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

Abstract

The invention discloses an equipment anomaly detection method and system based on multi-source heterogeneous data, and the method comprises the steps: obtaining a real-time multi-source monitoring data flow of production equipment in a smart park, carrying out the cross-modal data alignment processing of the real-time multi-source monitoring data flow, and generating a target monitoring data set, and executing a dynamic feature extraction operation in each edge computing node, generating a multi-dimensional equipment state feature set based on the target monitoring data set, inputting the multi-dimensional equipment state feature set into the trained anomaly detection integrated model, generating equipment anomaly probability distribution through a multi-level feature fusion strategy, and outputting the equipment anomaly probability distribution. Determining the abnormity type and abnormity confidence of the target equipment according to the equipment abnormity probability distribution, generating an equipment maintenance instruction set based on the abnormity type and the abnormity confidence, and sending the equipment maintenance instruction set to an equipment management terminal of the smart park to trigger abnormity processing operation; therefore, the automation degree and decision reliability of the equipment maintenance response of the smart park are improved.

Description

Technical Field

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

[0002] With the continuous development of industrial Internet of Things technology, the anomaly monitoring and analysis of industrial devices have become particularly important. However, traditional technologies mostly adopt single-dimensional data acquisition and centralized processing architectures, and realize anomaly early warning through preset thresholds or single-sensor data analysis, lacking a spatio-temporal alignment mechanism for cross-modal data, resulting in problems such as data spatio-temporal misalignment and lack of feature correlation in device state evaluation. At the same time, traditional feature extraction methods are mostly based on fixed sampling periods and static feature templates, and it is difficult to adapt to the dynamic changes of device operating states and the influence of environmental coupling, and it is easy to produce misjudgments when dealing with complex working conditions.

[0003] In the anomaly detection link, existing solutions mostly rely on a single detection model to perform linear analysis on structured data, and cannot effectively capture the non-linear correlation features between multi-source data, and the recognition accuracy of compound anomalies is limited. In addition, traditional systems usually adopt a cloud centralized decision-making mode, with relatively high data processing delays and lack of fine-grained parameter guidance for generating maintenance instructions, resulting in lagged anomaly responses and insufficient adaptability of maintenance operations. The above series of defects seriously restrict the real-time performance, accuracy and decision-making effectiveness of the device health management in smart parks, and how to improve the automation degree and decision-making reliability of the device maintenance response in smart parks has become a difficult problem to overcome at present. Summary of the Invention

[0004] The embodiments of the present invention provide a method and system for device anomaly detection based on multi-source heterogeneous data, which are used to improve the automation degree and decision-making reliability of the device maintenance response in smart parks.

[0005] In a first aspect, an embodiment of the present invention provides a device anomaly detection method based on multi-source heterogeneous data, which is applied to a device anomaly detection system. The method includes: obtaining a real-time multi-source monitoring data stream of production devices in a smart park, where the real-time multi-source monitoring data stream includes device operation status data, environmental sensing data, and device 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, and the cross-modal data alignment processing includes timestamp synchronization, data sampling frequency unification, and data format standardization; performing dynamic feature extraction operations in each edge computing node to generate a multi-dimensional device status feature set based on the target monitoring data set, and the multi-dimensional device status feature set includes device operation status features, environment-related features, and data anomaly fluctuation features; 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 level of the target device according to the device anomaly probability distribution; generating a device maintenance instruction set based on the anomaly type and the anomaly confidence level, where the device maintenance instruction set includes an anomaly device identifier, a maintenance priority, and maintenance operation parameters, and sending the device maintenance instruction set to the device management terminal of the smart park to trigger an anomaly handling operation.

[0006] In a second aspect, an embodiment of the present invention provides a device anomaly detection system, including: a processor; a storage device, on which a computer program is stored, when the computer program is executed by the processor, the processor implements any one of the device anomaly detection methods based on multi-source heterogeneous data.

[0007] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored, and 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.

[0008] 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, realizing a double improvement in the real-time performance and accuracy of device anomaly detection. Through a multi-level data integration mechanism of timestamp synchronization, frequency unification, and format standardization, the fragmentation problem of multi-source data streams in temporal correlation and semantic consistency is effectively overcome, enabling dynamic coupling analysis among device operating states, environmental parameters, and operation logs; Edge nodes perform dynamic feature extraction based on real-time data streams, not only capturing the operating characteristics of the devices themselves, but also integrating environmental correlation effects and abnormal fluctuation patterns to construct a device health portrait in a multi-dimensional feature space; The anomaly detection integration model maps the non-linear correlation between heterogeneous features into a probabilistic anomaly distribution through a multi-level feature fusion strategy, breaking through the limitations of a single detection model in representing complex anomaly patterns. The finally generated maintenance instruction set realizes closed-loop decision-making for anomaly location, priority determination, and operation parameters, improving the automation level and decision-making reliability of intelligent park device maintenance response. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 FIG. is a flowchart of a device anomaly detection method based on multi-source heterogeneous data provided by an embodiment of the present invention.

[0010] Figure 2 FIG. is a schematic diagram of the basic structure of a device anomaly detection system provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] To make the above objects, features, and advantages of the present invention more obvious and understandable, the embodiments of the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0012] See Figure 1 shown, this figure is a flowchart of a device anomaly detection method based on multi-source heterogeneous data provided by an embodiment of the present invention, and this method can be applied to a device anomaly detection system. As Figure 1 shown, this method may include step 101-step 105.

[0013] For ease of understanding, an embodiment of the present invention takes a six-axis collaborative robot (exemplary production equipment) in a new energy vehicle battery module assembly line in a vehicle enterprise production line park (exemplary intelligent park) as a specific application object for description.

[0014] Step 101: Obtain real-time multi-source monitoring data streams of production equipment in the intelligent park, where the real-time multi-source monitoring data streams include device operating state data, environmental sensing data, and device operation log data.

[0015] In an 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.

[0016] The device operation status data represents real-time parameters of the mechanical performance of the six-axis collaborative robot body, covering core operation parameters such as dynamic response indicators of the joint drive system, mechanical feedback signals of the end effector, motion trajectory deviation monitoring values, and temperature rise rates 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-amplitude oscillation waveform of the mechanical drive system in an ultra-high frequency mode, and the Hall effect sensor of the servo motor continuously records the current phase change of the rotor winding.

[0017] The environmental sensing data includes physical environment indicators such as air quality parameters, temperature and humidity gradient distribution, and electrostatic field strength in the assembly workshop, and realizes spatial continuous monitoring through a multi-spectral environmental sensor network deployed around the production line.

[0018] The device operation log data consists of structured operation sequences generated by a programmable logic controller, and details timing operation information such as robotic arm coordinate positioning instructions, workpiece clamping state change events, and safety protection system trigger records. Each log is attached with an accurate time stamp and device identity identifier.

[0019] In the actual application process, the data acquisition system adopts a redundant transmission protocol to ensure the integrity of multi-source data. For example, high-priority device status data is transmitted collaboratively through industrial Ethernet and Time-Sensitive Network (TSN), and at the same time, wireless Mesh network is used to backhaul environmental sensing data to optimize wiring costs. The real-time guarantee mechanism of the data stream is implemented based on the cache queue of the edge computing node. When the network bandwidth fluctuates, a data priority hierarchical transmission strategy is automatically triggered to ensure the low-latency transmission of key operation parameters.

[0020] Step 102: Perform cross-modal data alignment processing on the real-time multi-source monitoring data stream to generate a target monitoring data set. The cross-modal data alignment processing includes timestamp synchronization, data sampling frequency unification, and data format standardization.

[0021] In the embodiments of the present invention, cross-modal data alignment processing aims to solve the problem of inconsistent spatio-temporal benchmarks for multi-source heterogeneous data. Timestamp synchronization eliminates the time benchmark offset between devices through a precision clock synchronization protocol. Specifically, the IEEE 1588 Precision Time Protocol (PTP) is used to calibrate the clocks of all data sources at the nanosecond level, and a hardware timestamp marking module is deployed at the edge gateway to map the analog signal sampling time of the vibration sensor, the discrete acquisition period of the environmental sensor, and the generation time of the operation log to the global timeline. The data sampling frequency needs to be unified by performing dynamic resampling on data streams with different sampling frequencies: for high-frequency device status data, anti-aliasing filtering combined with a downsampling algorithm is used to generate an equivalent signal that matches the target frequency; for low-frequency environmental parameters, a continuous function model is constructed based on the cubic spline interpolation algorithm and then resampled to the target frequency. Data format normalization involves the unified representation conversion of multi-modal data. For example, the original analog voltage signal of the robotic arm joint angle sensor is converted into a radian value in the standard International System of Units, the multi-dimensional temperature field data of the 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 codes through a semantic parsing engine.

[0022] During the actual application process, the data quality verification module runs synchronously. It identifies abnormal data points caused by sensor failures through an outlier detection algorithm, and performs data repair or marked deletion based on the historical data distribution model, and finally generates a set of target monitoring data with spatio-temporal consistency.

[0023] Step 103: Perform dynamic feature extraction operations in each edge computing node, and generate a multi-dimensional device status feature set based on the set of target monitoring data. The multi-dimensional device status feature set includes device operation status features, environment correlation features, and data abnormal fluctuation features.

[0024] In the embodiment of the present invention, the dynamic feature extraction operation relies on the real-time processing capability of the edge computing architecture to implement multi-level feature engineering at the source of the data. The equipment operation status feature focuses on the quantitative characterization of the microscopic behavior of the mechanical system, such as extracting the frequency band energy entropy feature from the vibration signal through the wavelet packet transform algorithm to reveal the subtle changes in the meshing state of the harmonic reducer gear; using the Kalman filter to estimate the state of the servo motor current signal to generate the dynamic impedance feature that characterizes the health of the winding; based on the kinematic inverse solution model, the statistical distribution characteristics of the positioning error at the end of the robot arm are calculated to reflect the cumulative effect of the backlash of the transmission system. The environmental correlation feature reveals the coupling relationship between the equipment operation parameters and the external environmental variables, such as constructing the time-varying correlation matrix of the workshop temperature and humidity change rate and the motor heat dissipation efficiency, and analyzing the causal correlation model between the concentration of air particles and the wear rate of precision bearings. The data abnormal fluctuation feature uses an adaptive threshold detection algorithm to identify potential abnormal events, such as analyzing the transient mutation mode of the current signal through the sliding window statistic, and combining the hidden Markov model (HMM) to probabilistically model the duration and intensity of abnormal events.

[0025] 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.

[0026] In the specific scenario of the battery module assembly line for new energy vehicles, the operating status of the six-axis collaborative robot is directly related to the battery pack assembly quality and production line safety. Taking the battery cell stacking process as an example, the dynamic feature extraction of step 103 needs to pay special attention to the physical feature dimensions that are strongly related to the battery module assembly. In the equipment operation status feature layer, for the mechanical feedback signal of the end effector during the battery cell grasping process, the feature extraction module will construct a clamping force-displacement joint feature matrix: by real-time analysis of the pulsating waveform of the vacuum suction cup pressure sensor, combined with the micron-level positioning data of the contact displacement probe, the clamping force steady-state retention rate, contact surface pressure distribution uniformity and other characteristic parameters are extracted. These parameters are strongly correlated with the surface flatness of the lithium battery cell: when the clamping force fluctuation coefficient is detected to exceed the threshold, it may indicate that there is warping or diaphragm wrinkle defects on the battery cell surface. At the same time, for the motion trajectory characteristics of the six-axis robot arm of the welding station, the system will focus on analyzing the positioning repeatability accuracy of the TCP (tool center point) in three-dimensional space, and generate an attenuation index characterizing the dynamic accuracy of the robot arm by calculating the Hausdorff distance between the actual trajectory and the theoretical trajectory between adjacent welding points. 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.

[0027] In the environmental correlation feature dimension, due to the extreme sensitivity of battery module assembly to cleanliness, temperature, and humidity, feature engineering needs to deeply integrate the correlation between environmental sensing data and equipment operation parameters. For example, when it is monitored that the electrostatic field strength in the workshop exceeds 50 kV / m during the dry season, the system will dynamically activate the electrostatic protection correlation analysis model: by analyzing the charge transfer amount at the moment when the end effector of the robotic arm contacts the battery case and combining the spatial gradient distribution data of the environmental humidity sensor, a time-varying relationship model between the electrostatic accumulation rate and the equipment grounding impedance is constructed. When it is detected that the electrostatic dissipation time constant of a certain workstation is abnormally extended, the poor contact of the conductive brush or the coverage blind area of the ion blower can be accurately located. In addition, for the strict requirements of the dew point temperature in the lithium battery liquid injection process, the system will calculate the deviation between the environmental temperature and humidity sensor data and the set value of the sealed chamber dew point control system in real time, and reveal the quantitative influence coefficient of temperature fluctuation on the liquid injection accuracy through sliding window regression analysis, providing dynamic compensation parameters for subsequent anomaly detection.

[0028] In the data abnormal fluctuation feature layer, the specific process risk points of battery module assembly determine the targeted strategy for feature extraction. 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 a constant pressure. At this time, the system will establish a multi-physical field coupling monitoring model: by integrating the current ripple characteristics of the servo motor, the temperature gradient distribution data of the pressure head, and the step response curve of the pressure sensor, a thermal-mechanical coupling state feature vector is constructed. When it is detected that the current harmonic component in the pressure holding stage increases abnormally and is accompanied by a sudden local temperature rise, the mechanical jamming phenomenon caused by the deviation of the mold cavity positioning can be quickly identified. Such composite features are of great value for preventing the misalignment of the electrode sheets inside the battery module. Compared with the traditional single-dimensional threshold warning method, the detection timeliness is significantly improved.

[0029] Step 104: Input the multi-dimensional device state feature set into the trained anomaly detection integration model, generate the device anomaly probability distribution through a multi-level feature fusion strategy, and determine the anomaly type and anomaly confidence level of the target device according to the device anomaly probability distribution.

[0030] In the embodiment of the present invention, the anomaly detection integration model adopts a heterogeneous model fusion architecture to realize multi-source feature collaborative analysis. The bottom layer of the model is processed by a convolutional neural network (CNN) branch for vibration spectrum image features, and local anomaly patterns in the frequency domain are extracted through multiple convolutional kernels; the long short-term memory network (LSTM) branch analyzes the dynamic evolution law of time series features to capture the progressive deterioration trend of device performance; the graph neural network (GNN) branch models the topological association relationship between device components to identify the abnormal propagation path.

[0031] Furthermore, the multi-level feature fusion strategy dynamically assigns feature weights through the attention mechanism. For example, in the mechanical overload warning scenario, it increases the weight coefficient of the mechanical feature branch, and in the environmental interference scenario, it enhances the decision-making contribution of the environmental correlation features. The device abnormal probability distribution synchronously outputs the probability values of preset fault categories and the confidence scores of unknown abnormal types through a multi-task learning framework. The known fault category library covers typical industrial scenarios such as mechanical transmission failures, electrical system failures, and environmental interference anomalies. The abnormal confidence calculation fuses the prediction probability of the model and the similarity measurement results of historical fault examples, and uses the evidence reasoning theory to quantify the decision credibility under uncertain conditions. In addition, the model inference process introduces an interpretability enhancement module, which assists maintenance personnel in understanding the abnormal determination basis through feature importance ranking and decision path visualization techniques.

[0032] It is worth mentioning that for the above abnormal detection integrated model, the model's recognition ability for process-specific faults needs to be particularly optimized in the battery module assembly scenario. Taking the battery pack bolt tightening process as an example, the six-axis robot needs to complete the tightening operation of hundreds of connection points according to a preset torque curve. At this time, the model will adopt a multi-modal fusion detection strategy: the CNN branch focuses on analyzing the time-frequency diagram features of the servo motor current signal to capture the harmonic distortion mode caused by torque mutation; the LSTM network tracks the temporal correlation of the joint angle deviation of each axis during the tightening process to identify the progressive accuracy loss caused by tool wear; at the same time, the graph neural network models the device topology relationship between the robot body, the torque gun, and the vision positioning system, and automatically triggers an abnormal alarm for the assembly process chain when the multi-node collaborative error is detected to exceed the tolerance. For the battery module sealing detection scenario, the model will introduce a transfer learning mechanism to transfer the feature patterns of helium leak detection rate exceeding the standard events in historical examples to the current vibration signal analysis, and can achieve a quite high fault recall rate even in the initial stage of new model production.

[0033] In terms of specific fault discrimination, the system has established a fault knowledge graph covering the entire battery assembly process. For example, when abnormal vibrations occur during the Busbar laser welding by the robotic arm, 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 reading changes of the environmental oxygen concentration sensor. By calculating the influence weights of each feature node through the graph attention network, three typical faults can be accurately distinguished: if the high-frequency vibration feature is strongly correlated with the cooling water pressure fluctuation (weight > 0.7), it is determined as a cooling pipeline blockage; if the vibration mode is synchronized with the penetration data change and the oxygen concentration is abnormal, it is classified as a protective gas supply fault; when the vibration energy is concentrated in the mid-low frequency band and accompanied by abnormal joint temperature gradients, it is identified as a reducer lubrication failure. This fine-grained diagnosis ability greatly shortens the average fault location time.

[0034] Step 105: Generate a set of device maintenance instructions based on the exception type and the exception confidence level. The set of device maintenance instructions includes the exception device identifier, the maintenance priority, and the maintenance operation parameters, and send the set of device maintenance instructions to the device management terminal of the smart park to trigger an exception handling operation.

[0035] In the embodiment of the present invention, the generation of the set of device maintenance instructions follows the principles of hierarchical response and resource optimization. Specifically, the maintenance priority determination module comprehensively considers the exception confidence level, the device criticality index, and the production plan status. For example, for a core device failure with a high confidence level, the highest priority response is immediately triggered, while for an exception of an auxiliary device with a low confidence level, it is scheduled for processing during non-production periods. The exception device identifier adopts a multi-level coding system, integrating device physical location, function module number, and production line level information to ensure that the device positioning accuracy reaches the workstation level granularity. The maintenance operation parameter library is associated with the device digital twin model to automatically match the standard maintenance process, the list of special tool requirements, and the safety operation constraints corresponding to the failure mode.

[0036] Furthermore, the instruction distribution system adopts a publish-subscribe mode to achieve multi-terminal collaboration. For example, the detailed maintenance steps are pushed to the augmented reality (AR) inspection terminal through the OPCUA protocol, and at the same time, a spare part requisition request is sent to the material management system. The maintenance process implements closed-loop monitoring. When the operator executes the key steps, the operation compliance is verified in real time through Internet of Things devices. For example, a torque wrench sensor is used to confirm that the bolt tightening torque meets the process specifications. The maintenance result data is automatically fed back to the exception detection model for online parameter fine-tuning to continuously optimize the fault diagnosis accuracy.

[0037] Taking the detection of the deterioration of the positioning accuracy of the battery cell stacking robot as an example, the system will perform multi-dimensional decision-making and reasoning: First, analyze the root cause of the accuracy loss based on historical maintenance records: If the kinematic model parameter error shows a linear increase, a servo motor encoder calibration instruction is generated; if the joint clearance error shows a non-linear mutation, a harmonic reducer wear inspection work order is triggered; when it is detected that the environmental vibration exceeds the standard at the same time, it is recommended to perform the equipment base leveling operation first. The battery assembly-specific knowledge is preset in the maintenance operation parameter library. For example, when replacing the vacuum suction cup of the battery cell gripper, the system not only provides the standard torque parameters but also associates the battery cell model change records in the past three months at this workstation to dynamically adjust the suction cup material recommendation plan (such as silicone suction cups are suitable for prismatic battery cells, and polyurethane suction cups are more suitable for soft-pack battery cells).

[0038] In terms of maintaining priority decisions, the system adopts a dynamic weighted evaluation model. Take the case of multiple device alarms occurring simultaneously: When the clock deviation between the stacking robot and the vision positioning system reaches 10 ms due to bus transmission delay (which may cause a positioning error of 0.1 mm), the system will automatically adjust the response level according to the accuracy requirements of the battery cells in the current production batch (the energy storage battery allows an error of 0.3 mm, and the power battery needs to be controlled within 0.1 mm). For the power battery production line, such an anomaly will immediately trigger a yellow warning and schedule calibration; while in the energy storage battery production line, it is marked as an observation-level event. This differential management strategy significantly improves the utilization rate of maintenance resources at key workstations and reduces unnecessary downtime at the same time.

[0039] At the level of executing maintenance instructions, the system is deeply integrated with the digital twin system of the battery production line. When it is detected that the guide shaft of the module end plate pressing equipment is worn, the maintenance instructions not only include the standard operation procedures for replacing components, but also synchronously retrieve the three-dimensional point cloud model of the equipment and overlay and display the heat map of the key disassembly path on the AR terminal: specifically marking the dangerous operation areas that may affect adjacent high-voltage wire harnesses. For complex maintenance tasks that require multi-disciplinary collaboration (such as the optical path calibration of the laser welding head), the system will automatically decompose the task steps and synchronously push them to the terminals of technicians in different majors of machinery, optics, and electricity to ensure strict spatio-temporal matching of operations in each link.

[0040] In terms of the unique safety prevention and control in battery module assembly, the maintenance instruction system is deeply bound to the process safety parameters. For example, when the environmental sensor detects that the VOCs concentration at a certain workstation approaches 25% of the lower explosion limit, all maintenance operation instructions related to this area will automatically append third-level protection requirements: forcibly turn on the explosion-proof ventilation system, disable wireless communication devices, and limit the use of intrinsically safe tools. At the same time, the system will dynamically adjust the maintenance schedule, prioritize the replacement operation of mechanical components that may generate sparks, and arrange the risk processes during the low-production period of the shift handover.

[0041] With such a design, 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: The cross-modal data alignment processing effectively solves the spatio-temporal mismatch problem of multi-source heterogeneous data in the industrial field and ensures the basic quality of monitoring data; the dynamic feature extraction mechanism fully explores the multi-dimensional characterization information of the equipment operation state and enhances the identification ability of anomaly patterns; the anomaly detection integration model realizes robust diagnosis under complex working conditions through a multi-level feature fusion strategy and reduces the risk of false alarms and missed detections; the generation of intelligent maintenance instructions realizes the full-process automation from fault diagnosis to disposal execution, significantly shortening the equipment downtime, and thus can meet the stringent requirements of the smart park for equipment reliability and production continuity, providing reliable technical support for the intelligent manufacturing scenario.

[0042] In one embodiment, the cross-modal data alignment processing of the real-time multi-source monitoring data stream in step 102 to generate a target monitoring data set includes: Step 1021: Extract a first raw data stream from the device operating state data, where the first raw data stream includes device temperature, vibration amplitude, and current fluctuation parameters.

[0043] In the application scenario of a six-axis collaborative robot on a new energy vehicle battery module assembly line, the device temperature is specifically manifested as the surface temperature of the harmonic reducer housing, the winding temperature of the servo motor, and the contact surface temperature of the end effector. Among them, the winding temperature of the servo motor is obtained 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 state. The vibration amplitude data comes from triaxial acceleration sensors installed at 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 judge the abnormal backlash of the transmission system, and its original signal contains vibration waveforms in the 0 - 5 kHz frequency band. The current fluctuation parameter specifically refers to the effective value of the three-phase current of the servo driver and its ripple coefficient. For example, in the process of screwing the battery module, when the torque gun reaches the preset torque threshold, the instantaneous fluctuation amplitude of the B-phase current exceeds 15% of the reference value, which triggers 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 for grasping the battery cell (sampling value 78.3 °C), the RMS value of the vibration acceleration of the third-axis reducer (0.23 g), and the C-phase current ripple coefficient of the servo driver driving the rotating joint (12.8%).

[0044] Step 1022: Extract a second raw data stream from the environmental sensing data, where the second raw data stream includes environmental temperature, humidity, and dust concentration.

[0045] In the battery module assembly workshop, the environmental temperature is obtained by a distributed infrared temperature sensor array at a density of 1 monitoring point per square meter, and its value is accurate to 0.1 °C to match the sensitivity requirements of the lithium battery cell thermal expansion coefficient. The humidity data comes from a capacitive humidity sensor at the top of the assembly station. In the injection process, a strict range of 45% ± 2% RH needs to be maintained in the sealed cabin, and its original data contains a sampling sequence of 12 times per minute. The dust concentration is monitored in real time by a laser particle counter, especially paying attention to the number concentration of particulate matter with a particle size > 5 μm. When the concentration exceeds the ISO7 cleanroom standard during the module encapsulation process, the air purification system is triggered. Taking the Busbar welding station as an example, the specific parameters of the second raw data stream include: the environmental temperature of the welding area (23.5 °C), the local humidity (43.7% RH), and the concentration of particulate matter with a particle size of 5 - 10 μm in a 0.5 m³ space (182 particles / m³). These data are synchronously collected by an environmental sensor group installed at the base of the welding robot arm.

[0046] 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.

[0047] Among them, the equipment start and stop records are accurate to millisecond timestamps, recording the power state change events of the six-axis robot, the emergency stop command (event code E-Stop-002) during batch switching and its release time. The operation instruction sequence includes the G code execution record of the motion control system, such as the linear interpolation instruction G01X120.5Y-35.2Z78.3F1500 in the battery positioning stage, and its parameters are parsed as the spatial coordinates of the tool center point and the feed rate. The alarm history records are stored by severity level, including items such as servo overload alarm (code AL501) and visual positioning timeout (code AL312) and the device status snapshot when they occur. Taking the module hot pressing and shaping process as an example, the third original data stream includes: the hot press linkage start timestamp (2023-07-1514:23:17.532), the pressure control PID parameter adjustment instruction sequence (Kp=1.25, Ki=0.03, Kd=0.12), and three hydraulic overlimit warnings that occurred in the last 24 hours (code AL607).

[0048] 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, wherein the timestamp synchronization operation includes segmenting each original data stream based on a preset time window and filling in data of missing time segments.

[0049] During the data integration process of 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 5 sampling points, and the second sampling point of the third window lost due to network delay is supplemented by linear interpolation (timestamp 14:23:17.532-14:23:17.632). The ambient humidity data stream (sampling frequency 0.2Hz) is filled in each time window using the nearest neighbor interpolation method. For example, the humidity value of 43.2%RH at 14:23:17.600 is directly used in the 14:23:17.700 window. Discrete event data such as alarm records are matched to the corresponding window by timestamps, and a time series queue is established when multiple events occur in the same window. Taking the battery cell 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), scanning gun success signal (event ID: Scan-0987) and ambient dust concentration at that moment (155 pieces / m³).

[0050] Step 1025: Perform unified processing on the sampling frequency of the initially time-aligned data set to generate a target data set. The unified processing of the sampling frequency includes: downsampling the high-frequency data stream to a preset reference frequency and interpolating and complementing the low-frequency data stream to the preset reference frequency.

[0051] Among them, the preset reference frequency is set to 10 Hz to meet the response requirements of most control systems. For the original data stream of the joint vibration signal (sampling frequency 5 kHz), an anti-aliasing FIR filter is used for 100-fold downsampling, retaining the effective frequency band of 0 - 50 Hz. For example, the vibration acceleration data of the third axis outputs a valid value of 0.25 g every reference period (100 ms) after processing. The environmental temperature data stream (original sampling frequency 0.1 Hz) is upsampled to 10 Hz through cubic spline interpolation to generate a continuous temperature change curve within the period of 14:23:17.800 - 14:23:17.900. The discrete events in the operation instruction sequence are converted into pulse signals of 10 Hz. For example, a high-level mark lasting 100 ms is generated within the time window of receiving the G01 instruction. Taking the welding process as an example, the target data set after unified sampling frequency includes: the welding gun current (350 A ± 5 A) at intervals of every 100 ms, the synchronously interpolated environmental temperature (24.1 → 24.2 °C), and the safety door status signal resampled at 10 Hz.

[0052] 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.

[0053] It can be understood that the servo motor temperature data is converted from the originally collected Fahrenheit to standard Celsius. For example, the original value of 172.3 °F is converted to 78.0 °C. The vibration acceleration data is uniformly converted to the unit of g. The original voltage signal (2.34 V) is calibrated by the sensitivity formula a = V / 0.5 (V / g) to obtain 4.68 g. Data range normalization maps each parameter to the interval [0, 1]. For example, the current parameter is linearly normalized according to the rated value of the servo drive (30 A), and the actual value of 18 A is converted to 0.6. Outlier removal uses the dynamic threshold method. When the instantaneous value of the environmental dust concentration exceeds 3σ of the moving average, it is regarded as invalid data. For example, outliers > 500 / m³ appearing continuously for 3 cycles will be replaced by the median. Taking the liquid injection process as an example, the normalized target monitoring data set includes: the vacuum degree (0.85 atm → 0.85), the liquid injection rate (12.3 ml / s → 0.615), and the environmental humidity after removing outliers (45.2% RH → 0.452).

[0054] In one embodiment, performing the dynamic feature extraction operation in each edge computing node in step 103 and generating a multi-dimensional device status feature set based on the target monitoring data set includes: Step 1031: Extract a subset of device operating status data from the target monitoring data set, perform time-domain feature extraction on the subset of device operating status data, and generate device operating status features. The time-domain feature extraction includes mean calculation, variance calculation, and peak detection.

[0055] In the battery cell handling process, calculate the 10-second moving average of the vacuum suction cup servo motor current (14.2 A) to characterize the steady-state load level, and the variance (0.45 A²) reflects the torque fluctuation when grasping battery cells with different surface states. Peak detection identifies the extreme value of the vibration amplitude in the acceleration stage of the robotic arm. For example, the vibration acceleration of the fourth axis instantaneously reaches 1.2 g when turning at 90°. For the module pressing process, extract the peak-to-valley difference of the pressure sensor data (82.3 kN → 75.1 kN) as the backlash accumulation index. Taking the welding station as an example, the device operating status features include: the 30-second average of the laser power (980 W), the variance of the cooling water flow rate (0.12 L / min²), and the current peak value (1050 A) at the moment when the welding spot is completed.

[0056] Step 1032: Extract a subset of environmental sensing data from the target monitoring data set, perform environmental correlation analysis on the subset of environmental sensing data, and generate environmental correlation features. The environmental correlation analysis includes calculating the difference between the device temperature and the environmental temperature, and extracting the influence coefficient of humidity on the device vibration.

[0057] Exemplarily, calculate the time-varying difference curve between the servo motor winding temperature and the environmental temperature. When the difference exceeds 15 °C, trigger a heat dissipation anomaly warning. For example, the difference reaches 17.3 °C during continuous pressing operations. Establish an association model between humidity and the RMS value of joint vibration through multiple regression analysis, and obtain an influence coefficient that for every 10%RH increase in humidity, the vibration of the fifth axis increases by 0.05 g. In the liquid injection process, analyze the sensitivity coefficient of the environmental temperature fluctuation to the energy consumption of the vacuum pump as 0.38 kW / °C. Taking the cleaning station as an example, the environmental correlation features include: the temperature difference between the cleaning agent temperature and the environment (8.7 °C), the transfer coefficient of the air flow rate to the nozzle pressure (0.12 bar / (m / s)), and the attenuation factor of humidity on the static electricity elimination efficiency (-0.35% / RH).

[0058] Step 1033: Extract a subset of device operation log data from the target monitoring data set, perform abnormal fluctuation detection on the subset of device operation log data, and generate data abnormal fluctuation features. The abnormal fluctuation detection includes detecting sudden changes in the operation instruction frequency and associating and matching the alarm records with the operation parameters.

[0059] For example, monitor the moving standard deviation of the G-code instruction sending frequency, and mark it as an abnormal scheduling event when the standard deviation of the instruction frequency exceeds 3 times per second within 1 minute. Analyze the frequency of coordinate correction instructions before the occurrence of the visual positioning failure alarm (code AL312) through correlation analysis, and establish a correlation rule that when the correction instruction > 5 times per second, the alarm probability increases to 82%. In the module testing process, it is detected that the interval time between three consecutive cell internal resistance test instructions suddenly changes from the standard 5 seconds to 3.2 seconds, triggering the characteristic of abnormal fluctuation of the production beat. Taking the assembly line changeover stage as an example, the characteristics of abnormal data fluctuation include: the coefficient of variation of the fixture replacement instruction frequency (0.37), the correlation coefficient between the vacuum break alarm and the sucker wear degree (0.79), and the sudden change amount of the temperature gradient before and after the program restart event (ΔT = 4.2 °C).

[0060] Step 1034: Combine the device operation state characteristics, the environmental correlation characteristics, and the data abnormal fluctuation characteristics in chronological order to generate the multi-dimensional device state characteristic set.

[0061] At the final inspection station of the battery module, in the feature combination process, the average welding temperature (956 W), the environmental static electricity influence factor (0.67), and the code scanning failure fluctuation index (0.83) are aligned according to the millisecond-level time stamp to form a time series containing a 28-dimensional feature vector. Each feature vector corresponds to a complete state description of a specific process step. For example, the feature set at the moment when the cell stacking is completed includes: the variance of the grasping force (0.12 kN²), the environmental temperature difference (6.4 °C), and the number of retry times of the positioning instruction (2 times). The feature set establishes a cross-dimensional correlation relationship through time axis matching. When it is found that the phase difference between the peak value of the servo current and the environmental humidity fluctuation reaches 200 ms, it is marked as a potential condensate interference event. The finally generated multi-dimensional device state characteristic set provides a standardized input for the subsequent anomaly detection model, ensuring strict consistency of each feature dimension in the time domain and the spatial domain.

[0062] In an optional embodiment, the trained anomaly detection integration model is generated through the following steps: Step 201: Obtain 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 with marked abnormal types.

[0063] In the application scenario of a six-axis collaborative robot in a new energy vehicle battery module assembly line, the historical multi-source monitoring data set covers the production cycle data of 12 consecutive months. Among them, the normal equipment data samples contain the servo motor steady-state operation records of more than 5,000 hours, specifically manifested as the winding temperature fluctuation range ≤ ±3°C, and the joint vibration acceleration RMS value is stable in the range of 0.15 - 0.25g. The abnormal equipment data samples with marked abnormal types are classified and marked for historical fault events through an expert system, including 7 known faults such as harmonic reducer wear (type code F01), encoder signal drift (F02), etc. Each abnormal sample contains multi-modal data slices from 30 minutes before the fault occurrence to the shutdown moment. For example, the F01 class sample contains the temperature gradient data of the third-axis reducer temperature rising suddenly from 65°C to 89°C, the frequency domain feature that the energy of the 2.5kHz sideband in the vibration spectrum rises to 3.2 times the reference value, and the axial clearance over-limit alarm log (code AL704) at the corresponding moment.

[0064] Step 202: Perform cross-modal data alignment processing on the historical multi-source monitoring data set to generate a historical alignment data set, and extract a historical multi-dimensional device state feature set based on the historical alignment data set.

[0065] Among them, the cross-modal data alignment processing adopts the same time series synchronization protocol as the online monitoring system to align the servo motor current waveform data (sampling frequency 50kHz) with the trigger signal (millisecond-level event) of the visual positioning system to a unified time axis. For example, in the case of an abnormal cell positioning event, the time deviation between the moment of sudden current increase (14:23:17.532) and the time when the visual system sends a coordinate correction instruction is calibrated to within ±5ms. The extraction strategy of the historical multi-dimensional device state feature set is consistent with the online feature engineering. The features extracted for the F02 class abnormal samples include the average value of the encoder original pulse count (1423 counts / cycle), the influence coefficient of the ambient temperature on the counting error (0.08 counts / °C), and the operation frequency of 10 consecutive zero resets (3 times / minute).

[0066] Step 203: Divide the historical multi-dimensional device state feature set into a training set and a validation set, and construct an initial abnormal detection model set. The initial abnormal detection model set includes a first detection model based on the isolation forest, a second detection model based on the time series convolutional network, and a feature fusion model based on the attention mechanism.

[0067] For example, the training set contains 80% of historical data samples, which are stratified sampled according to the device serial number to ensure the balanced distribution of various types of faults. For example, in the training set, the proportion of samples of class F01 is 12.7% and the proportion of samples of class F02 is 9.3%. The validation set retains the remaining 20% of the data and includes 6 fault subtypes that have not participated in the training for testing the generalization ability of the model. The first detection model based on the isolation forest configures the tree depth parameter to 64, and uses the random subspace method to select 30% of the feature dimensions for splitting node calculation. The architecture of the second detection model of the temporal convolutional network contains 5 dilated convolutional layers, and the dilation coefficient sequence is [1, 2, 4, 8, 16], and the number of output channels of each layer is set to 128. The feature fusion model uses the multi-head attention mechanism and configures 4 independent attention heads to capture the feature associations of different fault modes.

[0068] Step 204: Train the initial anomaly detection model set through a cascaded training strategy to obtain a target anomaly detection model set. The cascaded training strategy includes: performing unsupervised training on the first detection model using the training set, using the output probability of the first detection model as the input feature of the second detection model, and jointly performing supervised training on the second detection model and the feature fusion model.

[0069] During the unsupervised training process of the first detection model, the isolation forest algorithm automatically identifies 5.7% of the outlier samples in the training set, and 83% of these samples are manually reviewed and confirmed as unlabeled potential fault events. The input feature dimension of the second detection model is extended to the original feature number + anomaly probability. For example, in the task of detecting the anomaly of the cell clamping force, the input feature increases from 28 dimensions to 29 dimensions (the added anomaly probability of class F01 output by the isolation forest). In the joint training stage, an alternating optimization strategy is adopted, and the output of the temporal convolutional network and the attention weights of the feature fusion model are updated synchronously through backpropagation. In the detection of class F03 faults (end effector vacuum leakage), the attention weight of the model to the suction cup pressure fluctuation feature is increased to 0.63.

[0070] Step 205: Perform performance evaluation on the target anomaly detection model set based on the validation set. When the anomaly type recognition accuracy exceeds the preset threshold, use the target anomaly detection model set as the anomaly detection integrated model that has completed training and deploy it to the edge computing node.

[0071] It can be understood that the performance evaluation metrics include multi-class accuracy (≥92%), unknown fault detection rate (≥85%), and false alarm rate (≤1.2 times / hour). In the verification test of the module packaging process, the recognition accuracy of the model for F05 type faults (thermal compression die positioning offset) reached 94.7%, and the detection rate for new fault types (laser welding head lens contamination) was 81.3%. When deployed to the edge computing node, the model inference latency was strictly controlled within 50 ms to ensure that the anomaly determination was completed within the robotic arm motion control cycle (100 ms). For example, at the Busbar welding station, the total processing time of the model from feature input to output of the maintenance instruction was 43 ms, meeting the real-time requirement.

[0072] In an optional embodiment, the generating the device anomaly probability distribution through the multi-level feature fusion strategy and determining the anomaly type and anomaly confidence level of the target device according to the device anomaly probability distribution in step 104 includes: Step 1041: Input the multi-dimensional device state feature set into the first detection model to generate a first anomaly probability distribution, where the first anomaly probability distribution includes the overall device anomaly probability and the local component anomaly probability.

[0073] In the battery cell stacking process, the first detection model calculates that the overall device anomaly probability is 0.87 (threshold 0.75) based on the isolation forest algorithm, and the local anomaly probability of the third joint component reaches 0.93. The local component anomaly probability is obtained by decomposing the feature contribution degree. For example, it is identified that the contribution degree of the 1.8 kHz component of the third axis vibration spectrum to the anomaly probability is 58%, and the contribution degree of the temperature gradient feature is 32%. In the servo overload warning scenario, the model outputs a local anomaly probability of 0.91 for the B-phase winding, and at the same time, the overall anomaly probability is 0.79, triggering a secondary warning response.

[0074] Step 1042: Input 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, where the second anomaly probability distribution includes the time series anomaly pattern and the prediction of the anomaly duration.

[0075] Among them, the time series convolutional network model captures the periodic fluctuation anomaly pattern of the servo current, detects 3 current peaks exceeding the threshold within 10 consecutive sampling periods (1 second), and predicts that the anomaly duration is in the range of 8 - 12 seconds. For the visual positioning drift problem, the model identifies the exponential growth pattern of the coordinate correction instruction sending frequency (R² = 0.91), and predicts that the system will reach the fault tolerance limit after 23 seconds. In the thermal compression process, the second anomaly probability distribution shows that the autocorrelation coefficient of the die temperature time series data drops to 0.12 (normal benchmark > 0.6), and it is predicted that the abnormal state will continue until the end of the current production batch.

[0076] Step 1043: Input the first abnormal probability distribution and the second abnormal probability distribution into the feature fusion model, and generate a fused abnormal probability distribution through an attention weight allocation mechanism.

[0077] For example, the multi-head attention mechanism assigns a weight of 0.55 to the component abnormal probability output by the first detection model and a weight of 0.45 to the time-series abnormal pattern. In the example of vacuum suction cup leakage, the feature fusion model detects a strong correlation between the local abnormal probability (0.88) and the duration prediction (>30 seconds), and raises the fused probability to 0.91. For the encoder signal drift problem, the model finds that the phase mutation feature in the time-series pattern is more diagnostically valuable than the outlier score of the isolation forest, so the attention weight of the time-series branch is raised to 0.68.

[0078] Step 1044: Perform threshold segmentation processing on the fused abnormal probability distribution to determine the abnormal type and abnormal confidence level of the target device. The threshold segmentation processing includes dynamic threshold adjustment and probability density clustering.

[0079] Among them, the dynamic threshold is adaptively adjusted according to the device operation stage. The alarm threshold is raised from 0.75 to 0.85 during the startup stage to suppress transient interference. The probability density clustering algorithm identifies two abnormal probability aggregation regions: the 0.72 - 0.78 interval corresponds to the F02 type of fault (confidence level 68%), and the 0.88 - 0.92 interval corresponds to the F01 type of fault (confidence level 93%). In the event of abnormal cell scanning code, the fused probability of 0.79 is classified as the F07 type of fault (visual system contamination), and the confidence level is calculated as (0.79 - 0.65) / (0.95 - 0.65) = 46.7%.

[0080] In an optional embodiment, step 105 of generating a device maintenance instruction set based on the abnormal type and the abnormal confidence level includes: Step 1051: Match the preset maintenance strategy library according to the abnormal type, and obtain the maintenance operation template and priority rule corresponding to the abnormal type.

[0081] Among them, the maintenance policy library establishes the mapping relationship between fault codes and standard operation procedures. For example, faults of type F01 are mapped to the "Preventive Replacement of Harmonic Reducers" template, which includes disassembly steps (12 items), cleaning specifications (5 items), and torque setting values (35 Nm ± 5%). The priority rule library defines the response levels for type F faults. Among them, faults of types F01 and F04 are set to the highest priority (response time limit < 15 minutes), and faults of type F07 are set to medium priority (response time limit < 2 hours). In the event of abnormal Busbar welding, the system calls the "Optical Path Calibration of Laser Welding Head" template, which includes 23 parameters such as the collimator adjustment tolerance (±0.02 mm) and the protection lens replacement cycle (4000 times).

[0082] Step 1052: Dynamically adjust the parameters in the maintenance operation template based on the abnormal confidence level to generate maintenance operation parameters. The dynamic adjustment includes that the higher the confidence level, the shorter the maintenance interval and the more the maintenance resource allocation.

[0083] When the confidence level of type F01 faults exceeds 90%, the lubrication maintenance interval of the harmonic reducer is shortened from the standard 2000 hours to 1500 hours, and the double amount of lubricating grease is automatically allocated (increased from 5 ml to 10 ml). For type F07 faults with a confidence level of 75%, the visual lens cleaning frequency is increased from once per shift to twice per shift, and the standby lighting source is activated. In the example of servo overheat warning, a confidence level of 85% triggers forced cooling measures: increase the cooling fan speed to 120% of the rated value and extend the standby cooling time from 5 minutes to 8 minutes.

[0084] Step 1053: Sort multiple abnormal devices according to the priority rules to generate a maintenance priority queue, and bind the maintenance priority queue to the maintenance operation parameters to generate the set of device maintenance instructions.

[0085] For example, the sorting algorithm comprehensively considers the fault level (F01 > F04 > F07), the confidence level weight (70%), and the production impact coefficient (30%). For example, when F01 (confidence level 92%) and F04 (confidence level 88%) occur simultaneously, the calculated comprehensive priority score is: 92×0.7 + 100×0.3 = 94.4 vs 88×0.7 + 95×0.3 = 89.9, determining that F01 is to be processed first. The generated set of device maintenance instructions includes: work order number (MT20230715 - 001), target device code (RB06 - J3), required spare parts list (1 set of harmonic reducer GHD - 203), and safety operation constraints (stand still for 5 minutes after power off). When multiple low - priority faults are detected during the change - over phase, the system automatically schedules the maintenance operations to be executed during the production gap to minimize the downtime loss.

[0086] In a preferred embodiment, the method further includes: Step 301: Real-time monitor the execution status of the device management terminal for the device maintenance instruction set. When detecting a delay or failure in the maintenance operation, 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 abnormal detection integrated model based on the historical multi-dimensional device status feature set to generate an updated abnormal detection integrated model; Push the updated abnormal detection integrated model to the edge computing node for model replacement, and regenerate the device maintenance instruction set.

[0087] In the maintenance scenario of a six-axis collaborative robot on a new energy vehicle battery module assembly line, when it is detected that the work order for replacing the harmonic reducer of the third-axis servo motor (work order number MT20230715-003) has timed out and not been completed, the system extracts the historical multi-dimensional device status feature set of this device in the past 72 hours, including the temperature gradient feature (0.78 °C / min), vibration energy entropy (1.23), and operation instruction retry frequency (4 times / hour). The incremental training process uses an online learning algorithm. On the basis of retaining 90% of the classification layer weights of the original model, 5 new feature dimensions are added for the F01 type of fault (wear of the harmonic reducer). After 200 iterations, the detection accuracy of the model for the same type of fault is increased from 83% to 91%. The updated model is pushed to the assembly line edge gateway through a secure transmission protocol to replace the original model file (version number v2.1.7 → v2.1.8), and a maintenance instruction including an adjustment of the lubricant filling amount (+15%) is regenerated.

[0088] In a preferred embodiment, the incremental training of the abnormal detection integrated model based on the historical multi-dimensional device status feature set to generate an updated abnormal detection integrated model includes: 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.

[0089] It can be understood that the preset time window is set from 30 minutes before the fault occurs to the current moment. For the vacuum suction cup leakage event (equipment code VP-06), the extracted feature segments include: the average suction cup pressure (-82 kPa → -65 kPa), the environmental humidity influence coefficient (0.34 → 0.51), and the vacuum break alarm frequency (0 → 3 times / minute). Among them, the equipment operation state feature segments are accurate to 100 ms intervals and contain 12 temperature-pressure correlation feature vectors; the environmental correlation feature segments are stored by work station area, and record the temperature and humidity gradients (ΔT = 2.3 °C / m, ΔRH = 4.1% / m) at the liquid injection work station when the leakage occurs.

[0090] Step 402: Perform event annotation processing on the equipment operation state feature segments, environmental correlation feature segments, and data abnormal fluctuation feature segments to generate an incremental training data set. The event annotation processing includes marking the feature segments in the continuous time period before the occurrence time of the delay or failure event as the abnormal category to be optimized.

[0091] In the example of the failure of the visual positioning system calibration, the feature data 5 minutes before the event occurs is marked as class F07 (visual pollution fault), specifically including: the standard deviation of the positioning error (0.12 mm → 0.35 mm), the environmental light intensity (850 lux → 210 lux), and the lens cleaning instruction interval (1200 seconds → not executed). The annotation process uses the sliding window method to generate 30 labeled samples with a step size of 10 seconds, among which the positive samples (precursors of faults) account for 67% and the negative samples (normal working conditions) account for 33%.

[0092] Step 403: Load the network parameters of the feature extraction layer from the anomaly detection integration 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 in a state to be updated.

[0093] Among them, the feature extraction layer retains the 5-layer dilated convolution structure (dilation coefficients [1, 2, 4, 8, 16]) of the temporal convolutional network and the head configuration (4 heads) of the attention mechanism, and the weight matrix is frozen in a read-only mode. The fully connected network parameters (dimension 28 → 7) of the classification layer are initialized to 85% of the original value, the learning rate is set to 0.001, and the momentum term is adjusted to 0.92. In the vacuum leakage detection task, the 128-dimensional embedding vector output by the feature extraction layer remains fixed, and the classification layer weights are set with an initial weight 3 times that of other features for the suction cup pressure feature (dimension 19).

[0094] Step 404: Input the incremental training data set into the classification layer to perform backpropagation optimization operations, and adjust the numerical distribution of the fully connected weight matrix through the gradient descent algorithm until the decrease amplitude of the classification loss function on the validation subset is less than the preset convergence threshold.

[0095] For example, in the optimization process, mini-batch gradient descent (batch_size = 32) is adopted. The cross-entropy loss function drops from 1.23 to 0.87 within 50 epochs, and the accuracy of the validation set increases by 12%. For faults of type F05 (mold positioning offset), the weight coefficient of the classification layer for the temperature-displacement coupling features is adjusted from 0.15 to 0.28. The convergence threshold is set such that the loss decrease amplitude ≤ 0.5% for 10 consecutive epochs, and the training is terminated when the loss value fluctuation range in the 47th - 50th epochs < 0.03.

[0096] Step 405: Recombine the optimized fully-connected weight matrix with the locked feature extraction layer parameters to generate an updated anomaly detection integrated model, and perform performance verification of the updated anomaly detection integrated model on the historical validation set. If the difference in the anomaly type recognition accuracy relative to that before the update is within the preset tolerance interval, confirm that the model update is completed.

[0097] Exemplarily, the combined model is tested on a historical validation set containing 1200 samples. The recall rate of faults of type F01 increases from 78% to 85%, and the unknown fault detection rate remains above 82%. The tolerance interval is set to ±3%. When the accuracy change of type F03 (vacuum leak) is +4.2%, an artificial review process is triggered, and after confirmation, the model version is upgraded (v2.1.8 → v2.1.9).

[0098] In an alternative embodiment, the method further includes: Step 501: Display the device anomaly probability distribution and the device maintenance instruction set in the visualization monitoring interface of the smart park.

[0099] It can be understood that the visualization interface adopts a hierarchical rendering technique, and the real-time status of the six-axis robot is superimposed and displayed on the workshop digital twin model: devices with an anomaly probability exceeding 75% are highlighted with a red pulse border, and devices under maintenance are displayed with a yellow progress bar. For example, the anomaly probability of type F01 for the welding robot RB03 (88%) is rendered with a red halo with a diameter of 30 cm, and its maintenance instruction (replace the harmonic reducer) is displayed in the form of a floating card showing the work order number, remaining time (23 minutes), and spare part inventory status (2 sets of GHD - 203 remaining).

[0100] Step 502: In response to the user's interaction operation on the selected device, dynamically display the multi-dimensional device status feature set, the anomaly type determination basis, and the maintenance operation execution progress of the selected device.

[0101] When the battery stack robot RB06 is clicked, the right panel of the interface synchronously displays: the temperature-vibration correlation matrix (28×28 heat map), the top 3 feature contribution degrees of the current abnormal type (F02, confidence level 89%) (encoder error 42%, environmental static electricity 35%, instruction frequency 23%), and the lubrication maintenance progress (63% completed). The perspective of the 3D model automatically focuses on the third joint component, and a CT scan comparison diagram of the worn parts inside the reducer is displayed with a semi-transparent effect.

[0102] Step 503: When it is detected that the user manually corrects the abnormal type or maintenance priority, record the correction record and trigger incremental training of the abnormal detection integration model.

[0103] For example, after the operation and maintenance personnel correct the abnormal type of RB12 from F04 (electrical fault) determined by the system to F01 (mechanical wear), the correction record (operator ID, timestamp, correction basis) is stored in the audit database, and at the same time, fine-tuning training of the feature extraction layer is triggered: add a confusion matrix correction item of F01→F04 to the original classification matrix. After 15 minutes of online training, the correction rate of the model for misjudgments of the same type is increased to 93%.

[0104] In an alternative embodiment, the dynamically displaying the multi-dimensional device state feature set of the device in step 502 includes: Step 5021: Extract the numerical sequence of the temperature change rate, the numerical sequence of the vibration spectrum energy, and the current fluctuation parameter sequence corresponding to the device operation state feature from the multi-dimensional device state feature set, generate a temperature-vibration-current superposition curve graph based on the unified time axis, and render it to the visualization monitoring interface.

[0105] Taking the module pressing device PP05 as an example, the horizontal axis of the curve graph is the time scale accurate to milliseconds (14:23:17.532 - 14:23:19.632), and the vertical axis respectively shows: the temperature change rate of the mold surface (-0.8℃ / s → +1.2℃ / s), the vibration energy of the hydraulic cylinder (152dB → 168dB), and the servo current fluctuation coefficient (8% → 23%). The occurrence time of the abnormal event (14:23:18.921) is marked with a vertical red line, and the corresponding current peak value (142% of the rated value) is highlighted as a pulse waveform in the curve.

[0106] Step 5022: Extract the environmental temperature distribution matrix, the humidity influence coefficient matrix, and the dust concentration correlation matrix corresponding to the environmental correlation feature from the multi-dimensional device state feature set, generate a three-dimensional heat map based on the mapping of the device physical position coordinates, and superimpose and mark the abnormal correlation areas exceeding the preset threshold in the three-dimensional heat map.

[0107] 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 where the humidity influence coefficient exceeds 0.6 (the northwest corner workstation) is surrounded by orange contour lines; the grids with a dust concentration correlation value > 0.75 (the conveyor belt interface) are displayed as red cubes, indicating that enhanced sealing treatment is required.

[0108] Step 5023: Extract the operation instruction frequency mutation time point, alarm record timestamp, and abnormal fluctuation period of the operating parameters 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 text fragments of the device operation logs at the corresponding moments on the event timeline.

[0109] For example, the event timeline of the Busbar welding robot WT07 shows: G01X120Y-35 command sent at 14:23:17.532; AL312 vision timeout alarm detected at 14:23:17.892; current fluctuation entered the abnormal range at 14:23:18.215. When clicking on the AL312 alarm mark, the associated log shows "The visual positioning reference point is lost, retried 3 times in total, it is recommended to clean the reflector".

[0110] Step 5024: Embed the temperature-vibration-current superposition curve graph, 3D thermal map, and interactive event timeline into the dynamic display panel of the visualization monitoring interface according to the preset layout. When it is detected that the user clicks on the abnormal association area or the marked point on the event timeline, trigger the highlighted display of the corresponding device operation log text fragment and the curve magnification display of the abnormal fluctuation period.

[0111] Among them, the interface layout adopts a three-column design: the left column (30% width) displays the device tree list; the middle column (50%) renders the 3D workshop model and thermal map; the right column (20%) displays the curve graph and timeline. When the user clicks on the high-temperature area of the hydraulic station (the red block in the 3D thermal map), the associated oil temperature curve segment (14:23:17.200 - 14:23:18.500) is automatically magnified to 200% scale, and at the same time, the log panel highlights "Hydraulic oil cooler fan failure (code AL455)".

[0112] In an optional but non-limiting embodiment, the method further includes: Receiving in real time the maintenance operation execution result returned by the device management terminal, and extracting the actual maintenance parameters and maintenance completion timestamp corresponding to the abnormal device identifier from the maintenance operation execution result; Based on the comparison of the actual maintenance parameters and the maintenance operation parameters, generating a maintenance effect evaluation index, and the maintenance effect evaluation index includes a parameter deviation rate, an operation delay duration, and an abnormal recurrence detection result; When the maintenance effect evaluation index exceeds the preset tolerance threshold, an optimization instruction for the anomaly detection integration model is triggered, and an incremental training sample set is generated based on the actual maintenance parameters and the anomaly recurrence detection result; Use the incremental training sample set to perform online fine-tuning on the feature fusion model in the anomaly detection integration model, update the weight coefficients in the attention weight allocation mechanism, and redeploy the fine-tuned feature fusion model to the corresponding edge computing node.

[0113] During the equipment maintenance execution phase, the system receives the maintenance operation execution result message returned by the equipment management terminal in real time through the Internet of Things protocol. The message contains key information such as the abnormal equipment identifier, actual maintenance parameters, and maintenance completion timestamp. The maintenance effect evaluation module compares the actual maintenance parameters with the original instruction parameters item by item, and calculates the parameter deviation rate and operation delay duration through a preset algorithm. For example, when 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 an anomaly recurrence is detected (such as the equipment vibration data not returning to the normal range), the system extracts multi-dimensional feature data within a specific time window before and after maintenance from the historical database, including temperature gradient, current fluctuation parameters, and alarm frequency in the operation log. These data are cleaned and labeled to form an incremental training sample set for optimizing 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 only dynamically adjusts 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 vibration spectrum energy distribution). The updated model is compressed and packaged, and is deployed through the rolling update mechanism of the edge node to ensure that the service interruption time is controlled within milliseconds. At the same time, the system regenerates the maintenance instruction based on the latest model, optimizes the parameter range, and adds an anomaly re-inspection rule to form a closed-loop optimization process from execution feedback to model iteration. The entire process realizes the dynamic adaptation of the maintenance strategy and the equipment state while ensuring real-time performance.

[0114] In an optional but non-limiting embodiment, the method further includes: Obtain the current maintenance resource status data of the smart park, where the current maintenance resource status data includes the number of available maintenance personnel, spare part inventory distribution, and equipment downtime window; Sort the equipment maintenance instruction set according to the maintenance priority from high to low, and dynamically adjust the resource allocation of the maintenance operation parameters in combination with the current maintenance resource status data to generate adjusted maintenance operation parameters; Generate a subset of maintenance instructions under resource constraints based on the adjusted maintenance operation parameters, where the subset of maintenance instructions under resource constraints includes the optimized result of spare part call path, the scheduling route of maintenance personnel, and the identification of parallel maintenance equipment; Send the subset of maintenance instructions under resource constraints to the device management terminal, and update the spare part inventory distribution and the location information of maintenance personnel in the current maintenance resource status data in real time.

[0115] For example, the system obtains the status of maintenance resources in the smart park in real time through a data interface, including the locations of on-duty personnel, the distribution of spare part inventories, and the production line shutdown plan. The maintenance scheduling engine combines the device priority queue and resource constraint conditions, and uses intelligent algorithms to generate an optimal scheduling plan. For high-priority faulty devices, the system preferentially allocates the nearest available professional maintenance team and plans the shortest path for spare part transportation. For example, for the task of replacing the harmonic reducer of a key device, which requires the cooperation of two technicians, the system screens the maintenance team with the required qualifications and the shortest distance from the personnel location data, and at the same time calculates the optimal transportation route from the central warehouse to the target work station. During the resource allocation process, the system monitors the change of spare part inventory in real time, and automatically triggers a replenishment warning when the inventory level is lower than the safety threshold. For maintenance tasks that need to be processed in parallel, the system establishes a dependency relationship model between devices to avoid resource conflicts. For example, when two work stations share the same type of spare parts, the system dynamically adjusts the outbound order to ensure that high-priority tasks are completed first. The generated subset of maintenance instructions includes a detailed resource call list, a personnel division plan, and time node requirements, and visualizes the real-time location of transportation vehicles and the maintenance progress through a 3D map. When unexpected situations (such as temporary absence of personnel or transportation delays) cause the original plan to be blocked, the system immediately activates the emergency plan, recalculates the alternative plan, and allocates spare resources to ensure that key maintenance tasks are completed on time. All scheduling instructions are synchronized to the terminal device in real time through the industrial communication protocol, and the maintenance kanban dynamically updates the task status, realizing the transparent management of the entire process of resource scheduling.

[0116] In an optional but non-limiting embodiment, the method further includes: Before the device management terminal performs a maintenance operation, generate predicted data on the operating status of the device after simulated maintenance based on the maintenance operation parameters, where the predicted data on the operating status of the device after simulated maintenance includes the simulated value of the temperature change trend, the simulated range of vibration amplitude, and the predicted interval of current stability; Match the predicted data on the operating status of the device after simulated maintenance with the historical normal device data samples to generate a maintenance operation risk level, where the maintenance operation risk level includes the probability of device overload, the sensitivity to environmental interference, and the possibility of secondary anomaly triggering; When the risk level of the maintenance operation exceeds the preset safety threshold, the maintenance interval duration or resource allocation ratio in the maintenance operation parameters is corrected to generate risk-mitigated maintenance operation parameters; Rebind the risk-mitigated maintenance operation parameters to the device maintenance instruction set, and trigger the device management terminal to perform the actual maintenance operation after passing the simulation verification.

[0117] Before the maintenance operation is executed, the system constructs a virtual mapping of the device state based on digital twin technology, and predicts the change of the operating parameters after maintenance through multi-physics field simulation. The prediction model comprehensively considers multi-dimensional factors such as mechanical transmission, thermodynamic conduction, and electrical control, and generates simulated values of key indicators such as the temperature change trend, vibration amplitude range, and current stability interval. For example, after replacing a certain component, the system predicts that the peak temperature of the servo motor will drop to a reasonable range, and the vibration energy distribution tends to be stable. Subsequently, the system matches the predicted data with the historical normal operating condition database, and evaluates the effectiveness of the maintenance plan through a similarity algorithm. When potential risks (such as insufficient heat dissipation efficiency or abnormal load fluctuations) are detected, the system automatically dynamically corrects the maintenance parameters. The correction strategies include extending the maintenance cycle, adding auxiliary measures, or adjusting the operation process. For example, for the identified heat dissipation risk, the system adjusts the operation mode of the cooling fan from intermittent start to continuous operation, and increases the temperature monitoring frequency. The corrected parameters need to be verified through secondary simulation to ensure that the risk indicators are reduced below the safety threshold. The device management terminal delays the execution of the actual maintenance operation after receiving the final instruction until the digital twin system outputs a verification passed signal. This mechanism effectively avoids secondary failures caused by defects in the maintenance plan, and at the same time balances the maintenance efficiency and the device reliability requirements. All correction records and verification results are synchronously stored in the knowledge base to provide data support for subsequent maintenance strategy optimization.

[0118] In an optional but non-limiting embodiment, the method further includes: Set an effective execution time window for each maintenance operation in the device maintenance instruction set, and the effective execution time window is dynamically adjusted based on the anomaly confidence level and the device type; Real-time monitor the maintenance operation start signal and completion signal of the device management terminal within the effective execution time window. If the completion signal is not detected within the time window, mark this maintenance operation as an event of timeout and non-completion; Based on the abnormal device identifier corresponding to the timeout and non-completion event, re-acquire the real-time multi-source monitoring data stream of the device, and trigger a new round of anomaly detection to generate an updated device anomaly probability distribution; Re-evaluate the original maintenance priority and maintenance operation parameters according to the updated device anomaly probability distribution, generate a corrected maintenance instruction set corresponding to the failure time window, and send it to the device management terminal.

[0119] In specific implementation, the system sets a dynamic effective execution time window for each maintenance instruction, and the length of the time window is automatically adjusted according to the fault type, confidence level, and equipment criticality. For example, the default time window for mechanical faults with high confidence is relatively short to ensure quick response; while environmental anomalies with low risk are assigned a longer processing cycle. The time window monitoring module tracks the start and completion status of maintenance operations in real time, and identifies timeout events through timestamp comparison. When it is detected that a task is not completed within the time window, the system immediately marks the event and triggers an abnormal re-evaluation process. The system re-collects the real-time operation data of the target device, and quickly generates an updated probability distribution through the anomaly detection model. For example, due to maintenance delay, the vibration parameters of a certain device continue to deteriorate, and after recalculation, the model raises its fault probability from the initial value to a higher level. Based on the latest evaluation results, the system re-orders the maintenance priorities and optimizes the resource allocation strategy. For faulty devices upgraded to an emergency state, the system breaks through the conventional scheduling rules and enables an alternative resource channel. For example, directly call spare parts of pre-research models in the laboratory or coordinate the intervention of an expert team across departments. The corrected instruction set is generated through a priority weighting algorithm and is accompanied by detailed execution constraint conditions (such as specific environmental parameter requirements). All adjustments are synchronized to the terminal device in real time, and the global monitoring interface prompts the change of task status in a highlighted form. At the same time, the system records the cause analysis and processing log of the timeout event, which is used to optimize the subsequent time window calculation model and gradually improve the accuracy of the maintenance plan.

[0120] In summary, through the collaborative optimization of cross-modal data alignment and edge computing dynamic feature extraction, the embodiment of the present invention constructs an end-to-end analysis architecture for multi-source heterogeneous monitoring data, achieving a double improvement in the real-time performance and accuracy of device anomaly detection. Through a multi-level data integration mechanism of timestamp synchronization, frequency unification, and format standardization, the problem of fragmentation in the temporal correlation and semantic consistency of multi-source data streams is effectively overcome, enabling a dynamic coupling analysis among the device operation status, environmental parameters, and operation logs; the edge node performs dynamic feature extraction based on the real-time data stream, not only capturing the operation characteristics of the device itself, but also integrating the environmental correlation effects and abnormal fluctuation patterns, constructing a device health portrait in a multi-dimensional feature space; the anomaly detection integrated model maps the non-linear correlation between heterogeneous features into a probabilistic anomaly distribution through a multi-level feature fusion strategy, breaking through the limitation of the single detection model in representing complex anomaly patterns. The finally generated maintenance instruction set realizes closed-loop decision-making for anomaly location, priority determination, and operation parameters, improving the automation degree and decision-making reliability of the device maintenance response in the smart park.

[0121] See Figure 2 As shown in the figure, 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: Processor 201; Storage device 202, on which a computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 implements any one of the device anomaly detection methods based on multi-source heterogeneous data.

[0122] On the above basis, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0123] It should be noted that the various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same and similar parts between the various embodiments, reference can be made to each other. 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 reference can be made to the description in the method part for the relevant parts.

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 of production equipment in the smart park, wherein the real-time multi-source monitoring data streams include equipment operation 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 association 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 according to the device anomaly probability distribution; Based on the exception type and the exception confidence, a device maintenance instruction set is generated, the device maintenance instruction set includes an abnormal device identification, 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.

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 original data stream from the device operation status data, wherein the first original data stream includes 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, wherein the third original data stream includes a device start and stop record, an operation instruction sequence, and an alarm history record; 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 in data of 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 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; 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, 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 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 operation instruction frequency mutation detection, and association matching between alarm records and operating parameters; The equipment operation status characteristics, the environment association characteristics and the data abnormal fluctuation characteristics are combined in chronological order to generate the multi-dimensional equipment status feature set.

4. The method according to claim 1, characterized in that: The trained anomaly detection integrated 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 data set to generate a historical aligned data set, and extracting a historical multi-dimensional device state feature set based on the historical aligned data set; 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; The target anomaly detection model set is obtained by training the initial anomaly detection model set through a cascade training strategy, wherein the cascade training strategy includes: 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 the 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 an edge computing node.

5. The method according to claim 4, characterized in that The generating of the device abnormality probability distribution by the multi-level feature fusion strategy, and determining the abnormality type and abnormality confidence of the target device according to the device abnormality probability distribution, includes: Inputting the multi-dimensional device state feature set into the first detection model to generate a first abnormality probability distribution, wherein the first abnormality probability distribution includes the overall abnormality probability of the device and the abnormality probability of a local component; Inputting the multi-dimensional device state 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; Inputting the first abnormal probability distribution and the second abnormal probability distribution into the feature fusion model, and generating a fused abnormal 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 a set of equipment maintenance instructions based on the exception type and the exception confidence level includes: According to the abnormality type, a preset maintenance strategy library is matched to obtain a maintenance operation template and a priority rule corresponding to the abnormality type; Dynamically adjusting the 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 is higher; 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 status 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 incrementally training 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 the abnormal device identification corresponding to the delay or failure event, and extracting the device operation status feature segment, environment-related 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 identification; Performing event labeling processing on the equipment 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 abnormal categories 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, and 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 back-propagation 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 updated anomaly detection integrated model is subjected to performance verification on the historical verification set. If the difference in anomaly type recognition accuracy relative 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 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; 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 multi-dimensional device status feature set of the device is dynamically displayed, including: 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 based on a unified time axis and render 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 equipment status feature set, generating a three-dimensional heat map based on the equipment physical position coordinate mapping, and superimposing and marking abnormal correlation areas exceeding a preset threshold in the three-dimensional heat map; 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 time 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