≤1.2 times / hour). During verification testing of the module packaging process, the model achieved 94.7% accuracy for F05 faults (hot press mold positioning deviation) and 81.3% for a newly added fault type (laser welding head lens contamination). When deployed on edge computing nodes, model inference latency is strictly controlled to less than 50ms, ensuring that anomaly detection is completed within the robot arm motion control cycle (100ms). For example, at the Busbar welding station, the model's total processing time from feature input to maintenance command output is 43ms, meeting real-time requirements.

[0078] In an optional embodiment, the step 104 of generating a device anomaly probability distribution by using a multi-level feature fusion strategy and determining the anomaly type and anomaly confidence of the target device according to the device anomaly probability distribution includes:

[0079] Step 1041: Input the multi-dimensional device status feature set into the first detection model to generate a first abnormality probability distribution, where the first abnormality probability distribution includes the overall abnormality probability of the device and the abnormality probability of local components.

[0080] During the cell stacking process, the first detection model, using the isolation forest algorithm, calculated an overall equipment abnormality probability of 0.87 (threshold 0.75), with a local abnormality probability of 0.93 for the third joint assembly. Local component abnormality probabilities were determined through feature contribution decomposition. For example, the 1.8kHz component of the third-axis vibration spectrum contributed 58% to the abnormality probability, while the temperature gradient feature contributed 32%. In the servo overload warning scenario, the model output a local abnormality probability of 0.91 for the B-phase winding and an overall abnormality probability of 0.79, triggering a secondary warning response.

[0081] Step 1042: Input the multi-dimensional device status feature set and the first abnormality probability distribution into the second detection model to generate a second abnormality probability distribution, wherein the second abnormality probability distribution includes a time series abnormality pattern and an abnormality duration prediction.

[0082] The time series convolutional network model captured the abnormal periodic fluctuation pattern of the servo current, detecting three current peaks exceeding the threshold within 10 consecutive sampling periods (1 second), and predicted that the abnormality would last between 8 and 12 seconds. Regarding the visual positioning drift issue, the model identified an exponential growth pattern in the frequency of coordinate correction instructions (R² = 0.91), predicting that the system would reach its fault tolerance limit in 23 seconds. During the hot pressing process, the second abnormality probability distribution showed that the autocorrelation coefficient of the mold temperature time series data dropped to 0.12 (the normal baseline is > 0.6), predicting that the abnormal state would persist until the end of the current production batch.

[0083] Step 1043: Input the first abnormality probability distribution and the second abnormality probability distribution into the feature fusion model, and generate a fused abnormality probability distribution through the attention weight allocation mechanism.

[0084] For example, the multi-head attention mechanism assigns a weight of 0.55 to the component anomaly probability output by the first detection model and a weight of 0.45 to the timing anomaly pattern. In the vacuum cup leak example, the feature fusion model detected a strong correlation between the local anomaly probability (0.88) and the duration prediction (>30 seconds), increasing the fusion probability to 0.91. To address encoder signal drift, the model discovered that phase mutation features in the timing pattern are more diagnostically valuable than the outlier score of the isolation forest, and therefore increased the attention weight of the timing branch to 0.68.

[0085] Step 1044: performing threshold segmentation processing on the fused abnormality probability distribution to determine the abnormality type and abnormality confidence of the target device. The threshold segmentation processing includes dynamic threshold adjustment and probability density clustering.

[0086] The dynamic threshold adaptively adjusts based on the device's operating phase. During startup, the alarm threshold is raised from 0.75 to 0.85 to suppress transient interference. The probability density clustering algorithm identified two abnormal probability clusters: the 0.72-0.78 range corresponds to F02 faults (68% confidence), and the 0.88-0.92 range corresponds to F01 faults (93% confidence). In the battery cell code scanning anomaly, the fusion probability of 0.79 was classified as an F07 fault (visual system contamination), with a calculated confidence of (0.79-0.65) / (0.95-0.65) = 46.7%.

[0087] In an optional embodiment, the generating of the device maintenance instruction set based on the anomaly type and the anomaly confidence level in step 105 includes:

[0088] Step 1051: Match the preset maintenance strategy library according to the abnormality type to obtain the maintenance operation template and priority rule corresponding to the abnormality type.

[0089] The maintenance strategy library establishes a mapping between fault codes and standard operating procedures. For example, an F01 fault is mapped to the "Harmonic Reducer Preventive Replacement" template, which includes disassembly steps (12 items), cleaning specifications (5 items), and a torque setting (35Nm ± 5%). The priority rule library defines the response level for F-class faults, with F01 and F04 faults assigned the highest priority (response time < 15 minutes) and F07 assigned a medium priority (response time < 2 hours). In the case of a Busbar welding anomaly, the system invokes the "Laser Welding Head Optical Path Calibration" template, which includes 23 parameters, such as the collimator adjustment tolerance (±0.02mm) and the protective lens replacement cycle (4000 times).

[0090] Step 1052: Dynamically adjust the parameters in the maintenance operation template based on the abnormality confidence to generate maintenance operation parameters. The dynamic adjustment includes shortening the maintenance interval and increasing the maintenance resource allocation as the confidence level is higher.

[0091] When the confidence level for an F01 fault exceeds 90%, the harmonic reducer lubrication maintenance interval is shortened from the standard 2,000 hours to 1,500 hours, and the system automatically dispenses double the amount of grease (from 5ml to 10ml). For an F07 fault with a confidence level of 75%, the frequency of visual lens cleaning is increased from once per shift to twice per shift, and the backup lighting source is activated. In the servo overheat warning example, a confidence level of 85% triggers forced cooling measures: increasing the cooling fan speed to 120% of the rated value and extending the standby cooling time from 5 minutes to 8 minutes.

[0092] Step 1053: Sort the multiple abnormal devices according to the priority rule to generate a maintenance priority queue, and bind the maintenance priority queue with the maintenance operation parameters to generate the device maintenance instruction set.

[0093] For example, the ranking algorithm comprehensively considers the fault level (F01>F04>F07), the confidence weight (70%), and the production impact coefficient (30%). For example, when F01 (92% confidence) and F04 (88% confidence) occur at the same time, the calculated comprehensive priority score is:

[0094] 92 × 0.7 + 100 × 0.3 = 94.4 vs 88 × 0.7 + 95 × 0.3 = 89.9, prioritizing F01. The generated equipment maintenance instruction set includes the work order number (MT20230715-001), target equipment code (RB06-J3), required spare parts list (1 set of harmonic reducers GHD-203), and safety operating constraints (5 minutes of idle time after power failure). When multiple low-priority faults are detected during the changeover phase, the system automatically schedules maintenance work during production breaks to minimize downtime losses.

[0095] In a preferred embodiment, the method further comprises:

[0096] Step 301: Monitor the execution status of the device maintenance instruction set by the device management terminal in real time. When a maintenance operation delay or failure is detected, trigger the following operations: obtain the abnormal device identifier corresponding to the delay or failure event, and extract the historical multi-dimensional device status feature set of the abnormal device based on the abnormal device identifier; incrementally train the anomaly detection integrated model based on the historical multi-dimensional device status feature set to generate an updated anomaly detection integrated model; push the updated anomaly detection integrated model to the edge computing node for model replacement, and regenerate the device maintenance instruction set.

[0097] In a six-axis collaborative robot maintenance scenario on a new energy vehicle battery module assembly line, when a work order (work order number MT20230715-003) for replacing the harmonic reducer on the third-axis servo motor (work order number MT20230715-003) was detected to have timed out and not completed, the system extracted a set of historical multi-dimensional device status features from the previous 72 hours. These features included temperature gradient characteristics (0.78°C / min), vibration energy entropy (1.23), and the frequency of operation command retries (4 times / hour). The incremental training process employed an online learning algorithm. While retaining 90% of the classification layer weights from the original model, five new feature dimensions were added for F01 faults (harmonic reducer wear). After 200 iterations, the model's detection accuracy for this type of fault increased from 83% to 91%. The updated model was pushed to the assembly line's edge gateway via a secure transmission protocol, replacing the original model file (version number v2.1.7 → v2.1.8) and regenerating maintenance instructions including a lubricant fill adjustment (+15%).

[0098] In a preferred embodiment, the incremental training of the anomaly detection integrated model based on the historical multi-dimensional device status feature set to generate an updated anomaly detection integrated model includes:

[0099] Step 401: Obtain the abnormal device identifier corresponding to the delay or failure event, and extract the device operation status feature segment, environment association feature segment and data abnormal fluctuation feature segment of the abnormal device within a preset time window from the historical multi-dimensional device status feature set based on the abnormal device identifier.

[0100] As you can understand, the preset time window is set from 30 minutes before the fault occurs to the current moment. For the vacuum suction cup leak event (device code VP-06), the extracted feature segments include: the average suction cup pressure (-82kPa → -65kPa), the environmental humidity influence coefficient (0.34 → 0.51), and the vacuum breach alarm frequency (0 → 3 times / minute). The device operating status feature segments are accurate to 100ms intervals and contain 12 temperature-pressure correlation feature vectors. The environmental correlation feature segments are stored by workstation partition, recording the temperature and humidity gradient of the filling station at the time of the leak (ΔT = 2.3°C / m, ΔRH = 4.1% / m).

[0101] Step 402: Perform event labeling processing on the device operation status feature segments, environment-related feature segments, and data abnormal fluctuation feature segments to generate an incremental training data set. The event labeling processing includes marking the feature segments of the continuous time period before the delay or failure event occurs as the abnormality category to be optimized.

[0102] In the example of a visual positioning system calibration failure, the characteristic data from the five minutes prior to the event was labeled as F07 (visual contamination failure). Specifically, the data included the positioning error standard deviation (0.12mm → 0.35mm), ambient light intensity (850 lux → 210 lux), and the lens cleaning instruction interval (1200 seconds → not executed). The labeling process used a sliding window method with a 10-second step size to generate 30 labeled samples, of which 67% were positive samples (precursors of failure) and 33% were negative samples (normal operating conditions).

[0103] Step 403: Load the network parameters of the feature extraction layer from the anomaly detection integrated model and lock them in a non-updated state. At the same time, load the fully connected weight matrix of the classification layer and set it to a pending update state.

[0104] The feature extraction layer retains the 5-layer dilated convolutional structure (dilation coefficients [1, 2, 4, 8, 16]) of the temporal convolutional network and the number of attention heads (4 heads). The weight matrix is ​​frozen in read-only mode. The fully connected network parameters of the classification layer (dimension 28 → 7) are initialized to 85% of their original values, the learning rate is set to 0.001, and the momentum term is adjusted to 0.92. For the vacuum leak detection task, the 128-dimensional embedding vector output by the feature extraction layer remains fixed. The classification layer weights for the suction cup pressure feature (dimension 19) are set to three times the initial weights of other features.

[0105] Step 404: Input the incremental training data set into the classification layer to perform a back-propagation optimization operation, and adjust the numerical distribution of the fully connected weight matrix through the gradient descent algorithm until the decrease of the classification loss function on the validation subset is less than a preset convergence threshold.

[0106] For example, the optimization process used mini-batch gradient descent (batch_size=32), and the cross-entropy loss function decreased from 1.23 to 0.87 within 50 epochs, improving validation set accuracy by 12%. For F05 faults (mold positioning deviation), the weight coefficient of the classification layer for the temperature-displacement coupling feature was adjusted from 0.15 to 0.28. The convergence threshold was set to a loss decrease of ≤0.5% over 10 consecutive epochs. Training was terminated when the loss fluctuation range from epochs 47 to 50 was less than 0.03.

[0107] Step 405: Recombining the optimized fully connected weight matrix with the locked feature extraction layer parameters to generate an updated anomaly detection integrated model, and performing performance verification on the updated anomaly detection integrated model on the historical verification set. If the difference in anomaly type recognition accuracy compared to the value before the update is within the preset tolerance range, the model update is confirmed to be complete.

[0108] For example, the combined model was tested on a historical validation set of 1,200 samples. The recall rate for F01 faults increased from 78% to 85%, and the detection rate for unknown faults remained above 82%. With a tolerance of ±3%, a manual review process was triggered when the accuracy for F03 (vacuum leaks) changed by +4.2%. After confirmation, the model version upgrade (v2.1.8 to v2.1.9) was completed.

[0109] In an alternative embodiment, the method further comprises:

[0110] Step 501: Display the equipment abnormality probability distribution and equipment maintenance instruction set in the visual monitoring interface of the smart park.

[0111] As you can see, the visualization interface uses layered rendering technology to overlay the real-time status of the six-axis robots on the workshop's digital twin model. Devices with an abnormality probability exceeding 75% are highlighted with a pulsating red border, while devices undergoing maintenance display a yellow progress bar. For example, the F01 abnormality probability (88%) of the welding robot RB03 is rendered with a 30cm-diameter red halo. Its maintenance instruction (replace the harmonic reducer) appears as a floating card displaying the work order number, remaining time (23 minutes), and spare parts inventory status (two GHD-203 sets remaining).

[0112] Step 502: In response to the user's interactive operation on the selected device, dynamically display the multi-dimensional device status feature set, abnormality type determination basis and maintenance operation execution progress of the selected device.

[0113] When clicking on the RB06 battery stacking robot, the right panel displays the following: the temperature-vibration correlation matrix (28×28 heat map), the top three feature contributions (encoder error 42%, ambient static electricity 35%, and command frequency 23%) for the current anomaly type (F02, 89% confidence), and the lubrication maintenance progress (63% complete). The 3D model automatically focuses on the third joint assembly, displaying a semi-transparent CT scan comparison of the wear areas inside the reducer.

[0114] Step 503: When it is detected that the user manually modifies the anomaly type or maintenance priority, the modification record is recorded and the incremental training of the anomaly detection integrated model is triggered.

[0115] For example, after the operation and maintenance personnel corrected the abnormality type of RB12 from F04 (electrical failure) determined by the system to F01 (mechanical wear), they stored the correction record (operator ID, timestamp, correction basis) in the audit database and triggered fine-tuning training of the feature extraction layer: adding the confusion matrix correction item F01→F04 to the original classification matrix. After 15 minutes of online training, the model's correction rate for similar misjudgments increased to 93%.

[0116] In an alternative embodiment, the dynamic display of the multi-dimensional device status feature set of the device in step 502 includes:

[0117] Step 5021: Extract the temperature change rate numerical sequence, vibration spectrum energy numerical sequence and current fluctuation parameter sequence corresponding to the equipment operation status characteristics from the multi-dimensional equipment status feature set, generate a temperature-vibration-current superposition curve chart based on a unified time axis and render it to the visual monitoring interface.

[0118] Taking the PP05 module press equipment as an example, the horizontal axis of the graph shows the millisecond-accurate time scale (14:23:17.532-14:23:19.632), while the vertical axis shows the mold surface temperature change rate (-0.8°C / s → +1.2°C / s), the hydraulic cylinder vibration energy (152dB → 168dB), and the servo current fluctuation coefficient (8% → 23%). The time of the abnormal event (14:23:18.921) is marked with a vertical red line, and the corresponding current peak (142% of the rated value) is highlighted in the graph as a pulse waveform.

[0119] Step 5022: Extract the ambient temperature distribution matrix, humidity influence coefficient matrix, and dust concentration correlation matrix corresponding to the environmental correlation features from the multi-dimensional device status feature set, generate a three-dimensional thermal map based on the device physical position coordinate mapping, and superimpose and mark abnormal correlation areas that exceed the preset threshold in the three-dimensional thermal map.

[0120] In the battery filling workshop, the temperature distribution matrix is ​​rendered with a grid accuracy of 0.5m×0.5m. The highest temperature area (24.8°C) is located in the southeast corner of the sealed cabin; the area with a humidity influence coefficient exceeding 0.6 (the northwest corner workstation) is surrounded by orange contour lines; the grid with a dust concentration correlation value greater than 0.75 (at the conveyor belt interface) is displayed as a red cube, indicating that enhanced sealing treatment is needed.

[0121] Step 5023: Extract the operation instruction frequency mutation time point, alarm record timestamp and operation parameter abnormal fluctuation period corresponding to the data abnormal fluctuation feature from the multi-dimensional device status feature set, generate an interactive event timeline based on the time alignment strategy, and associate and display the device operation log text fragment at the corresponding moment on the event timeline.

[0122] For example, the event timeline for the Busbar welding robot WT07 shows the following: 14:23:17.532 sent the G01X120Y-35 command; 14:23:17.892 detected the AL312 vision timeout alarm; and 14:23:18.215 the current fluctuation entered the abnormal range. When the AL312 alarm marker is clicked, the associated log displays "Vision positioning reference point lost, cumulative retries 3 times, recommended to clean the reflector."

[0123] Step 5024: The temperature-vibration-current superposition curve graph, three-dimensional thermal map and interactive event timeline are embedded in the dynamic display panel of the visual monitoring interface according to a preset layout. When it is detected that the user clicks on the abnormal correlation area or the marked point in the event timeline, the corresponding device operation log text segment is highlighted and the curve of the abnormal fluctuation period is enlarged and displayed.

[0124] The interface layout uses a three-column design: the left column (30% width) displays the device tree; the middle column (50%) renders the 3D workshop model and heat map; and the right column (20%) displays a graph and timeline. When a user clicks a high-temperature area in the hydraulic station (the red area in the 3D heat map), the associated oil temperature curve segment (14:23:17.200-14:23:18.500) automatically zooms in to 200%, and the log panel highlights "Hydraulic Oil Cooler Fan Fault (Code AL455)."

[0125] In an optional but non-limiting embodiment, the method further comprises:

[0126] Receiving the maintenance operation execution result returned by the device management terminal in real time, and extracting the actual maintenance parameters and maintenance completion timestamp corresponding to the abnormal device identifier from the maintenance operation execution result;

[0127] Based on the difference comparison between the actual maintenance parameters and the maintenance operation parameters, a maintenance effect evaluation index is generated, wherein the maintenance effect evaluation index includes a parameter deviation rate, an operation delay time, and an abnormal recurrence detection result;

[0128] When the maintenance effect evaluation index exceeds a preset tolerance threshold, an optimization instruction of the anomaly detection integrated model is triggered to generate an incremental training sample set based on the actual maintenance parameters and the anomaly recurrence detection result;

[0129] The incremental training sample set is used to perform online fine-tuning on the feature fusion model in the anomaly detection integrated model, the weight coefficient in the attention weight allocation mechanism is updated, and the fine-tuned feature fusion model is redeployed to the corresponding edge computing node.

[0130] During the equipment maintenance execution phase, the system receives maintenance operation result messages from the equipment management terminal in real time via the IoT protocol. These messages contain key information such as the abnormal equipment identifier, actual maintenance parameters, and the maintenance completion timestamp. The maintenance effectiveness evaluation module compares the actual maintenance parameters against the original instruction parameters item by item, calculating the parameter deviation rate and operation delay duration using a preset algorithm. For example, if the lubricant filling amount deviates from the preset range or the maintenance completion time exceeds the planned window, the system automatically triggers an optimization instruction. During the difference comparison process, if a recurrence of an anomaly is detected (such as equipment vibration data failing to return to a normal range), the system extracts multi-dimensional feature data from the historical database within a specific time window before and after maintenance, including temperature gradients, current fluctuation parameters, and alarm frequency from the operation log. This data is cleaned and annotated to form an incremental training sample set, which is used to optimize the feature fusion mechanism of the anomaly detection model. During the model fine-tuning phase, the system locks the network parameters of the feature extraction layer and dynamically adjusts only the weight matrix of the classification layer. Through iterative optimization using the backpropagation algorithm, the model gradually improves its sensitivity to key features (such as the vibration spectrum energy distribution). The updated model is compressed and packaged, then deployed via a rolling update mechanism on edge nodes, ensuring service interruptions are kept to the millisecond level. Simultaneously, the system regenerates maintenance instructions based on the latest model, optimizing parameter ranges and adding exception rechecking rules, forming a closed-loop optimization process from execution feedback to model iteration. This entire process dynamically adapts maintenance strategies to device status while ensuring real-time performance.

[0131] In an optional but non-limiting embodiment, the method further comprises:

[0132] Obtaining current maintenance resource status data of the smart park, the current maintenance resource status data including the number of available maintenance personnel, spare parts inventory distribution, and equipment downtime windows;

[0133] sorting the equipment maintenance instruction set from high to low according to the maintenance priority, and dynamically adjusting the maintenance operation parameters in resource allocation in combination with the current maintenance resource status data to generate adjusted maintenance operation parameters;

[0134] generating a maintenance instruction subset under resource constraints based on the adjusted maintenance operation parameters, wherein the maintenance instruction subset under resource constraints includes a spare parts call path optimization result, a maintenance personnel scheduling route, and a parallel maintenance equipment identifier;

[0135] The maintenance instruction subset under the resource constraint is sent to the equipment management terminal, and the spare parts inventory distribution and maintenance personnel location information in the current maintenance resource status data are updated in real time.

[0136] For example, the system uses a data interface to obtain real-time information about the smart campus' maintenance resource status, including the location of on-duty personnel, spare parts inventory distribution, and production line downtime plans. The maintenance scheduling engine combines equipment priority queues with resource constraints and employs intelligent algorithms to generate the optimal scheduling plan. For high-priority faulty equipment, the system prioritizes the nearest available specialized maintenance team and plans the shortest possible route for spare parts transportation. For example, if replacing a harmonic reducer for a critical piece of equipment requires collaboration between two technicians, the system uses personnel location data to identify the qualified and closest maintenance team and simultaneously calculates the optimal transportation route from the central warehouse to the target workstation. During resource allocation, the system monitors spare parts inventory changes in real time and automatically triggers replenishment alerts when inventory levels fall below a safe threshold. For maintenance tasks that require parallel processing, the system models inter-equipment dependencies to avoid resource conflicts. For example, if two workstations share the same spare part model, the system dynamically adjusts the order of dispatch to ensure that the higher-priority task is completed first. The generated maintenance instruction subset includes a detailed resource call list, personnel division plan, and timeline requirements. A 3D map visualizes the real-time location of transport vehicles and maintenance progress. When unexpected situations (such as temporary staff absences or transport delays) disrupt the original plan, the system immediately activates the contingency plan, recalculates alternatives, and allocates backup resources to ensure that critical maintenance tasks are completed on time. All scheduling instructions are synchronized to terminal devices in real time via industrial communication protocols, and task status is dynamically updated on the maintenance dashboard, achieving transparent management of the entire resource scheduling process.

[0137] In an optional but non-limiting embodiment, the method further comprises:

[0138] Before the equipment management terminal performs a maintenance operation, generating equipment operation status prediction data after simulated maintenance based on the maintenance operation parameters, the equipment operation status prediction data after simulated maintenance including a temperature change trend simulation value, a vibration amplitude simulation range, and a current stability prediction interval;

[0139] Perform similarity matching on the predicted equipment operating status data after the simulated maintenance and historical normal equipment data samples to generate a maintenance operation risk level, which includes equipment overload probability, environmental interference sensitivity, and secondary abnormality triggering possibility;

[0140] When the maintenance operation risk level exceeds a preset safety threshold, the maintenance interval duration or resource allocation ratio in the maintenance operation parameters is modified to generate maintenance operation parameters after risk mitigation;

[0141] The maintenance operation parameters after risk mitigation are re-bound to the equipment maintenance instruction set, and the equipment management terminal is triggered to perform actual maintenance operations after simulation verification passes.

[0142] Before performing maintenance operations, the system uses digital twin technology to construct a virtual map of the equipment's state and predict post-maintenance changes in operating parameters through multi-physics simulation. The predictive model comprehensively considers multiple factors, including mechanical transmission, thermodynamic conduction, and electrical control, generating simulated values ​​for key indicators such as temperature trends, vibration amplitude ranges, and current stability intervals. For example, after replacing a component, the system predicts that the servo motor's peak temperature will drop to a reasonable range and the vibration energy distribution will stabilize. The system then performs pattern matching on the predicted data against a database of historical normal operating conditions and uses a similarity algorithm to evaluate the effectiveness of the maintenance plan. When potential risks are detected (such as insufficient heat dissipation efficiency or abnormal load fluctuations), the system automatically and dynamically adjusts maintenance parameters. Correction strategies include extending the maintenance cycle, adding auxiliary measures, or adjusting operational procedures. For example, in response to identified heat dissipation risks, the system adjusts the cooling fan's operating mode from intermittent to continuous operation and increases the frequency of temperature monitoring. These corrected parameters are then verified through a secondary simulation to ensure that risk indicators fall below safety thresholds. After receiving the final command, the device management terminal delays the actual maintenance operation until the digital twin system outputs a verification pass signal. This mechanism effectively avoids secondary failures caused by maintenance plan defects while balancing maintenance efficiency and equipment reliability. All correction records and verification results are stored in the knowledge base, providing data support for subsequent maintenance strategy optimization.

[0143] In an optional but non-limiting embodiment, the method further comprises:

[0144] Setting an effective execution time window for each maintenance operation in the equipment maintenance instruction set, wherein the effective execution time window is dynamically adjusted based on the abnormality confidence and the equipment type;

[0145] monitoring in real time the maintenance operation start signal and completion signal of the device management terminal within the effective execution time window, and marking the maintenance operation as a timeout and incomplete event if no completion signal is detected within the time window;

[0146] Based on the abnormal device identifier corresponding to the timeout incomplete event, re-acquire the real-time multi-source monitoring data stream of the device and trigger a new round of abnormality detection to generate an updated device abnormality probability distribution;

[0147] The original maintenance priority and maintenance operation parameters are re-evaluated according to the updated equipment abnormality probability distribution, and a revised maintenance instruction set corresponding to the failure time window is generated and sent to the equipment management terminal.

[0148] In implementation, the system sets a dynamic effective execution time window for each maintenance instruction. The window length automatically adjusts based on the fault type, confidence level, and equipment criticality. For example, high-confidence mechanical faults are assigned a shorter default time window to ensure a rapid response, while low-risk environmental anomalies are assigned a longer processing period. The time window monitoring module tracks the initiation and completion status of maintenance operations in real time, identifying timeout events through timestamp comparison. When a task is detected not to be completed within the window, the system immediately flags the event and triggers an anomaly reassessment process. The system recollects real-time operating data from the target equipment and rapidly generates an updated probability distribution using an anomaly detection model. For example, if a piece of equipment's vibration parameters continue to deteriorate due to delayed maintenance, the model recalculates the failure probability from its initial value to a higher level. Based on the updated assessment results, the system re-prioritizes maintenance and optimizes resource allocation strategies. For equipment that escalates to an emergency status, the system overrides standard scheduling rules and activates alternative resource channels. For example, pre-researched spare parts from the laboratory may be directly deployed, or a cross-departmental expert team may intervene. The revised instruction set is generated using a priority-weighted algorithm and includes detailed execution constraints (such as specific environmental parameter requirements). All adjustments are synchronized to the terminal device in real time, and the global monitoring interface highlights task status changes. The system also records the cause analysis and processing logs of timeout events, which are used to optimize subsequent time window calculation models and gradually improve the accuracy of maintenance plans.

[0149] In summary, the embodiment of the present invention constructs an end-to-end analysis architecture for multi-source heterogeneous monitoring data through the collaborative optimization of cross-modal data alignment and edge computing dynamic feature extraction, achieving a dual improvement in the real-time performance and accuracy of equipment anomaly detection. Through a multi-level data integration mechanism with timestamp synchronization, frequency unification and format standardization, the problem of separation between multi-source data streams in temporal correlation and semantic consistency is effectively overcome, so that a dynamic coupling analysis is formed between the equipment operating status, environmental parameters and operation logs; the edge node performs dynamic feature extraction based on the real-time data stream, which not only captures the equipment's own operating characteristics, but also integrates environmental correlation effects and abnormal fluctuation patterns, and constructs a health portrait of the equipment in a multi-dimensional feature space; the anomaly detection integration model uses a multi-level feature fusion strategy to map the nonlinear correlation between heterogeneous features into a probabilistic anomaly distribution, breaking through the limitations of a single detection model's ability to represent complex anomaly patterns. The maintenance instruction set finally generated realizes closed-loop decision-making on anomaly location, priority determination and operating parameters, improving the automation level and decision reliability of the maintenance response of smart park equipment.

[0150] See also Figure 2 As shown in FIG. 1 , this figure is a schematic diagram of the basic structure of a device anomaly detection system 200 provided by an embodiment of the present invention. The device anomaly detection system 200 includes:

[0151] Processor 201;

[0152] a storage device 202 having a computer program 2020 stored thereon;

[0153] When the computer program 2020 is executed by the processor 201, the processor 201 implements any of the device anomaly detection methods based on multi-source heterogeneous data.

[0154] Based on the above, a readable storage medium is provided, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0155] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

Claims

1. A device anomaly detection method based on multi-source heterogeneous data, characterized in that: include: Acquire real-time multi-source monitoring data streams for production equipment within the smart park, including equipment operating status data, environmental sensor data, and equipment operation log data; Performing cross-modal data alignment processing on the real-time multi-source monitoring data stream to generate a target monitoring data set, wherein the cross-modal data alignment processing includes timestamp synchronization, data sampling frequency unification, and data format normalization; Performing a dynamic feature extraction operation in each edge computing node to generate a multi-dimensional device state feature set based on the target monitoring data set, wherein the multi-dimensional device state feature set includes device operation state features, environment-related features, and data abnormal fluctuation features; Inputting the multi-dimensional device status feature set into the trained anomaly detection integrated model, generating a device anomaly probability distribution through a multi-level feature fusion strategy, and determining the anomaly type and anomaly confidence of the target device based on the device anomaly probability distribution; Generate a device maintenance instruction set based on the exception type and the exception confidence, the device maintenance instruction set including the abnormal device identifier, maintenance priority, and maintenance operation parameters, and send the device maintenance instruction set to the device management terminal of the smart park to trigger an exception handling operation; The method further comprises: Before the equipment management terminal performs a maintenance operation, generating equipment operation status prediction data after simulated maintenance based on the maintenance operation parameters, the equipment operation status prediction data after simulated maintenance including a temperature change trend simulation value, a vibration amplitude simulation range, and a current stability prediction interval; Perform similarity matching on the predicted equipment operating status data after the simulated maintenance and historical normal equipment data samples to generate a maintenance operation risk level, which includes equipment overload probability, environmental interference sensitivity, and secondary abnormality triggering possibility; When the maintenance operation risk level exceeds a preset safety threshold, the maintenance interval duration or resource allocation ratio in the maintenance operation parameters is modified to generate maintenance operation parameters after risk mitigation; The maintenance operation parameters after risk mitigation are re-bound to the equipment maintenance instruction set, and the equipment management terminal is triggered to perform actual maintenance operations after simulation verification is passed.

2. The method according to claim 1, characterized in that The performing cross-modal data alignment processing on the real-time multi-source monitoring data stream to generate a target monitoring data set includes: Extracting a first raw data stream from the device operating status data, the first raw data stream including device temperature, vibration amplitude, and current fluctuation parameters; Extracting a second raw data stream from the environmental sensor data, wherein the second raw data stream includes environmental temperature, humidity, and dust concentration; Extracting a third original data stream from the device operation log data, the third original data stream including device start and stop records, operation instruction sequences, and alarm history records; Performing a timestamp synchronization operation on the first original data stream, the second original data stream, and the third original data stream to generate a time-aligned initial data set, wherein the timestamp synchronization operation includes: segmenting each original data stream based on a preset time window and filling data in missing time segments; Performing sampling frequency unification processing on the time-aligned initial data set to generate a target data set, wherein the sampling frequency unification processing includes: downsampling the high-frequency data stream to a preset reference frequency, and interpolating and completing the low-frequency data stream to the preset reference frequency; The target data set is subjected to format normalization processing to generate the target monitoring data set, wherein the format normalization processing includes data unit conversion, data range normalization and outlier removal.

3. The method according to claim 2, characterized in that The performing of a dynamic feature extraction operation in each edge computing node to generate a multi-dimensional device state feature set based on the target monitoring data set includes: Extracting a subset of device operating status data from the target monitoring data set, performing time domain feature extraction on the subset of device operating status data to generate device operating status features, wherein the time domain feature extraction includes mean calculation, variance calculation, and peak detection; Extracting a subset of environmental sensor data from the target monitoring data set, performing environmental correlation analysis on the subset of environmental sensor data to generate environmental correlation features, wherein the environmental correlation analysis includes calculating the difference between the device temperature and the ambient temperature and extracting the influence coefficient of humidity on device vibration; Extracting a subset of device operation log data from the target monitoring data set, performing abnormal fluctuation detection on the subset of device operation log data, and generating data abnormal fluctuation features, wherein the abnormal fluctuation detection includes detecting sudden changes in operation instruction frequency and correlating alarm records with operating parameters; The device operation status features, the environment-related features, and the data abnormal fluctuation features are combined in chronological order to generate the multi-dimensional device status feature set.

4. The method according to claim 1, wherein The trained anomaly detection ensemble model is generated by the following steps: Acquire a historical multi-source monitoring data set, wherein the historical multi-source monitoring data set includes normal device data samples and abnormal device data samples marked with abnormal types; Performing cross-modal data alignment processing on the historical multi-source monitoring dataset to generate a historical aligned dataset, and extracting a historical multi-dimensional device state feature set based on the historical aligned dataset; Dividing the historical multi-dimensional device status feature set into a training set and a validation set, and constructing an initial anomaly detection model set, wherein the initial anomaly detection model set includes a first detection model based on an isolation forest, a second detection model based on a temporal convolutional network, and a feature fusion model based on an attention mechanism; Training the initial anomaly detection model set using a cascade training strategy to obtain a target anomaly detection model set, the cascade training strategy comprising: performing unsupervised training on the first detection model using the training set, using the output probability of the first detection model as an input feature of a second detection model, and combining the second detection model with the feature fusion model for supervised training; The performance of the target anomaly detection model set is evaluated based on the validation set. When the accuracy of anomaly type recognition exceeds a preset threshold, the target anomaly detection model set is used as a trained anomaly detection integrated model and deployed to the edge computing node.

5. The method according to claim 4, characterized in that The method of generating a device anomaly probability distribution through a multi-level feature fusion strategy and determining the anomaly type and anomaly confidence of the target device based on the device anomaly probability distribution includes: Inputting the multi-dimensional device status feature set into the first detection model to generate a first abnormality probability distribution, the first abnormality probability distribution including the overall abnormality probability of the device and the abnormality probability of local components; Inputting the multi-dimensional device state feature set and the first anomaly probability distribution into the second detection model to generate a second anomaly probability distribution, the second anomaly probability distribution including a time series anomaly pattern and an anomaly duration prediction; Inputting the first anomaly probability distribution and the second anomaly probability distribution into the feature fusion model, and generating a fused anomaly probability distribution through an attention weight allocation mechanism; Threshold segmentation processing is performed on the fused abnormal probability distribution to determine the abnormal type and abnormal confidence of the target device, and the threshold segmentation processing includes dynamic threshold adjustment and probability density clustering.

6. The method according to claim 1, characterized in that The generating of a device maintenance instruction set based on the exception type and the exception confidence level includes: Matching a preset maintenance strategy library according to the abnormality type to obtain a maintenance operation template and priority rules corresponding to the abnormality type; Dynamically adjusting parameters in the maintenance operation template based on the abnormality confidence to generate maintenance operation parameters, wherein the dynamic adjustment includes shortening the maintenance interval and increasing the maintenance resource allocation as the confidence level increases; A plurality of abnormal devices are sorted according to the priority rule to generate a maintenance priority queue, and the maintenance priority queue is bound to a maintenance operation parameter to generate the device maintenance instruction set.

7. The method according to claim 1, characterized in that The method further comprises: Monitor the execution status of the device maintenance instruction set by the device management terminal in real time, and trigger the following operations when a maintenance operation delay or failure is detected: Obtaining an abnormal device identifier corresponding to a delay or failure event, and extracting a historical multi-dimensional device status feature set of the abnormal device based on the abnormal device identifier; Incrementally training the anomaly detection integrated model based on the historical multi-dimensional device state feature set to generate an updated anomaly detection integrated model; The updated anomaly detection integrated model is pushed to the edge computing node for model replacement, and the device maintenance instruction set is regenerated.

8. The method according to claim 7, characterized in that The incremental training of the anomaly detection integrated model based on the historical multi-dimensional device status feature set to generate an updated anomaly detection integrated model includes: Obtaining an abnormal device identifier corresponding to the delay or failure event, and extracting, based on the abnormal device identifier, a device operation status feature segment, an environment-related feature segment, and a data abnormal fluctuation feature segment of the abnormal device within a preset time window from the historical multi-dimensional device status feature set; Performing event labeling processing on the device operation status feature segments, environment-related feature segments, and data abnormal fluctuation feature segments to generate an incremental training data set, wherein the event labeling processing includes marking the feature segments of the continuous time period before the delay or failure event occurs as the abnormality category to be optimized; Loading the network parameters of the feature extraction layer from the anomaly detection integrated model and locking them in a non-updated state, while loading the fully connected weight matrix of the classification layer and setting it to a pending update state; Inputting the incremental training data set into the classification layer to perform a backpropagation optimization operation, and adjusting the numerical distribution of the fully connected weight matrix by a gradient descent algorithm until the decrease of the classification loss function on the validation subset is less than a preset convergence threshold; The optimized fully connected weight matrix is ​​recombined with the locked feature extraction layer parameters to generate an updated anomaly detection integrated model, and the performance of the updated anomaly detection integrated model is verified on the historical verification set. If the difference in anomaly type recognition accuracy compared to the value before the update is within the preset tolerance range, the model update is confirmed to be complete.

9. The method according to claim 1, characterized in that The method further comprises: Displaying the equipment abnormality probability distribution and equipment maintenance instruction set in the visual monitoring interface of the smart park; In response to user interaction with a selected device, dynamically display a multi-dimensional device status feature set, anomaly type determination criteria, and maintenance operation execution progress for the selected device; When it is detected that the user manually corrects the anomaly type or maintenance priority, a correction record is recorded and incremental training of the anomaly detection integrated model is triggered; The dynamic display of the multi-dimensional device status feature set of the device includes: Extracting the temperature change rate numerical sequence, vibration spectrum energy numerical sequence, and current fluctuation parameter sequence corresponding to the equipment operation status characteristics from the multi-dimensional equipment status feature set, generating a temperature-vibration-current superposition curve based on a unified time axis, and rendering it to the visual monitoring interface; Extracting the ambient temperature distribution matrix, humidity influence coefficient matrix, and dust concentration correlation matrix corresponding to the environmental correlation features from the multi-dimensional device status feature set, generating a three-dimensional heat map based on the device physical location coordinate mapping, and superimposing and marking abnormal correlation areas exceeding a preset threshold on the three-dimensional heat map; Extracting the operation instruction frequency mutation time point, alarm record timestamp, and operation parameter abnormal fluctuation period corresponding to the data abnormal fluctuation feature from the multi-dimensional device status feature set, generating an interactive event timeline based on a time alignment strategy, and displaying the device operation log text fragment at the corresponding moment on the event timeline; The temperature-vibration-current superposition curve graph, three-dimensional thermal map and interactive event timeline are embedded in the dynamic display panel of the visual monitoring interface according to a preset layout. When it is detected that the user clicks on the abnormal correlation area or the marked point in the event timeline, the corresponding device operation log text segment is highlighted and the curve of the abnormal fluctuation period is enlarged and displayed.

10. A device anomaly detection system, characterized in that: include: processor; A storage device having a computer program stored thereon, wherein when the computer program is executed by the processor, the processor implements the device anomaly detection method based on multi-source heterogeneous data as described in any one of claims 1 to 9.

Images