Wiring terminal assembly equipment fault early warning method and system based on machine learning

Abstract

The invention discloses a machine learning-based wiring terminal assembly equipment fault early warning method and system. The method comprises the following steps: collecting process parameter data, state data and image data of each process of a wiring terminal assembly process through a sensor and a computer vision device, performing process marking, and performing data preprocessing based on an application mode of the computer vision device adapted to an equipment life cycle; extracting feature vectors of each process to construct a process state model; the process state model comprises feature vector distribution templates of a normal state and a fault state; identifying fault types through a pre-constructed machine learning classification model and establishing a mapping relation between faults and processes; monitoring process parameter abnormity in real time through a multi-layer analysis model, determining a fault responsibility process, and outputting a fault positioning result and early warning information; according to the invention, accurate positioning of the fault process of the wiring terminal assembly equipment can be realized, the accuracy and efficiency of fault diagnosis are improved, and the downtime and the maintenance cost are reduced.

Description

Technical Field

[0001] This application relates to the field of fault early warning technology, specifically to a fault early warning method and system for terminal assembly equipment based on machine learning. Background Technology

[0002] As a key component in electrical connections, the assembly quality of terminal blocks directly affects the stability and safety of electrical systems. With the development of industrial automation, terminal block assembly equipment is increasingly used on production lines, improving production efficiency and product quality. However, the stable operation of this equipment is crucial for ensuring the continuity of the production line and product quality. Therefore, how to effectively maintain terminal block assembly equipment and promptly identify and resolve potential faults has become an urgent problem to be solved in the industry.

[0003] In existing technologies, the maintenance of terminal block assembly equipment mainly relies on two modes. One is periodic maintenance, which involves comprehensive inspection and maintenance of the equipment at preset time intervals, including cleaning, lubrication, and tightening, to ensure the normal operation of the equipment. The other is post-fault repair, where maintenance personnel diagnose and repair the equipment based on the symptoms when obvious faults occur. Some production lines have introduced equipment monitoring systems. Traditional monitoring methods mainly rely on direct sensor data, such as temperature, vibration, and current, to determine the operating status of the equipment. These systems typically use fixed thresholds to determine whether a fault has occurred. Fault diagnosis largely depends on the experience of maintenance personnel, who determine the cause of the fault by observing the symptoms and inspecting equipment components.

[0004] However, these existing technologies have many shortcomings. The model of periodic maintenance and post-failure repair is a reactive approach, lacking real-time monitoring of equipment status. Traditional monitoring methods ignore equipment status information contained in product quality data and rely on fixed thresholds, failing to capture the gradual degradation process of equipment performance. Data analysis methods are simplistic, unable to identify complex faults, lacking adaptability, and struggling to cope with changes in production conditions and equipment aging. Furthermore, existing systems suffer from "information silos," with quality inspection systems and equipment maintenance systems operating independently, hindering data fusion and collaborative analysis. Fault diagnosis relies on the experience of maintenance personnel, lacking systematic analysis supported by data, resulting in low diagnostic efficiency and insufficient accuracy. Summary of the Invention

[0005] To accurately locate faulty processes in terminal block assembly equipment, improve the accuracy and efficiency of fault diagnosis, and reduce downtime and maintenance costs, this application provides a fault early warning method and system for terminal block assembly equipment based on machine learning.

[0006] In a first aspect, this application provides a fault early warning method for terminal block assembly equipment based on machine learning, including: The process parameters, status data, and image data of each step in the terminal assembly process are collected using sensors and computer vision devices, and the process is marked. Extract feature vectors from each process to construct a process state model; the process state model includes feature vector distribution templates for normal and fault states. Fault types are identified by pre-built machine learning classification models, and a mapping relationship between faults and processes is established. A multi-layer analysis model is used to monitor abnormal process parameters in real time and determine the responsible process for the fault, outputting fault location results and early warning information. The multi-layer analysis model includes a process status monitoring layer, a fault type identification layer, and a responsibility location layer. The process status monitoring layer determines abnormal process data based on the constructed process status model. The fault type identification layer determines the fault type of the abnormal process data based on a machine learning classification model. The responsibility location layer determines the final responsible process based on the mapping relationship between faults and processes, combined with the logical relationship between processes.

[0007] By adopting the above scheme, comprehensive collection and labeling of different types of data are carried out to provide basic data for subsequent analysis; an accurate process status model is constructed using feature vectors, which helps to accurately identify process status; machine learning is used to accurately identify fault types and establish mapping relationships to clearly locate faults; and a multi-layer analysis model is used to monitor and locate faulty processes in real time, thereby effectively improving the accuracy and timeliness of fault diagnosis and reducing downtime and maintenance costs.

[0008] Preferred options also include: Based on the quantity of materials in the feeding process, the number of crimping operations in the crimping process, and the coupling characteristics of each process, the terminal assembly scenarios are divided into simple, medium, and complex scenarios, in descending order of complexity. The acquisition devices and frequencies for acquiring process parameters, status data, and image data of each step in the assembly process of the terminal block are configured differently according to different scenarios. The more complex the scenario, the higher the accuracy and frequency of the corresponding acquisition devices. Differentiated adaptation is made for each process of feature vector extraction dimension, machine learning classification model complexity, fault library and mapping rules for different scenarios. The more complex the scenario, the higher the corresponding feature vector extraction dimension, the higher the fault recognition model complexity, the more fault types in the fault library and the more complex the mapping rules. The analysis levels of the multi-level analysis model are configured according to different scenarios, and the lower the scenario complexity, the lower the application level of the analysis level model.

[0009] By adopting the above solution, differentiated configuration and adaptation can be carried out according to different terminal assembly scenarios, enabling more accurate data collection, fault type identification, and determination of responsible processes. This effectively improves the accuracy and adaptability of fault early warning and reduces the fault misjudgment rate caused by complex scenarios.

[0010] Preferred options also include: Based on the full life cycle of the terminal assembly equipment, the current life cycle of the terminal assembly equipment is determined in real time; based on the current life cycle, the application mode of the computer vision device is adapted, including the first application mode adapted to the initial stage, the second application mode adapted to the stable stage, and the third application mode adapted to the decline stage. The first application mode includes: using Canny edge detection to extract the edges of the positioning platform baseline, mold closing surface, and terminal pins; expanding the edge contour through morphological expansion for mold gaps; fitting a straight line using Hough linear transform for the positioning platform baseline edge and calculating the angle between the fitted line and the standard baseline; filling the micro-holes in the mold gap using morphological reconstruction. The second application mode includes: using Sobel edge detection to extract the terminal contour and performing template matching with the pre-stored standard posture contour; calculating the pixel displacement of the contact area between the pressure head and the terminal using optical flow for the pressing time-series images acquired by the high-speed line scan camera; removing reflective spots on the terminal surface using morphological opening operation, and then extracting scratches and dents using Laplacian edge detection. The third application mode includes: segmenting the wear area using Otsu thresholding, and then filling the micro-holes in the wear area using morphological closing operation; calculating the area ratio, average depth, and maximum width of the segmented wear area; and calculating the wear area offset at each time point for the time-series images using template matching.

[0011] By adopting the above scheme, the application mode of the computer vision device is adaptively adjusted according to the different life cycle stages of the terminal assembly equipment, thereby accurately capturing key features at different operating stages of the equipment and improving the accuracy of fault identification at different stages.

[0012] Preferred options also include: Extract the time sequence feature vectors of each process to construct a process state model; the process state model also includes a feature vector distribution template for potential faults. Predict fault types by pre-built machine learning classification models and establish a mapping relationship between potential faults and processes; The system uses a multi-layered analysis model to predict abnormal process parameters and the responsible process for failures in real time, outputting the predicted fault location results and early warning information. The process status monitoring layer also predicts and monitors abnormal process data based on the constructed process status model. The fault type identification layer predicts the fault type of abnormal process data based on a machine learning classification model. The responsibility location layer predicts the final responsible process based on the mapping relationship between potential faults and processes, combined with the logical relationship between processes.

[0013] By adopting the above scheme, potential faults can be predicted, and the mapping relationship between potential faults and processes can be established in advance. The predicted fault location results and early warning information can be output before the fault actually occurs, thereby improving the foresight of fault early warning and reducing the risk of fault occurrence.

[0014] Preferred options also include: For the new terminal block assembly equipment, the multi-layer analysis model built using the same type of terminal block assembly equipment is synchronously migrated to the new terminal block assembly equipment, and the model is trained and parameters are set using historical process data without difference. In response to the cycle switching of the terminal assembly equipment, the constructed process state model, machine learning classification model, and multi-level analysis model are replaced with corresponding cycle models; the corresponding cycle model replacement means that for each cycle, a corresponding process state model, machine learning classification model, and multi-level analysis model are constructed. Federated learning transfer was performed on multiple terminal block assembly devices of the same model and cycle to complete the model iteration of the multi-layer analysis model.

[0015] By adopting the above scheme, the multi-layer analysis model can be quickly applied to new equipment, reducing the construction time and cost of the fault early warning model for new equipment; it can adapt the model to the state changes of the equipment in different cycles, improving the accuracy of fault early warning; and it can also improve the performance and adaptability of the multi-layer analysis model through federated learning transfer, further ensuring the stable operation of the terminal assembly equipment.

[0016] Preferred options also include: The establishment of the mapping relationship between faults and processes adopts a hierarchical mapping, including process anomaly layer mapping, equipment fault layer mapping, and process data anomaly layer mapping. Each level contains the feature description and judgment rules of each fault. The mapping relationship between each level is established through unique coding and association weight. The association weight is obtained through statistical calculation of historical data.

[0017] By adopting the above scheme, hierarchical mapping can clearly present the relationship between faults and processes and equipment, clarify the characteristic descriptions, judgment rules and responsible processes of each fault, and make the mapping more accurate by combining association weights.

[0018] Preferred options also include: Obtain historical fault case library, troubleshooting rule library and troubleshooting solution library, and calculate the priority score of each type of fault by comprehensively considering fault urgency, occurrence frequency and processing efficiency, and obtain the priority ranking of fault types of different processes of terminal assembly equipment; In the process of real-time monitoring of abnormal process parameters and identification of faulty processes through a multi-layer analysis model, the fault location results and early warning information are output according to the priority of fault types.

[0019] By adopting the above scheme, obtaining the fault priority ranking and outputting the results accordingly, relevant personnel can prioritize the handling of important faults, further improving the accuracy and efficiency of fault diagnosis for terminal block assembly equipment.

[0020] Secondly, this application provides a fault early warning system for terminal block assembly equipment based on machine learning, comprising: The process data acquisition and marking module is used to acquire process parameter data, status data, and image data of each process in the terminal assembly process through sensors and computer vision devices, and to mark the process. The process status model construction module is used to extract feature vectors of each process to construct a process status model; the process status model includes feature vector distribution templates for normal and fault states. The process fault identification and mapping module is used to identify fault types and establish a mapping relationship between faults and processes through a pre-built machine learning classification model; The real-time process fault monitoring module is used to monitor abnormal process parameters in real time and determine the responsible process for the fault through a multi-layer analysis model, and output fault location results and early warning information. The multi-layer analysis model includes a process status monitoring layer, a fault type identification layer, and a responsibility location layer. The process status monitoring layer determines the abnormality of monitored process data based on the constructed process status model. The fault type identification layer determines the fault type of the abnormal process data based on a machine learning classification model. The responsibility location layer determines the final responsible process based on the mapping relationship between faults and processes, combined with the logical relationship between processes.

[0021] By adopting the above scheme, data from each process in the terminal assembly process are collected and marked, a process status model containing feature vector distribution templates of normal and fault states is constructed, fault types are accurately identified and a mapping relationship between faults and processes is established, abnormal process parameters are monitored in real time and the responsible process for the fault is determined, and then detailed fault location results and early warning information are output.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: 1. Utilize the process data acquisition and labeling module to comprehensively collect and label process data, providing a rich and accurate data foundation; construct a process status model containing feature vector distribution templates for normal and fault states, clearly defining normal and fault states; use a machine learning classification model to identify fault types and establish a mapping relationship between faults and processes, accurately locating the source of faults; monitor and determine the fault-responsible process in real time through a multi-layer analysis model and output results and early warning information, promptly discovering faults and providing decision-making basis, improving the timeliness, accuracy, and reliability of fault early warning for terminal assembly equipment; 2. By adopting scenario-specific differentiated configurations and adapting the application mode of computer vision devices according to the equipment lifecycle, the fault types and responsible processes are predicted. Model migration, replacement, and iteration are performed, and a hierarchical mapping is used to establish the fault-process mapping relationship and determine the fault priority order, thereby improving the accuracy and efficiency of fault diagnosis for terminal block assembly equipment and ultimately achieving precise location of faulty processes in terminal block assembly equipment. Attached Figure Description

[0025] Figure 1 This is a flowchart of the machine learning-based fault early warning method for terminal assembly equipment described in a specific embodiment; Figure 2 This is a schematic diagram of the fault early warning system for terminal assembly equipment based on machine learning described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a fault early warning method for terminal assembly equipment based on machine learning. This method analyzes the process parameters and quality data in the assembly process to achieve accurate location and early warning of fault steps in the assembly process. Specifically, it includes steps such as process data acquisition and marking, process status model construction, process fault identification and mapping, and real-time monitoring of process faults.

[0028] The implementation steps of this method are described in detail below through a specific embodiment. This embodiment targets an automated assembly production line for terminal blocks, including key processes such as feeding, positioning, crimping, and inspection. The feeding process is responsible for the loading and conveying of components such as buttons, springs, conductive sheets, and housings; the positioning process is responsible for the precise positioning of each component at the assembly station; the crimping process is responsible for the mechanical connection and fixation between components; and the inspection process is responsible for verifying the assembly quality. The entire process is executed sequentially, realizing a complete logical chain from data acquisition to fault location.

[0029] S1. Collect process parameter data, status data and image data of each step in the terminal assembly process, and mark the steps.

[0030] Specifically, data from each step of the terminal assembly process is collected using sensors and computer vision devices. Sensors can be of varying measurement precision, such as photoelectric sensors, level sensors, displacement sensors, and pressure sensors. The computer vision devices can be microprocessor-based, and can include monocular cameras, binocular cameras, and laser-vision composite vision devices.

[0031] Accordingly, the process parameters and status data collected for each process include: First, process parameters of the process execution device in the terminal assembly equipment: For each process, record its input conditions (such as component specifications, upstream process status), control parameters (such as pressure setpoint, positioning coordinates), execution actions (such as cylinder action, motor operation), and output results (such as crimping completion signal, detection results); at the same time, record the logical relationships between processes; organize this information into a structured process mapping table, which includes fields such as process name, parameter list, preceding process and following process, and store it in the database.

[0032] Secondly, the assembly operation status data of the process execution devices in the terminal assembly equipment includes: Vibration sensors, photoelectric sensors, and material counters deployed in the feeding process to collect parameters such as feeding frequency, terminal posture deviation, and material supply continuity; laser displacement sensors and servo motor encoders deployed in the positioning process to collect data such as positioning coordinate error, positioning completion time, and positioning mechanism operating temperature; pressure sensors, temperature sensors, and displacement sensors deployed in the crimping process to collect key parameters such as peak crimping force, crimping depth, mold temperature, and crimping time; and continuity testers, insulation resistance testers, and visual inspection cameras deployed in the inspection process to collect quality and status data such as terminal continuity resistance, insulation resistance, and appearance defects. An appropriate sampling frequency is set, and the data storage format is configured as a time-series database, with each record containing a timestamp, sensor ID, value, and status flag.

[0033] Accordingly, image data collected using computer vision devices at each stage includes: monitoring the vibratory feeder outlet or conveyor belt during the feeding stage, collecting image data including the number, orientation, spacing, and surface defects of terminals; monitoring the precise placement of terminals into the correct position on the crimping mold during the positioning stage, collecting image data including terminal position and mold assembly; monitoring the deformation of terminals and wires during the crimping process during the crimping stage, collecting image data including crimping depth, wire position, and terminal deformation; and conducting final appearance and quality inspection of the crimped terminals during the inspection stage, collecting image data including terminal appearance defects, geometric dimensions, and color. To better extract subsequent feature data, the image data can be preprocessed, including: adaptive lighting adjustment, background modeling and updating, edge detection, and morphological operations.

[0034] Finally, process tags are added to the collected data, including process ID, operation type, start time and end time. The start and end times of the process are determined by the process switching signal of the PLC. Product information (such as batch number and serial number read from the barcode scanner) and the correlation information between processes (such as the output coordinate value of the positioning process affecting the alignment accuracy of the crimping process) are also recorded to form a structured dataset with process context. The table structure includes fields such as timestamp, process ID, parameter name, parameter value, product ID and associated process.

[0035] S2. Extract feature vectors for each process to construct a process state model.

[0036] Specifically, based on the labeled process data, features reflecting the state of each process are extracted, and fault modes and their process affiliations are learned and identified, thereby constructing a process state model.

[0037] First, considering that feature vector extraction uses different calculation methods for different processes, the feature extraction content for different processes will be specifically described below. From the perspective of sensor feature extraction: For the feeding process, the focus is on extracting time-domain features (such as the average and standard deviation of feeding frequency) and abnormal features (such as the number of times material gets stuck), including features such as the feeding position change rate, feeding cycle, and jamming frequency. For the positioning process, the focus is on extracting statistical features of coordinate deviation (such as the maximum deviation and deviation fluctuation rate) and dynamic features (such as the positioning speed change rate), including features such as positioning accuracy, repeatability error, and positioning time. For the crimping process, the focus is on extracting waveform features of the crimping force (such as peak value, rise time, and settling time) and temperature features (such as the average temperature and temperature gradient), including features such as the maximum pressure value, pressure rise rate, crimping duration, and pressure curve shape parameters. For the inspection process, the focus is on extracting quality index features (such as the average conduction resistance and minimum insulation resistance) and defect features (such as the area of ​​appearance defects and defect type), including the pass rate of each inspection item, the distribution of defect types, and inspection stability.

[0038] In addition, statistical characteristics such as mean, standard deviation, coefficient of variation, and trend slope are calculated for the control parameters of all processes. For time-series data, a fixed-length sliding window is used to calculate the trend and fluctuation pattern of the parameters within the window. These characteristics are organized into feature vectors, one group per process, and stored in a feature data table. The table structure includes fields such as time window, process ID, feature name, and feature value.

[0039] From the perspective of image feature extraction, for the feeding process, features such as the number of terminals, terminal posture, and position distribution are extracted; for the positioning process, features such as terminal position deviation, terminal rotation deviation, and mold status are extracted; for the crimping process, features such as crimping depth, wire position, burr detection, and terminal deformation curve are extracted; and for the inspection process, features such as appearance defects, geometric dimensions, and color detection are extracted.

[0040] For each process, visual features extracted and sensor features are fused, including feature stitching and temporal feature fusion, to obtain the final multi-dimensional feature vector extracted for each process.

[0041] Secondly, a process state model is constructed. This model includes feature vector distribution templates for normal and fault states. For example, a process state model is constructed using normal state modeling and anomaly threshold definition: First, based on historical qualified production data, the normal distribution pattern of each process's feature vectors is fitted using a Gaussian mixture model to determine the core parameters such as the mean and variance of the normal state features; then, a 3D model is used... The criteria define the abnormal thresholds for each characteristic, forming a process state model with a binary criterion of normal and abnormal. Furthermore, comprehensive statistical analysis can be performed using environmental data and equipment operating age data to determine dynamic thresholds under different environments and equipment operating ages, corresponding to the constructed process state models.

[0042] The feature templates of normal and various fault states are organized into a state model and stored in a state template table in the database. The table structure includes fields such as state type (normal / fault type), process ID, and feature vector. By constructing the process state model, it is possible to quickly determine whether the current production state deviates from the health baseline and to initially indicate which process's signal is abnormal.

[0043] S3. Identify fault types and establish a mapping relationship between faults and processes by pre-constructing machine learning classification models.

[0044] Specifically, to better identify fault types, a lightweight or integrated machine learning classification model can be constructed. Common fault types in the terminal assembly process are pre-organized, and each fault type is labeled accordingly. Then, the pre-processed process feature vector is used as input, and the fault type label is used as output. A multi-model ensemble classifier can be constructed using decision tree algorithm, random forest, support vector machine, gradient boosting tree.

[0045] The decision tree algorithm is employed, with process feature vectors as input and fault types as output. The decision tree construction process includes preparing the training dataset (feature vectors and corresponding fault labels); selecting the feature with the highest information gain as the splitting node; recursively constructing subtrees until a stopping condition is met; and pruning the tree to avoid overfitting. Random forests are used to filter key features (such as pressure fluctuations and positioning deviations as key features of poor contact), support vector machines are used to handle high-dimensional nonlinear features, and gradient boosting trees are used to improve the recognition accuracy of small-sample faults. Five-fold cross-validation is used to optimize model parameters, ensuring the overall model recognition accuracy reaches the preset accuracy.

[0046] Based on the identification results of the above fault classification model and historical data statistics, an association rule mining algorithm (such as the Apriori algorithm) is used to establish a mapping relationship library between fault types and processes. The steps include: setting support and confidence thresholds; by analyzing the distribution of abnormal processes corresponding to the occurrence of various faults, calculating the association confidence and support between fault types and processes, selecting high-confidence association rules as the mapping basis, generating frequent itemsets (frequently occurring abnormal combinations of parameters), and generating association rules from frequent itemsets.

[0047] Furthermore, to better and more intuitively display the mapping relationship, a three-layer mapping structure is created: the top layer is the assembly flowchart, containing all process nodes and connections; the middle layer is the process details, containing the parameter list and normal range for each process; and the bottom layer is the fault type library, containing the characteristic description and judgment rules for each fault. For each fault type, graph neural network algorithms or causal analysis rules can be used to determine the fault source, clarifying its primary responsible process and secondary influencing processes. For example, the primary responsible process for "poor crimping" is the crimping process, and the secondary influencing processes are the positioning and inspection processes. Fault diagnosis rules are written, such as: IF crimping force curve peak is below threshold AND crimping time exceeds normal range AND feeding status is normal THEN Fault process = crimping process AND fault type = insufficient pressure. The rule includes a condition part (parameter judgment) and a conclusion part (fault location). For fuzzy cases, a confidence calculation method is set, such as calculating the confidence of fault location based on the number of matching rules and rule weights.

[0048] Finally, the identification and mapping results are saved as model files and rule bases. The model files contain decision tree structures (nodes, splitting conditions, and leaf node categories), random deep forest models, support vector machine models, etc., while the rule bases contain a list of association rules (preconditions, conclusions, and confidence levels).

[0049] S4. Real-time monitoring of abnormal process parameters and determination of the responsible process for the fault are achieved through a multi-layer analysis model, and the fault location results and early warning information are output.

[0050] First, a multi-layered analysis model is constructed as the fault location model. The multi-layered analysis model includes a process status monitoring layer, a fault type identification layer, and a responsibility location layer; a three-layer progressive analysis architecture of "process status monitoring layer - fault type identification layer - responsibility location layer" is adopted to achieve accurate location of faulty processes.

[0051] Specifically, the process status monitoring layer identifies anomalies in monitored process data based on the constructed process status model. This involves a preliminary assessment of the process data to determine if a fault exists, and then using the aforementioned process status model to check if the real-time monitored process data exceeds normal thresholds and identify any anomalies. The fault type identification layer utilizes the pre-built machine learning classification model for fault type identification to determine the specific fault type of the process data anomaly. The responsibility location layer determines the final responsible process based on the mapping relationship between faults and processes, combined with the logical relationships between processes. These three layers are integrated into a unified location system, with information exchanged between layers through data interfaces.

[0052] Secondly, develop early warning strategies and adjust thresholds. For example, set three early warning levels: Attention, Warning, and Emergency. The specific levels are determined based on fault frequency, number of faults, fault type, and the criticality of the fault location process. Furthermore, to improve early warning capabilities, the constructed process status model can be further optimized by setting multi-level thresholds. Abnormal process data can be categorized into different levels to determine the severity of faults. The early warning level classification can then be optimized by combining fault frequency, number of faults, fault type, and the criticality of the fault location process.

[0053] Finally, real-time fault location and early warning information output are performed.

[0054] Specifically, real-time process monitoring and fault location are performed. Assembly process data is collected from sensor networks and computer vision networks according to a preset sampling period. The collected raw data is preprocessed, and feature vectors are extracted from the preprocessed data. These feature vectors are then input into a multi-layer analysis model for three-layer analysis: First, the status monitoring layer detects whether parameters are abnormal; if an anomaly is found, the fault type identification layer determines the specific fault type; finally, the process responsibility location layer identifies the fault source process. The location results are recorded in a fault log table, including timestamps, abnormal processes, fault types, relevant parameters, and confidence levels.

[0055] The output presents the fault location analysis results in a structured format, providing clear fault location and detailed information. A visual interface displays the assembly process status, using different colors to indicate the health status of each process and intuitively showing the location of abnormalities. For the located faulty process, the system provides parameter trend charts, displaying historical changes and current values ​​of key parameters. It can also generate a fault location report, including a fault summary, detailed analysis, and historical comparisons, facilitating subsequent analysis and recording.

[0056] In a specific embodiment, through scenario segmentation and differentiated configuration, data collection, model building, and fault monitoring can be performed more effectively based on the characteristics of different assembly scenarios; the method also includes: Terminal block assembly scenarios are categorized based on the quantity of materials in the feeding process, the number of crimping operations in the crimping process, and the coupling characteristics of each process (the strength of coupling can be determined based on expert experience). These scenarios are classified into simple, medium, and complex scenarios, in descending order of complexity. For example, in a simple scenario, the quantity of materials is 1, the number of crimping operations is 1, and the coupling coefficient between processes is less than 0.3, indicating weak coupling. In a medium scenario, the quantity of materials is 2, the number of crimping operations is 2, and the processes involve sequential and combined crimping, with a coupling coefficient between processes ranging from 0.4 to 0.7, indicating strong coupling. In a complex scenario, the quantity of materials is greater than or equal to 3, the number of crimping operations is greater than or equal to 3, and the processes involve multiple rounds of sequential and final combined crimping, with a coupling coefficient greater than 0.8, indicating strong coupling.

[0057] The acquisition devices and frequencies for process parameters, status data, and image data of each step in the terminal assembly process are configured differently for different scenarios. The more complex the scenario, the higher the accuracy and frequency of the acquisition devices are configured. For example, in simple scenarios, the key data acquisition sensors and vision devices are configured for the process, and the acquisition frequency of other non-key data is reduced from 100Hz to 20Hz. In medium scenarios, high-precision acquisition sensors (with a spacing of ±0.02mm before acquisition and merging), ordinary precision sensors, and binocular vision devices are used for the sequence preset interaction parameters. In complex scenarios, basic acquisition sensors, high-precision sensors for acquiring process preset interaction parameters, and laser-vision composite vision devices are configured, and the acquisition frequency is increased for coupled processes.

[0058] The feature vector extraction dimensions, machine learning classification model complexity, fault database, and mapping rules are adapted to different scenarios. The more complex the scenario, the higher the feature vector extraction dimension, the higher the fault recognition model complexity, the more fault types in the fault database, and the more complex the mapping rules. For example, in simple scenarios, only the basic features and core statistical features of each process are extracted, a lightweight machine learning classification model (decision tree classifier and linear regression localization) is used, only common fault types in the assembly process are considered, and a fixed mapping rule is used for fault-process mapping. In medium-scale scenarios, in addition to extracting the basic features and core statistical features of each process, attention is also paid to the correlation features between processes, and an integrated machine learning classification model (integrated decision tree classifier, random...) is used. In addition to considering common assembly process failure types, the system also adds failure types related to merging operations, such as alignment deviation and pressure mismatch between material A and material B. Furthermore, it uses the Apriori algorithm to mine the association rules between failures and processes for failure-process mapping. In complex scenarios, in addition to extracting the above features, it also focuses on spatial features (such as the three-dimensional coordinate deviation of multiple material alignment) and coupling features (such as the association between process switching interval and pressing quality). It adopts an integrated machine learning classification model (integrated decision tree classifier, random forest, Transformer, etc.). In addition to considering common assembly process failure types, it adds all failure types such as multi-round pressing fatigue and material compatibility failures. Furthermore, it uses reinforcement learning to dynamically generate mapping rules for failure-process mapping.

[0059] The analysis hierarchy of the multi-level analysis model is configured according to different scenarios, and the lower the scenario complexity, the lower the application level of the analysis hierarchy model. Considering that simple scenarios are linear processes, only the process status monitoring layer and the fault type identification layer can be applied. Then, by verifying whether the abnormal parameters of each process match the fault one by one, the process with the highest matching degree is identified as the main responsible process to determine the final responsibility location, which can replace the application of the entire hierarchical analysis model and improve efficiency. For medium-sized scenarios, the process status monitoring layer and the fault type identification layer can be applied. Based on the abnormal coupling characteristics, the processes suspected of fault coupling are identified, and the contribution of each process to the fault is calculated. The process with the highest contribution degree is identified as the main responsible process. For complex scenarios, the hierarchical analysis model is applied in its entirety.

[0060] In one specific embodiment, different computer vision device application modes are adapted according to different lifecycle stages of the device, which can better adapt to the characteristics of the device at different stages; the method also includes: Considering the additional influencing factors such as mold wear that occur with the use of terminal block assembly equipment, an optimized computer vision device is used to improve fault location and thus optimize the matching of image acquisition data and feature extraction for the current cycle. First, based on the pre-defined full lifecycle of the terminal block assembly equipment, the current lifecycle of the equipment is determined in real time, including the initial, stable, and declining phases.

[0061] Secondly, the application mode of the computer vision device is adapted based on the current lifecycle. Specifically, for the initial stage of terminal assembly equipment, the main root causes of failures at this stage are generally installation deviations, improper mold gaps, and sensor calibration deviations. The corresponding visually measurable features include positioning platform tilt, terminal pin alignment misalignment, and mold fitting gaps. Therefore, a first application mode adapted to the initial stage is set up. The first application mode includes: using Canny edge detection (low threshold 50, high threshold 150) to extract the positioning platform baseline, mold mating surface, and terminal pin edges; for mold gaps (usually 0.01-0.05mm), the edge contour is expanded through morphological dilation (structural element 3×3 rectangular kernel), changing the gap from a single pixel line to a 3-pixel line; for the positioning platform baseline edge, Hough linear transformation is used to fit a straight line, and the angle between the fitted line and the standard baseline is calculated (accuracy 0.01°) to quantify the platform horizontal deviation; for the mold gap, morphological reconstruction (using the edge detection result as the marked image and a 3×3 cross kernel as the structural element) is used to fill the tiny holes in the gap; and then the gap width is calculated through pixel distance (accuracy 0.001mm). Correspondingly, subsequent feature extraction can focus on the straightness of the positioning baseline, the average gap of the mold closing surface, and the alignment offset between the terminal pins and the mold. Moreover, the initial visual features are more direct in representing the installation deviation, so the weight of visual features can be set to 0.6, and the weight of sensor features (such as level sensor and gap sensor data) can be set to 0.4, for subsequent feature fusion weight settings.

[0062] During the stabilization period of terminal block assembly equipment, the main root causes of failures include material batch fluctuations, environmental interference (vibration / lighting), and occasional jamming. Corresponding visually measurable features include scratches on the terminal surface, indentations on the crimping surface, and instantaneous deviations in the feeding posture. Therefore, a second application mode adapted to the stabilization period is established. This second application mode includes: using Sobel edge detection to extract the terminal contour and performing template matching with a pre-stored standard posture contour (normalized correlation coefficient threshold 0.95); a matching degree below the threshold is considered an abnormal posture. For crimping time-series images acquired by a high-speed linear scan camera, optical flow is used to calculate the pixel displacement of the contact area between the crimping head and the terminal to extract the displacement peak (reflecting crimping depth) and displacement fluctuation rate (reflecting crimping stability). Morphological opening operations (5×5 circular kernel for structuring elements) are used to remove reflective spots on the terminal surface, and Laplacian edge detection is used to extract scratches and indentations. Accordingly, subsequent feature extraction can focus on the feeding posture matching degree, crimping displacement peak, crimping displacement fluctuation rate, and surface defect area. The stable-period failure is caused by both environmental interference and equipment malfunction. A dynamic weighting strategy is adopted: if there is no interference (normal lighting / vibration), the weights of visual and sensor features are both 0.5, and they are fused through Bayesian inference; if there is interference and the interference source is lighting, the visual weight drops to 0.3 and the sensor weight rises to 0.7; if the interference source is vibration, the visual weight rises to 0.7 and the sensor weight drops to 0.3.

[0063] For terminal block assembly equipment in its degradation phase, the main root causes of failure are pressure head wear, guide rail aging, and seal leakage and contamination. Corresponding visually measurable features include pitting on the pressure head surface, scratches on the guide rail, residual terminal contamination, and abnormal crimping deformation. Therefore, a third application mode adapted to the degradation phase is established. This third application mode includes: segmenting the wear area (such as pressure head pitting and guide rail scratches) using Otsu thresholding, and then filling the micro-holes within the wear area using morphological closing operations; calculating the area ratio, average depth, and maximum width of the segmented wear area; and using template matching (with the initial reference image as a template) to calculate the wear area offset at each time point for time-series images, reflecting the wear expansion trend. Accordingly, subsequent feature extraction can focus on the wear area ratio; average wear depth; maximum wear width; circumferential expansion of the wear area; and terminal deformation deviation, etc., resulting in more accurate visual representation of wear / contamination during the degradation phase, with visual feature weights gradually increasing as the wear degree deepens.

[0064] In one specific embodiment, considering the technical effect of not only detecting and diagnosing existing faults but also predicting potential faults and locating their possible positions, the method further includes: Extract time-series feature vectors for each process to construct a process status model; the process status model also includes a feature vector distribution template for potential faults; specifically, for the extracted time-series features of each process, introduce time series analysis, use a sliding window to calculate the statistics of the features, fuse the time-series feature data, construct a health index to quantify the health status of the equipment, and use a degradation model (such as an exponential model or a linear model) to fit the degradation process of the health index, thereby obtaining a feature vector distribution template for potential risks (that will exceed the health index threshold within a preset time) in the health status.

[0065] Fault types are predicted by pre-constructing a machine learning classification model, and a mapping relationship between potential faults and processes is established. Correspondingly, common potential fault types in the terminal assembly process (such as critical wear of the pressure head, precursors to guide rail lubrication failure, etc.) are compiled, and each potential fault type is labeled accordingly. Then, the pre-processed process feature vectors are used as input, and the potential fault type labels are used as output. A multi-model ensemble classifier is constructed using decision tree algorithms, random forests, support vector machines, and gradient boosting trees. Based on the identification results of the aforementioned potential fault classification model and historical data statistics, an association rule mining algorithm (such as the Apriori algorithm) is used to establish a mapping relationship library between potential fault types and processes.

[0066] The system uses a multi-layered analysis model to predict abnormal process parameters and the responsible process for failures in real time, outputting the predicted fault location results and early warning information. The process status monitoring layer also predicts and monitors abnormal process data based on the constructed process status model. The fault type identification layer predicts the fault type of abnormal process data based on a machine learning classification model. The responsibility location layer predicts the final responsible process based on the mapping relationship between potential faults and processes, combined with the logical relationship between processes.

[0067] This involves performing real-time process monitoring, fault prediction, and location. Assembly process data is collected from sensor networks and computer vision networks according to a preset sampling period. The collected raw data is preprocessed, and feature vectors are extracted from the preprocessed data. The feature vectors are then input into a multi-layer analysis model to perform three-layer analysis: First, the status monitoring layer predicts whether the parameters will be abnormal; if an abnormality is predicted, the potential fault type identification layer predicts the fault type; and finally, the process responsibility location layer predicts the fault source process.

[0068] In a specific embodiment, model migration and replacement can quickly establish suitable models for new equipment and equipment after cycle switching, reducing the time and cost of model building. Federated learning transfer can integrate data from multiple devices to further optimize the model and improve its accuracy and adaptability. To better adapt fault warnings to the needs of different devices and different stages, the method further includes: for the new terminal assembly equipment, a multi-layer analysis model built using the same model of terminal assembly equipment is synchronously migrated to the new terminal assembly equipment, and model training and parameter setting are performed using real-time monitoring data of identical process parameters (such as establishing individual benchmarks with 50 qualified data points).

[0069] In response to the cycle switching of terminal block assembly equipment, the constructed process state model, machine learning classification model, and multi-level analysis model undergo corresponding cycle model replacement. This corresponding cycle model replacement means that for each cycle, a corresponding process state model, machine learning classification model, and multi-level analysis model are constructed. As described in the previous embodiments, the terminal block assembly equipment in different cycles can affect fault diagnosis. Therefore, parameter fine-tuning is required to correct the judgments of each model. For example, the threshold values ​​for judging normal and fault states in the process state model differ in different cycle stages. Therefore, different process state models are constructed for different cycles to adapt to the judgment of whether the process data is abnormal in the corresponding cycle. Similarly, the machine learning classification model and multi-level analysis model also need parameter adjustments for different cycle stages to accurately locate faults, and corresponding machine learning classification models and multi-level analysis models are set for different stages.

[0070] For multiple terminal block assembly devices of the same model and production cycle, federated learning transfer is performed to complete the model iteration of the multi-layer analysis model. Taking the machine learning classification model within the terminal block assembly device as an example, sub-models are trained locally on each device, and federated learning aggregates parameters to update the main model. The sub-models synchronize with the main model, realizing data sharing and collaborative optimization among multiple devices.

[0071] In a specific embodiment, to more accurately locate process errors and determine the specific causes of those errors, a unique association is established between abstract faults and concrete processes, equipment components, and process parameters. This allows for the identification of specific fault sources and causes, thus better assisting in fault repair. The method further includes: The mapping relationship between faults and processes adopts a hierarchical mapping. That is, based on the creation of a three-layer mapping relationship, the bottom layer is a fault type library and it is further divided into layers, including process abnormality layer mapping, equipment fault layer mapping, and process data abnormality layer mapping. Each layer mapping contains the feature description and judgment rules of each fault.

[0072] The process anomaly mapping includes feature descriptions and judgment rules for each type of fault, such as poor conductivity (fault type), corresponding to insufficient crimping pressure (primary responsible process) and excessive positioning deviation (secondary responsible process). Based on this, equipment fault mapping is performed, including feature descriptions and judgment rules for each type of fault (rule mining based on the known correlation between process anomalies and equipment anomalies). This identifies the abnormal equipment corresponding to the abnormal process in the fault mapping, such as insufficient crimping pressure (process anomaly) - pressure head wear (equipment anomaly) and pressure sensor drift (equipment anomaly). Further, process data anomaly mapping is performed, including feature descriptions and judgment rules for each type of fault (rule mining based on the known correlation between process anomalies, equipment anomalies, and process data). This identifies the anomalies in the process data corresponding to the abnormal equipment in the fault mapping, such as pressure sensor drift - excessive pressure fluctuation. The final output mapping chain is: poor conductivity - insufficient crimping pressure - pressure head wear - abnormal peak crimping force, accurately locating the root cause. Each level establishes a mapping relationship through a unique code and association weights, which are obtained through statistical calculations of historical data.

[0073] Furthermore, considering that multiple faults may occur during the real-time assembly process of the terminal block assembly equipment, and it is impossible to locate and handle each fault simultaneously, it is necessary to integrate historical fault cases, troubleshooting experience, and process standards. This involves setting rule priorities and intelligently matching recommendations to optimize the troubleshooting logic. The method also includes: The system acquires historical fault case databases, troubleshooting rule databases, and troubleshooting solution databases. It also calculates priority scores for each type of fault by comprehensively considering fault urgency, occurrence frequency, and processing efficiency, and obtains the priority ranking of fault types for different processes in the terminal assembly equipment.

[0074] The historical fault case library stores basic information about historical faults (fault code, occurrence time), related data (hierarchical mapping chain, parameter curve), processing procedures (troubleshooting steps, time consumption), and effect verification (resolution rate, recurrence rate). Each case is marked with three major indicators: occurrence frequency, scope of impact, and processing cost. The troubleshooting rule library extracts troubleshooting rules based on the case library and stores troubleshooting solutions for related faults. The troubleshooting solution library stores troubleshooting solutions.

[0075] In the process of real-time monitoring of abnormal process parameters and identification of faulty processes through a multi-layer analysis model, the fault location results and early warning information are output according to the priority of fault types.

[0076] like Figure 2 As shown, this application provides a fault early warning system for terminal block assembly equipment based on machine learning, specifically including: The process data acquisition and marking module 101 is used to acquire process parameter data, status data and image data of each process in the terminal assembly process through sensors and computer vision devices, and to mark the process. The process status model construction module 102 is used to extract feature vectors of each process to construct a process status model; the process status model includes feature vector distribution templates for normal state and fault state. The process fault identification and mapping module 103 is used to identify fault types and establish a mapping relationship between faults and processes through a pre-built machine learning classification model; The real-time process fault monitoring module 104 is used to monitor abnormal process parameters in real time through a multi-layer analysis model, determine the responsible process for the fault, and output fault location results and early warning information.

[0077] In one specific embodiment, the system further includes: The scenario division module 105 is used to divide the terminal assembly scenario according to the material quantity in the feeding process, the number of crimping processes in the crimping process, and the coupling characteristics of each process in the terminal assembly process. The scenario division includes simple, medium and complex scenarios in descending order of scenario complexity. The process data acquisition and marking module 101 is also used to configure the acquisition devices and frequencies for process parameters, status data, and image data of each process in the terminal assembly process according to different scenarios. The more complex the scenario, the higher the accuracy and frequency of the corresponding acquisition devices. The process status model construction module 102 is also used to adapt the feature vector extraction dimension of each process according to different scenarios. The process fault identification and mapping module 103 is also used to adapt the model complexity of the machine learning classification model and the fault library and mapping rules according to different scenarios. The more complex the scenario, the higher the feature vector extraction dimension, the higher the fault identification model complexity, the more fault types in the fault library, and the more complex the mapping rules. The process fault real-time monitoring module 104 is also used to configure the analysis level of the multi-layer analysis model according to different scenarios. The lower the scenario complexity, the lower the application level of the analysis level model.

[0078] In one specific embodiment, the system further includes: The process data acquisition and optimization module 106 is used to determine the current life cycle of the terminal assembly equipment in real time according to the full life cycle of the divided terminal assembly equipment; and to adapt the application mode of the computer vision device based on the current life cycle, including the first application mode adapted to the initial stage, the second application mode adapted to the stable stage, and the third application mode adapted to the decline stage.

[0079] In one specific embodiment, the system further includes: The process state model construction module 102 is also used to extract the time sequence feature vectors of each process to construct the process state model; the process state model also includes a feature vector distribution template for potential faults. The process fault identification and mapping module 103 is also used to predict fault types and establish a mapping relationship between potential faults and processes by pre-built machine learning classification models; The real-time process fault monitoring module 104 is also used to predict abnormal process parameters and the responsible process for the fault in real time through a multi-layer analysis model, and output the predicted fault location results and early warning information; the process status monitoring layer also predicts and monitors abnormal process data based on the constructed process status model; the fault type identification layer predicts the fault type of abnormal process data based on a machine learning classification model; the responsibility location layer predicts the final responsible process based on the mapping relationship between potential faults and processes, combined with the logical relationship between processes.

[0080] In one specific embodiment, the system further includes: The process fault monitoring application module 107 is also used to synchronously migrate the multi-layer analysis model built using the same type of terminal assembly equipment to the new terminal assembly equipment, and to use historical process data without differences for model training and parameter setting; when the terminal assembly equipment undergoes a cycle switch, the constructed process state model, machine learning classification model, and multi-layer analysis model are replaced with corresponding cycle models; the corresponding cycle model replacement means that for each cycle, a corresponding cycle process state model, machine learning classification model, and multi-layer analysis model are constructed; federated learning transfer is performed for multiple terminal assembly equipment of the same type and cycle to complete the model iteration of the multi-layer analysis model.

[0081] In one specific embodiment, the system further includes: The real-time process fault monitoring module 104 is also used to acquire historical fault case library, troubleshooting rule library and troubleshooting solution library, and calculate the priority score of each type of fault by comprehensively considering the fault urgency, occurrence frequency and processing efficiency, and obtain the priority ranking of fault types of different processes of the terminal assembly equipment; in the process of real-time monitoring of abnormal process parameters and determining the fault responsible process through multi-layer analysis model, the fault location results and early warning information are output according to the priority ranking of fault types.

[0082] This application also discloses a computer-readable storage medium.

[0083] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the aforementioned machine learning-based terminal assembly equipment fault warning method. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] This application also discloses a computer device.

[0085] Specifically, the computer device includes a memory and a processor, and the memory stores a computer program that can be loaded and executed by the processor to perform the aforementioned machine learning-based fault prediction method for terminal assembly equipment.

[0086] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A fault early warning method for terminal block assembly equipment based on machine learning, characterized in that, include: The process parameters, status data, and image data of each step in the terminal assembly process are collected using sensors and computer vision devices, and the process is marked. Extract feature vectors from each process to construct a process state model; the process state model includes feature vector distribution templates for normal and fault states. Fault types are identified by pre-built machine learning classification models, and a mapping relationship between faults and processes is established. A multi-layer analysis model is used to monitor abnormal process parameters in real time and determine the responsible process for the fault, outputting fault location results and early warning information. The multi-layer analysis model includes a process status monitoring layer, a fault type identification layer, and a responsibility location layer. The process status monitoring layer determines abnormal process data based on the constructed process status model. The fault type identification layer determines the fault type of the abnormal process data based on a machine learning classification model. The responsibility location layer determines the final responsible process based on the mapping relationship between faults and processes, combined with the logical relationship between processes.

2. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: Based on the quantity of materials in the feeding process, the number of crimping operations in the crimping process, and the coupling characteristics of each process, the terminal assembly scenarios are divided into simple, medium, and complex scenarios, in descending order of complexity. The acquisition devices and frequencies for acquiring process parameters, status data, and image data of each step in the assembly process of the terminal block are configured differently according to different scenarios. The more complex the scenario, the higher the accuracy and frequency of the corresponding acquisition devices. Differentiated adaptation is made for each process of feature vector extraction dimension, machine learning classification model complexity, fault library and mapping rules for different scenarios. The more complex the scenario, the higher the corresponding feature vector extraction dimension, the higher the fault recognition model complexity, the more fault types in the fault library and the more complex the mapping rules. The analysis levels of the multi-level analysis model are configured according to different scenarios, and the lower the scenario complexity, the lower the application level of the analysis level model.

3. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: Based on the full life cycle of the terminal assembly equipment, the current life cycle of the terminal assembly equipment is determined in real time; based on the current life cycle, the application mode of the computer vision device is adapted, including the first application mode adapted to the initial stage, the second application mode adapted to the stable stage, and the third application mode adapted to the decline stage. The first application mode includes: using Canny edge detection to extract the edges of the positioning platform baseline, mold closing surface, and terminal pins; expanding the edge contour through morphological expansion for mold gaps; fitting a straight line using Hough linear transform for the positioning platform baseline edge and calculating the angle between the fitted line and the standard baseline; filling the micro-holes in the mold gap using morphological reconstruction. The second application mode includes: using Sobel edge detection to extract the terminal contour and performing template matching with the pre-stored standard posture contour; calculating the pixel displacement of the contact area between the pressure head and the terminal using optical flow for the pressing time-series images acquired by the high-speed line scan camera; removing reflective spots on the terminal surface using morphological opening operation, and then extracting scratches and dents using Laplacian edge detection. The third application mode includes: segmenting the wear area using Otsu thresholding, and then filling the micro-holes in the wear area using morphological closing operation; calculating the area ratio, average depth, and maximum width of the segmented wear area; and calculating the wear area offset at each time point for the time-series images using template matching.

4. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: Extract the time sequence feature vectors of each process to construct a process state model; the process state model also includes a feature vector distribution template for potential faults. Predict fault types by pre-built machine learning classification models and establish a mapping relationship between potential faults and processes; The system uses a multi-layered analysis model to predict abnormal process parameters and the responsible process for failures in real time, outputting the predicted fault location results and early warning information. The process status monitoring layer also predicts and monitors abnormal process data based on the constructed process status model. The fault type identification layer predicts the fault type of abnormal process data based on a machine learning classification model. The responsibility location layer predicts the final responsible process based on the mapping relationship between potential faults and processes, combined with the logical relationship between processes.

5. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: For the new terminal block assembly equipment, the multi-layer analysis model built using the same type of terminal block assembly equipment is synchronously migrated to the new terminal block assembly equipment, and the model is trained and parameters are set using historical process data without difference. In response to the cycle switching of the terminal assembly equipment, the constructed process state model, machine learning classification model, and multi-level analysis model are replaced with corresponding cycle models; the corresponding cycle model replacement means that for each cycle, a corresponding process state model, machine learning classification model, and multi-level analysis model are constructed. Federated learning transfer was performed on multiple terminal block assembly devices of the same model and cycle to complete the model iteration of the multi-layer analysis model.

6. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: The establishment of the mapping relationship between faults and processes adopts a hierarchical mapping, including process anomaly layer mapping, equipment fault layer mapping, and process data anomaly layer mapping. Each level contains the feature description and judgment rules of each fault, and each level establishes a mapping relationship through a unique code and association weight; the association weight is obtained through historical data statistical calculation.

7. The fault early warning method for terminal assembly equipment based on machine learning according to claim 1, characterized in that, Also includes: Obtain historical fault case library, troubleshooting rule library and troubleshooting solution library, and calculate the priority score of each type of fault by comprehensively considering fault urgency, occurrence frequency and processing efficiency, and obtain the priority ranking of fault types of different processes of terminal assembly equipment; In the process of real-time monitoring of abnormal process parameters and identification of faulty processes through a multi-layer analysis model, the fault location results and early warning information are output according to the priority of fault types.

8. A fault early warning system for terminal assembly equipment based on machine learning, characterized in that, include: The process data acquisition and marking module is used to acquire process parameter data, status data, and image data of each process in the terminal assembly process through sensors and computer vision devices, and to mark the process. The process status model construction module is used to extract feature vectors of each process to construct a process status model; the process status model includes feature vector distribution templates for normal and fault states. The process fault identification and mapping module is used to identify fault types and establish a mapping relationship between faults and processes through a pre-built machine learning classification model; The real-time process fault monitoring module is used to monitor abnormal process parameters in real time and determine the responsible process for the fault through a multi-layer analysis model, and output fault location results and early warning information. The multi-layer analysis model includes a process status monitoring layer, a fault type identification layer, and a responsibility location layer. The process status monitoring layer determines the abnormality of monitored process data based on the constructed process status model. The fault type identification layer determines the fault type of the abnormal process data based on a machine learning classification model. The responsibility location layer determines the final responsible process based on the mapping relationship between faults and processes, combined with the logical relationship between processes.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images