A product board fault prediction and classification method based on electrical performance big data

Abstract

This invention discloses a method for product board fault prediction and classification based on big data of electrical performance, specifically relating to the field of electronic engineering. It addresses the problem of existing technologies struggling to uniformly utilize multi-source electrical performance data and maintenance records for future fault prediction and classification. The method involves aggregating various information in a management system to obtain an electrical performance timeline, extracting electrical performance feature sequences from the timeline, and associating them with process information and maintenance results to form a lifecycle dataset. Based on process information and operational stages, fault prediction and classification models are trained in groups to construct a model library. Upon receiving new board electrical performance data, the method updates features, combines stage degradation comparison parameters and batch anomaly aggregation parameters, and uses a coefficient learning model to assess the risk impact, dynamically adjusting the risk level and fault type results output by the selected fault prediction model. Finally, it records actual faults and maintenance results, compares and adjusts the results, and updates the lifecycle dataset and model library to generate production operation and maintenance management decision-making information.

Description

Technical Field

[0001] This invention relates to the field of electronic engineering, and more specifically, to a method for predicting and classifying product circuit board faults based on big data on electrical performance. Background Technology

[0002] In the production and maintenance of electronic products, product circuit boards generate a large amount of electrical performance-related data at multiple stages, including incoming material inspection, production testing, aging tests, and field operation. Simultaneously, rework, replacement, and quality analysis processes also generate corresponding fault records and handling information. Enterprises typically use production management systems, quality management systems, and maintenance management platforms to centrally manage the inspection results, repair status, and operational status of these circuit boards. The aim is to leverage big data on electrical performance to identify high-risk circuit boards in advance and manage them according to fault type, thereby facilitating maintenance planning, optimizing production schedules, and allocating spare parts. Existing technologies, such as CN109738782A, "Method and System for Circuit Board Fault Detection Based on Measurement Data and Machine Learning," can utilize historical measurement data and repair results to diagnose faults in circuit boards already identified as defective or under inspection, improving repair efficiency and diagnostic accuracy to a certain extent. Such solutions have already been applied in engineering practice.

[0003] However, existing technologies still present a crucial technical problem closely related to the title of this invention: they primarily target individual circuit boards that have already exhibited abnormalities or entered the repair phase. They focus on providing current fault assessments based on existing measurement data, making it difficult to integrate and utilize the large amounts of electrical performance data and repair data accumulated across different stages within a management system. Furthermore, they struggle to provide timely fault predictions and fault type classifications for the entire circuit board group. More specifically, existing solutions typically only indicate the current fault type or qualification status of a particular circuit board. They cannot predict the likelihood and possible fault types of a board that will subsequently fail while it is still operating normally or has just completed testing, based on multi-source electrical performance data, and thus cannot provide production and maintenance management systems with decision-making basis that can be directly used for production scheduling adjustments, planned maintenance, and spare parts management. Therefore, the core technical problem this invention aims to solve is how to propose a product circuit board fault prediction and classification method based on large amounts of electrical performance data. This method should utilize production test data, operational monitoring data, and repair records within the same hardware and software framework to output circuit board fault risk predictions and fault type classifications that can be directly accessed by the management system.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a product board fault prediction and classification method based on big data of electrical performance. This method involves aggregating various information in a management system to obtain an electrical performance timeline, extracting electrical performance feature sequences from the timeline and associating them with process information and maintenance results to form a lifecycle dataset, training fault prediction and classification models based on process information and operational stages to construct a model library, receiving new board electrical performance data, updating features, combining stage degradation comparison parameters and batch anomaly aggregation parameters, using a coefficient learning model to assess risk impact, dynamically adjusting the risk level and fault type results output by the selected fault prediction model, recording actual faults and maintenance results, comparing and adjusting the results, updating the lifecycle dataset and model library, and generating production operation and maintenance management decision information to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: S1. In the management system, summarize the test data, running data and maintenance records according to the board identifier, and generate the corresponding electrical performance timeline; S2. Extract electrical performance feature sequences from the electrical performance timeline, and associate the electrical performance feature sequences with process information and maintenance results to form a lifecycle dataset; S3. Group the lifecycle dataset according to process information and operation stage, train a fault prediction classification model based on the electrical performance characteristic sequence and maintenance results of each group, and build a model library; S4. After receiving the electrical performance data of the new board, update the features, combine the changes in its own stage and the abnormal characteristics of the batch, and evaluate its risk impact through the learning model, and dynamically adjust the risk level and fault type results output by the selected fault prediction model. S5. Record the actual faults and maintenance results of the new board, compare the actual results with the corrected fault risk level and fault type prediction results, and update the lifecycle dataset and model library based on the comparison results to generate production operation and maintenance management decision information.

[0007] Furthermore, step S1 includes summarizing test data, operation data, and maintenance records using the board's unique identifier as the primary key; extracting timestamps, process nodes, operation stage markers, and multidimensional electrical performance measurement values ​​to form a preliminary record set; constructing an electrical performance timeline by sorting the timestamps in ascending order; inserting state transition records; establishing a unified data structure; performing time alignment; calculating missing values ​​or nearest neighbor resampling selection values ​​through linear interpolation to ensure the completeness of multidimensional electrical performance measurement values; and storing the electrical performance timeline, process information, and operation stage.

[0008] Furthermore, step S2 includes segmenting the electrical performance timeline according to the operating phase label based on the operating phase division rules, serializing the multidimensional electrical performance measurement values ​​through a sliding time window to form a segment sequence, extracting the electrical performance curve shape change, the duration of rising and falling periods, the duration of exceeding the range, and the relative change relationship of channels from the segment sequence and concatenating them to form an electrical performance feature sequence, associating it with process information and maintenance results to generate life cycle records, and combining all board life cycle records to form a life cycle dataset.

[0009] Furthermore, step S3 includes grouping the lifecycle dataset according to process batch, usage environment, and operation stage; using the electrical performance feature sequence within the group as input and maintenance results as supervision labels; selecting a recurrent neural network model or random forest ensemble learning model based on sequence length or feature dimension to map features to fault risk level and fault type probability distribution; registering the model and recording the process batch range, usage environment range, and operation stage conditions to build a model library.

[0010] Further, step S4 includes receiving new board test data and running data, inserting electrical performance timelines and updating electrical performance feature sequences, calculating stage degradation comparison parameters by comparing the pattern difference ratio point by point between historical stage change sequences and current stage change sequences, summing the absolute difference exceeding the limit difference ratio and dividing by the number of channels, calculating batch anomaly aggregation parameters by comparing the similarity of anomaly pattern vectors to select similar board sets and statistically analyzing the failure ratio, inputting coefficients to learn the model and outputting adaptive adjustment coefficients, selecting a fault prediction classification model from the model library to generate a basic risk score sequence and a basic confidence distribution, amplifying and compressing the basic risk score and moving the boundary score to obtain the risk level by using adaptive adjustment coefficients to offset the confidence distribution and normalize it to select the main judgment fault type and alternative fault types to obtain the fault type prediction result.

[0011] Furthermore, the calculation of the phase degradation comparison parameters selects the time alignment of the electrical performance characteristic sequences of the historical operation phase and the current operation phase. For each key electrical performance channel, the characteristic points at the same position are taken out to calculate the change direction to form the historical phase change sequence and the current phase change sequence. The statistical pattern difference ratio is compared point by point, and the absolute difference of the statistical over-limit ratio is taken as the over-limit difference ratio. The normalized value is obtained by summing by channel and dividing by the number of channels.

[0012] Furthermore, the batch anomaly clustering parameter calculation involves extracting the anomaly pattern vector of each board within the batch group, comparing the target board's anomaly pattern vector with other board vectors, calculating the number of intersection elements divided by the number of union elements as the similarity, selecting the boards with the highest similarity to form a set of similar boards, and counting the number of faulty boards in the set divided by the total number to obtain the parameter value.

[0013] Furthermore, step S5 includes receiving operational data and maintenance records to add lifecycle records, comparing the predicted results of the adjusted fault risk level and fault type with the actual fault calculation deviation and consistency, writing back the lifecycle dataset to trigger retraining of the fault prediction classification model and coefficient learning model, and using the updated lifecycle dataset and model library to generate production scheduling adjustment information, maintenance plans and spare parts preparation strategies.

[0014] Furthermore, the key electrical performance channels are extracted from multidimensional electrical performance measurements based on circuit design specifications and historical maintenance records. The correlation score of each channel in the lifecycle dataset is calculated, and the linear correlation between the measured values ​​and the severity of the fault is analyzed by Pearson correlation coefficient analysis. Channels with correlation scores higher than a preset threshold are selected.

[0015] The technical effects and advantages of the product board fault prediction and classification method based on big data of electrical performance of this invention are as follows: The method of this invention combines the construction of electrical performance timelines with feature sequence extraction to achieve unified integration of multi-source data, forming a lifecycle dataset to support group model training. This ensures that the fault prediction classification model captures the dynamic patterns of the board throughout its entire lifecycle from production to operation and maintenance. As a result, potential risks can be identified and classified in advance during the normal operation of the board. This surpasses the limitations of existing technologies that are limited to anomaly diagnosis, and provides directly callable decision information to optimize production scheduling and spare parts configuration, significantly reducing production interruptions and economic losses caused by sudden failures.

[0016] Furthermore, this method utilizes the calculation and coefficient adjustment mechanism of stage degradation comparison parameters and batch anomaly aggregation parameters to dynamically correct the prediction output. After comparing actual faults, it updates the model library through incremental retraining, forming an adaptive closed-loop process. This improves the predictive adaptability to process variations and environmental changes, supports enterprises in achieving data-driven operation and maintenance management, extends the service life of circuit boards, and improves the overall system reliability. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a product board fault prediction and classification method based on big data of electrical performance, according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: Figure 1This invention presents a method for predicting and classifying product circuit board faults based on big data on electrical performance, comprising: S1. In the management system, summarize the test data, running data and maintenance records according to the board identifier, and generate the corresponding electrical performance timeline.

[0020] S2. Extract electrical performance feature sequences from the electrical performance timeline, and associate the electrical performance feature sequences with process information and maintenance results to form a lifecycle dataset.

[0021] S3. Group the lifecycle dataset according to process information and operation stage, train a fault prediction classification model based on the electrical performance characteristic sequence and maintenance results of each group, and build a model library.

[0022] S4. After receiving the electrical performance data of the new board, update the features, combine the changes in its own stage and the abnormal characteristics of the batch with the learning model to assess its risk impact, and dynamically adjust the risk level and fault type results output by the selected fault prediction model.

[0023] S5. Record the actual faults and maintenance results of the new board, compare the actual results with the corrected fault risk level and fault type prediction results, and update the lifecycle dataset and model library based on the comparison results to generate production operation and maintenance management decision information.

[0024] Product circuit boards generate massive amounts of electrical performance data throughout the electronic product lifecycle. This data is scattered across testing platforms, online monitoring systems, and maintenance systems, forming isolated records that cannot directly reveal potential failure modes. Without systematic aggregation and time-series integration, this data is difficult to transform into predictive insights, thus limiting the optimization potential of production operations. Aggregating multi-source data using the unique identifier of the circuit board as the primary key and constructing a continuous electrical performance timeline can capture the dynamic changes of the circuit board from production to operation, providing a solid foundation for subsequent feature extraction and fault model training. This process not only unifies the data format but also enhances the depth of analysis through state transition records, ensuring that predictive models can assess risks based on complete historical data and achieve proactive management of the circuit board group.

[0025] S1-1 Original Record Summary Processing.

[0026] Using the board's unique identifier as the primary key, the system aggregates and processes test data from the testing platform, operational data from the online monitoring system, and maintenance records from the maintenance system. First, each data source is scanned to verify the existence of the board's unique identifier, and all records are associated with the corresponding board entity. Next, measurement results from incoming material inspection, production testing, and aging tests are extracted from the test data, including timestamps, process nodes, operational stage markers, and multi-dimensional electrical performance measurements such as voltage, current, power, and temperature. Real-time monitoring values ​​from the operational data are extracted, including identical fields. Finally, fault occurrence time, maintenance measures, and handling information are extracted from the maintenance records, including fault type tags and maintenance time.

[0027] For missing data, data is filled according to predefined mapping rules. For example, if a timestamp is missing, the default timestamp from the data source or the average timestamp of adjacent records is used to calculate the filler value. For inconsistent formats, conversion is performed according to predefined mapping rules. For example, the local time format of the test platform is converted to a Coordinated Universal Time (UTC) timestamp, or string representations of electrical performance measurements are converted to floating-point numbers. During the aggregation process, all extracted records are integrated into a single set, grouped by the board's unique identifier to form a preliminary record set. This set contains all relevant fields, providing input for subsequent sorting.

[0028] S1-2 constructs the electrical performance timeline.

[0029] From the initial record set, all records are sorted in ascending order by timestamp to construct an electrical performance timeline with time as the main axis. First, the timestamp of each record is validated, removing outliers such as future timestamps or negative values. Then, the set is sorted using a quicksort algorithm, with the timestamp as the sort key. If timestamps are the same, they are sorted according to the priority of the data source type, defined as test data above operational data, and operational data above maintenance records. This priority is determined based on the order of data generation to ensure that production data is presented first.

[0030] After sorting, the ordered records are concatenated into an electrical performance timeline. Each node corresponds to a specific time point and includes a unique board identifier, timestamp, data source type, process information marker, and multi-dimensional electrical performance measurement values. Between adjacent time points, the time difference is calculated. If the process information marker or operational phase marker differs, a state transition record is inserted, indicating the transition type (e.g., process change, operation mode conversion, or operational mode change), and recording the values ​​before and after the transition. This concatenation is achieved through a linked list structure, ensuring the temporal integrity of the timeline.

[0031] S1-3 Establishment of a unified data structure.

[0032] A unified data structure is established for each time point in the electrical performance timeline, including at least a unique board identifier, timestamp, data source type, process information marker, and multidimensional electrical performance measurements. Each time point is represented as a dictionary object, with the board identifier as a fixed key-value pair, the timestamp represented as a floating-point number with millisecond precision, the data source type distinguished by enumerated values ​​for test platform, online monitoring system, or maintenance system, the process information marker including batch number, usage environment, and process node, and the multidimensional electrical performance measurements stored as a vector, arranged in channel order: voltage, current, power, and temperature values. During the establishment process, for state transitions between adjacent time points, the time difference is calculated and the change is determined: if the process information markers are different, an intermediate record is inserted, indicating the transition type and the values ​​before and after; if the markers for the operational phases are different, similar processing is performed. This structuring is achieved by traversing the timeline and applying template filling to ensure consistent formatting for each time point.

[0033] S1-4 Time Alignment Processing.

[0034] Time alignment is performed on data sources with different sampling frequencies. This is achieved through linear interpolation or nearest-neighbor resampling to ensure that multidimensional electrical performance measurements from different pins and measurement points have a complete vector on the same time scale. First, the sampling frequency of all data sources is calculated, and the highest frequency is taken as the reference sampling frequency. Then, for low-frequency data sources, linear interpolation is used to calculate values ​​at missing time points: two adjacent time points are designated as an earlier and a later timestamp, corresponding to electrical performance measurements at the earlier and later timestamps, respectively; the interpolation value at the intermediate time point is the measurement at the earlier timestamp + (the measurement at the later timestamp - the measurement at the earlier timestamp) × (the intermediate time minus the earlier timestamp) / (the later timestamp minus the earlier timestamp). For high-frequency data sources, nearest-neighbor resampling is used, selecting the value from the most recent time point. During the alignment process, multidimensional electrical performance measurements are calculated independently for each channel to avoid cross-channel interference, and the position of the state transition record is updated. This alignment is achieved by iterating the timeline and inserting calculated values ​​to ensure continuity.

[0035] S1-5 Processing Completed and Stored.

[0036] After processing, a continuous electrical performance timeline depicting the changes in the board's electrical performance over time is obtained and stored uniformly with the corresponding process information and operational stages. The electrical performance timeline is serialized into a database table format, with the data structure of each time point as a row record and the board's unique identifier as an index. Process information markers and operational stage markers are embedded in the structure of each time point. The operational stage markers are divided into startup stage, stable operation stage, load fluctuation stage, and shutdown stage, and are associated with state transition events. A distributed database is used for storage, supporting large-scale data queries, and efficiency is optimized through timestamp indexing. Simultaneously, metadata including data source coverage, alignment method, and checksum is recorded, and data integrity is verified through periodic comparisons using this metadata. This storage is implemented through batch insert operations, forming a reusable data asset.

[0037] The completed electrical performance timeline integrates multi-source data from the circuit boards, forming a continuous time-series view. This foundation directly supports subsequent steps such as feature sequence extraction and lifecycle dataset generation. Through time alignment and state transition recording, the timeline not only reflects the dynamic changes in electrical performance but also captures the correlation between process and operational phases, ensuring that the fault prediction model can utilize complete historical information to output risk level and type results. This output can be accessed by the management system for production scheduling adjustments and maintenance plans, enabling early intervention and resource optimization for circuit board faults.

[0038] Step S2, based on segmentation and sliding window serialization using operational phase labels, extracts features such as curve shape, time duration, and channel relationships to form an electrical performance feature sequence. This sequence is then correlated with process information and maintenance results to generate a complete lifecycle record combined into a dataset. This process bridges time-series data to trainable resources, supporting step S3 in building adaptive models through grouping, enabling group prediction and hierarchical management of potential faults.

[0039] S2-1 Operational phase segmentation.

[0040] Based on pre-configured operational phase segmentation rules, the electrical performance timeline is segmented according to operational phase labels such as start-up phase, stable operation phase, load fluctuation phase, and shutdown phase.

[0041] First, the entire electrical performance timeline is traversed, examining the operational phase markers at each point. These markers originate from embedded process information markers and state transition records. Then, the time point sequence is matched according to rules. For example, when the operational phase marker changes from none to the startup phase, that point is used as the starting point for a segment. When the marker switches to the stable operation phase, the previous segment ends and is marked as the startup phase. Rules include time threshold judgments (determining automatic switching by comparing the duration within a segment with a preset upper limit) and state transition event triggers (directly defining segment boundaries by detecting process changes or operational mode changes). During segmentation, each segment is assigned a unique segment identifier, and the start timestamp, end timestamp, and number of time points within the segment are recorded. If boundaries overlap, the boundary is adjusted primarily based on state transition events; if a time point is missed, it is inserted into the nearest segment. After segmentation, a segment sequence is formed, with each segment corresponding to an operational phase label, serving as input for time segment serialization.

[0042] S2-2 time segment serialization.

[0043] For each time segment within an operational phase, multi-dimensional electrical performance measurements such as voltage, current, power, and temperature are serialized using a sliding time window, forming a series of segment sequences with a fixed order and length. First, a fixed window length and step size are selected. The window length is determined based on a proportion of the total segment duration, and the step size is a fixed fraction of the window length to cover variations within the segment. Then, starting from the segment's beginning timestamp, the sliding window collects multi-dimensional electrical performance measurements at all time points within each window, arranging them into a vector sequence in chronological order, ensuring each segment sequence has the same length. If the window is insufficient at the end, zero-padding is used to extend the sequence, or it is truncated to a standard length, but the step size is adjusted to preserve complete data. During serialization, multi-dimensional values ​​are separated by channel; for example, voltage channels are independently serialized into sub-sequences to preserve channel independence. After processing, each operational phase corresponds to multiple segment sequences, which serve as input for feature extraction.

[0044] S2-3 Feature extraction and concatenation.

[0045] For each segment sequence, features reflecting changes in the board's operating status are extracted. For example, the shape changes of the electrical performance curve are recorded within the same window, and the time-domain sequence is converted into frequency-domain components through Fourier transform to analyze the main frequency and amplitude. The duration of rising and falling periods is determined by detecting changes in the slope sign of adjacent points and accumulating the number of consecutive points with positive or negative slopes as the duration. The duration exceeding the process setting range is determined by comparing each point value with preset upper and lower limits and accumulating the number of consecutive exceedance points, which is then converted into time units. The relative changes between each electrical performance channel are calculated using the covariance matrix to form a symmetric matrix representing the correlation.

[0046] These features are concatenated in chronological order to form the electrical performance feature sequence of the board at the corresponding stage. First, features are independently calculated for each segment sequence to form feature vectors. Then, these vectors are concatenated in segment order to form a one-dimensional sequence, the length of which is the number of segments multiplied by the feature dimension of each segment. During extraction, a predefined feature list is used to cover all channels, and padding is used to align if the dimensions are inconsistent. After concatenation, the electrical performance feature sequence serves as a compressed representation of that stage and is used as input for the correlation operation.

[0047] S2-4 associates and generates lifecycle records.

[0048] The electrical performance characteristic sequence is associated with the board's process information and the fault type label, maintenance time, and maintenance measures in the maintenance record to generate a complete life cycle record for each board.

[0049] First, the unique identifier of the circuit board is matched, and the corresponding process information tags, including batch number and usage environment, are retrieved from the electrical performance timeline. Then, fault type tags such as short circuit or overheating, maintenance timestamps, and action descriptions such as component replacement are extracted from the maintenance records. These elements are appended to the end of the electrical performance feature sequence for each operating stage. By expanding the sequence dimension, the process information tags are encoded as numerical vector appended, the fault type tags are converted into category index appended, and the maintenance time and action are appended as string vectors, forming a unified extended vector.

[0050] During association, ensure time consistency by appending only records whose maintenance time is later than the sequence end timestamp. If no match is found, fill the extension with a zero vector or an empty string. After generation, each lifecycle record contains a unique board identifier, process information, operating stage, electrical performance characteristic sequence, and maintenance result, serving as an independent unit.

[0051] S2-5 are combined to form a lifecycle dataset.

[0052] The lifecycle records of all circuit boards are combined to form a lifecycle dataset that can be directly used for modeling. Each record contains a unique board identifier, process information, operating stage, corresponding electrical performance characteristic sequence, and final maintenance result. First, all circuit boards are traversed, their lifecycle records are collected, sorted by board identifier, and then stored in a database table, with one record per row. During combination, the completeness of the records is verified; if a field is missing, the corresponding value is supplemented by matching the board identifier from the electrical performance timeline or maintenance records. The dataset is in a structured table format, with each column corresponding to a field; for example, the electrical performance characteristic sequence column is stored as serialized binary data. After combination, this lifecycle dataset serves as the input for step S3.

[0053] The lifecycle dataset output in step S2 is organized by unique board identifier. Each record integrates electrical performance feature sequences, process information, operating stages, and maintenance results to form a structured modeling foundation. This data is directly imported into step S3 to train the fault prediction and classification model by batch environment grouping. Through feature refinement and result correlation, the dataset reveals degradation trajectories and commonalities in anomalies, ensuring that the model outputs accurate risk levels and fault types. This supports dynamic adjustments to production scheduling, maintenance plans, and spare parts strategies, driving closed-loop optimization of board operation and maintenance.

[0054] Although the lifecycle dataset provides board features and result records through step S2, it cannot cope with stage variations and group differences if the model is not trained according to process batches and usage environments, resulting in a lack of adaptability in the prediction model. Step S3 involves traversing the records to extract process information and operational stage labels for grouping, separating the input sequence from the supervision label, selecting a recurrent neural network or random forest structure to map features to risk level and type probability, and optimizing parameters to build a model library. This process integrates historical bias adjustments to ensure that the model outputs personalized predictions for batch environments, supporting the dynamic risk assessment in step S4, and enabling early classification of board failures and optimization of operation and maintenance.

[0055] S3-1 Grouping Processing.

[0056] Based on the process batch, usage environment, and operational stage in the process information, boards of the same type are grouped. Boards belonging to the same process batch, similar usage environment, and the same operational stage are grouped into the same analysis group. First, all records in the lifecycle dataset are traversed, and the process information of each record is extracted, including the process batch number, usage environment description, and operational stage label, such as startup stage or stable operation stage. Then, the usage environment description is converted into a numerical vector, keywords are quantified using a word embedding model, and K-means clustering is applied to determine similar environment clusters. Records matching the process batch number, environment cluster label, and operational stage label are assigned to the corresponding group. A unique group identifier is assigned to each group, and the number of records within the group is counted. If the number of records in a group is less than the preset minimum sample size, the similarity between groups is calculated, and records are integrated by merging the group with the nearest cluster center. After grouping, multiple analysis groups are formed, each containing a subset of lifecycle records, which serve as input for model training.

[0057] The process of converting the usage environment description into a numerical vector, quantifying keywords using a word embedding model, and then applying K-means clustering to determine similar environment clusters, assigning records matching the process batch number, environment cluster label, and operating stage label to the corresponding group, is as follows: First, for each record in the lifecycle dataset, the usage environment description field is extracted. This field is a text string containing keywords describing the board's operating environment, such as high temperature, high humidity, or industrial environment. Then, the usage environment description is split into a keyword list, and individual words are separated using word segmentation, for example, using spaces or punctuation as delimiters, forming a keyword sequence. Subsequently, a word embedding model such as Word2Vec is applied to quantify the keyword sequence. This model is pre-trained on relevant domain corpora, mapping each keyword to a fixed-dimensional vector, such as a 300-dimensional floating-point vector, representing the semantic embedding of the keyword. The vector is calculated through the model's hidden layer, ensuring that similar keyword vectors are close in Euclidean space. The average of all keyword vectors is taken to obtain the overall numerical vector of the usage environment description, which serves as the quantified representation of the environment. Next, the K-means clustering algorithm is applied to the set of environmental numerical vectors for all records. First, K cluster centers are initialized by randomly selecting K vectors as starting points. Then, each vector is iteratively assigned to the nearest cluster center using Euclidean distance. The cluster centers are updated to the mean of the vectors within the cluster until the cluster assignment is stable or the number of iterations reaches a preset upper limit. The value of K is preset based on the dataset size or determined using the elbow method, thereby generating environmental cluster labels. Each cluster label is an integer identifier representing a group of similar usage environments. Finally, the lifecycle dataset is traversed. For each record, its process batch number is checked to see if it equals the batch number defined for the group, if the environmental cluster label matches the group's cluster identifier, and if the operating stage label is the same as the group's stage label. If all matches, the record is assigned to the corresponding analysis group by appending it to the group's record list, ensuring that the records within a group only contain a subset that meets all the conditions.

[0058] S3-2 model training.

[0059] For each group of lifecycle records, the electrical performance feature sequence is used as input, and the fault category, fault severity, and fault occurrence status from the maintenance records are used as supervision labels. A supervised learning algorithm is then used to train the corresponding fault prediction classification model. First, the input and labels are separated from the group records. The input is a matrix of electrical performance feature sequences, with each row corresponding to a sequence vector of one record. The labels are multi-dimensional vectors, including one-hot encoding of the fault category, a numerical value for severity, and a binary label indicating fault occurrence. Then, the model structure is selected based on the group characteristics. For example, if the sequence length exceeds a preset standard, a recurrent neural network model is chosen to handle time dependencies; if the feature dimension exceeds a preset standard, a random forest ensemble learning model is chosen to handle non-linear relationships, mapping the electrical performance feature sequences to fault risk levels and fault type probability distributions. During training, the cross-entropy loss function is used to calculate the deviation between the predicted probability distribution and the actual labels. Gradient descent is used to optimize and adjust the model parameters, iterating until the loss converges. During training, the group records are divided into training and validation sets proportionally, and an early stopping mechanism is applied to monitor the validation loss. After training, each group obtains an optimized fault prediction classification model, which serves as the registration basis.

[0060] For recurrent neural network models, in product board fault prediction scenarios, this model combines the processing of the time dependencies of electrical performance feature sequences with big data on electrical performance to predict the likelihood of subsequent board faults. Specifically, the model input is an electrical performance feature sequence, where each sequence is a two-dimensional tensor with the dimension being the number of time steps multiplied by the feature dimension. For example, the number of time steps corresponds to the number of segment sequences, and the feature dimension corresponds to extracted shape changes, duration of time periods, etc. The output is a softmax probability distribution of fault risk level and a fault type probability distribution. The multidimensional vector represents the risk level (e.g., low, medium, high) and the fault type (e.g., short circuit or overheating). The model structure includes an input layer that receives the sequence tensor, a long short-term memory layer as a hidden layer to process temporal dependencies (this layer has 128 units and uses a bidirectional structure to capture the preceding and following context), a fully connected layer for feature fusion (this fully connected layer has 64 neurons and uses the ReLU activation function), and finally, an output layer that uses the softmax function to generate the probability distribution. The necessary training steps include initializing model parameters, such as the weight matrix using Xavier initialization, setting the Adam optimizer with a learning rate of 0.001, a batch size of 32, calculating predictions through forward propagation, updating gradients through backpropagation, iterating through 50 epochs, and calculating the loss on the validation set after each epoch for adjustment. The model is integrated with the scenario by using electrical performance feature sequences as input to capture the temporal pattern of the board from normal to degradation, and the output directly corresponds to maintenance record labels, ensuring accurate prediction of the actual fault distribution of the associated board.

[0061] For the random forest ensemble learning model, in the scenario of product board failure prediction, this model integrates multiple decision trees to process the nonlinear relationship of high-dimensional electrical performance feature sequences, and combines it with multi-source big data to achieve risk classification of board groups. Specifically, the model input is a flattened electrical performance feature sequence vector, with the dimension being the concatenation of all segment features; the output is the category probability and failure type probability of the failure risk level, with multiple labels corresponding to severity and category. The model structure includes multiple decision trees, each tree based on Gini impurity split nodes, with a maximum tree depth of 20 and a forest of 100 trees. Subsets of each tree are constructed through bootstrapping, and feature subsets are randomly selected for splitting to reduce correlation. The necessary training steps are to randomly sample a subset from the training set, construct each tree independently, recursively split nodes until the purity of the leaf nodes reaches the target or the depth limit is reached, aggregate all tree predictions, and obtain the final output through majority voting or average probability. The minimum number of sample splits is set to 2, and the minimum number of leaf samples is set to 1. The model and scenario are combined to capture the relationship between relative changes between channels and the duration of exceeding limits by inputting high-dimensional features, and the output is mapped to maintenance result labels to ensure that the classification reflects the abnormal clustering of board batches and stage degradation.

[0062] S3-3 Model Registration and Model Library Construction.

[0063] After training, a corresponding fault prediction classification model is registered for each group, and a model library is established. The model library records the process batch range, usage environment range, and operating stage conditions corresponding to each model. First, a unique model identifier is generated for each model, and its parameters are serialized into a file for storage. Then, an entry is created in the model library database, including the model identifier, process batch range (e.g., continuous intervals of batch numbers), usage environment range (e.g., a list of quantization cluster boundaries), and operating stage conditions (e.g., a list of enumerated labels). These ranges are based on the group definition, and similar variations are covered by expanding boundary values. During registration, model performance is verified. Accuracy is calculated as the proportion of correct predictions, and recall is calculated as the fault detection coverage proportion by retaining the test set. If the performance is lower than the preset standard, retraining is triggered using incremental samples. After construction, the model library serves as a lookup table, supporting model matching based on process information and operating stage.

[0064] The model library constructed in step S3 registers models by group range, forming query resources. These resources are directly supplied to step S4 for feature updates, where suitable models are selected to adjust the output, ensuring that risk levels and fault types reflect stage degradation and batch anomalies. Through supervised training and performance validation, the model library captures sequence dependencies and high-dimensional relationships, improving prediction robustness. This supports step S5 in updating the dataset and library by comparing actual results, achieving closed-loop learning to drive the generation of production scheduling information, maintenance plans, and spare parts strategies.

[0065] Although the model library is completed in step S3, providing a group-based fault prediction classification model, it cannot integrate the dynamic impact of stage degradation and batch anomalies without real-time feature updates and risk parameter assessments for new board data. This results in the output failing to reflect the actual risk evolution of individuals and groups. Step S4 receives test data and operational data, inserts them into the electrical performance timeline, updates the electrical performance feature sequence, calculates stage degradation comparison parameters by comparing pattern differences and exceedance differences point-by-point with historical and current sequences, calculates batch anomaly clustering parameters by statistically analyzing the fault proportion through anomaly pattern vector similarity, outputs adaptive adjustment coefficients through a coefficient learning model, selects models from the model library to generate basic results, and adjusts the risk level through scoring amplification / compression and boundary shifting, and the fault type through confidence shifting. This process bridges historical models with new data, ensuring that predictions adapt to production and maintenance scenarios, supports actual fault comparison in step S5, and achieves closed-loop optimization to improve the accuracy of board prediction.

[0066] S4-1 Data Reception and Feature Update.

[0067] When the test data and operation data of the new board are received, the new data is first inserted into the electrical performance timeline of the board in accordance with step S1.

[0068] First, match the unique identifier of the board, extract the timestamp, process node, operation stage mark and multi-dimensional electrical performance measurement value of the new data. After verifying that the timestamp has no future value or negative value, insert it into the electrical performance timeline position in ascending order of timestamp, update the state transition record between adjacent time points, mark the transition type such as process change or operation mode change, and record the values ​​before and after.

[0069] Then, the inserted timeline is time-aligned, and missing point values ​​are calculated through linear interpolation to ensure that the multidimensional electrical performance measurements are complete on a uniform scale. After insertion, the timeline is re-segmented according to the same operating phase division rules as in step S2. For each operating phase's time segment, a sliding time window is applied to serialize the multidimensional electrical performance measurements such as voltage, current, power, and temperature, forming a fixed-length segment sequence.

[0070] Then, for each segment sequence, the shape change of the electrical performance curve is extracted and analyzed using Fourier transform to determine the frequency components; the duration of the rising and falling periods is counted by the change in the slope sign; the duration exceeding the process setting range is accumulated by threshold comparison; and the relative change relationship of the channels is calculated using the covariance matrix to form a feature vector. These features are then concatenated in chronological order to form an updated electrical performance feature sequence. This sequence serves as the basis for subsequent parameter calculations.

[0071] After updating the electrical performance characteristic sequence, two parameters are calculated for the board to quantify its individual degradation trend and group risk clustering, thereby providing a basis for subsequent risk adjustment and avoiding bias caused by ignoring time-series changes and batch commonalities in the basic prediction.

[0072] First, the calculation of the stage degradation comparison parameters is based on the electrical performance characteristic sequence of the same board at different operating stages. The aim is to capture the evolution of the board's electrical performance pattern from history to the present through self-comparison, so as to ensure that the evaluation reflects the internal degradation dynamics.

[0073] Secondly, the calculation of batch anomaly clustering parameters is based on a group of boards in the same process batch and under the same usage environment. The aim is to evaluate the relative risk position of the target board in the same batch by comparing anomaly patterns, and to ensure that the evaluation incorporates the anomaly distribution of the group.

[0074] S4-2 stage degradation comparison parameter calculation.

[0075] In the electrical performance timeline of the board, a historical operating phase and a current operating phase are selected. The electrical performance feature sequences corresponding to these two phases are time-aligned to ensure that the two sequences have the same length on the same time scale. Nearest neighbor resampling is used to fill missing positions with the nearest value. For each key electrical performance channel, feature points at the same position are extracted in both the historical and current phases. The direction of change between adjacent feature points is calculated, with rising trends marked as positive one, falling trends as negative one, and essentially unchanged trends as zero, forming the historical phase change sequence and the current phase change sequence. Subsequently, the two change sequences at corresponding positions of the same channel are compared point by point. The number of inconsistent change directions divided by the total number of points is used as the mode difference ratio for that channel. Simultaneously, the number of feature points exceeding the normal operating range for that channel in both phases is divided by the total number of points to obtain the over-limit ratio. The absolute value of the difference between the two over-limit ratios is taken as the over-limit difference ratio for that channel.

[0076] The key electrical performance channels are determined based on the circuit design specifications and historical maintenance records of the product board. First, all channels, such as voltage, current, power, and temperature channels, are extracted from multi-dimensional electrical performance measurements. Then, the correlation score of each channel in the life cycle dataset is calculated. The linear correlation between the channel measurement value and the severity of the fault is analyzed by Pearson correlation coefficient. Channels with correlation scores higher than a preset threshold are selected as key electrical performance channels to ensure that the selected channels directly reflect the board degradation mode and cover the main fault types.

[0077] The modality difference ratio and the over-limit difference ratio of each channel are added together, summed for all channels, and then divided by the number of channels to obtain a normalized dimensionless value, which serves as the stage degradation contrast parameter. This parameter is calculated as: Stage Degradation Contrast Parameter = (Modality Difference Ratio of All Channels + Sum of Absolute Values ​​of Over-Limit Ratios in Historical Stages and Over-Limit Ratios in the Current Stage) / Number of Channels. This parameter serves as the input to the coefficient learning model.

[0078] S4-3 batch anomaly aggregation parameter calculation.

[0079] Within this batch group, abnormal patterns are extracted from the electrical performance characteristic sequence of each board. Based on predefined patterns such as voltage spikes detected by peak value, ripple anomalies filtered by frequency domain, and waveform distortion at specific stages matched by morphological matching, each abnormal pattern is marked as true or false in an abnormal pattern vector, ensuring that each board in the same batch corresponds to an abnormal pattern vector composed of multiple abnormal pattern markers. For the target board, its abnormal pattern vector is extracted and then compared sequentially with the abnormal pattern vectors of other boards in the same batch. The ratio of the number of intersection elements to the number of union elements is calculated as the abnormal pattern similarity. The Jaccard coefficient formula is used: Abnormal pattern similarity = Number of intersection elements of the target board's abnormal pattern vector element set / Number of union elements of the other board's abnormal pattern vector element sets, where all elements are Boolean values. All similarities are sorted in descending order, and boards with the highest pre-defined proportions are selected to form a similar board set. Subsequently, the number of boards that actually failed in the similar board set is counted and divided by the total number of boards in the set, serving as the batch anomaly aggregation parameter. This parameter serves as the input to the coefficient learning model.

[0080] S4-4 coefficient learning and basic prediction.

[0081] The stage degradation comparison parameters and batch anomaly aggregation parameters are input into a pre-trained coefficient learning model. This model uses these two parameters as input from historical data and the deviation between the basic prediction results and actual fault results as the training signal. Through supervised learning, it obtains a mapping relationship from the two-parameter space to a coefficient space. The model structure is a multilayer perceptron, including an input layer that receives parameter vectors, a hidden layer that processes nonlinear mappings, and an output layer that generates coefficient values. The coefficient learning model adaptively adjusts the coefficients for new boards. Subsequently, based on the board's process information and operating stage, a corresponding fault prediction classification model is selected from the model library. Using the updated electrical performance feature sequence as input, it generates basic fault risk level and fault type prediction results. Specifically, the model library is queried to match the process batch range, usage environment range, and operating stage conditions. The matching model is loaded, and the electrical performance feature sequence is input into the model for forward propagation. A basic risk score sequence is calculated as a numerical vector representing the risk level, and the basic confidence distribution of each candidate fault type is used as a probability vector. These basic results serve as the input for adjustment.

[0082] The coefficient learning model employs a multilayer perceptron structure. In product board failure prediction, it maps stage degradation comparison parameters and batch anomaly aggregation parameters to adaptive adjustment coefficients, thereby correcting the basic prediction bias. The specific structure includes an input layer receiving two parameter vectors of dimension 2, connected to two hidden layers with 64 neurons per layer. The ReLU activation function is used to handle non-linear relationships. Fully connected layers between hidden layers capture interactions between parameters. Finally, the output layer is a single neuron, using a linear activation function to output the coefficient value. The entire model calculates the mapping from input to output through forward propagation.

[0083] The training steps are as follows: First, samples are extracted from the historical lifecycle dataset. Each sample input is a vector of stage degradation comparison parameters and batch anomaly aggregation parameters. The label represents the degree of deviation between the basic prediction result and the actual failure result, and the deviation is calculated using mean squared error. The model parameters are initialized using the Xavier method, with the learning rate set to a preset small value, such as 0.001, for the Adam optimizer, and the batch size is 32. The training is iterated for 50 epochs, with forward propagation to calculate the prediction coefficients and backpropagation to update the gradient. An early stopping mechanism is used to monitor the validation set loss to prevent overfitting. The training set and validation set are split in an 8:2 ratio to ensure that the parameters are saved after the model converges.

[0084] The coefficient learning model quantifies the self-degradation trend of the board by inputting the stage degradation comparison parameter and the batch anomaly aggregation parameter to quantify the group risk aggregation. It outputs an adaptive adjustment coefficient as a scalar value to amplify or compress the basic risk score and offset the failure type confidence distribution. This ensures that the model's internal correlation reflects the correspondence between historical deviations and actual failures, allowing technicians to directly use the updated electrical performance feature sequence and model library to generate adjustment results and achieve dynamic assessment of the risk of new boards.

[0085] S4-5 Adjustment and Result Writing.

[0086] The basic risk score sequence is obtained from the fault prediction classification model. The adaptive adjustment coefficient is output by the coefficient learning model based on the stage degradation comparison parameter and the batch anomaly aggregation parameter. The basic risk score sequence is uniformly amplified by multiplication or compressed by division according to the adaptive adjustment coefficient. Fixed upward and downward shifts are set for the boundary scores of each preset risk level. When the coefficient is greater than one, indicating an underestimation trend, the coefficient is multiplied by the basic risk score, and scores close to the boundary value are added to the upward shift and moved towards the high-risk side. When the coefficient is less than one, indicating an overestimation trend, the coefficient is multiplied by the basic risk score, and scores close to the boundary value are subtracted from the downward shift and moved towards the low-risk side, resulting in a new risk score sequence and corresponding risk levels.

[0087] The basic confidence distribution of each candidate fault type is obtained from the fault prediction classification model. An adaptive adjustment coefficient changes the proportion of different fault type sets in the confidence distribution. Based on historical data, a set of fault types with high correlation to the stage degradation comparison parameter and a set of fault types with high correlation to the batch anomaly aggregation parameter are pre-labeled through correlation analysis. When the coefficient learning model determines that the stage degradation comparison parameter's information contribution is dominant based on the two parameters, the adaptive adjustment coefficient is multiplied by a preset factor as the adjustment magnitude, increasing the confidence value in the stage degradation-related fault type set and correspondingly decreasing other confidence values. When the batch anomaly aggregation parameter's information contribution is dominant, the confidence value in the batch anomaly-related fault type set is increased and other confidence values ​​are decreased. After adjustment, the confidence values ​​of all fault types are normalized by dividing the sum by the total confidence value. The fault type with the highest confidence value is selected as the primary fault type, and the fault type with the second highest confidence value and a difference from the highest confidence value within a preset range is recorded as a candidate fault type. The adjusted fault risk level and fault type prediction results are written into the risk assessment information database using a unique board identifier as the current record for that board.

[0088] When the information contribution of the batch anomaly cluster parameters is determined to be the main factor, the confidence value of the batch anomaly-related fault type set is increased while the confidence values ​​of other values ​​are decreased. The specific process is as follows: 1) The coefficient learning model, while outputting adaptive adjustment coefficients, calculates the contribution scores of two parameters to determine the dominant information contribution. Specifically, the hidden layer of the model applies independent weights to the stage degradation comparison parameter and the batch anomaly aggregation parameter, calculating their respective contribution scores as the result of multiplying the absolute value of the parameter by its corresponding weight. If the contribution score of the batch anomaly aggregation parameter is higher than that of the stage degradation comparison parameter, then the information contribution of the batch anomaly aggregation parameter is determined to be dominant.

[0089] 2) When the information contribution of the batch anomaly cluster parameter is determined to be the main factor, the process of increasing and decreasing is as follows: First, the adjustment range is calculated as the adaptive adjustment coefficient multiplied by the preset scaling factor; then, for the confidence value of each candidate fault type in the set of batch anomaly-related fault types, the new confidence value is equal to the original confidence value plus the original confidence value multiplied by the adjustment range; for the confidence values ​​of other fault types, the new confidence value is equal to the original confidence value minus the original confidence value multiplied by the adjustment range; after all confidence values ​​are updated, the confidence values ​​of all candidate fault types are normalized by dividing them by their sum to ensure that the sum of the confidence distributions is 1.

[0090] The adjusted fault risk level and fault type prediction results generated in step S4 are integrated with parameter evaluation to form dynamic risk information. This output is directly fed into step S5 to record the deviation between actual faults and maintenance results, triggering updates to the lifecycle dataset and model library. Through coefficient adjustment and confidence normalization, the prediction results capture degradation trends and anomaly commonalities, ensuring that hierarchical management adapts to batch environment variations. This supports the generation of production and maintenance decision-making information such as production scheduling adjustments, maintenance plans, and spare parts preparation strategies, enabling continuous intervention and resource allocation optimization for board group risks.

[0091] Although the adjusted fault risk level and fault type prediction results are generated in step S4, without recording actual faults and comparing and updating the model, the bias cannot be corrected, leading to increasingly inaccurate predictions. Step S5 continuously receives operational data and maintenance records, adding lifecycle records, comparing the absolute difference risk bias between the prediction and actual calculations with the consistency of Jaccard similarity types. The dataset is written back, triggering retraining of the fault prediction classification model and coefficient learning model based on the cumulative upper limit or batch size. Gradient descent is used to optimize parameters and update the bias mapping using recent samples. This process forms a closed loop, ensuring the model adapts to process changes, supports decision information generation, and achieves dynamic hierarchical management of board risks.

[0092] S5-1 Data Reception and Appending Lifecycle Records.

[0093] During the board's continued operation, it continuously receives operational data and subsequent maintenance records related to the board. When the board malfunctions or undergoes maintenance, the time of the malfunction, the type of malfunction, the scope of impact, and the maintenance measures are recorded and appended to the board's lifecycle record. First, the board's unique identifier is matched. Timestamps, multi-dimensional electrical performance measurements, and operational phase markers are extracted from the operational data. Fault type tags (e.g., short circuit or overheating), scope of impact descriptions (e.g., local failure or global interruption), and maintenance measures (e.g., component replacement or circuit rewiring) are extracted from the maintenance records. After verifying that the timestamps are normal, they are appended to the maintenance result field of the lifecycle record in chronological order. New information is appended to the end by expanding the record vector to form a complete sequence. During appending, the temporal consistency with existing electrical performance characteristic sequences is verified; if there is overlap, the measurements are merged. This record serves as input for comparison.

[0094] S5-2 compares the calculated deviations and degree of consistency.

[0095] For each board that experiences an actual failure, the predicted failure risk level and failure type recorded in step S4 are read and compared with the actual failure situation. The deviation between the predicted risk level and the actual failure severity, as well as the consistency between the predicted failure type and the actual failure type, are calculated. First, the adjusted failure risk level numerical value, the primary failure type, and the set of alternative failure types for the board's unique identifier are retrieved from the risk assessment information database and compared with the numerical value of the actual failure severity and the set of failure type labels. The absolute difference is used to calculate the risk deviation as the absolute number of the predicted value minus the actual value. The Jaccard similarity is used to calculate the type consistency as the size of the intersection of the predicted failure type set and the actual failure type set divided by the size of the union, where the type set is a set of label strings. During the calculation, for multi-type predictions, alternatives are included in the expanded set. This deviation and consistency are used as fields written back.

[0096] S5-3 writes back the record and triggers retraining.

[0097] The aforementioned deviations and consistency levels are written back to the corresponding board's records in the lifecycle dataset. Based on the updated lifecycle records, retraining or incremental training of the relevant fault prediction classification model and coefficient learning model is triggered periodically or according to a strategy. First, the lifecycle dataset row is located using the board's unique identifier. Deviation is appended as a numerical field and consistency as a proportion field to the extended column of the maintenance result, integrating them into the record structure through column appending. Triggering strategies include calculating when the cumulative average deviation exceeds a preset limit using a rolling window, monitoring when the number of newly added records reaches the batch size using a counter, and periodically scanning daily to check the proportion of new records using timestamp filtering. If any condition is met, the training module is activated, selecting the model for the affected group for updating. This trigger prepares retraining samples.

[0098] S5-4 retraining process.

[0099] During retraining, samples from recent process batches and usage environments are selected first to maintain the consistency between the model and the current production process and operating environment. The correspondence between the stage degradation comparison parameters and batch anomaly aggregation parameters and prediction deviations is updated based on the newly added life cycle records, so that the calculation of the adaptive adjustment coefficients can reflect the latest fault distribution.

[0100] First, the most recent timestamp records are filtered from the lifecycle dataset, grouped by process batch range and usage environment range to form sample subsets. For the fault prediction classification model, the electrical performance feature sequence of the input subset is used, with supervised labels such as fault category and severity, and parameters are iteratively adjusted through gradient descent to optimize the cross-entropy loss. For the coefficient learning model, the stage degradation contrast parameters and batch anomaly clustering parameters of the input subset are used, labeled as the degree of deviation. The weights are incrementally updated using the mean squared error loss function, and the mapping relationship is corrected through batch forward and backward propagation. The process uses learning rate decay to prevent oscillations and batch processing of samples to manage computational load. This update improves the model's adaptability.

[0101] S5-5 Updated Applications and Decision Generation.

[0102] After the update, the new lifecycle dataset and model library are adopted for subsequent fault prediction and classification of new boards, thereby generating production scheduling adjustment information, maintenance plans, and spare parts preparation strategies for production and operation and maintenance management. First, the updated lifecycle dataset replaces the original one, serving as the new input for grouping in step S3, and the model library is rebuilt to integrate the new parameters. Then, for subsequent new boards, the new model is called through step S4 to output the adjusted fault risk level and fault type prediction results. Based on the results, decisions are generated: a maintenance plan list is generated by sorting boards by risk level; spare parts demand is calculated through type statistics; and production scheduling is adjusted according to batch anomaly aggregation parameters, such as delaying the production of high-risk batches. During generation, the information is organized into structured tables and output to the management platform to achieve dynamic management of board group risks.

[0103] Step S5 updates the lifecycle dataset and model library, integrating deviation feedback to form adaptive resources. These resources are directly applied to subsequent new board predictions, ensuring the method's continued effectiveness under large-scale electrical performance data. Through incremental training and decision table output, the update process improves prediction robustness and supports production scheduling adjustments in production and maintenance management, such as high-risk batch delays, maintenance plan lists sorted by level, and spare parts preparation strategies statistically analyzed by type. This enables preventative optimization and resource allocation for board group risks.

[0104] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0105] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0106] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for predicting and classifying product circuit board faults based on big data on electrical performance, characterized in that, Including the following steps: S1. In the management system, summarize the test data, running data and maintenance records according to the board identifier, and generate the corresponding electrical performance timeline; S2. Extract electrical performance feature sequences from the electrical performance timeline, and associate the electrical performance feature sequences with process information and maintenance results to form a lifecycle dataset; S3. Group the lifecycle dataset according to process information and operation stage, train a fault prediction classification model based on the electrical performance characteristic sequence and maintenance results of each group, and build a model library; S4. After receiving the electrical performance data of the new board, update the features, combine the changes in its own stage and the abnormal characteristics of the batch, and evaluate its risk impact through the learning model, and dynamically adjust the risk level and fault type results output by the selected fault prediction model. S5. Record the actual faults and maintenance results of the new board, compare the actual results with the corrected fault risk level and fault type prediction results, and update the lifecycle dataset and model library based on the comparison results to generate production operation and maintenance management decision information.

2. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 1, characterized in that: Step S1 includes summarizing test data, operation data, and maintenance records using the board's unique identifier as the primary key; extracting timestamps, process nodes, operation stage markers, and multidimensional electrical performance measurement values ​​to form a preliminary record set; constructing an electrical performance timeline by sorting the timestamps in ascending order; inserting state transition records; establishing a unified data structure; performing time alignment; calculating missing values ​​or selecting values ​​by nearest neighbor resampling through linear interpolation to ensure the completeness of multidimensional electrical performance measurement values; and storing the electrical performance timeline, process information, and operation stage.

3. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 2, characterized in that: Step S2 includes segmenting the electrical performance timeline according to the operating phase label based on the operating phase division rules, serializing the multidimensional electrical performance measurement values ​​through a sliding time window to form a segment sequence, extracting the electrical performance curve shape change, the duration of rising and falling periods, the duration of exceeding the range, and the relative change relationship of channels from the segment sequence, and concatenating them to form an electrical performance feature sequence, which is then associated with process information and maintenance results to generate life cycle records, and all board life cycle records are combined to form a life cycle dataset.

4. The product board fault prediction and classification method based on electrical performance big data as described in claim 3, characterized in that: Step S3 includes grouping the lifecycle dataset according to process batch, usage environment, and operation stage; using the electrical performance feature sequence within the group as input and maintenance results as supervision labels; selecting a recurrent neural network model or random forest ensemble learning model based on sequence length or feature dimension to map features to fault risk level and fault type probability distribution; registering the model and recording the process batch range, usage environment range, and operation stage conditions to build a model library.

5. The product board fault prediction and classification method based on electrical performance big data according to claim 4, characterized in that: Step S4 includes receiving new board test data and running data, inserting the electrical performance timeline and updating the electrical performance feature sequence; calculating the stage degradation comparison parameter by comparing the pattern difference ratio point by point between the historical stage change sequence and the current stage change sequence, summing the absolute difference exceeding the limit difference ratio and dividing by the number of channels; calculating the batch anomaly aggregation parameter by comparing the similarity of the anomaly pattern vector and selecting the set of similar boards to statistically analyze the failure ratio; inputting coefficients to learn the model and outputting adaptive adjustment coefficients; selecting a fault prediction classification model from the model library to generate a basic risk score sequence and a basic confidence distribution; amplifying and compressing the basic risk score and moving the boundary score to obtain the risk level by using adaptive adjustment coefficients; and shifting the confidence distribution by adaptive adjustment coefficients and normalizing to select the main judgment fault type and alternative fault types to obtain the fault type prediction result.

6. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 5, characterized in that: The calculation of phase degradation comparison parameters selects the time alignment of the electrical performance characteristic sequences of the historical operation phase and the current operation phase. For each key electrical performance channel, the same position feature points are taken out to calculate the change direction to form the historical phase change sequence and the current phase change sequence. The statistical pattern difference ratio is compared point by point, and the absolute difference of the statistical over-limit ratio is taken as the over-limit difference ratio. The normalized value is obtained by summing by channel and dividing by the number of channels.

7. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 5, characterized in that: The batch anomaly clustering parameter calculation involves extracting the anomaly pattern vector of each board within the batch group, comparing the target board's anomaly pattern vector with other board vectors, calculating the number of intersection elements divided by the number of union elements as the similarity, selecting the boards with the highest similarity to form a set of similar boards, and counting the number of faulty boards in the set divided by the total number to obtain the parameter value.

8. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 5, characterized in that: Step S5 includes receiving operational data and maintenance records, adding lifecycle records, comparing the predicted results of the adjusted fault risk level and fault type with the actual fault calculation deviation and consistency, writing back the lifecycle dataset to trigger retraining of the fault prediction classification model and coefficient learning model, and using the updated lifecycle dataset and model library to generate production scheduling adjustment information, maintenance plans and spare parts preparation strategies.

9. The method for product board fault prediction and classification based on big data of electrical performance as described in claim 6, characterized in that: The critical electrical performance channels are extracted from multidimensional electrical performance measurements based on circuit design specifications and historical maintenance records. The correlation score of each channel in the lifecycle dataset is calculated. The linear correlation between the measured value and the severity of the fault is analyzed by Pearson correlation coefficient analysis, and channels with correlation scores higher than a preset threshold are selected.

Images